userns: Convert ext4 to user kuid/kgid where appropriate
[linux-3.10.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  64-bit file support on 64-bit platforms by Jakub Jelinek
16  *      (jj@sunsite.ms.mff.cuni.cz)
17  *
18  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
19  */
20
21 #include <linux/fs.h>
22 #include <linux/time.h>
23 #include <linux/jbd2.h>
24 #include <linux/highuid.h>
25 #include <linux/pagemap.h>
26 #include <linux/quotaops.h>
27 #include <linux/string.h>
28 #include <linux/buffer_head.h>
29 #include <linux/writeback.h>
30 #include <linux/pagevec.h>
31 #include <linux/mpage.h>
32 #include <linux/namei.h>
33 #include <linux/uio.h>
34 #include <linux/bio.h>
35 #include <linux/workqueue.h>
36 #include <linux/kernel.h>
37 #include <linux/printk.h>
38 #include <linux/slab.h>
39 #include <linux/ratelimit.h>
40
41 #include "ext4_jbd2.h"
42 #include "xattr.h"
43 #include "acl.h"
44 #include "truncate.h"
45
46 #include <trace/events/ext4.h>
47
48 #define MPAGE_DA_EXTENT_TAIL 0x01
49
50 static inline int ext4_begin_ordered_truncate(struct inode *inode,
51                                               loff_t new_size)
52 {
53         trace_ext4_begin_ordered_truncate(inode, new_size);
54         /*
55          * If jinode is zero, then we never opened the file for
56          * writing, so there's no need to call
57          * jbd2_journal_begin_ordered_truncate() since there's no
58          * outstanding writes we need to flush.
59          */
60         if (!EXT4_I(inode)->jinode)
61                 return 0;
62         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
63                                                    EXT4_I(inode)->jinode,
64                                                    new_size);
65 }
66
67 static void ext4_invalidatepage(struct page *page, unsigned long offset);
68 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
69                                    struct buffer_head *bh_result, int create);
70 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
71 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
72 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
73 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
74 static int ext4_discard_partial_page_buffers_no_lock(handle_t *handle,
75                 struct inode *inode, struct page *page, loff_t from,
76                 loff_t length, int flags);
77
78 /*
79  * Test whether an inode is a fast symlink.
80  */
81 static int ext4_inode_is_fast_symlink(struct inode *inode)
82 {
83         int ea_blocks = EXT4_I(inode)->i_file_acl ?
84                 (inode->i_sb->s_blocksize >> 9) : 0;
85
86         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
87 }
88
89 /*
90  * Restart the transaction associated with *handle.  This does a commit,
91  * so before we call here everything must be consistently dirtied against
92  * this transaction.
93  */
94 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
95                                  int nblocks)
96 {
97         int ret;
98
99         /*
100          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
101          * moment, get_block can be called only for blocks inside i_size since
102          * page cache has been already dropped and writes are blocked by
103          * i_mutex. So we can safely drop the i_data_sem here.
104          */
105         BUG_ON(EXT4_JOURNAL(inode) == NULL);
106         jbd_debug(2, "restarting handle %p\n", handle);
107         up_write(&EXT4_I(inode)->i_data_sem);
108         ret = ext4_journal_restart(handle, nblocks);
109         down_write(&EXT4_I(inode)->i_data_sem);
110         ext4_discard_preallocations(inode);
111
112         return ret;
113 }
114
115 /*
116  * Called at the last iput() if i_nlink is zero.
117  */
118 void ext4_evict_inode(struct inode *inode)
119 {
120         handle_t *handle;
121         int err;
122
123         trace_ext4_evict_inode(inode);
124
125         ext4_ioend_wait(inode);
126
127         if (inode->i_nlink) {
128                 /*
129                  * When journalling data dirty buffers are tracked only in the
130                  * journal. So although mm thinks everything is clean and
131                  * ready for reaping the inode might still have some pages to
132                  * write in the running transaction or waiting to be
133                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
134                  * (via truncate_inode_pages()) to discard these buffers can
135                  * cause data loss. Also even if we did not discard these
136                  * buffers, we would have no way to find them after the inode
137                  * is reaped and thus user could see stale data if he tries to
138                  * read them before the transaction is checkpointed. So be
139                  * careful and force everything to disk here... We use
140                  * ei->i_datasync_tid to store the newest transaction
141                  * containing inode's data.
142                  *
143                  * Note that directories do not have this problem because they
144                  * don't use page cache.
145                  */
146                 if (ext4_should_journal_data(inode) &&
147                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
148                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
149                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
150
151                         jbd2_log_start_commit(journal, commit_tid);
152                         jbd2_log_wait_commit(journal, commit_tid);
153                         filemap_write_and_wait(&inode->i_data);
154                 }
155                 truncate_inode_pages(&inode->i_data, 0);
156                 goto no_delete;
157         }
158
159         if (!is_bad_inode(inode))
160                 dquot_initialize(inode);
161
162         if (ext4_should_order_data(inode))
163                 ext4_begin_ordered_truncate(inode, 0);
164         truncate_inode_pages(&inode->i_data, 0);
165
166         if (is_bad_inode(inode))
167                 goto no_delete;
168
169         handle = ext4_journal_start(inode, ext4_blocks_for_truncate(inode)+3);
170         if (IS_ERR(handle)) {
171                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
172                 /*
173                  * If we're going to skip the normal cleanup, we still need to
174                  * make sure that the in-core orphan linked list is properly
175                  * cleaned up.
176                  */
177                 ext4_orphan_del(NULL, inode);
178                 goto no_delete;
179         }
180
181         if (IS_SYNC(inode))
182                 ext4_handle_sync(handle);
183         inode->i_size = 0;
184         err = ext4_mark_inode_dirty(handle, inode);
185         if (err) {
186                 ext4_warning(inode->i_sb,
187                              "couldn't mark inode dirty (err %d)", err);
188                 goto stop_handle;
189         }
190         if (inode->i_blocks)
191                 ext4_truncate(inode);
192
193         /*
194          * ext4_ext_truncate() doesn't reserve any slop when it
195          * restarts journal transactions; therefore there may not be
196          * enough credits left in the handle to remove the inode from
197          * the orphan list and set the dtime field.
198          */
199         if (!ext4_handle_has_enough_credits(handle, 3)) {
200                 err = ext4_journal_extend(handle, 3);
201                 if (err > 0)
202                         err = ext4_journal_restart(handle, 3);
203                 if (err != 0) {
204                         ext4_warning(inode->i_sb,
205                                      "couldn't extend journal (err %d)", err);
206                 stop_handle:
207                         ext4_journal_stop(handle);
208                         ext4_orphan_del(NULL, inode);
209                         goto no_delete;
210                 }
211         }
212
213         /*
214          * Kill off the orphan record which ext4_truncate created.
215          * AKPM: I think this can be inside the above `if'.
216          * Note that ext4_orphan_del() has to be able to cope with the
217          * deletion of a non-existent orphan - this is because we don't
218          * know if ext4_truncate() actually created an orphan record.
219          * (Well, we could do this if we need to, but heck - it works)
220          */
221         ext4_orphan_del(handle, inode);
222         EXT4_I(inode)->i_dtime  = get_seconds();
223
224         /*
225          * One subtle ordering requirement: if anything has gone wrong
226          * (transaction abort, IO errors, whatever), then we can still
227          * do these next steps (the fs will already have been marked as
228          * having errors), but we can't free the inode if the mark_dirty
229          * fails.
230          */
231         if (ext4_mark_inode_dirty(handle, inode))
232                 /* If that failed, just do the required in-core inode clear. */
233                 ext4_clear_inode(inode);
234         else
235                 ext4_free_inode(handle, inode);
236         ext4_journal_stop(handle);
237         return;
238 no_delete:
239         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
240 }
241
242 #ifdef CONFIG_QUOTA
243 qsize_t *ext4_get_reserved_space(struct inode *inode)
244 {
245         return &EXT4_I(inode)->i_reserved_quota;
246 }
247 #endif
248
249 /*
250  * Calculate the number of metadata blocks need to reserve
251  * to allocate a block located at @lblock
252  */
253 static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
254 {
255         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
256                 return ext4_ext_calc_metadata_amount(inode, lblock);
257
258         return ext4_ind_calc_metadata_amount(inode, lblock);
259 }
260
261 /*
262  * Called with i_data_sem down, which is important since we can call
263  * ext4_discard_preallocations() from here.
264  */
265 void ext4_da_update_reserve_space(struct inode *inode,
266                                         int used, int quota_claim)
267 {
268         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
269         struct ext4_inode_info *ei = EXT4_I(inode);
270
271         spin_lock(&ei->i_block_reservation_lock);
272         trace_ext4_da_update_reserve_space(inode, used, quota_claim);
273         if (unlikely(used > ei->i_reserved_data_blocks)) {
274                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
275                          "with only %d reserved data blocks",
276                          __func__, inode->i_ino, used,
277                          ei->i_reserved_data_blocks);
278                 WARN_ON(1);
279                 used = ei->i_reserved_data_blocks;
280         }
281
282         /* Update per-inode reservations */
283         ei->i_reserved_data_blocks -= used;
284         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
285         percpu_counter_sub(&sbi->s_dirtyclusters_counter,
286                            used + ei->i_allocated_meta_blocks);
287         ei->i_allocated_meta_blocks = 0;
288
289         if (ei->i_reserved_data_blocks == 0) {
290                 /*
291                  * We can release all of the reserved metadata blocks
292                  * only when we have written all of the delayed
293                  * allocation blocks.
294                  */
295                 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
296                                    ei->i_reserved_meta_blocks);
297                 ei->i_reserved_meta_blocks = 0;
298                 ei->i_da_metadata_calc_len = 0;
299         }
300         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
301
302         /* Update quota subsystem for data blocks */
303         if (quota_claim)
304                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
305         else {
306                 /*
307                  * We did fallocate with an offset that is already delayed
308                  * allocated. So on delayed allocated writeback we should
309                  * not re-claim the quota for fallocated blocks.
310                  */
311                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
312         }
313
314         /*
315          * If we have done all the pending block allocations and if
316          * there aren't any writers on the inode, we can discard the
317          * inode's preallocations.
318          */
319         if ((ei->i_reserved_data_blocks == 0) &&
320             (atomic_read(&inode->i_writecount) == 0))
321                 ext4_discard_preallocations(inode);
322 }
323
324 static int __check_block_validity(struct inode *inode, const char *func,
325                                 unsigned int line,
326                                 struct ext4_map_blocks *map)
327 {
328         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
329                                    map->m_len)) {
330                 ext4_error_inode(inode, func, line, map->m_pblk,
331                                  "lblock %lu mapped to illegal pblock "
332                                  "(length %d)", (unsigned long) map->m_lblk,
333                                  map->m_len);
334                 return -EIO;
335         }
336         return 0;
337 }
338
339 #define check_block_validity(inode, map)        \
340         __check_block_validity((inode), __func__, __LINE__, (map))
341
342 /*
343  * Return the number of contiguous dirty pages in a given inode
344  * starting at page frame idx.
345  */
346 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
347                                     unsigned int max_pages)
348 {
349         struct address_space *mapping = inode->i_mapping;
350         pgoff_t index;
351         struct pagevec pvec;
352         pgoff_t num = 0;
353         int i, nr_pages, done = 0;
354
355         if (max_pages == 0)
356                 return 0;
357         pagevec_init(&pvec, 0);
358         while (!done) {
359                 index = idx;
360                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
361                                               PAGECACHE_TAG_DIRTY,
362                                               (pgoff_t)PAGEVEC_SIZE);
363                 if (nr_pages == 0)
364                         break;
365                 for (i = 0; i < nr_pages; i++) {
366                         struct page *page = pvec.pages[i];
367                         struct buffer_head *bh, *head;
368
369                         lock_page(page);
370                         if (unlikely(page->mapping != mapping) ||
371                             !PageDirty(page) ||
372                             PageWriteback(page) ||
373                             page->index != idx) {
374                                 done = 1;
375                                 unlock_page(page);
376                                 break;
377                         }
378                         if (page_has_buffers(page)) {
379                                 bh = head = page_buffers(page);
380                                 do {
381                                         if (!buffer_delay(bh) &&
382                                             !buffer_unwritten(bh))
383                                                 done = 1;
384                                         bh = bh->b_this_page;
385                                 } while (!done && (bh != head));
386                         }
387                         unlock_page(page);
388                         if (done)
389                                 break;
390                         idx++;
391                         num++;
392                         if (num >= max_pages) {
393                                 done = 1;
394                                 break;
395                         }
396                 }
397                 pagevec_release(&pvec);
398         }
399         return num;
400 }
401
402 /*
403  * Sets the BH_Da_Mapped bit on the buffer heads corresponding to the given map.
404  */
405 static void set_buffers_da_mapped(struct inode *inode,
406                                    struct ext4_map_blocks *map)
407 {
408         struct address_space *mapping = inode->i_mapping;
409         struct pagevec pvec;
410         int i, nr_pages;
411         pgoff_t index, end;
412
413         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
414         end = (map->m_lblk + map->m_len - 1) >>
415                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
416
417         pagevec_init(&pvec, 0);
418         while (index <= end) {
419                 nr_pages = pagevec_lookup(&pvec, mapping, index,
420                                           min(end - index + 1,
421                                               (pgoff_t)PAGEVEC_SIZE));
422                 if (nr_pages == 0)
423                         break;
424                 for (i = 0; i < nr_pages; i++) {
425                         struct page *page = pvec.pages[i];
426                         struct buffer_head *bh, *head;
427
428                         if (unlikely(page->mapping != mapping) ||
429                             !PageDirty(page))
430                                 break;
431
432                         if (page_has_buffers(page)) {
433                                 bh = head = page_buffers(page);
434                                 do {
435                                         set_buffer_da_mapped(bh);
436                                         bh = bh->b_this_page;
437                                 } while (bh != head);
438                         }
439                         index++;
440                 }
441                 pagevec_release(&pvec);
442         }
443 }
444
445 /*
446  * The ext4_map_blocks() function tries to look up the requested blocks,
447  * and returns if the blocks are already mapped.
448  *
449  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
450  * and store the allocated blocks in the result buffer head and mark it
451  * mapped.
452  *
453  * If file type is extents based, it will call ext4_ext_map_blocks(),
454  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
455  * based files
456  *
457  * On success, it returns the number of blocks being mapped or allocate.
458  * if create==0 and the blocks are pre-allocated and uninitialized block,
459  * the result buffer head is unmapped. If the create ==1, it will make sure
460  * the buffer head is mapped.
461  *
462  * It returns 0 if plain look up failed (blocks have not been allocated), in
463  * that case, buffer head is unmapped
464  *
465  * It returns the error in case of allocation failure.
466  */
467 int ext4_map_blocks(handle_t *handle, struct inode *inode,
468                     struct ext4_map_blocks *map, int flags)
469 {
470         int retval;
471
472         map->m_flags = 0;
473         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
474                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
475                   (unsigned long) map->m_lblk);
476         /*
477          * Try to see if we can get the block without requesting a new
478          * file system block.
479          */
480         down_read((&EXT4_I(inode)->i_data_sem));
481         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
482                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
483                                              EXT4_GET_BLOCKS_KEEP_SIZE);
484         } else {
485                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
486                                              EXT4_GET_BLOCKS_KEEP_SIZE);
487         }
488         up_read((&EXT4_I(inode)->i_data_sem));
489
490         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
491                 int ret = check_block_validity(inode, map);
492                 if (ret != 0)
493                         return ret;
494         }
495
496         /* If it is only a block(s) look up */
497         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
498                 return retval;
499
500         /*
501          * Returns if the blocks have already allocated
502          *
503          * Note that if blocks have been preallocated
504          * ext4_ext_get_block() returns the create = 0
505          * with buffer head unmapped.
506          */
507         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
508                 return retval;
509
510         /*
511          * When we call get_blocks without the create flag, the
512          * BH_Unwritten flag could have gotten set if the blocks
513          * requested were part of a uninitialized extent.  We need to
514          * clear this flag now that we are committed to convert all or
515          * part of the uninitialized extent to be an initialized
516          * extent.  This is because we need to avoid the combination
517          * of BH_Unwritten and BH_Mapped flags being simultaneously
518          * set on the buffer_head.
519          */
520         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
521
522         /*
523          * New blocks allocate and/or writing to uninitialized extent
524          * will possibly result in updating i_data, so we take
525          * the write lock of i_data_sem, and call get_blocks()
526          * with create == 1 flag.
527          */
528         down_write((&EXT4_I(inode)->i_data_sem));
529
530         /*
531          * if the caller is from delayed allocation writeout path
532          * we have already reserved fs blocks for allocation
533          * let the underlying get_block() function know to
534          * avoid double accounting
535          */
536         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
537                 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
538         /*
539          * We need to check for EXT4 here because migrate
540          * could have changed the inode type in between
541          */
542         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
543                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
544         } else {
545                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
546
547                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
548                         /*
549                          * We allocated new blocks which will result in
550                          * i_data's format changing.  Force the migrate
551                          * to fail by clearing migrate flags
552                          */
553                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
554                 }
555
556                 /*
557                  * Update reserved blocks/metadata blocks after successful
558                  * block allocation which had been deferred till now. We don't
559                  * support fallocate for non extent files. So we can update
560                  * reserve space here.
561                  */
562                 if ((retval > 0) &&
563                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
564                         ext4_da_update_reserve_space(inode, retval, 1);
565         }
566         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
567                 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
568
569                 /* If we have successfully mapped the delayed allocated blocks,
570                  * set the BH_Da_Mapped bit on them. Its important to do this
571                  * under the protection of i_data_sem.
572                  */
573                 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
574                         set_buffers_da_mapped(inode, map);
575         }
576
577         up_write((&EXT4_I(inode)->i_data_sem));
578         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
579                 int ret = check_block_validity(inode, map);
580                 if (ret != 0)
581                         return ret;
582         }
583         return retval;
584 }
585
586 /* Maximum number of blocks we map for direct IO at once. */
587 #define DIO_MAX_BLOCKS 4096
588
589 static int _ext4_get_block(struct inode *inode, sector_t iblock,
590                            struct buffer_head *bh, int flags)
591 {
592         handle_t *handle = ext4_journal_current_handle();
593         struct ext4_map_blocks map;
594         int ret = 0, started = 0;
595         int dio_credits;
596
597         map.m_lblk = iblock;
598         map.m_len = bh->b_size >> inode->i_blkbits;
599
600         if (flags && !handle) {
601                 /* Direct IO write... */
602                 if (map.m_len > DIO_MAX_BLOCKS)
603                         map.m_len = DIO_MAX_BLOCKS;
604                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
605                 handle = ext4_journal_start(inode, dio_credits);
606                 if (IS_ERR(handle)) {
607                         ret = PTR_ERR(handle);
608                         return ret;
609                 }
610                 started = 1;
611         }
612
613         ret = ext4_map_blocks(handle, inode, &map, flags);
614         if (ret > 0) {
615                 map_bh(bh, inode->i_sb, map.m_pblk);
616                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
617                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
618                 ret = 0;
619         }
620         if (started)
621                 ext4_journal_stop(handle);
622         return ret;
623 }
624
625 int ext4_get_block(struct inode *inode, sector_t iblock,
626                    struct buffer_head *bh, int create)
627 {
628         return _ext4_get_block(inode, iblock, bh,
629                                create ? EXT4_GET_BLOCKS_CREATE : 0);
630 }
631
632 /*
633  * `handle' can be NULL if create is zero
634  */
635 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
636                                 ext4_lblk_t block, int create, int *errp)
637 {
638         struct ext4_map_blocks map;
639         struct buffer_head *bh;
640         int fatal = 0, err;
641
642         J_ASSERT(handle != NULL || create == 0);
643
644         map.m_lblk = block;
645         map.m_len = 1;
646         err = ext4_map_blocks(handle, inode, &map,
647                               create ? EXT4_GET_BLOCKS_CREATE : 0);
648
649         if (err < 0)
650                 *errp = err;
651         if (err <= 0)
652                 return NULL;
653         *errp = 0;
654
655         bh = sb_getblk(inode->i_sb, map.m_pblk);
656         if (!bh) {
657                 *errp = -EIO;
658                 return NULL;
659         }
660         if (map.m_flags & EXT4_MAP_NEW) {
661                 J_ASSERT(create != 0);
662                 J_ASSERT(handle != NULL);
663
664                 /*
665                  * Now that we do not always journal data, we should
666                  * keep in mind whether this should always journal the
667                  * new buffer as metadata.  For now, regular file
668                  * writes use ext4_get_block instead, so it's not a
669                  * problem.
670                  */
671                 lock_buffer(bh);
672                 BUFFER_TRACE(bh, "call get_create_access");
673                 fatal = ext4_journal_get_create_access(handle, bh);
674                 if (!fatal && !buffer_uptodate(bh)) {
675                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
676                         set_buffer_uptodate(bh);
677                 }
678                 unlock_buffer(bh);
679                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
680                 err = ext4_handle_dirty_metadata(handle, inode, bh);
681                 if (!fatal)
682                         fatal = err;
683         } else {
684                 BUFFER_TRACE(bh, "not a new buffer");
685         }
686         if (fatal) {
687                 *errp = fatal;
688                 brelse(bh);
689                 bh = NULL;
690         }
691         return bh;
692 }
693
694 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
695                                ext4_lblk_t block, int create, int *err)
696 {
697         struct buffer_head *bh;
698
699         bh = ext4_getblk(handle, inode, block, create, err);
700         if (!bh)
701                 return bh;
702         if (buffer_uptodate(bh))
703                 return bh;
704         ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
705         wait_on_buffer(bh);
706         if (buffer_uptodate(bh))
707                 return bh;
708         put_bh(bh);
709         *err = -EIO;
710         return NULL;
711 }
712
713 static int walk_page_buffers(handle_t *handle,
714                              struct buffer_head *head,
715                              unsigned from,
716                              unsigned to,
717                              int *partial,
718                              int (*fn)(handle_t *handle,
719                                        struct buffer_head *bh))
720 {
721         struct buffer_head *bh;
722         unsigned block_start, block_end;
723         unsigned blocksize = head->b_size;
724         int err, ret = 0;
725         struct buffer_head *next;
726
727         for (bh = head, block_start = 0;
728              ret == 0 && (bh != head || !block_start);
729              block_start = block_end, bh = next) {
730                 next = bh->b_this_page;
731                 block_end = block_start + blocksize;
732                 if (block_end <= from || block_start >= to) {
733                         if (partial && !buffer_uptodate(bh))
734                                 *partial = 1;
735                         continue;
736                 }
737                 err = (*fn)(handle, bh);
738                 if (!ret)
739                         ret = err;
740         }
741         return ret;
742 }
743
744 /*
745  * To preserve ordering, it is essential that the hole instantiation and
746  * the data write be encapsulated in a single transaction.  We cannot
747  * close off a transaction and start a new one between the ext4_get_block()
748  * and the commit_write().  So doing the jbd2_journal_start at the start of
749  * prepare_write() is the right place.
750  *
751  * Also, this function can nest inside ext4_writepage() ->
752  * block_write_full_page(). In that case, we *know* that ext4_writepage()
753  * has generated enough buffer credits to do the whole page.  So we won't
754  * block on the journal in that case, which is good, because the caller may
755  * be PF_MEMALLOC.
756  *
757  * By accident, ext4 can be reentered when a transaction is open via
758  * quota file writes.  If we were to commit the transaction while thus
759  * reentered, there can be a deadlock - we would be holding a quota
760  * lock, and the commit would never complete if another thread had a
761  * transaction open and was blocking on the quota lock - a ranking
762  * violation.
763  *
764  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
765  * will _not_ run commit under these circumstances because handle->h_ref
766  * is elevated.  We'll still have enough credits for the tiny quotafile
767  * write.
768  */
769 static int do_journal_get_write_access(handle_t *handle,
770                                        struct buffer_head *bh)
771 {
772         int dirty = buffer_dirty(bh);
773         int ret;
774
775         if (!buffer_mapped(bh) || buffer_freed(bh))
776                 return 0;
777         /*
778          * __block_write_begin() could have dirtied some buffers. Clean
779          * the dirty bit as jbd2_journal_get_write_access() could complain
780          * otherwise about fs integrity issues. Setting of the dirty bit
781          * by __block_write_begin() isn't a real problem here as we clear
782          * the bit before releasing a page lock and thus writeback cannot
783          * ever write the buffer.
784          */
785         if (dirty)
786                 clear_buffer_dirty(bh);
787         ret = ext4_journal_get_write_access(handle, bh);
788         if (!ret && dirty)
789                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
790         return ret;
791 }
792
793 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
794                    struct buffer_head *bh_result, int create);
795 static int ext4_write_begin(struct file *file, struct address_space *mapping,
796                             loff_t pos, unsigned len, unsigned flags,
797                             struct page **pagep, void **fsdata)
798 {
799         struct inode *inode = mapping->host;
800         int ret, needed_blocks;
801         handle_t *handle;
802         int retries = 0;
803         struct page *page;
804         pgoff_t index;
805         unsigned from, to;
806
807         trace_ext4_write_begin(inode, pos, len, flags);
808         /*
809          * Reserve one block more for addition to orphan list in case
810          * we allocate blocks but write fails for some reason
811          */
812         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
813         index = pos >> PAGE_CACHE_SHIFT;
814         from = pos & (PAGE_CACHE_SIZE - 1);
815         to = from + len;
816
817 retry:
818         handle = ext4_journal_start(inode, needed_blocks);
819         if (IS_ERR(handle)) {
820                 ret = PTR_ERR(handle);
821                 goto out;
822         }
823
824         /* We cannot recurse into the filesystem as the transaction is already
825          * started */
826         flags |= AOP_FLAG_NOFS;
827
828         page = grab_cache_page_write_begin(mapping, index, flags);
829         if (!page) {
830                 ext4_journal_stop(handle);
831                 ret = -ENOMEM;
832                 goto out;
833         }
834         *pagep = page;
835
836         if (ext4_should_dioread_nolock(inode))
837                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
838         else
839                 ret = __block_write_begin(page, pos, len, ext4_get_block);
840
841         if (!ret && ext4_should_journal_data(inode)) {
842                 ret = walk_page_buffers(handle, page_buffers(page),
843                                 from, to, NULL, do_journal_get_write_access);
844         }
845
846         if (ret) {
847                 unlock_page(page);
848                 page_cache_release(page);
849                 /*
850                  * __block_write_begin may have instantiated a few blocks
851                  * outside i_size.  Trim these off again. Don't need
852                  * i_size_read because we hold i_mutex.
853                  *
854                  * Add inode to orphan list in case we crash before
855                  * truncate finishes
856                  */
857                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
858                         ext4_orphan_add(handle, inode);
859
860                 ext4_journal_stop(handle);
861                 if (pos + len > inode->i_size) {
862                         ext4_truncate_failed_write(inode);
863                         /*
864                          * If truncate failed early the inode might
865                          * still be on the orphan list; we need to
866                          * make sure the inode is removed from the
867                          * orphan list in that case.
868                          */
869                         if (inode->i_nlink)
870                                 ext4_orphan_del(NULL, inode);
871                 }
872         }
873
874         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
875                 goto retry;
876 out:
877         return ret;
878 }
879
880 /* For write_end() in data=journal mode */
881 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
882 {
883         if (!buffer_mapped(bh) || buffer_freed(bh))
884                 return 0;
885         set_buffer_uptodate(bh);
886         return ext4_handle_dirty_metadata(handle, NULL, bh);
887 }
888
889 static int ext4_generic_write_end(struct file *file,
890                                   struct address_space *mapping,
891                                   loff_t pos, unsigned len, unsigned copied,
892                                   struct page *page, void *fsdata)
893 {
894         int i_size_changed = 0;
895         struct inode *inode = mapping->host;
896         handle_t *handle = ext4_journal_current_handle();
897
898         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
899
900         /*
901          * No need to use i_size_read() here, the i_size
902          * cannot change under us because we hold i_mutex.
903          *
904          * But it's important to update i_size while still holding page lock:
905          * page writeout could otherwise come in and zero beyond i_size.
906          */
907         if (pos + copied > inode->i_size) {
908                 i_size_write(inode, pos + copied);
909                 i_size_changed = 1;
910         }
911
912         if (pos + copied >  EXT4_I(inode)->i_disksize) {
913                 /* We need to mark inode dirty even if
914                  * new_i_size is less that inode->i_size
915                  * bu greater than i_disksize.(hint delalloc)
916                  */
917                 ext4_update_i_disksize(inode, (pos + copied));
918                 i_size_changed = 1;
919         }
920         unlock_page(page);
921         page_cache_release(page);
922
923         /*
924          * Don't mark the inode dirty under page lock. First, it unnecessarily
925          * makes the holding time of page lock longer. Second, it forces lock
926          * ordering of page lock and transaction start for journaling
927          * filesystems.
928          */
929         if (i_size_changed)
930                 ext4_mark_inode_dirty(handle, inode);
931
932         return copied;
933 }
934
935 /*
936  * We need to pick up the new inode size which generic_commit_write gave us
937  * `file' can be NULL - eg, when called from page_symlink().
938  *
939  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
940  * buffers are managed internally.
941  */
942 static int ext4_ordered_write_end(struct file *file,
943                                   struct address_space *mapping,
944                                   loff_t pos, unsigned len, unsigned copied,
945                                   struct page *page, void *fsdata)
946 {
947         handle_t *handle = ext4_journal_current_handle();
948         struct inode *inode = mapping->host;
949         int ret = 0, ret2;
950
951         trace_ext4_ordered_write_end(inode, pos, len, copied);
952         ret = ext4_jbd2_file_inode(handle, inode);
953
954         if (ret == 0) {
955                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
956                                                         page, fsdata);
957                 copied = ret2;
958                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
959                         /* if we have allocated more blocks and copied
960                          * less. We will have blocks allocated outside
961                          * inode->i_size. So truncate them
962                          */
963                         ext4_orphan_add(handle, inode);
964                 if (ret2 < 0)
965                         ret = ret2;
966         } else {
967                 unlock_page(page);
968                 page_cache_release(page);
969         }
970
971         ret2 = ext4_journal_stop(handle);
972         if (!ret)
973                 ret = ret2;
974
975         if (pos + len > inode->i_size) {
976                 ext4_truncate_failed_write(inode);
977                 /*
978                  * If truncate failed early the inode might still be
979                  * on the orphan list; we need to make sure the inode
980                  * is removed from the orphan list in that case.
981                  */
982                 if (inode->i_nlink)
983                         ext4_orphan_del(NULL, inode);
984         }
985
986
987         return ret ? ret : copied;
988 }
989
990 static int ext4_writeback_write_end(struct file *file,
991                                     struct address_space *mapping,
992                                     loff_t pos, unsigned len, unsigned copied,
993                                     struct page *page, void *fsdata)
994 {
995         handle_t *handle = ext4_journal_current_handle();
996         struct inode *inode = mapping->host;
997         int ret = 0, ret2;
998
999         trace_ext4_writeback_write_end(inode, pos, len, copied);
1000         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1001                                                         page, fsdata);
1002         copied = ret2;
1003         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1004                 /* if we have allocated more blocks and copied
1005                  * less. We will have blocks allocated outside
1006                  * inode->i_size. So truncate them
1007                  */
1008                 ext4_orphan_add(handle, inode);
1009
1010         if (ret2 < 0)
1011                 ret = ret2;
1012
1013         ret2 = ext4_journal_stop(handle);
1014         if (!ret)
1015                 ret = ret2;
1016
1017         if (pos + len > inode->i_size) {
1018                 ext4_truncate_failed_write(inode);
1019                 /*
1020                  * If truncate failed early the inode might still be
1021                  * on the orphan list; we need to make sure the inode
1022                  * is removed from the orphan list in that case.
1023                  */
1024                 if (inode->i_nlink)
1025                         ext4_orphan_del(NULL, inode);
1026         }
1027
1028         return ret ? ret : copied;
1029 }
1030
1031 static int ext4_journalled_write_end(struct file *file,
1032                                      struct address_space *mapping,
1033                                      loff_t pos, unsigned len, unsigned copied,
1034                                      struct page *page, void *fsdata)
1035 {
1036         handle_t *handle = ext4_journal_current_handle();
1037         struct inode *inode = mapping->host;
1038         int ret = 0, ret2;
1039         int partial = 0;
1040         unsigned from, to;
1041         loff_t new_i_size;
1042
1043         trace_ext4_journalled_write_end(inode, pos, len, copied);
1044         from = pos & (PAGE_CACHE_SIZE - 1);
1045         to = from + len;
1046
1047         BUG_ON(!ext4_handle_valid(handle));
1048
1049         if (copied < len) {
1050                 if (!PageUptodate(page))
1051                         copied = 0;
1052                 page_zero_new_buffers(page, from+copied, to);
1053         }
1054
1055         ret = walk_page_buffers(handle, page_buffers(page), from,
1056                                 to, &partial, write_end_fn);
1057         if (!partial)
1058                 SetPageUptodate(page);
1059         new_i_size = pos + copied;
1060         if (new_i_size > inode->i_size)
1061                 i_size_write(inode, pos+copied);
1062         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1063         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1064         if (new_i_size > EXT4_I(inode)->i_disksize) {
1065                 ext4_update_i_disksize(inode, new_i_size);
1066                 ret2 = ext4_mark_inode_dirty(handle, inode);
1067                 if (!ret)
1068                         ret = ret2;
1069         }
1070
1071         unlock_page(page);
1072         page_cache_release(page);
1073         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1074                 /* if we have allocated more blocks and copied
1075                  * less. We will have blocks allocated outside
1076                  * inode->i_size. So truncate them
1077                  */
1078                 ext4_orphan_add(handle, inode);
1079
1080         ret2 = ext4_journal_stop(handle);
1081         if (!ret)
1082                 ret = ret2;
1083         if (pos + len > inode->i_size) {
1084                 ext4_truncate_failed_write(inode);
1085                 /*
1086                  * If truncate failed early the inode might still be
1087                  * on the orphan list; we need to make sure the inode
1088                  * is removed from the orphan list in that case.
1089                  */
1090                 if (inode->i_nlink)
1091                         ext4_orphan_del(NULL, inode);
1092         }
1093
1094         return ret ? ret : copied;
1095 }
1096
1097 /*
1098  * Reserve a single cluster located at lblock
1099  */
1100 static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1101 {
1102         int retries = 0;
1103         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1104         struct ext4_inode_info *ei = EXT4_I(inode);
1105         unsigned int md_needed;
1106         int ret;
1107
1108         /*
1109          * recalculate the amount of metadata blocks to reserve
1110          * in order to allocate nrblocks
1111          * worse case is one extent per block
1112          */
1113 repeat:
1114         spin_lock(&ei->i_block_reservation_lock);
1115         md_needed = EXT4_NUM_B2C(sbi,
1116                                  ext4_calc_metadata_amount(inode, lblock));
1117         trace_ext4_da_reserve_space(inode, md_needed);
1118         spin_unlock(&ei->i_block_reservation_lock);
1119
1120         /*
1121          * We will charge metadata quota at writeout time; this saves
1122          * us from metadata over-estimation, though we may go over by
1123          * a small amount in the end.  Here we just reserve for data.
1124          */
1125         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1126         if (ret)
1127                 return ret;
1128         /*
1129          * We do still charge estimated metadata to the sb though;
1130          * we cannot afford to run out of free blocks.
1131          */
1132         if (ext4_claim_free_clusters(sbi, md_needed + 1, 0)) {
1133                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1134                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1135                         yield();
1136                         goto repeat;
1137                 }
1138                 return -ENOSPC;
1139         }
1140         spin_lock(&ei->i_block_reservation_lock);
1141         ei->i_reserved_data_blocks++;
1142         ei->i_reserved_meta_blocks += md_needed;
1143         spin_unlock(&ei->i_block_reservation_lock);
1144
1145         return 0;       /* success */
1146 }
1147
1148 static void ext4_da_release_space(struct inode *inode, int to_free)
1149 {
1150         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1151         struct ext4_inode_info *ei = EXT4_I(inode);
1152
1153         if (!to_free)
1154                 return;         /* Nothing to release, exit */
1155
1156         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1157
1158         trace_ext4_da_release_space(inode, to_free);
1159         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1160                 /*
1161                  * if there aren't enough reserved blocks, then the
1162                  * counter is messed up somewhere.  Since this
1163                  * function is called from invalidate page, it's
1164                  * harmless to return without any action.
1165                  */
1166                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1167                          "ino %lu, to_free %d with only %d reserved "
1168                          "data blocks", inode->i_ino, to_free,
1169                          ei->i_reserved_data_blocks);
1170                 WARN_ON(1);
1171                 to_free = ei->i_reserved_data_blocks;
1172         }
1173         ei->i_reserved_data_blocks -= to_free;
1174
1175         if (ei->i_reserved_data_blocks == 0) {
1176                 /*
1177                  * We can release all of the reserved metadata blocks
1178                  * only when we have written all of the delayed
1179                  * allocation blocks.
1180                  * Note that in case of bigalloc, i_reserved_meta_blocks,
1181                  * i_reserved_data_blocks, etc. refer to number of clusters.
1182                  */
1183                 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
1184                                    ei->i_reserved_meta_blocks);
1185                 ei->i_reserved_meta_blocks = 0;
1186                 ei->i_da_metadata_calc_len = 0;
1187         }
1188
1189         /* update fs dirty data blocks counter */
1190         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1191
1192         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1193
1194         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1195 }
1196
1197 static void ext4_da_page_release_reservation(struct page *page,
1198                                              unsigned long offset)
1199 {
1200         int to_release = 0;
1201         struct buffer_head *head, *bh;
1202         unsigned int curr_off = 0;
1203         struct inode *inode = page->mapping->host;
1204         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1205         int num_clusters;
1206
1207         head = page_buffers(page);
1208         bh = head;
1209         do {
1210                 unsigned int next_off = curr_off + bh->b_size;
1211
1212                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1213                         to_release++;
1214                         clear_buffer_delay(bh);
1215                         clear_buffer_da_mapped(bh);
1216                 }
1217                 curr_off = next_off;
1218         } while ((bh = bh->b_this_page) != head);
1219
1220         /* If we have released all the blocks belonging to a cluster, then we
1221          * need to release the reserved space for that cluster. */
1222         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1223         while (num_clusters > 0) {
1224                 ext4_fsblk_t lblk;
1225                 lblk = (page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits)) +
1226                         ((num_clusters - 1) << sbi->s_cluster_bits);
1227                 if (sbi->s_cluster_ratio == 1 ||
1228                     !ext4_find_delalloc_cluster(inode, lblk, 1))
1229                         ext4_da_release_space(inode, 1);
1230
1231                 num_clusters--;
1232         }
1233 }
1234
1235 /*
1236  * Delayed allocation stuff
1237  */
1238
1239 /*
1240  * mpage_da_submit_io - walks through extent of pages and try to write
1241  * them with writepage() call back
1242  *
1243  * @mpd->inode: inode
1244  * @mpd->first_page: first page of the extent
1245  * @mpd->next_page: page after the last page of the extent
1246  *
1247  * By the time mpage_da_submit_io() is called we expect all blocks
1248  * to be allocated. this may be wrong if allocation failed.
1249  *
1250  * As pages are already locked by write_cache_pages(), we can't use it
1251  */
1252 static int mpage_da_submit_io(struct mpage_da_data *mpd,
1253                               struct ext4_map_blocks *map)
1254 {
1255         struct pagevec pvec;
1256         unsigned long index, end;
1257         int ret = 0, err, nr_pages, i;
1258         struct inode *inode = mpd->inode;
1259         struct address_space *mapping = inode->i_mapping;
1260         loff_t size = i_size_read(inode);
1261         unsigned int len, block_start;
1262         struct buffer_head *bh, *page_bufs = NULL;
1263         int journal_data = ext4_should_journal_data(inode);
1264         sector_t pblock = 0, cur_logical = 0;
1265         struct ext4_io_submit io_submit;
1266
1267         BUG_ON(mpd->next_page <= mpd->first_page);
1268         memset(&io_submit, 0, sizeof(io_submit));
1269         /*
1270          * We need to start from the first_page to the next_page - 1
1271          * to make sure we also write the mapped dirty buffer_heads.
1272          * If we look at mpd->b_blocknr we would only be looking
1273          * at the currently mapped buffer_heads.
1274          */
1275         index = mpd->first_page;
1276         end = mpd->next_page - 1;
1277
1278         pagevec_init(&pvec, 0);
1279         while (index <= end) {
1280                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1281                 if (nr_pages == 0)
1282                         break;
1283                 for (i = 0; i < nr_pages; i++) {
1284                         int commit_write = 0, skip_page = 0;
1285                         struct page *page = pvec.pages[i];
1286
1287                         index = page->index;
1288                         if (index > end)
1289                                 break;
1290
1291                         if (index == size >> PAGE_CACHE_SHIFT)
1292                                 len = size & ~PAGE_CACHE_MASK;
1293                         else
1294                                 len = PAGE_CACHE_SIZE;
1295                         if (map) {
1296                                 cur_logical = index << (PAGE_CACHE_SHIFT -
1297                                                         inode->i_blkbits);
1298                                 pblock = map->m_pblk + (cur_logical -
1299                                                         map->m_lblk);
1300                         }
1301                         index++;
1302
1303                         BUG_ON(!PageLocked(page));
1304                         BUG_ON(PageWriteback(page));
1305
1306                         /*
1307                          * If the page does not have buffers (for
1308                          * whatever reason), try to create them using
1309                          * __block_write_begin.  If this fails,
1310                          * skip the page and move on.
1311                          */
1312                         if (!page_has_buffers(page)) {
1313                                 if (__block_write_begin(page, 0, len,
1314                                                 noalloc_get_block_write)) {
1315                                 skip_page:
1316                                         unlock_page(page);
1317                                         continue;
1318                                 }
1319                                 commit_write = 1;
1320                         }
1321
1322                         bh = page_bufs = page_buffers(page);
1323                         block_start = 0;
1324                         do {
1325                                 if (!bh)
1326                                         goto skip_page;
1327                                 if (map && (cur_logical >= map->m_lblk) &&
1328                                     (cur_logical <= (map->m_lblk +
1329                                                      (map->m_len - 1)))) {
1330                                         if (buffer_delay(bh)) {
1331                                                 clear_buffer_delay(bh);
1332                                                 bh->b_blocknr = pblock;
1333                                         }
1334                                         if (buffer_da_mapped(bh))
1335                                                 clear_buffer_da_mapped(bh);
1336                                         if (buffer_unwritten(bh) ||
1337                                             buffer_mapped(bh))
1338                                                 BUG_ON(bh->b_blocknr != pblock);
1339                                         if (map->m_flags & EXT4_MAP_UNINIT)
1340                                                 set_buffer_uninit(bh);
1341                                         clear_buffer_unwritten(bh);
1342                                 }
1343
1344                                 /*
1345                                  * skip page if block allocation undone and
1346                                  * block is dirty
1347                                  */
1348                                 if (ext4_bh_delay_or_unwritten(NULL, bh))
1349                                         skip_page = 1;
1350                                 bh = bh->b_this_page;
1351                                 block_start += bh->b_size;
1352                                 cur_logical++;
1353                                 pblock++;
1354                         } while (bh != page_bufs);
1355
1356                         if (skip_page)
1357                                 goto skip_page;
1358
1359                         if (commit_write)
1360                                 /* mark the buffer_heads as dirty & uptodate */
1361                                 block_commit_write(page, 0, len);
1362
1363                         clear_page_dirty_for_io(page);
1364                         /*
1365                          * Delalloc doesn't support data journalling,
1366                          * but eventually maybe we'll lift this
1367                          * restriction.
1368                          */
1369                         if (unlikely(journal_data && PageChecked(page)))
1370                                 err = __ext4_journalled_writepage(page, len);
1371                         else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
1372                                 err = ext4_bio_write_page(&io_submit, page,
1373                                                           len, mpd->wbc);
1374                         else if (buffer_uninit(page_bufs)) {
1375                                 ext4_set_bh_endio(page_bufs, inode);
1376                                 err = block_write_full_page_endio(page,
1377                                         noalloc_get_block_write,
1378                                         mpd->wbc, ext4_end_io_buffer_write);
1379                         } else
1380                                 err = block_write_full_page(page,
1381                                         noalloc_get_block_write, mpd->wbc);
1382
1383                         if (!err)
1384                                 mpd->pages_written++;
1385                         /*
1386                          * In error case, we have to continue because
1387                          * remaining pages are still locked
1388                          */
1389                         if (ret == 0)
1390                                 ret = err;
1391                 }
1392                 pagevec_release(&pvec);
1393         }
1394         ext4_io_submit(&io_submit);
1395         return ret;
1396 }
1397
1398 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd)
1399 {
1400         int nr_pages, i;
1401         pgoff_t index, end;
1402         struct pagevec pvec;
1403         struct inode *inode = mpd->inode;
1404         struct address_space *mapping = inode->i_mapping;
1405
1406         index = mpd->first_page;
1407         end   = mpd->next_page - 1;
1408         while (index <= end) {
1409                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1410                 if (nr_pages == 0)
1411                         break;
1412                 for (i = 0; i < nr_pages; i++) {
1413                         struct page *page = pvec.pages[i];
1414                         if (page->index > end)
1415                                 break;
1416                         BUG_ON(!PageLocked(page));
1417                         BUG_ON(PageWriteback(page));
1418                         block_invalidatepage(page, 0);
1419                         ClearPageUptodate(page);
1420                         unlock_page(page);
1421                 }
1422                 index = pvec.pages[nr_pages - 1]->index + 1;
1423                 pagevec_release(&pvec);
1424         }
1425         return;
1426 }
1427
1428 static void ext4_print_free_blocks(struct inode *inode)
1429 {
1430         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1431         struct super_block *sb = inode->i_sb;
1432
1433         ext4_msg(sb, KERN_CRIT, "Total free blocks count %lld",
1434                EXT4_C2B(EXT4_SB(inode->i_sb),
1435                         ext4_count_free_clusters(inode->i_sb)));
1436         ext4_msg(sb, KERN_CRIT, "Free/Dirty block details");
1437         ext4_msg(sb, KERN_CRIT, "free_blocks=%lld",
1438                (long long) EXT4_C2B(EXT4_SB(inode->i_sb),
1439                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1440         ext4_msg(sb, KERN_CRIT, "dirty_blocks=%lld",
1441                (long long) EXT4_C2B(EXT4_SB(inode->i_sb),
1442                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1443         ext4_msg(sb, KERN_CRIT, "Block reservation details");
1444         ext4_msg(sb, KERN_CRIT, "i_reserved_data_blocks=%u",
1445                  EXT4_I(inode)->i_reserved_data_blocks);
1446         ext4_msg(sb, KERN_CRIT, "i_reserved_meta_blocks=%u",
1447                EXT4_I(inode)->i_reserved_meta_blocks);
1448         return;
1449 }
1450
1451 /*
1452  * mpage_da_map_and_submit - go through given space, map them
1453  *       if necessary, and then submit them for I/O
1454  *
1455  * @mpd - bh describing space
1456  *
1457  * The function skips space we know is already mapped to disk blocks.
1458  *
1459  */
1460 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
1461 {
1462         int err, blks, get_blocks_flags;
1463         struct ext4_map_blocks map, *mapp = NULL;
1464         sector_t next = mpd->b_blocknr;
1465         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
1466         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
1467         handle_t *handle = NULL;
1468
1469         /*
1470          * If the blocks are mapped already, or we couldn't accumulate
1471          * any blocks, then proceed immediately to the submission stage.
1472          */
1473         if ((mpd->b_size == 0) ||
1474             ((mpd->b_state  & (1 << BH_Mapped)) &&
1475              !(mpd->b_state & (1 << BH_Delay)) &&
1476              !(mpd->b_state & (1 << BH_Unwritten))))
1477                 goto submit_io;
1478
1479         handle = ext4_journal_current_handle();
1480         BUG_ON(!handle);
1481
1482         /*
1483          * Call ext4_map_blocks() to allocate any delayed allocation
1484          * blocks, or to convert an uninitialized extent to be
1485          * initialized (in the case where we have written into
1486          * one or more preallocated blocks).
1487          *
1488          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
1489          * indicate that we are on the delayed allocation path.  This
1490          * affects functions in many different parts of the allocation
1491          * call path.  This flag exists primarily because we don't
1492          * want to change *many* call functions, so ext4_map_blocks()
1493          * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
1494          * inode's allocation semaphore is taken.
1495          *
1496          * If the blocks in questions were delalloc blocks, set
1497          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
1498          * variables are updated after the blocks have been allocated.
1499          */
1500         map.m_lblk = next;
1501         map.m_len = max_blocks;
1502         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
1503         if (ext4_should_dioread_nolock(mpd->inode))
1504                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
1505         if (mpd->b_state & (1 << BH_Delay))
1506                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
1507
1508         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
1509         if (blks < 0) {
1510                 struct super_block *sb = mpd->inode->i_sb;
1511
1512                 err = blks;
1513                 /*
1514                  * If get block returns EAGAIN or ENOSPC and there
1515                  * appears to be free blocks we will just let
1516                  * mpage_da_submit_io() unlock all of the pages.
1517                  */
1518                 if (err == -EAGAIN)
1519                         goto submit_io;
1520
1521                 if (err == -ENOSPC && ext4_count_free_clusters(sb)) {
1522                         mpd->retval = err;
1523                         goto submit_io;
1524                 }
1525
1526                 /*
1527                  * get block failure will cause us to loop in
1528                  * writepages, because a_ops->writepage won't be able
1529                  * to make progress. The page will be redirtied by
1530                  * writepage and writepages will again try to write
1531                  * the same.
1532                  */
1533                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
1534                         ext4_msg(sb, KERN_CRIT,
1535                                  "delayed block allocation failed for inode %lu "
1536                                  "at logical offset %llu with max blocks %zd "
1537                                  "with error %d", mpd->inode->i_ino,
1538                                  (unsigned long long) next,
1539                                  mpd->b_size >> mpd->inode->i_blkbits, err);
1540                         ext4_msg(sb, KERN_CRIT,
1541                                 "This should not happen!! Data will be lost\n");
1542                         if (err == -ENOSPC)
1543                                 ext4_print_free_blocks(mpd->inode);
1544                 }
1545                 /* invalidate all the pages */
1546                 ext4_da_block_invalidatepages(mpd);
1547
1548                 /* Mark this page range as having been completed */
1549                 mpd->io_done = 1;
1550                 return;
1551         }
1552         BUG_ON(blks == 0);
1553
1554         mapp = &map;
1555         if (map.m_flags & EXT4_MAP_NEW) {
1556                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
1557                 int i;
1558
1559                 for (i = 0; i < map.m_len; i++)
1560                         unmap_underlying_metadata(bdev, map.m_pblk + i);
1561
1562                 if (ext4_should_order_data(mpd->inode)) {
1563                         err = ext4_jbd2_file_inode(handle, mpd->inode);
1564                         if (err) {
1565                                 /* Only if the journal is aborted */
1566                                 mpd->retval = err;
1567                                 goto submit_io;
1568                         }
1569                 }
1570         }
1571
1572         /*
1573          * Update on-disk size along with block allocation.
1574          */
1575         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
1576         if (disksize > i_size_read(mpd->inode))
1577                 disksize = i_size_read(mpd->inode);
1578         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
1579                 ext4_update_i_disksize(mpd->inode, disksize);
1580                 err = ext4_mark_inode_dirty(handle, mpd->inode);
1581                 if (err)
1582                         ext4_error(mpd->inode->i_sb,
1583                                    "Failed to mark inode %lu dirty",
1584                                    mpd->inode->i_ino);
1585         }
1586
1587 submit_io:
1588         mpage_da_submit_io(mpd, mapp);
1589         mpd->io_done = 1;
1590 }
1591
1592 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1593                 (1 << BH_Delay) | (1 << BH_Unwritten))
1594
1595 /*
1596  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1597  *
1598  * @mpd->lbh - extent of blocks
1599  * @logical - logical number of the block in the file
1600  * @bh - bh of the block (used to access block's state)
1601  *
1602  * the function is used to collect contig. blocks in same state
1603  */
1604 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1605                                    sector_t logical, size_t b_size,
1606                                    unsigned long b_state)
1607 {
1608         sector_t next;
1609         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
1610
1611         /*
1612          * XXX Don't go larger than mballoc is willing to allocate
1613          * This is a stopgap solution.  We eventually need to fold
1614          * mpage_da_submit_io() into this function and then call
1615          * ext4_map_blocks() multiple times in a loop
1616          */
1617         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
1618                 goto flush_it;
1619
1620         /* check if thereserved journal credits might overflow */
1621         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
1622                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1623                         /*
1624                          * With non-extent format we are limited by the journal
1625                          * credit available.  Total credit needed to insert
1626                          * nrblocks contiguous blocks is dependent on the
1627                          * nrblocks.  So limit nrblocks.
1628                          */
1629                         goto flush_it;
1630                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1631                                 EXT4_MAX_TRANS_DATA) {
1632                         /*
1633                          * Adding the new buffer_head would make it cross the
1634                          * allowed limit for which we have journal credit
1635                          * reserved. So limit the new bh->b_size
1636                          */
1637                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1638                                                 mpd->inode->i_blkbits;
1639                         /* we will do mpage_da_submit_io in the next loop */
1640                 }
1641         }
1642         /*
1643          * First block in the extent
1644          */
1645         if (mpd->b_size == 0) {
1646                 mpd->b_blocknr = logical;
1647                 mpd->b_size = b_size;
1648                 mpd->b_state = b_state & BH_FLAGS;
1649                 return;
1650         }
1651
1652         next = mpd->b_blocknr + nrblocks;
1653         /*
1654          * Can we merge the block to our big extent?
1655          */
1656         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
1657                 mpd->b_size += b_size;
1658                 return;
1659         }
1660
1661 flush_it:
1662         /*
1663          * We couldn't merge the block to our extent, so we
1664          * need to flush current  extent and start new one
1665          */
1666         mpage_da_map_and_submit(mpd);
1667         return;
1668 }
1669
1670 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1671 {
1672         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1673 }
1674
1675 /*
1676  * This function is grabs code from the very beginning of
1677  * ext4_map_blocks, but assumes that the caller is from delayed write
1678  * time. This function looks up the requested blocks and sets the
1679  * buffer delay bit under the protection of i_data_sem.
1680  */
1681 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1682                               struct ext4_map_blocks *map,
1683                               struct buffer_head *bh)
1684 {
1685         int retval;
1686         sector_t invalid_block = ~((sector_t) 0xffff);
1687
1688         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1689                 invalid_block = ~0;
1690
1691         map->m_flags = 0;
1692         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1693                   "logical block %lu\n", inode->i_ino, map->m_len,
1694                   (unsigned long) map->m_lblk);
1695         /*
1696          * Try to see if we can get the block without requesting a new
1697          * file system block.
1698          */
1699         down_read((&EXT4_I(inode)->i_data_sem));
1700         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1701                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1702         else
1703                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1704
1705         if (retval == 0) {
1706                 /*
1707                  * XXX: __block_prepare_write() unmaps passed block,
1708                  * is it OK?
1709                  */
1710                 /* If the block was allocated from previously allocated cluster,
1711                  * then we dont need to reserve it again. */
1712                 if (!(map->m_flags & EXT4_MAP_FROM_CLUSTER)) {
1713                         retval = ext4_da_reserve_space(inode, iblock);
1714                         if (retval)
1715                                 /* not enough space to reserve */
1716                                 goto out_unlock;
1717                 }
1718
1719                 /* Clear EXT4_MAP_FROM_CLUSTER flag since its purpose is served
1720                  * and it should not appear on the bh->b_state.
1721                  */
1722                 map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
1723
1724                 map_bh(bh, inode->i_sb, invalid_block);
1725                 set_buffer_new(bh);
1726                 set_buffer_delay(bh);
1727         }
1728
1729 out_unlock:
1730         up_read((&EXT4_I(inode)->i_data_sem));
1731
1732         return retval;
1733 }
1734
1735 /*
1736  * This is a special get_blocks_t callback which is used by
1737  * ext4_da_write_begin().  It will either return mapped block or
1738  * reserve space for a single block.
1739  *
1740  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1741  * We also have b_blocknr = -1 and b_bdev initialized properly
1742  *
1743  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1744  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1745  * initialized properly.
1746  */
1747 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1748                                   struct buffer_head *bh, int create)
1749 {
1750         struct ext4_map_blocks map;
1751         int ret = 0;
1752
1753         BUG_ON(create == 0);
1754         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1755
1756         map.m_lblk = iblock;
1757         map.m_len = 1;
1758
1759         /*
1760          * first, we need to know whether the block is allocated already
1761          * preallocated blocks are unmapped but should treated
1762          * the same as allocated blocks.
1763          */
1764         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1765         if (ret <= 0)
1766                 return ret;
1767
1768         map_bh(bh, inode->i_sb, map.m_pblk);
1769         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1770
1771         if (buffer_unwritten(bh)) {
1772                 /* A delayed write to unwritten bh should be marked
1773                  * new and mapped.  Mapped ensures that we don't do
1774                  * get_block multiple times when we write to the same
1775                  * offset and new ensures that we do proper zero out
1776                  * for partial write.
1777                  */
1778                 set_buffer_new(bh);
1779                 set_buffer_mapped(bh);
1780         }
1781         return 0;
1782 }
1783
1784 /*
1785  * This function is used as a standard get_block_t calback function
1786  * when there is no desire to allocate any blocks.  It is used as a
1787  * callback function for block_write_begin() and block_write_full_page().
1788  * These functions should only try to map a single block at a time.
1789  *
1790  * Since this function doesn't do block allocations even if the caller
1791  * requests it by passing in create=1, it is critically important that
1792  * any caller checks to make sure that any buffer heads are returned
1793  * by this function are either all already mapped or marked for
1794  * delayed allocation before calling  block_write_full_page().  Otherwise,
1795  * b_blocknr could be left unitialized, and the page write functions will
1796  * be taken by surprise.
1797  */
1798 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
1799                                    struct buffer_head *bh_result, int create)
1800 {
1801         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
1802         return _ext4_get_block(inode, iblock, bh_result, 0);
1803 }
1804
1805 static int bget_one(handle_t *handle, struct buffer_head *bh)
1806 {
1807         get_bh(bh);
1808         return 0;
1809 }
1810
1811 static int bput_one(handle_t *handle, struct buffer_head *bh)
1812 {
1813         put_bh(bh);
1814         return 0;
1815 }
1816
1817 static int __ext4_journalled_writepage(struct page *page,
1818                                        unsigned int len)
1819 {
1820         struct address_space *mapping = page->mapping;
1821         struct inode *inode = mapping->host;
1822         struct buffer_head *page_bufs;
1823         handle_t *handle = NULL;
1824         int ret = 0;
1825         int err;
1826
1827         ClearPageChecked(page);
1828         page_bufs = page_buffers(page);
1829         BUG_ON(!page_bufs);
1830         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
1831         /* As soon as we unlock the page, it can go away, but we have
1832          * references to buffers so we are safe */
1833         unlock_page(page);
1834
1835         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1836         if (IS_ERR(handle)) {
1837                 ret = PTR_ERR(handle);
1838                 goto out;
1839         }
1840
1841         BUG_ON(!ext4_handle_valid(handle));
1842
1843         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1844                                 do_journal_get_write_access);
1845
1846         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1847                                 write_end_fn);
1848         if (ret == 0)
1849                 ret = err;
1850         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1851         err = ext4_journal_stop(handle);
1852         if (!ret)
1853                 ret = err;
1854
1855         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
1856         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1857 out:
1858         return ret;
1859 }
1860
1861 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
1862 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
1863
1864 /*
1865  * Note that we don't need to start a transaction unless we're journaling data
1866  * because we should have holes filled from ext4_page_mkwrite(). We even don't
1867  * need to file the inode to the transaction's list in ordered mode because if
1868  * we are writing back data added by write(), the inode is already there and if
1869  * we are writing back data modified via mmap(), no one guarantees in which
1870  * transaction the data will hit the disk. In case we are journaling data, we
1871  * cannot start transaction directly because transaction start ranks above page
1872  * lock so we have to do some magic.
1873  *
1874  * This function can get called via...
1875  *   - ext4_da_writepages after taking page lock (have journal handle)
1876  *   - journal_submit_inode_data_buffers (no journal handle)
1877  *   - shrink_page_list via pdflush (no journal handle)
1878  *   - grab_page_cache when doing write_begin (have journal handle)
1879  *
1880  * We don't do any block allocation in this function. If we have page with
1881  * multiple blocks we need to write those buffer_heads that are mapped. This
1882  * is important for mmaped based write. So if we do with blocksize 1K
1883  * truncate(f, 1024);
1884  * a = mmap(f, 0, 4096);
1885  * a[0] = 'a';
1886  * truncate(f, 4096);
1887  * we have in the page first buffer_head mapped via page_mkwrite call back
1888  * but other buffer_heads would be unmapped but dirty (dirty done via the
1889  * do_wp_page). So writepage should write the first block. If we modify
1890  * the mmap area beyond 1024 we will again get a page_fault and the
1891  * page_mkwrite callback will do the block allocation and mark the
1892  * buffer_heads mapped.
1893  *
1894  * We redirty the page if we have any buffer_heads that is either delay or
1895  * unwritten in the page.
1896  *
1897  * We can get recursively called as show below.
1898  *
1899  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1900  *              ext4_writepage()
1901  *
1902  * But since we don't do any block allocation we should not deadlock.
1903  * Page also have the dirty flag cleared so we don't get recurive page_lock.
1904  */
1905 static int ext4_writepage(struct page *page,
1906                           struct writeback_control *wbc)
1907 {
1908         int ret = 0, commit_write = 0;
1909         loff_t size;
1910         unsigned int len;
1911         struct buffer_head *page_bufs = NULL;
1912         struct inode *inode = page->mapping->host;
1913
1914         trace_ext4_writepage(page);
1915         size = i_size_read(inode);
1916         if (page->index == size >> PAGE_CACHE_SHIFT)
1917                 len = size & ~PAGE_CACHE_MASK;
1918         else
1919                 len = PAGE_CACHE_SIZE;
1920
1921         /*
1922          * If the page does not have buffers (for whatever reason),
1923          * try to create them using __block_write_begin.  If this
1924          * fails, redirty the page and move on.
1925          */
1926         if (!page_has_buffers(page)) {
1927                 if (__block_write_begin(page, 0, len,
1928                                         noalloc_get_block_write)) {
1929                 redirty_page:
1930                         redirty_page_for_writepage(wbc, page);
1931                         unlock_page(page);
1932                         return 0;
1933                 }
1934                 commit_write = 1;
1935         }
1936         page_bufs = page_buffers(page);
1937         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
1938                               ext4_bh_delay_or_unwritten)) {
1939                 /*
1940                  * We don't want to do block allocation, so redirty
1941                  * the page and return.  We may reach here when we do
1942                  * a journal commit via journal_submit_inode_data_buffers.
1943                  * We can also reach here via shrink_page_list but it
1944                  * should never be for direct reclaim so warn if that
1945                  * happens
1946                  */
1947                 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
1948                                                                 PF_MEMALLOC);
1949                 goto redirty_page;
1950         }
1951         if (commit_write)
1952                 /* now mark the buffer_heads as dirty and uptodate */
1953                 block_commit_write(page, 0, len);
1954
1955         if (PageChecked(page) && ext4_should_journal_data(inode))
1956                 /*
1957                  * It's mmapped pagecache.  Add buffers and journal it.  There
1958                  * doesn't seem much point in redirtying the page here.
1959                  */
1960                 return __ext4_journalled_writepage(page, len);
1961
1962         if (buffer_uninit(page_bufs)) {
1963                 ext4_set_bh_endio(page_bufs, inode);
1964                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
1965                                             wbc, ext4_end_io_buffer_write);
1966         } else
1967                 ret = block_write_full_page(page, noalloc_get_block_write,
1968                                             wbc);
1969
1970         return ret;
1971 }
1972
1973 /*
1974  * This is called via ext4_da_writepages() to
1975  * calculate the total number of credits to reserve to fit
1976  * a single extent allocation into a single transaction,
1977  * ext4_da_writpeages() will loop calling this before
1978  * the block allocation.
1979  */
1980
1981 static int ext4_da_writepages_trans_blocks(struct inode *inode)
1982 {
1983         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
1984
1985         /*
1986          * With non-extent format the journal credit needed to
1987          * insert nrblocks contiguous block is dependent on
1988          * number of contiguous block. So we will limit
1989          * number of contiguous block to a sane value
1990          */
1991         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
1992             (max_blocks > EXT4_MAX_TRANS_DATA))
1993                 max_blocks = EXT4_MAX_TRANS_DATA;
1994
1995         return ext4_chunk_trans_blocks(inode, max_blocks);
1996 }
1997
1998 /*
1999  * write_cache_pages_da - walk the list of dirty pages of the given
2000  * address space and accumulate pages that need writing, and call
2001  * mpage_da_map_and_submit to map a single contiguous memory region
2002  * and then write them.
2003  */
2004 static int write_cache_pages_da(struct address_space *mapping,
2005                                 struct writeback_control *wbc,
2006                                 struct mpage_da_data *mpd,
2007                                 pgoff_t *done_index)
2008 {
2009         struct buffer_head      *bh, *head;
2010         struct inode            *inode = mapping->host;
2011         struct pagevec          pvec;
2012         unsigned int            nr_pages;
2013         sector_t                logical;
2014         pgoff_t                 index, end;
2015         long                    nr_to_write = wbc->nr_to_write;
2016         int                     i, tag, ret = 0;
2017
2018         memset(mpd, 0, sizeof(struct mpage_da_data));
2019         mpd->wbc = wbc;
2020         mpd->inode = inode;
2021         pagevec_init(&pvec, 0);
2022         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2023         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2024
2025         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2026                 tag = PAGECACHE_TAG_TOWRITE;
2027         else
2028                 tag = PAGECACHE_TAG_DIRTY;
2029
2030         *done_index = index;
2031         while (index <= end) {
2032                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2033                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2034                 if (nr_pages == 0)
2035                         return 0;
2036
2037                 for (i = 0; i < nr_pages; i++) {
2038                         struct page *page = pvec.pages[i];
2039
2040                         /*
2041                          * At this point, the page may be truncated or
2042                          * invalidated (changing page->mapping to NULL), or
2043                          * even swizzled back from swapper_space to tmpfs file
2044                          * mapping. However, page->index will not change
2045                          * because we have a reference on the page.
2046                          */
2047                         if (page->index > end)
2048                                 goto out;
2049
2050                         *done_index = page->index + 1;
2051
2052                         /*
2053                          * If we can't merge this page, and we have
2054                          * accumulated an contiguous region, write it
2055                          */
2056                         if ((mpd->next_page != page->index) &&
2057                             (mpd->next_page != mpd->first_page)) {
2058                                 mpage_da_map_and_submit(mpd);
2059                                 goto ret_extent_tail;
2060                         }
2061
2062                         lock_page(page);
2063
2064                         /*
2065                          * If the page is no longer dirty, or its
2066                          * mapping no longer corresponds to inode we
2067                          * are writing (which means it has been
2068                          * truncated or invalidated), or the page is
2069                          * already under writeback and we are not
2070                          * doing a data integrity writeback, skip the page
2071                          */
2072                         if (!PageDirty(page) ||
2073                             (PageWriteback(page) &&
2074                              (wbc->sync_mode == WB_SYNC_NONE)) ||
2075                             unlikely(page->mapping != mapping)) {
2076                                 unlock_page(page);
2077                                 continue;
2078                         }
2079
2080                         wait_on_page_writeback(page);
2081                         BUG_ON(PageWriteback(page));
2082
2083                         if (mpd->next_page != page->index)
2084                                 mpd->first_page = page->index;
2085                         mpd->next_page = page->index + 1;
2086                         logical = (sector_t) page->index <<
2087                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2088
2089                         if (!page_has_buffers(page)) {
2090                                 mpage_add_bh_to_extent(mpd, logical,
2091                                                        PAGE_CACHE_SIZE,
2092                                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2093                                 if (mpd->io_done)
2094                                         goto ret_extent_tail;
2095                         } else {
2096                                 /*
2097                                  * Page with regular buffer heads,
2098                                  * just add all dirty ones
2099                                  */
2100                                 head = page_buffers(page);
2101                                 bh = head;
2102                                 do {
2103                                         BUG_ON(buffer_locked(bh));
2104                                         /*
2105                                          * We need to try to allocate
2106                                          * unmapped blocks in the same page.
2107                                          * Otherwise we won't make progress
2108                                          * with the page in ext4_writepage
2109                                          */
2110                                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2111                                                 mpage_add_bh_to_extent(mpd, logical,
2112                                                                        bh->b_size,
2113                                                                        bh->b_state);
2114                                                 if (mpd->io_done)
2115                                                         goto ret_extent_tail;
2116                                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2117                                                 /*
2118                                                  * mapped dirty buffer. We need
2119                                                  * to update the b_state
2120                                                  * because we look at b_state
2121                                                  * in mpage_da_map_blocks.  We
2122                                                  * don't update b_size because
2123                                                  * if we find an unmapped
2124                                                  * buffer_head later we need to
2125                                                  * use the b_state flag of that
2126                                                  * buffer_head.
2127                                                  */
2128                                                 if (mpd->b_size == 0)
2129                                                         mpd->b_state = bh->b_state & BH_FLAGS;
2130                                         }
2131                                         logical++;
2132                                 } while ((bh = bh->b_this_page) != head);
2133                         }
2134
2135                         if (nr_to_write > 0) {
2136                                 nr_to_write--;
2137                                 if (nr_to_write == 0 &&
2138                                     wbc->sync_mode == WB_SYNC_NONE)
2139                                         /*
2140                                          * We stop writing back only if we are
2141                                          * not doing integrity sync. In case of
2142                                          * integrity sync we have to keep going
2143                                          * because someone may be concurrently
2144                                          * dirtying pages, and we might have
2145                                          * synced a lot of newly appeared dirty
2146                                          * pages, but have not synced all of the
2147                                          * old dirty pages.
2148                                          */
2149                                         goto out;
2150                         }
2151                 }
2152                 pagevec_release(&pvec);
2153                 cond_resched();
2154         }
2155         return 0;
2156 ret_extent_tail:
2157         ret = MPAGE_DA_EXTENT_TAIL;
2158 out:
2159         pagevec_release(&pvec);
2160         cond_resched();
2161         return ret;
2162 }
2163
2164
2165 static int ext4_da_writepages(struct address_space *mapping,
2166                               struct writeback_control *wbc)
2167 {
2168         pgoff_t index;
2169         int range_whole = 0;
2170         handle_t *handle = NULL;
2171         struct mpage_da_data mpd;
2172         struct inode *inode = mapping->host;
2173         int pages_written = 0;
2174         unsigned int max_pages;
2175         int range_cyclic, cycled = 1, io_done = 0;
2176         int needed_blocks, ret = 0;
2177         long desired_nr_to_write, nr_to_writebump = 0;
2178         loff_t range_start = wbc->range_start;
2179         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2180         pgoff_t done_index = 0;
2181         pgoff_t end;
2182         struct blk_plug plug;
2183
2184         trace_ext4_da_writepages(inode, wbc);
2185
2186         /*
2187          * No pages to write? This is mainly a kludge to avoid starting
2188          * a transaction for special inodes like journal inode on last iput()
2189          * because that could violate lock ordering on umount
2190          */
2191         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2192                 return 0;
2193
2194         /*
2195          * If the filesystem has aborted, it is read-only, so return
2196          * right away instead of dumping stack traces later on that
2197          * will obscure the real source of the problem.  We test
2198          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2199          * the latter could be true if the filesystem is mounted
2200          * read-only, and in that case, ext4_da_writepages should
2201          * *never* be called, so if that ever happens, we would want
2202          * the stack trace.
2203          */
2204         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2205                 return -EROFS;
2206
2207         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2208                 range_whole = 1;
2209
2210         range_cyclic = wbc->range_cyclic;
2211         if (wbc->range_cyclic) {
2212                 index = mapping->writeback_index;
2213                 if (index)
2214                         cycled = 0;
2215                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2216                 wbc->range_end  = LLONG_MAX;
2217                 wbc->range_cyclic = 0;
2218                 end = -1;
2219         } else {
2220                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2221                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2222         }
2223
2224         /*
2225          * This works around two forms of stupidity.  The first is in
2226          * the writeback code, which caps the maximum number of pages
2227          * written to be 1024 pages.  This is wrong on multiple
2228          * levels; different architectues have a different page size,
2229          * which changes the maximum amount of data which gets
2230          * written.  Secondly, 4 megabytes is way too small.  XFS
2231          * forces this value to be 16 megabytes by multiplying
2232          * nr_to_write parameter by four, and then relies on its
2233          * allocator to allocate larger extents to make them
2234          * contiguous.  Unfortunately this brings us to the second
2235          * stupidity, which is that ext4's mballoc code only allocates
2236          * at most 2048 blocks.  So we force contiguous writes up to
2237          * the number of dirty blocks in the inode, or
2238          * sbi->max_writeback_mb_bump whichever is smaller.
2239          */
2240         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2241         if (!range_cyclic && range_whole) {
2242                 if (wbc->nr_to_write == LONG_MAX)
2243                         desired_nr_to_write = wbc->nr_to_write;
2244                 else
2245                         desired_nr_to_write = wbc->nr_to_write * 8;
2246         } else
2247                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2248                                                            max_pages);
2249         if (desired_nr_to_write > max_pages)
2250                 desired_nr_to_write = max_pages;
2251
2252         if (wbc->nr_to_write < desired_nr_to_write) {
2253                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2254                 wbc->nr_to_write = desired_nr_to_write;
2255         }
2256
2257 retry:
2258         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2259                 tag_pages_for_writeback(mapping, index, end);
2260
2261         blk_start_plug(&plug);
2262         while (!ret && wbc->nr_to_write > 0) {
2263
2264                 /*
2265                  * we  insert one extent at a time. So we need
2266                  * credit needed for single extent allocation.
2267                  * journalled mode is currently not supported
2268                  * by delalloc
2269                  */
2270                 BUG_ON(ext4_should_journal_data(inode));
2271                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2272
2273                 /* start a new transaction*/
2274                 handle = ext4_journal_start(inode, needed_blocks);
2275                 if (IS_ERR(handle)) {
2276                         ret = PTR_ERR(handle);
2277                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2278                                "%ld pages, ino %lu; err %d", __func__,
2279                                 wbc->nr_to_write, inode->i_ino, ret);
2280                         blk_finish_plug(&plug);
2281                         goto out_writepages;
2282                 }
2283
2284                 /*
2285                  * Now call write_cache_pages_da() to find the next
2286                  * contiguous region of logical blocks that need
2287                  * blocks to be allocated by ext4 and submit them.
2288                  */
2289                 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
2290                 /*
2291                  * If we have a contiguous extent of pages and we
2292                  * haven't done the I/O yet, map the blocks and submit
2293                  * them for I/O.
2294                  */
2295                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2296                         mpage_da_map_and_submit(&mpd);
2297                         ret = MPAGE_DA_EXTENT_TAIL;
2298                 }
2299                 trace_ext4_da_write_pages(inode, &mpd);
2300                 wbc->nr_to_write -= mpd.pages_written;
2301
2302                 ext4_journal_stop(handle);
2303
2304                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2305                         /* commit the transaction which would
2306                          * free blocks released in the transaction
2307                          * and try again
2308                          */
2309                         jbd2_journal_force_commit_nested(sbi->s_journal);
2310                         ret = 0;
2311                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2312                         /*
2313                          * Got one extent now try with rest of the pages.
2314                          * If mpd.retval is set -EIO, journal is aborted.
2315                          * So we don't need to write any more.
2316                          */
2317                         pages_written += mpd.pages_written;
2318                         ret = mpd.retval;
2319                         io_done = 1;
2320                 } else if (wbc->nr_to_write)
2321                         /*
2322                          * There is no more writeout needed
2323                          * or we requested for a noblocking writeout
2324                          * and we found the device congested
2325                          */
2326                         break;
2327         }
2328         blk_finish_plug(&plug);
2329         if (!io_done && !cycled) {
2330                 cycled = 1;
2331                 index = 0;
2332                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2333                 wbc->range_end  = mapping->writeback_index - 1;
2334                 goto retry;
2335         }
2336
2337         /* Update index */
2338         wbc->range_cyclic = range_cyclic;
2339         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2340                 /*
2341                  * set the writeback_index so that range_cyclic
2342                  * mode will write it back later
2343                  */
2344                 mapping->writeback_index = done_index;
2345
2346 out_writepages:
2347         wbc->nr_to_write -= nr_to_writebump;
2348         wbc->range_start = range_start;
2349         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2350         return ret;
2351 }
2352
2353 #define FALL_BACK_TO_NONDELALLOC 1
2354 static int ext4_nonda_switch(struct super_block *sb)
2355 {
2356         s64 free_blocks, dirty_blocks;
2357         struct ext4_sb_info *sbi = EXT4_SB(sb);
2358
2359         /*
2360          * switch to non delalloc mode if we are running low
2361          * on free block. The free block accounting via percpu
2362          * counters can get slightly wrong with percpu_counter_batch getting
2363          * accumulated on each CPU without updating global counters
2364          * Delalloc need an accurate free block accounting. So switch
2365          * to non delalloc when we are near to error range.
2366          */
2367         free_blocks  = EXT4_C2B(sbi,
2368                 percpu_counter_read_positive(&sbi->s_freeclusters_counter));
2369         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2370         if (2 * free_blocks < 3 * dirty_blocks ||
2371                 free_blocks < (dirty_blocks + EXT4_FREECLUSTERS_WATERMARK)) {
2372                 /*
2373                  * free block count is less than 150% of dirty blocks
2374                  * or free blocks is less than watermark
2375                  */
2376                 return 1;
2377         }
2378         /*
2379          * Even if we don't switch but are nearing capacity,
2380          * start pushing delalloc when 1/2 of free blocks are dirty.
2381          */
2382         if (free_blocks < 2 * dirty_blocks)
2383                 writeback_inodes_sb_if_idle(sb, WB_REASON_FS_FREE_SPACE);
2384
2385         return 0;
2386 }
2387
2388 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2389                                loff_t pos, unsigned len, unsigned flags,
2390                                struct page **pagep, void **fsdata)
2391 {
2392         int ret, retries = 0;
2393         struct page *page;
2394         pgoff_t index;
2395         struct inode *inode = mapping->host;
2396         handle_t *handle;
2397
2398         index = pos >> PAGE_CACHE_SHIFT;
2399
2400         if (ext4_nonda_switch(inode->i_sb)) {
2401                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2402                 return ext4_write_begin(file, mapping, pos,
2403                                         len, flags, pagep, fsdata);
2404         }
2405         *fsdata = (void *)0;
2406         trace_ext4_da_write_begin(inode, pos, len, flags);
2407 retry:
2408         /*
2409          * With delayed allocation, we don't log the i_disksize update
2410          * if there is delayed block allocation. But we still need
2411          * to journalling the i_disksize update if writes to the end
2412          * of file which has an already mapped buffer.
2413          */
2414         handle = ext4_journal_start(inode, 1);
2415         if (IS_ERR(handle)) {
2416                 ret = PTR_ERR(handle);
2417                 goto out;
2418         }
2419         /* We cannot recurse into the filesystem as the transaction is already
2420          * started */
2421         flags |= AOP_FLAG_NOFS;
2422
2423         page = grab_cache_page_write_begin(mapping, index, flags);
2424         if (!page) {
2425                 ext4_journal_stop(handle);
2426                 ret = -ENOMEM;
2427                 goto out;
2428         }
2429         *pagep = page;
2430
2431         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2432         if (ret < 0) {
2433                 unlock_page(page);
2434                 ext4_journal_stop(handle);
2435                 page_cache_release(page);
2436                 /*
2437                  * block_write_begin may have instantiated a few blocks
2438                  * outside i_size.  Trim these off again. Don't need
2439                  * i_size_read because we hold i_mutex.
2440                  */
2441                 if (pos + len > inode->i_size)
2442                         ext4_truncate_failed_write(inode);
2443         }
2444
2445         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2446                 goto retry;
2447 out:
2448         return ret;
2449 }
2450
2451 /*
2452  * Check if we should update i_disksize
2453  * when write to the end of file but not require block allocation
2454  */
2455 static int ext4_da_should_update_i_disksize(struct page *page,
2456                                             unsigned long offset)
2457 {
2458         struct buffer_head *bh;
2459         struct inode *inode = page->mapping->host;
2460         unsigned int idx;
2461         int i;
2462
2463         bh = page_buffers(page);
2464         idx = offset >> inode->i_blkbits;
2465
2466         for (i = 0; i < idx; i++)
2467                 bh = bh->b_this_page;
2468
2469         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2470                 return 0;
2471         return 1;
2472 }
2473
2474 static int ext4_da_write_end(struct file *file,
2475                              struct address_space *mapping,
2476                              loff_t pos, unsigned len, unsigned copied,
2477                              struct page *page, void *fsdata)
2478 {
2479         struct inode *inode = mapping->host;
2480         int ret = 0, ret2;
2481         handle_t *handle = ext4_journal_current_handle();
2482         loff_t new_i_size;
2483         unsigned long start, end;
2484         int write_mode = (int)(unsigned long)fsdata;
2485
2486         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2487                 switch (ext4_inode_journal_mode(inode)) {
2488                 case EXT4_INODE_ORDERED_DATA_MODE:
2489                         return ext4_ordered_write_end(file, mapping, pos,
2490                                         len, copied, page, fsdata);
2491                 case EXT4_INODE_WRITEBACK_DATA_MODE:
2492                         return ext4_writeback_write_end(file, mapping, pos,
2493                                         len, copied, page, fsdata);
2494                 default:
2495                         BUG();
2496                 }
2497         }
2498
2499         trace_ext4_da_write_end(inode, pos, len, copied);
2500         start = pos & (PAGE_CACHE_SIZE - 1);
2501         end = start + copied - 1;
2502
2503         /*
2504          * generic_write_end() will run mark_inode_dirty() if i_size
2505          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2506          * into that.
2507          */
2508
2509         new_i_size = pos + copied;
2510         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
2511                 if (ext4_da_should_update_i_disksize(page, end)) {
2512                         down_write(&EXT4_I(inode)->i_data_sem);
2513                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2514                                 /*
2515                                  * Updating i_disksize when extending file
2516                                  * without needing block allocation
2517                                  */
2518                                 if (ext4_should_order_data(inode))
2519                                         ret = ext4_jbd2_file_inode(handle,
2520                                                                    inode);
2521
2522                                 EXT4_I(inode)->i_disksize = new_i_size;
2523                         }
2524                         up_write(&EXT4_I(inode)->i_data_sem);
2525                         /* We need to mark inode dirty even if
2526                          * new_i_size is less that inode->i_size
2527                          * bu greater than i_disksize.(hint delalloc)
2528                          */
2529                         ext4_mark_inode_dirty(handle, inode);
2530                 }
2531         }
2532         ret2 = generic_write_end(file, mapping, pos, len, copied,
2533                                                         page, fsdata);
2534         copied = ret2;
2535         if (ret2 < 0)
2536                 ret = ret2;
2537         ret2 = ext4_journal_stop(handle);
2538         if (!ret)
2539                 ret = ret2;
2540
2541         return ret ? ret : copied;
2542 }
2543
2544 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2545 {
2546         /*
2547          * Drop reserved blocks
2548          */
2549         BUG_ON(!PageLocked(page));
2550         if (!page_has_buffers(page))
2551                 goto out;
2552
2553         ext4_da_page_release_reservation(page, offset);
2554
2555 out:
2556         ext4_invalidatepage(page, offset);
2557
2558         return;
2559 }
2560
2561 /*
2562  * Force all delayed allocation blocks to be allocated for a given inode.
2563  */
2564 int ext4_alloc_da_blocks(struct inode *inode)
2565 {
2566         trace_ext4_alloc_da_blocks(inode);
2567
2568         if (!EXT4_I(inode)->i_reserved_data_blocks &&
2569             !EXT4_I(inode)->i_reserved_meta_blocks)
2570                 return 0;
2571
2572         /*
2573          * We do something simple for now.  The filemap_flush() will
2574          * also start triggering a write of the data blocks, which is
2575          * not strictly speaking necessary (and for users of
2576          * laptop_mode, not even desirable).  However, to do otherwise
2577          * would require replicating code paths in:
2578          *
2579          * ext4_da_writepages() ->
2580          *    write_cache_pages() ---> (via passed in callback function)
2581          *        __mpage_da_writepage() -->
2582          *           mpage_add_bh_to_extent()
2583          *           mpage_da_map_blocks()
2584          *
2585          * The problem is that write_cache_pages(), located in
2586          * mm/page-writeback.c, marks pages clean in preparation for
2587          * doing I/O, which is not desirable if we're not planning on
2588          * doing I/O at all.
2589          *
2590          * We could call write_cache_pages(), and then redirty all of
2591          * the pages by calling redirty_page_for_writepage() but that
2592          * would be ugly in the extreme.  So instead we would need to
2593          * replicate parts of the code in the above functions,
2594          * simplifying them because we wouldn't actually intend to
2595          * write out the pages, but rather only collect contiguous
2596          * logical block extents, call the multi-block allocator, and
2597          * then update the buffer heads with the block allocations.
2598          *
2599          * For now, though, we'll cheat by calling filemap_flush(),
2600          * which will map the blocks, and start the I/O, but not
2601          * actually wait for the I/O to complete.
2602          */
2603         return filemap_flush(inode->i_mapping);
2604 }
2605
2606 /*
2607  * bmap() is special.  It gets used by applications such as lilo and by
2608  * the swapper to find the on-disk block of a specific piece of data.
2609  *
2610  * Naturally, this is dangerous if the block concerned is still in the
2611  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2612  * filesystem and enables swap, then they may get a nasty shock when the
2613  * data getting swapped to that swapfile suddenly gets overwritten by
2614  * the original zero's written out previously to the journal and
2615  * awaiting writeback in the kernel's buffer cache.
2616  *
2617  * So, if we see any bmap calls here on a modified, data-journaled file,
2618  * take extra steps to flush any blocks which might be in the cache.
2619  */
2620 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2621 {
2622         struct inode *inode = mapping->host;
2623         journal_t *journal;
2624         int err;
2625
2626         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2627                         test_opt(inode->i_sb, DELALLOC)) {
2628                 /*
2629                  * With delalloc we want to sync the file
2630                  * so that we can make sure we allocate
2631                  * blocks for file
2632                  */
2633                 filemap_write_and_wait(mapping);
2634         }
2635
2636         if (EXT4_JOURNAL(inode) &&
2637             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2638                 /*
2639                  * This is a REALLY heavyweight approach, but the use of
2640                  * bmap on dirty files is expected to be extremely rare:
2641                  * only if we run lilo or swapon on a freshly made file
2642                  * do we expect this to happen.
2643                  *
2644                  * (bmap requires CAP_SYS_RAWIO so this does not
2645                  * represent an unprivileged user DOS attack --- we'd be
2646                  * in trouble if mortal users could trigger this path at
2647                  * will.)
2648                  *
2649                  * NB. EXT4_STATE_JDATA is not set on files other than
2650                  * regular files.  If somebody wants to bmap a directory
2651                  * or symlink and gets confused because the buffer
2652                  * hasn't yet been flushed to disk, they deserve
2653                  * everything they get.
2654                  */
2655
2656                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2657                 journal = EXT4_JOURNAL(inode);
2658                 jbd2_journal_lock_updates(journal);
2659                 err = jbd2_journal_flush(journal);
2660                 jbd2_journal_unlock_updates(journal);
2661
2662                 if (err)
2663                         return 0;
2664         }
2665
2666         return generic_block_bmap(mapping, block, ext4_get_block);
2667 }
2668
2669 static int ext4_readpage(struct file *file, struct page *page)
2670 {
2671         trace_ext4_readpage(page);
2672         return mpage_readpage(page, ext4_get_block);
2673 }
2674
2675 static int
2676 ext4_readpages(struct file *file, struct address_space *mapping,
2677                 struct list_head *pages, unsigned nr_pages)
2678 {
2679         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2680 }
2681
2682 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
2683 {
2684         struct buffer_head *head, *bh;
2685         unsigned int curr_off = 0;
2686
2687         if (!page_has_buffers(page))
2688                 return;
2689         head = bh = page_buffers(page);
2690         do {
2691                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
2692                                         && bh->b_private) {
2693                         ext4_free_io_end(bh->b_private);
2694                         bh->b_private = NULL;
2695                         bh->b_end_io = NULL;
2696                 }
2697                 curr_off = curr_off + bh->b_size;
2698                 bh = bh->b_this_page;
2699         } while (bh != head);
2700 }
2701
2702 static void ext4_invalidatepage(struct page *page, unsigned long offset)
2703 {
2704         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2705
2706         trace_ext4_invalidatepage(page, offset);
2707
2708         /*
2709          * free any io_end structure allocated for buffers to be discarded
2710          */
2711         if (ext4_should_dioread_nolock(page->mapping->host))
2712                 ext4_invalidatepage_free_endio(page, offset);
2713         /*
2714          * If it's a full truncate we just forget about the pending dirtying
2715          */
2716         if (offset == 0)
2717                 ClearPageChecked(page);
2718
2719         if (journal)
2720                 jbd2_journal_invalidatepage(journal, page, offset);
2721         else
2722                 block_invalidatepage(page, offset);
2723 }
2724
2725 static int ext4_releasepage(struct page *page, gfp_t wait)
2726 {
2727         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2728
2729         trace_ext4_releasepage(page);
2730
2731         WARN_ON(PageChecked(page));
2732         if (!page_has_buffers(page))
2733                 return 0;
2734         if (journal)
2735                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
2736         else
2737                 return try_to_free_buffers(page);
2738 }
2739
2740 /*
2741  * ext4_get_block used when preparing for a DIO write or buffer write.
2742  * We allocate an uinitialized extent if blocks haven't been allocated.
2743  * The extent will be converted to initialized after the IO is complete.
2744  */
2745 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
2746                    struct buffer_head *bh_result, int create)
2747 {
2748         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
2749                    inode->i_ino, create);
2750         return _ext4_get_block(inode, iblock, bh_result,
2751                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
2752 }
2753
2754 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
2755                             ssize_t size, void *private, int ret,
2756                             bool is_async)
2757 {
2758         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
2759         ext4_io_end_t *io_end = iocb->private;
2760         struct workqueue_struct *wq;
2761         unsigned long flags;
2762         struct ext4_inode_info *ei;
2763
2764         /* if not async direct IO or dio with 0 bytes write, just return */
2765         if (!io_end || !size)
2766                 goto out;
2767
2768         ext_debug("ext4_end_io_dio(): io_end 0x%p "
2769                   "for inode %lu, iocb 0x%p, offset %llu, size %zd\n",
2770                   iocb->private, io_end->inode->i_ino, iocb, offset,
2771                   size);
2772
2773         iocb->private = NULL;
2774
2775         /* if not aio dio with unwritten extents, just free io and return */
2776         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2777                 ext4_free_io_end(io_end);
2778 out:
2779                 if (is_async)
2780                         aio_complete(iocb, ret, 0);
2781                 inode_dio_done(inode);
2782                 return;
2783         }
2784
2785         io_end->offset = offset;
2786         io_end->size = size;
2787         if (is_async) {
2788                 io_end->iocb = iocb;
2789                 io_end->result = ret;
2790         }
2791         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
2792
2793         /* Add the io_end to per-inode completed aio dio list*/
2794         ei = EXT4_I(io_end->inode);
2795         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
2796         list_add_tail(&io_end->list, &ei->i_completed_io_list);
2797         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
2798
2799         /* queue the work to convert unwritten extents to written */
2800         queue_work(wq, &io_end->work);
2801 }
2802
2803 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
2804 {
2805         ext4_io_end_t *io_end = bh->b_private;
2806         struct workqueue_struct *wq;
2807         struct inode *inode;
2808         unsigned long flags;
2809
2810         if (!test_clear_buffer_uninit(bh) || !io_end)
2811                 goto out;
2812
2813         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
2814                 ext4_msg(io_end->inode->i_sb, KERN_INFO,
2815                          "sb umounted, discard end_io request for inode %lu",
2816                          io_end->inode->i_ino);
2817                 ext4_free_io_end(io_end);
2818                 goto out;
2819         }
2820
2821         /*
2822          * It may be over-defensive here to check EXT4_IO_END_UNWRITTEN now,
2823          * but being more careful is always safe for the future change.
2824          */
2825         inode = io_end->inode;
2826         ext4_set_io_unwritten_flag(inode, io_end);
2827
2828         /* Add the io_end to per-inode completed io list*/
2829         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
2830         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
2831         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
2832
2833         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
2834         /* queue the work to convert unwritten extents to written */
2835         queue_work(wq, &io_end->work);
2836 out:
2837         bh->b_private = NULL;
2838         bh->b_end_io = NULL;
2839         clear_buffer_uninit(bh);
2840         end_buffer_async_write(bh, uptodate);
2841 }
2842
2843 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
2844 {
2845         ext4_io_end_t *io_end;
2846         struct page *page = bh->b_page;
2847         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
2848         size_t size = bh->b_size;
2849
2850 retry:
2851         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
2852         if (!io_end) {
2853                 pr_warn_ratelimited("%s: allocation fail\n", __func__);
2854                 schedule();
2855                 goto retry;
2856         }
2857         io_end->offset = offset;
2858         io_end->size = size;
2859         /*
2860          * We need to hold a reference to the page to make sure it
2861          * doesn't get evicted before ext4_end_io_work() has a chance
2862          * to convert the extent from written to unwritten.
2863          */
2864         io_end->page = page;
2865         get_page(io_end->page);
2866
2867         bh->b_private = io_end;
2868         bh->b_end_io = ext4_end_io_buffer_write;
2869         return 0;
2870 }
2871
2872 /*
2873  * For ext4 extent files, ext4 will do direct-io write to holes,
2874  * preallocated extents, and those write extend the file, no need to
2875  * fall back to buffered IO.
2876  *
2877  * For holes, we fallocate those blocks, mark them as uninitialized
2878  * If those blocks were preallocated, we mark sure they are splited, but
2879  * still keep the range to write as uninitialized.
2880  *
2881  * The unwrritten extents will be converted to written when DIO is completed.
2882  * For async direct IO, since the IO may still pending when return, we
2883  * set up an end_io call back function, which will do the conversion
2884  * when async direct IO completed.
2885  *
2886  * If the O_DIRECT write will extend the file then add this inode to the
2887  * orphan list.  So recovery will truncate it back to the original size
2888  * if the machine crashes during the write.
2889  *
2890  */
2891 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
2892                               const struct iovec *iov, loff_t offset,
2893                               unsigned long nr_segs)
2894 {
2895         struct file *file = iocb->ki_filp;
2896         struct inode *inode = file->f_mapping->host;
2897         ssize_t ret;
2898         size_t count = iov_length(iov, nr_segs);
2899
2900         loff_t final_size = offset + count;
2901         if (rw == WRITE && final_size <= inode->i_size) {
2902                 /*
2903                  * We could direct write to holes and fallocate.
2904                  *
2905                  * Allocated blocks to fill the hole are marked as uninitialized
2906                  * to prevent parallel buffered read to expose the stale data
2907                  * before DIO complete the data IO.
2908                  *
2909                  * As to previously fallocated extents, ext4 get_block
2910                  * will just simply mark the buffer mapped but still
2911                  * keep the extents uninitialized.
2912                  *
2913                  * for non AIO case, we will convert those unwritten extents
2914                  * to written after return back from blockdev_direct_IO.
2915                  *
2916                  * for async DIO, the conversion needs to be defered when
2917                  * the IO is completed. The ext4 end_io callback function
2918                  * will be called to take care of the conversion work.
2919                  * Here for async case, we allocate an io_end structure to
2920                  * hook to the iocb.
2921                  */
2922                 iocb->private = NULL;
2923                 EXT4_I(inode)->cur_aio_dio = NULL;
2924                 if (!is_sync_kiocb(iocb)) {
2925                         ext4_io_end_t *io_end =
2926                                 ext4_init_io_end(inode, GFP_NOFS);
2927                         if (!io_end)
2928                                 return -ENOMEM;
2929                         io_end->flag |= EXT4_IO_END_DIRECT;
2930                         iocb->private = io_end;
2931                         /*
2932                          * we save the io structure for current async
2933                          * direct IO, so that later ext4_map_blocks()
2934                          * could flag the io structure whether there
2935                          * is a unwritten extents needs to be converted
2936                          * when IO is completed.
2937                          */
2938                         EXT4_I(inode)->cur_aio_dio = iocb->private;
2939                 }
2940
2941                 ret = __blockdev_direct_IO(rw, iocb, inode,
2942                                          inode->i_sb->s_bdev, iov,
2943                                          offset, nr_segs,
2944                                          ext4_get_block_write,
2945                                          ext4_end_io_dio,
2946                                          NULL,
2947                                          DIO_LOCKING);
2948                 if (iocb->private)
2949                         EXT4_I(inode)->cur_aio_dio = NULL;
2950                 /*
2951                  * The io_end structure takes a reference to the inode,
2952                  * that structure needs to be destroyed and the
2953                  * reference to the inode need to be dropped, when IO is
2954                  * complete, even with 0 byte write, or failed.
2955                  *
2956                  * In the successful AIO DIO case, the io_end structure will be
2957                  * desctroyed and the reference to the inode will be dropped
2958                  * after the end_io call back function is called.
2959                  *
2960                  * In the case there is 0 byte write, or error case, since
2961                  * VFS direct IO won't invoke the end_io call back function,
2962                  * we need to free the end_io structure here.
2963                  */
2964                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
2965                         ext4_free_io_end(iocb->private);
2966                         iocb->private = NULL;
2967                 } else if (ret > 0 && ext4_test_inode_state(inode,
2968                                                 EXT4_STATE_DIO_UNWRITTEN)) {
2969                         int err;
2970                         /*
2971                          * for non AIO case, since the IO is already
2972                          * completed, we could do the conversion right here
2973                          */
2974                         err = ext4_convert_unwritten_extents(inode,
2975                                                              offset, ret);
2976                         if (err < 0)
2977                                 ret = err;
2978                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
2979                 }
2980                 return ret;
2981         }
2982
2983         /* for write the the end of file case, we fall back to old way */
2984         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2985 }
2986
2987 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
2988                               const struct iovec *iov, loff_t offset,
2989                               unsigned long nr_segs)
2990 {
2991         struct file *file = iocb->ki_filp;
2992         struct inode *inode = file->f_mapping->host;
2993         ssize_t ret;
2994
2995         /*
2996          * If we are doing data journalling we don't support O_DIRECT
2997          */
2998         if (ext4_should_journal_data(inode))
2999                 return 0;
3000
3001         trace_ext4_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
3002         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3003                 ret = ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3004         else
3005                 ret = ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3006         trace_ext4_direct_IO_exit(inode, offset,
3007                                 iov_length(iov, nr_segs), rw, ret);
3008         return ret;
3009 }
3010
3011 /*
3012  * Pages can be marked dirty completely asynchronously from ext4's journalling
3013  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3014  * much here because ->set_page_dirty is called under VFS locks.  The page is
3015  * not necessarily locked.
3016  *
3017  * We cannot just dirty the page and leave attached buffers clean, because the
3018  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3019  * or jbddirty because all the journalling code will explode.
3020  *
3021  * So what we do is to mark the page "pending dirty" and next time writepage
3022  * is called, propagate that into the buffers appropriately.
3023  */
3024 static int ext4_journalled_set_page_dirty(struct page *page)
3025 {
3026         SetPageChecked(page);
3027         return __set_page_dirty_nobuffers(page);
3028 }
3029
3030 static const struct address_space_operations ext4_ordered_aops = {
3031         .readpage               = ext4_readpage,
3032         .readpages              = ext4_readpages,
3033         .writepage              = ext4_writepage,
3034         .write_begin            = ext4_write_begin,
3035         .write_end              = ext4_ordered_write_end,
3036         .bmap                   = ext4_bmap,
3037         .invalidatepage         = ext4_invalidatepage,
3038         .releasepage            = ext4_releasepage,
3039         .direct_IO              = ext4_direct_IO,
3040         .migratepage            = buffer_migrate_page,
3041         .is_partially_uptodate  = block_is_partially_uptodate,
3042         .error_remove_page      = generic_error_remove_page,
3043 };
3044
3045 static const struct address_space_operations ext4_writeback_aops = {
3046         .readpage               = ext4_readpage,
3047         .readpages              = ext4_readpages,
3048         .writepage              = ext4_writepage,
3049         .write_begin            = ext4_write_begin,
3050         .write_end              = ext4_writeback_write_end,
3051         .bmap                   = ext4_bmap,
3052         .invalidatepage         = ext4_invalidatepage,
3053         .releasepage            = ext4_releasepage,
3054         .direct_IO              = ext4_direct_IO,
3055         .migratepage            = buffer_migrate_page,
3056         .is_partially_uptodate  = block_is_partially_uptodate,
3057         .error_remove_page      = generic_error_remove_page,
3058 };
3059
3060 static const struct address_space_operations ext4_journalled_aops = {
3061         .readpage               = ext4_readpage,
3062         .readpages              = ext4_readpages,
3063         .writepage              = ext4_writepage,
3064         .write_begin            = ext4_write_begin,
3065         .write_end              = ext4_journalled_write_end,
3066         .set_page_dirty         = ext4_journalled_set_page_dirty,
3067         .bmap                   = ext4_bmap,
3068         .invalidatepage         = ext4_invalidatepage,
3069         .releasepage            = ext4_releasepage,
3070         .direct_IO              = ext4_direct_IO,
3071         .is_partially_uptodate  = block_is_partially_uptodate,
3072         .error_remove_page      = generic_error_remove_page,
3073 };
3074
3075 static const struct address_space_operations ext4_da_aops = {
3076         .readpage               = ext4_readpage,
3077         .readpages              = ext4_readpages,
3078         .writepage              = ext4_writepage,
3079         .writepages             = ext4_da_writepages,
3080         .write_begin            = ext4_da_write_begin,
3081         .write_end              = ext4_da_write_end,
3082         .bmap                   = ext4_bmap,
3083         .invalidatepage         = ext4_da_invalidatepage,
3084         .releasepage            = ext4_releasepage,
3085         .direct_IO              = ext4_direct_IO,
3086         .migratepage            = buffer_migrate_page,
3087         .is_partially_uptodate  = block_is_partially_uptodate,
3088         .error_remove_page      = generic_error_remove_page,
3089 };
3090
3091 void ext4_set_aops(struct inode *inode)
3092 {
3093         switch (ext4_inode_journal_mode(inode)) {
3094         case EXT4_INODE_ORDERED_DATA_MODE:
3095                 if (test_opt(inode->i_sb, DELALLOC))
3096                         inode->i_mapping->a_ops = &ext4_da_aops;
3097                 else
3098                         inode->i_mapping->a_ops = &ext4_ordered_aops;
3099                 break;
3100         case EXT4_INODE_WRITEBACK_DATA_MODE:
3101                 if (test_opt(inode->i_sb, DELALLOC))
3102                         inode->i_mapping->a_ops = &ext4_da_aops;
3103                 else
3104                         inode->i_mapping->a_ops = &ext4_writeback_aops;
3105                 break;
3106         case EXT4_INODE_JOURNAL_DATA_MODE:
3107                 inode->i_mapping->a_ops = &ext4_journalled_aops;
3108                 break;
3109         default:
3110                 BUG();
3111         }
3112 }
3113
3114
3115 /*
3116  * ext4_discard_partial_page_buffers()
3117  * Wrapper function for ext4_discard_partial_page_buffers_no_lock.
3118  * This function finds and locks the page containing the offset
3119  * "from" and passes it to ext4_discard_partial_page_buffers_no_lock.
3120  * Calling functions that already have the page locked should call
3121  * ext4_discard_partial_page_buffers_no_lock directly.
3122  */
3123 int ext4_discard_partial_page_buffers(handle_t *handle,
3124                 struct address_space *mapping, loff_t from,
3125                 loff_t length, int flags)
3126 {
3127         struct inode *inode = mapping->host;
3128         struct page *page;
3129         int err = 0;
3130
3131         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3132                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3133         if (!page)
3134                 return -ENOMEM;
3135
3136         err = ext4_discard_partial_page_buffers_no_lock(handle, inode, page,
3137                 from, length, flags);
3138
3139         unlock_page(page);
3140         page_cache_release(page);
3141         return err;
3142 }
3143
3144 /*
3145  * ext4_discard_partial_page_buffers_no_lock()
3146  * Zeros a page range of length 'length' starting from offset 'from'.
3147  * Buffer heads that correspond to the block aligned regions of the
3148  * zeroed range will be unmapped.  Unblock aligned regions
3149  * will have the corresponding buffer head mapped if needed so that
3150  * that region of the page can be updated with the partial zero out.
3151  *
3152  * This function assumes that the page has already been  locked.  The
3153  * The range to be discarded must be contained with in the given page.
3154  * If the specified range exceeds the end of the page it will be shortened
3155  * to the end of the page that corresponds to 'from'.  This function is
3156  * appropriate for updating a page and it buffer heads to be unmapped and
3157  * zeroed for blocks that have been either released, or are going to be
3158  * released.
3159  *
3160  * handle: The journal handle
3161  * inode:  The files inode
3162  * page:   A locked page that contains the offset "from"
3163  * from:   The starting byte offset (from the begining of the file)
3164  *         to begin discarding
3165  * len:    The length of bytes to discard
3166  * flags:  Optional flags that may be used:
3167  *
3168  *         EXT4_DISCARD_PARTIAL_PG_ZERO_UNMAPPED
3169  *         Only zero the regions of the page whose buffer heads
3170  *         have already been unmapped.  This flag is appropriate
3171  *         for updateing the contents of a page whose blocks may
3172  *         have already been released, and we only want to zero
3173  *         out the regions that correspond to those released blocks.
3174  *
3175  * Returns zero on sucess or negative on failure.
3176  */
3177 static int ext4_discard_partial_page_buffers_no_lock(handle_t *handle,
3178                 struct inode *inode, struct page *page, loff_t from,
3179                 loff_t length, int flags)
3180 {
3181         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
3182         unsigned int offset = from & (PAGE_CACHE_SIZE-1);
3183         unsigned int blocksize, max, pos;
3184         ext4_lblk_t iblock;
3185         struct buffer_head *bh;
3186         int err = 0;
3187
3188         blocksize = inode->i_sb->s_blocksize;
3189         max = PAGE_CACHE_SIZE - offset;
3190
3191         if (index != page->index)
3192                 return -EINVAL;
3193
3194         /*
3195          * correct length if it does not fall between
3196          * 'from' and the end of the page
3197          */
3198         if (length > max || length < 0)
3199                 length = max;
3200
3201         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3202
3203         if (!page_has_buffers(page))
3204                 create_empty_buffers(page, blocksize, 0);
3205
3206         /* Find the buffer that contains "offset" */
3207         bh = page_buffers(page);
3208         pos = blocksize;
3209         while (offset >= pos) {
3210                 bh = bh->b_this_page;
3211                 iblock++;
3212                 pos += blocksize;
3213         }
3214
3215         pos = offset;
3216         while (pos < offset + length) {
3217                 unsigned int end_of_block, range_to_discard;
3218
3219                 err = 0;
3220
3221                 /* The length of space left to zero and unmap */
3222                 range_to_discard = offset + length - pos;
3223
3224                 /* The length of space until the end of the block */
3225                 end_of_block = blocksize - (pos & (blocksize-1));
3226
3227                 /*
3228                  * Do not unmap or zero past end of block
3229                  * for this buffer head
3230                  */
3231                 if (range_to_discard > end_of_block)
3232                         range_to_discard = end_of_block;
3233
3234
3235                 /*
3236                  * Skip this buffer head if we are only zeroing unampped
3237                  * regions of the page
3238                  */
3239                 if (flags & EXT4_DISCARD_PARTIAL_PG_ZERO_UNMAPPED &&
3240                         buffer_mapped(bh))
3241                                 goto next;
3242
3243                 /* If the range is block aligned, unmap */
3244                 if (range_to_discard == blocksize) {
3245                         clear_buffer_dirty(bh);
3246                         bh->b_bdev = NULL;
3247                         clear_buffer_mapped(bh);
3248                         clear_buffer_req(bh);
3249                         clear_buffer_new(bh);
3250                         clear_buffer_delay(bh);
3251                         clear_buffer_unwritten(bh);
3252                         clear_buffer_uptodate(bh);
3253                         zero_user(page, pos, range_to_discard);
3254                         BUFFER_TRACE(bh, "Buffer discarded");
3255                         goto next;
3256                 }
3257
3258                 /*
3259                  * If this block is not completely contained in the range
3260                  * to be discarded, then it is not going to be released. Because
3261                  * we need to keep this block, we need to make sure this part
3262                  * of the page is uptodate before we modify it by writeing
3263                  * partial zeros on it.
3264                  */
3265                 if (!buffer_mapped(bh)) {
3266                         /*
3267                          * Buffer head must be mapped before we can read
3268                          * from the block
3269                          */
3270                         BUFFER_TRACE(bh, "unmapped");
3271                         ext4_get_block(inode, iblock, bh, 0);
3272                         /* unmapped? It's a hole - nothing to do */
3273                         if (!buffer_mapped(bh)) {
3274                                 BUFFER_TRACE(bh, "still unmapped");
3275                                 goto next;
3276                         }
3277                 }
3278
3279                 /* Ok, it's mapped. Make sure it's up-to-date */
3280                 if (PageUptodate(page))
3281                         set_buffer_uptodate(bh);
3282
3283                 if (!buffer_uptodate(bh)) {
3284                         err = -EIO;
3285                         ll_rw_block(READ, 1, &bh);
3286                         wait_on_buffer(bh);
3287                         /* Uhhuh. Read error. Complain and punt.*/
3288                         if (!buffer_uptodate(bh))
3289                                 goto next;
3290                 }
3291
3292                 if (ext4_should_journal_data(inode)) {
3293                         BUFFER_TRACE(bh, "get write access");
3294                         err = ext4_journal_get_write_access(handle, bh);
3295                         if (err)
3296                                 goto next;
3297                 }
3298
3299                 zero_user(page, pos, range_to_discard);
3300
3301                 err = 0;
3302                 if (ext4_should_journal_data(inode)) {
3303                         err = ext4_handle_dirty_metadata(handle, inode, bh);
3304                 } else
3305                         mark_buffer_dirty(bh);
3306
3307                 BUFFER_TRACE(bh, "Partial buffer zeroed");
3308 next:
3309                 bh = bh->b_this_page;
3310                 iblock++;
3311                 pos += range_to_discard;
3312         }
3313
3314         return err;
3315 }
3316
3317 int ext4_can_truncate(struct inode *inode)
3318 {
3319         if (S_ISREG(inode->i_mode))
3320                 return 1;
3321         if (S_ISDIR(inode->i_mode))
3322                 return 1;
3323         if (S_ISLNK(inode->i_mode))
3324                 return !ext4_inode_is_fast_symlink(inode);
3325         return 0;
3326 }
3327
3328 /*
3329  * ext4_punch_hole: punches a hole in a file by releaseing the blocks
3330  * associated with the given offset and length
3331  *
3332  * @inode:  File inode
3333  * @offset: The offset where the hole will begin
3334  * @len:    The length of the hole
3335  *
3336  * Returns: 0 on sucess or negative on failure
3337  */
3338
3339 int ext4_punch_hole(struct file *file, loff_t offset, loff_t length)
3340 {
3341         struct inode *inode = file->f_path.dentry->d_inode;
3342         if (!S_ISREG(inode->i_mode))
3343                 return -EOPNOTSUPP;
3344
3345         if (!ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3346                 /* TODO: Add support for non extent hole punching */
3347                 return -EOPNOTSUPP;
3348         }
3349
3350         if (EXT4_SB(inode->i_sb)->s_cluster_ratio > 1) {
3351                 /* TODO: Add support for bigalloc file systems */
3352                 return -EOPNOTSUPP;
3353         }
3354
3355         return ext4_ext_punch_hole(file, offset, length);
3356 }
3357
3358 /*
3359  * ext4_truncate()
3360  *
3361  * We block out ext4_get_block() block instantiations across the entire
3362  * transaction, and VFS/VM ensures that ext4_truncate() cannot run
3363  * simultaneously on behalf of the same inode.
3364  *
3365  * As we work through the truncate and commit bits of it to the journal there
3366  * is one core, guiding principle: the file's tree must always be consistent on
3367  * disk.  We must be able to restart the truncate after a crash.
3368  *
3369  * The file's tree may be transiently inconsistent in memory (although it
3370  * probably isn't), but whenever we close off and commit a journal transaction,
3371  * the contents of (the filesystem + the journal) must be consistent and
3372  * restartable.  It's pretty simple, really: bottom up, right to left (although
3373  * left-to-right works OK too).
3374  *
3375  * Note that at recovery time, journal replay occurs *before* the restart of
3376  * truncate against the orphan inode list.
3377  *
3378  * The committed inode has the new, desired i_size (which is the same as
3379  * i_disksize in this case).  After a crash, ext4_orphan_cleanup() will see
3380  * that this inode's truncate did not complete and it will again call
3381  * ext4_truncate() to have another go.  So there will be instantiated blocks
3382  * to the right of the truncation point in a crashed ext4 filesystem.  But
3383  * that's fine - as long as they are linked from the inode, the post-crash
3384  * ext4_truncate() run will find them and release them.
3385  */
3386 void ext4_truncate(struct inode *inode)
3387 {
3388         trace_ext4_truncate_enter(inode);
3389
3390         if (!ext4_can_truncate(inode))
3391                 return;
3392
3393         ext4_clear_inode_flag(inode, EXT4_INODE_EOFBLOCKS);
3394
3395         if (inode->i_size == 0 && !test_opt(inode->i_sb, NO_AUTO_DA_ALLOC))
3396                 ext4_set_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
3397
3398         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3399                 ext4_ext_truncate(inode);
3400         else
3401                 ext4_ind_truncate(inode);
3402
3403         trace_ext4_truncate_exit(inode);
3404 }
3405
3406 /*
3407  * ext4_get_inode_loc returns with an extra refcount against the inode's
3408  * underlying buffer_head on success. If 'in_mem' is true, we have all
3409  * data in memory that is needed to recreate the on-disk version of this
3410  * inode.
3411  */
3412 static int __ext4_get_inode_loc(struct inode *inode,
3413                                 struct ext4_iloc *iloc, int in_mem)
3414 {
3415         struct ext4_group_desc  *gdp;
3416         struct buffer_head      *bh;
3417         struct super_block      *sb = inode->i_sb;
3418         ext4_fsblk_t            block;
3419         int                     inodes_per_block, inode_offset;
3420
3421         iloc->bh = NULL;
3422         if (!ext4_valid_inum(sb, inode->i_ino))
3423                 return -EIO;
3424
3425         iloc->block_group = (inode->i_ino - 1) / EXT4_INODES_PER_GROUP(sb);
3426         gdp = ext4_get_group_desc(sb, iloc->block_group, NULL);
3427         if (!gdp)
3428                 return -EIO;
3429
3430         /*
3431          * Figure out the offset within the block group inode table
3432          */
3433         inodes_per_block = EXT4_SB(sb)->s_inodes_per_block;
3434         inode_offset = ((inode->i_ino - 1) %
3435                         EXT4_INODES_PER_GROUP(sb));
3436         block = ext4_inode_table(sb, gdp) + (inode_offset / inodes_per_block);
3437         iloc->offset = (inode_offset % inodes_per_block) * EXT4_INODE_SIZE(sb);
3438
3439         bh = sb_getblk(sb, block);
3440         if (!bh) {
3441                 EXT4_ERROR_INODE_BLOCK(inode, block,
3442                                        "unable to read itable block");
3443                 return -EIO;
3444         }
3445         if (!buffer_uptodate(bh)) {
3446                 lock_buffer(bh);
3447
3448                 /*
3449                  * If the buffer has the write error flag, we have failed
3450                  * to write out another inode in the same block.  In this
3451                  * case, we don't have to read the block because we may
3452                  * read the old inode data successfully.
3453                  */
3454                 if (buffer_write_io_error(bh) && !buffer_uptodate(bh))
3455                         set_buffer_uptodate(bh);
3456
3457                 if (buffer_uptodate(bh)) {
3458                         /* someone brought it uptodate while we waited */
3459                         unlock_buffer(bh);
3460                         goto has_buffer;
3461                 }
3462
3463                 /*
3464                  * If we have all information of the inode in memory and this
3465                  * is the only valid inode in the block, we need not read the
3466                  * block.
3467                  */
3468                 if (in_mem) {
3469                         struct buffer_head *bitmap_bh;
3470                         int i, start;
3471
3472                         start = inode_offset & ~(inodes_per_block - 1);
3473
3474                         /* Is the inode bitmap in cache? */
3475                         bitmap_bh = sb_getblk(sb, ext4_inode_bitmap(sb, gdp));
3476                         if (!bitmap_bh)
3477                                 goto make_io;
3478
3479                         /*
3480                          * If the inode bitmap isn't in cache then the
3481                          * optimisation may end up performing two reads instead
3482                          * of one, so skip it.
3483                          */
3484                         if (!buffer_uptodate(bitmap_bh)) {
3485                                 brelse(bitmap_bh);
3486                                 goto make_io;
3487                         }
3488                         for (i = start; i < start + inodes_per_block; i++) {
3489                                 if (i == inode_offset)
3490                                         continue;
3491                                 if (ext4_test_bit(i, bitmap_bh->b_data))
3492                                         break;
3493                         }
3494                         brelse(bitmap_bh);
3495                         if (i == start + inodes_per_block) {
3496                                 /* all other inodes are free, so skip I/O */
3497                                 memset(bh->b_data, 0, bh->b_size);
3498                                 set_buffer_uptodate(bh);
3499                                 unlock_buffer(bh);
3500                                 goto has_buffer;
3501                         }
3502                 }
3503
3504 make_io:
3505                 /*
3506                  * If we need to do any I/O, try to pre-readahead extra
3507                  * blocks from the inode table.
3508                  */
3509                 if (EXT4_SB(sb)->s_inode_readahead_blks) {
3510                         ext4_fsblk_t b, end, table;
3511                         unsigned num;
3512
3513                         table = ext4_inode_table(sb, gdp);
3514                         /* s_inode_readahead_blks is always a power of 2 */
3515                         b = block & ~(EXT4_SB(sb)->s_inode_readahead_blks-1);
3516                         if (table > b)
3517                                 b = table;
3518                         end = b + EXT4_SB(sb)->s_inode_readahead_blks;
3519                         num = EXT4_INODES_PER_GROUP(sb);
3520                         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3521                                        EXT4_FEATURE_RO_COMPAT_GDT_CSUM))
3522                                 num -= ext4_itable_unused_count(sb, gdp);
3523                         table += num / inodes_per_block;
3524                         if (end > table)
3525                                 end = table;
3526                         while (b <= end)
3527                                 sb_breadahead(sb, b++);
3528                 }
3529
3530                 /*
3531                  * There are other valid inodes in the buffer, this inode
3532                  * has in-inode xattrs, or we don't have this inode in memory.
3533                  * Read the block from disk.
3534                  */
3535                 trace_ext4_load_inode(inode);
3536                 get_bh(bh);
3537                 bh->b_end_io = end_buffer_read_sync;
3538                 submit_bh(READ | REQ_META | REQ_PRIO, bh);
3539                 wait_on_buffer(bh);
3540                 if (!buffer_uptodate(bh)) {
3541                         EXT4_ERROR_INODE_BLOCK(inode, block,
3542                                                "unable to read itable block");
3543                         brelse(bh);
3544                         return -EIO;
3545                 }
3546         }
3547 has_buffer:
3548         iloc->bh = bh;
3549         return 0;
3550 }
3551
3552 int ext4_get_inode_loc(struct inode *inode, struct ext4_iloc *iloc)
3553 {
3554         /* We have all inode data except xattrs in memory here. */
3555         return __ext4_get_inode_loc(inode, iloc,
3556                 !ext4_test_inode_state(inode, EXT4_STATE_XATTR));
3557 }
3558
3559 void ext4_set_inode_flags(struct inode *inode)
3560 {
3561         unsigned int flags = EXT4_I(inode)->i_flags;
3562
3563         inode->i_flags &= ~(S_SYNC|S_APPEND|S_IMMUTABLE|S_NOATIME|S_DIRSYNC);
3564         if (flags & EXT4_SYNC_FL)
3565                 inode->i_flags |= S_SYNC;
3566         if (flags & EXT4_APPEND_FL)
3567                 inode->i_flags |= S_APPEND;
3568         if (flags & EXT4_IMMUTABLE_FL)
3569                 inode->i_flags |= S_IMMUTABLE;
3570         if (flags & EXT4_NOATIME_FL)
3571                 inode->i_flags |= S_NOATIME;
3572         if (flags & EXT4_DIRSYNC_FL)
3573                 inode->i_flags |= S_DIRSYNC;
3574 }
3575
3576 /* Propagate flags from i_flags to EXT4_I(inode)->i_flags */
3577 void ext4_get_inode_flags(struct ext4_inode_info *ei)
3578 {
3579         unsigned int vfs_fl;
3580         unsigned long old_fl, new_fl;
3581
3582         do {
3583                 vfs_fl = ei->vfs_inode.i_flags;
3584                 old_fl = ei->i_flags;
3585                 new_fl = old_fl & ~(EXT4_SYNC_FL|EXT4_APPEND_FL|
3586                                 EXT4_IMMUTABLE_FL|EXT4_NOATIME_FL|
3587                                 EXT4_DIRSYNC_FL);
3588                 if (vfs_fl & S_SYNC)
3589                         new_fl |= EXT4_SYNC_FL;
3590                 if (vfs_fl & S_APPEND)
3591                         new_fl |= EXT4_APPEND_FL;
3592                 if (vfs_fl & S_IMMUTABLE)
3593                         new_fl |= EXT4_IMMUTABLE_FL;
3594                 if (vfs_fl & S_NOATIME)
3595                         new_fl |= EXT4_NOATIME_FL;
3596                 if (vfs_fl & S_DIRSYNC)
3597                         new_fl |= EXT4_DIRSYNC_FL;
3598         } while (cmpxchg(&ei->i_flags, old_fl, new_fl) != old_fl);
3599 }
3600
3601 static blkcnt_t ext4_inode_blocks(struct ext4_inode *raw_inode,
3602                                   struct ext4_inode_info *ei)
3603 {
3604         blkcnt_t i_blocks ;
3605         struct inode *inode = &(ei->vfs_inode);
3606         struct super_block *sb = inode->i_sb;
3607
3608         if (EXT4_HAS_RO_COMPAT_FEATURE(sb,
3609                                 EXT4_FEATURE_RO_COMPAT_HUGE_FILE)) {
3610                 /* we are using combined 48 bit field */
3611                 i_blocks = ((u64)le16_to_cpu(raw_inode->i_blocks_high)) << 32 |
3612                                         le32_to_cpu(raw_inode->i_blocks_lo);
3613                 if (ext4_test_inode_flag(inode, EXT4_INODE_HUGE_FILE)) {
3614                         /* i_blocks represent file system block size */
3615                         return i_blocks  << (inode->i_blkbits - 9);
3616                 } else {
3617                         return i_blocks;
3618                 }
3619         } else {
3620                 return le32_to_cpu(raw_inode->i_blocks_lo);
3621         }
3622 }
3623
3624 struct inode *ext4_iget(struct super_block *sb, unsigned long ino)
3625 {
3626         struct ext4_iloc iloc;
3627         struct ext4_inode *raw_inode;
3628         struct ext4_inode_info *ei;
3629         struct inode *inode;
3630         journal_t *journal = EXT4_SB(sb)->s_journal;
3631         long ret;
3632         int block;
3633         uid_t i_uid;
3634         gid_t i_gid;
3635
3636         inode = iget_locked(sb, ino);
3637         if (!inode)
3638                 return ERR_PTR(-ENOMEM);
3639         if (!(inode->i_state & I_NEW))
3640                 return inode;
3641
3642         ei = EXT4_I(inode);
3643         iloc.bh = NULL;
3644
3645         ret = __ext4_get_inode_loc(inode, &iloc, 0);
3646         if (ret < 0)
3647                 goto bad_inode;
3648         raw_inode = ext4_raw_inode(&iloc);
3649         inode->i_mode = le16_to_cpu(raw_inode->i_mode);
3650         i_uid = (uid_t)le16_to_cpu(raw_inode->i_uid_low);
3651         i_gid = (gid_t)le16_to_cpu(raw_inode->i_gid_low);
3652         if (!(test_opt(inode->i_sb, NO_UID32))) {
3653                 i_uid |= le16_to_cpu(raw_inode->i_uid_high) << 16;
3654                 i_gid |= le16_to_cpu(raw_inode->i_gid_high) << 16;
3655         }
3656         i_uid_write(inode, i_uid);
3657         i_gid_write(inode, i_gid);
3658         set_nlink(inode, le16_to_cpu(raw_inode->i_links_count));
3659
3660         ext4_clear_state_flags(ei);     /* Only relevant on 32-bit archs */
3661         ei->i_dir_start_lookup = 0;
3662         ei->i_dtime = le32_to_cpu(raw_inode->i_dtime);
3663         /* We now have enough fields to check if the inode was active or not.
3664          * This is needed because nfsd might try to access dead inodes
3665          * the test is that same one that e2fsck uses
3666          * NeilBrown 1999oct15
3667          */
3668         if (inode->i_nlink == 0) {
3669                 if (inode->i_mode == 0 ||
3670                     !(EXT4_SB(inode->i_sb)->s_mount_state & EXT4_ORPHAN_FS)) {
3671                         /* this inode is deleted */
3672                         ret = -ESTALE;
3673                         goto bad_inode;
3674                 }
3675                 /* The only unlinked inodes we let through here have
3676                  * valid i_mode and are being read by the orphan
3677                  * recovery code: that's fine, we're about to complete
3678                  * the process of deleting those. */
3679         }
3680         ei->i_flags = le32_to_cpu(raw_inode->i_flags);
3681         inode->i_blocks = ext4_inode_blocks(raw_inode, ei);
3682         ei->i_file_acl = le32_to_cpu(raw_inode->i_file_acl_lo);
3683         if (EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_64BIT))
3684                 ei->i_file_acl |=
3685                         ((__u64)le16_to_cpu(raw_inode->i_file_acl_high)) << 32;
3686         inode->i_size = ext4_isize(raw_inode);
3687         ei->i_disksize = inode->i_size;
3688 #ifdef CONFIG_QUOTA
3689         ei->i_reserved_quota = 0;
3690 #endif
3691         inode->i_generation = le32_to_cpu(raw_inode->i_generation);
3692         ei->i_block_group = iloc.block_group;
3693         ei->i_last_alloc_group = ~0;
3694         /*
3695          * NOTE! The in-memory inode i_data array is in little-endian order
3696          * even on big-endian machines: we do NOT byteswap the block numbers!
3697          */
3698         for (block = 0; block < EXT4_N_BLOCKS; block++)
3699                 ei->i_data[block] = raw_inode->i_block[block];
3700         INIT_LIST_HEAD(&ei->i_orphan);
3701
3702         /*
3703          * Set transaction id's of transactions that have to be committed
3704          * to finish f[data]sync. We set them to currently running transaction
3705          * as we cannot be sure that the inode or some of its metadata isn't
3706          * part of the transaction - the inode could have been reclaimed and
3707          * now it is reread from disk.
3708          */
3709         if (journal) {
3710                 transaction_t *transaction;
3711                 tid_t tid;
3712
3713                 read_lock(&journal->j_state_lock);
3714                 if (journal->j_running_transaction)
3715                         transaction = journal->j_running_transaction;
3716                 else
3717                         transaction = journal->j_committing_transaction;
3718                 if (transaction)
3719                         tid = transaction->t_tid;
3720                 else
3721                         tid = journal->j_commit_sequence;
3722                 read_unlock(&journal->j_state_lock);
3723                 ei->i_sync_tid = tid;
3724                 ei->i_datasync_tid = tid;
3725         }
3726
3727         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3728                 ei->i_extra_isize = le16_to_cpu(raw_inode->i_extra_isize);
3729                 if (EXT4_GOOD_OLD_INODE_SIZE + ei->i_extra_isize >
3730                     EXT4_INODE_SIZE(inode->i_sb)) {
3731                         ret = -EIO;
3732                         goto bad_inode;
3733                 }
3734                 if (ei->i_extra_isize == 0) {
3735                         /* The extra space is currently unused. Use it. */
3736                         ei->i_extra_isize = sizeof(struct ext4_inode) -
3737                                             EXT4_GOOD_OLD_INODE_SIZE;
3738                 } else {
3739                         __le32 *magic = (void *)raw_inode +
3740                                         EXT4_GOOD_OLD_INODE_SIZE +
3741                                         ei->i_extra_isize;
3742                         if (*magic == cpu_to_le32(EXT4_XATTR_MAGIC))
3743                                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
3744                 }
3745         } else
3746                 ei->i_extra_isize = 0;
3747
3748         EXT4_INODE_GET_XTIME(i_ctime, inode, raw_inode);
3749         EXT4_INODE_GET_XTIME(i_mtime, inode, raw_inode);
3750         EXT4_INODE_GET_XTIME(i_atime, inode, raw_inode);
3751         EXT4_EINODE_GET_XTIME(i_crtime, ei, raw_inode);
3752
3753         inode->i_version = le32_to_cpu(raw_inode->i_disk_version);
3754         if (EXT4_INODE_SIZE(inode->i_sb) > EXT4_GOOD_OLD_INODE_SIZE) {
3755                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
3756                         inode->i_version |=
3757                         (__u64)(le32_to_cpu(raw_inode->i_version_hi)) << 32;
3758         }
3759
3760         ret = 0;
3761         if (ei->i_file_acl &&
3762             !ext4_data_block_valid(EXT4_SB(sb), ei->i_file_acl, 1)) {
3763                 EXT4_ERROR_INODE(inode, "bad extended attribute block %llu",
3764                                  ei->i_file_acl);
3765                 ret = -EIO;
3766                 goto bad_inode;
3767         } else if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
3768                 if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3769                     (S_ISLNK(inode->i_mode) &&
3770                      !ext4_inode_is_fast_symlink(inode)))
3771                         /* Validate extent which is part of inode */
3772                         ret = ext4_ext_check_inode(inode);
3773         } else if (S_ISREG(inode->i_mode) || S_ISDIR(inode->i_mode) ||
3774                    (S_ISLNK(inode->i_mode) &&
3775                     !ext4_inode_is_fast_symlink(inode))) {
3776                 /* Validate block references which are part of inode */
3777                 ret = ext4_ind_check_inode(inode);
3778         }
3779         if (ret)
3780                 goto bad_inode;
3781
3782         if (S_ISREG(inode->i_mode)) {
3783                 inode->i_op = &ext4_file_inode_operations;
3784                 inode->i_fop = &ext4_file_operations;
3785                 ext4_set_aops(inode);
3786         } else if (S_ISDIR(inode->i_mode)) {
3787                 inode->i_op = &ext4_dir_inode_operations;
3788                 inode->i_fop = &ext4_dir_operations;
3789         } else if (S_ISLNK(inode->i_mode)) {
3790                 if (ext4_inode_is_fast_symlink(inode)) {
3791                         inode->i_op = &ext4_fast_symlink_inode_operations;
3792                         nd_terminate_link(ei->i_data, inode->i_size,
3793                                 sizeof(ei->i_data) - 1);
3794                 } else {
3795                         inode->i_op = &ext4_symlink_inode_operations;
3796                         ext4_set_aops(inode);
3797                 }
3798         } else if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode) ||
3799               S_ISFIFO(inode->i_mode) || S_ISSOCK(inode->i_mode)) {
3800                 inode->i_op = &ext4_special_inode_operations;
3801                 if (raw_inode->i_block[0])
3802                         init_special_inode(inode, inode->i_mode,
3803                            old_decode_dev(le32_to_cpu(raw_inode->i_block[0])));
3804                 else
3805                         init_special_inode(inode, inode->i_mode,
3806                            new_decode_dev(le32_to_cpu(raw_inode->i_block[1])));
3807         } else {
3808                 ret = -EIO;
3809                 EXT4_ERROR_INODE(inode, "bogus i_mode (%o)", inode->i_mode);
3810                 goto bad_inode;
3811         }
3812         brelse(iloc.bh);
3813         ext4_set_inode_flags(inode);
3814         unlock_new_inode(inode);
3815         return inode;
3816
3817 bad_inode:
3818         brelse(iloc.bh);
3819         iget_failed(inode);
3820         return ERR_PTR(ret);
3821 }
3822
3823 static int ext4_inode_blocks_set(handle_t *handle,
3824                                 struct ext4_inode *raw_inode,
3825                                 struct ext4_inode_info *ei)
3826 {
3827         struct inode *inode = &(ei->vfs_inode);
3828         u64 i_blocks = inode->i_blocks;
3829         struct super_block *sb = inode->i_sb;
3830
3831         if (i_blocks <= ~0U) {
3832                 /*
3833                  * i_blocks can be represnted in a 32 bit variable
3834                  * as multiple of 512 bytes
3835                  */
3836                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
3837                 raw_inode->i_blocks_high = 0;
3838                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3839                 return 0;
3840         }
3841         if (!EXT4_HAS_RO_COMPAT_FEATURE(sb, EXT4_FEATURE_RO_COMPAT_HUGE_FILE))
3842                 return -EFBIG;
3843
3844         if (i_blocks <= 0xffffffffffffULL) {
3845                 /*
3846                  * i_blocks can be represented in a 48 bit variable
3847                  * as multiple of 512 bytes
3848                  */
3849                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
3850                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3851                 ext4_clear_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3852         } else {
3853                 ext4_set_inode_flag(inode, EXT4_INODE_HUGE_FILE);
3854                 /* i_block is stored in file system block size */
3855                 i_blocks = i_blocks >> (inode->i_blkbits - 9);
3856                 raw_inode->i_blocks_lo   = cpu_to_le32(i_blocks);
3857                 raw_inode->i_blocks_high = cpu_to_le16(i_blocks >> 32);
3858         }
3859         return 0;
3860 }
3861
3862 /*
3863  * Post the struct inode info into an on-disk inode location in the
3864  * buffer-cache.  This gobbles the caller's reference to the
3865  * buffer_head in the inode location struct.
3866  *
3867  * The caller must have write access to iloc->bh.
3868  */
3869 static int ext4_do_update_inode(handle_t *handle,
3870                                 struct inode *inode,
3871                                 struct ext4_iloc *iloc)
3872 {
3873         struct ext4_inode *raw_inode = ext4_raw_inode(iloc);
3874         struct ext4_inode_info *ei = EXT4_I(inode);
3875         struct buffer_head *bh = iloc->bh;
3876         int err = 0, rc, block;
3877         uid_t i_uid;
3878         gid_t i_gid;
3879
3880         /* For fields not not tracking in the in-memory inode,
3881          * initialise them to zero for new inodes. */
3882         if (ext4_test_inode_state(inode, EXT4_STATE_NEW))
3883                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
3884
3885         ext4_get_inode_flags(ei);
3886         raw_inode->i_mode = cpu_to_le16(inode->i_mode);
3887         i_uid = i_uid_read(inode);
3888         i_gid = i_gid_read(inode);
3889         if (!(test_opt(inode->i_sb, NO_UID32))) {
3890                 raw_inode->i_uid_low = cpu_to_le16(low_16_bits(i_uid));
3891                 raw_inode->i_gid_low = cpu_to_le16(low_16_bits(i_gid));
3892 /*
3893  * Fix up interoperability with old kernels. Otherwise, old inodes get
3894  * re-used with the upper 16 bits of the uid/gid intact
3895  */
3896                 if (!ei->i_dtime) {
3897                         raw_inode->i_uid_high =
3898                                 cpu_to_le16(high_16_bits(i_uid));
3899                         raw_inode->i_gid_high =
3900                                 cpu_to_le16(high_16_bits(i_gid));
3901                 } else {
3902                         raw_inode->i_uid_high = 0;
3903                         raw_inode->i_gid_high = 0;
3904                 }
3905         } else {
3906                 raw_inode->i_uid_low = cpu_to_le16(fs_high2lowuid(i_uid));
3907                 raw_inode->i_gid_low = cpu_to_le16(fs_high2lowgid(i_gid));
3908                 raw_inode->i_uid_high = 0;
3909                 raw_inode->i_gid_high = 0;
3910         }
3911         raw_inode->i_links_count = cpu_to_le16(inode->i_nlink);
3912
3913         EXT4_INODE_SET_XTIME(i_ctime, inode, raw_inode);
3914         EXT4_INODE_SET_XTIME(i_mtime, inode, raw_inode);
3915         EXT4_INODE_SET_XTIME(i_atime, inode, raw_inode);
3916         EXT4_EINODE_SET_XTIME(i_crtime, ei, raw_inode);
3917
3918         if (ext4_inode_blocks_set(handle, raw_inode, ei))
3919                 goto out_brelse;
3920         raw_inode->i_dtime = cpu_to_le32(ei->i_dtime);
3921         raw_inode->i_flags = cpu_to_le32(ei->i_flags & 0xFFFFFFFF);
3922         if (EXT4_SB(inode->i_sb)->s_es->s_creator_os !=
3923             cpu_to_le32(EXT4_OS_HURD))
3924                 raw_inode->i_file_acl_high =
3925                         cpu_to_le16(ei->i_file_acl >> 32);
3926         raw_inode->i_file_acl_lo = cpu_to_le32(ei->i_file_acl);
3927         ext4_isize_set(raw_inode, ei->i_disksize);
3928         if (ei->i_disksize > 0x7fffffffULL) {
3929                 struct super_block *sb = inode->i_sb;
3930                 if (!EXT4_HAS_RO_COMPAT_FEATURE(sb,
3931                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE) ||
3932                                 EXT4_SB(sb)->s_es->s_rev_level ==
3933                                 cpu_to_le32(EXT4_GOOD_OLD_REV)) {
3934                         /* If this is the first large file
3935                          * created, add a flag to the superblock.
3936                          */
3937                         err = ext4_journal_get_write_access(handle,
3938                                         EXT4_SB(sb)->s_sbh);
3939                         if (err)
3940                                 goto out_brelse;
3941                         ext4_update_dynamic_rev(sb);
3942                         EXT4_SET_RO_COMPAT_FEATURE(sb,
3943                                         EXT4_FEATURE_RO_COMPAT_LARGE_FILE);
3944                         ext4_handle_sync(handle);
3945                         err = ext4_handle_dirty_super(handle, sb);
3946                 }
3947         }
3948         raw_inode->i_generation = cpu_to_le32(inode->i_generation);
3949         if (S_ISCHR(inode->i_mode) || S_ISBLK(inode->i_mode)) {
3950                 if (old_valid_dev(inode->i_rdev)) {
3951                         raw_inode->i_block[0] =
3952                                 cpu_to_le32(old_encode_dev(inode->i_rdev));
3953                         raw_inode->i_block[1] = 0;
3954                 } else {
3955                         raw_inode->i_block[0] = 0;
3956                         raw_inode->i_block[1] =
3957                                 cpu_to_le32(new_encode_dev(inode->i_rdev));
3958                         raw_inode->i_block[2] = 0;
3959                 }
3960         } else
3961                 for (block = 0; block < EXT4_N_BLOCKS; block++)
3962                         raw_inode->i_block[block] = ei->i_data[block];
3963
3964         raw_inode->i_disk_version = cpu_to_le32(inode->i_version);
3965         if (ei->i_extra_isize) {
3966                 if (EXT4_FITS_IN_INODE(raw_inode, ei, i_version_hi))
3967                         raw_inode->i_version_hi =
3968                         cpu_to_le32(inode->i_version >> 32);
3969                 raw_inode->i_extra_isize = cpu_to_le16(ei->i_extra_isize);
3970         }
3971
3972         BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
3973         rc = ext4_handle_dirty_metadata(handle, NULL, bh);
3974         if (!err)
3975                 err = rc;
3976         ext4_clear_inode_state(inode, EXT4_STATE_NEW);
3977
3978         ext4_update_inode_fsync_trans(handle, inode, 0);
3979 out_brelse:
3980         brelse(bh);
3981         ext4_std_error(inode->i_sb, err);
3982         return err;
3983 }
3984
3985 /*
3986  * ext4_write_inode()
3987  *
3988  * We are called from a few places:
3989  *
3990  * - Within generic_file_write() for O_SYNC files.
3991  *   Here, there will be no transaction running. We wait for any running
3992  *   trasnaction to commit.
3993  *
3994  * - Within sys_sync(), kupdate and such.
3995  *   We wait on commit, if tol to.
3996  *
3997  * - Within prune_icache() (PF_MEMALLOC == true)
3998  *   Here we simply return.  We can't afford to block kswapd on the
3999  *   journal commit.
4000  *
4001  * In all cases it is actually safe for us to return without doing anything,
4002  * because the inode has been copied into a raw inode buffer in
4003  * ext4_mark_inode_dirty().  This is a correctness thing for O_SYNC and for
4004  * knfsd.
4005  *
4006  * Note that we are absolutely dependent upon all inode dirtiers doing the
4007  * right thing: they *must* call mark_inode_dirty() after dirtying info in
4008  * which we are interested.
4009  *
4010  * It would be a bug for them to not do this.  The code:
4011  *
4012  *      mark_inode_dirty(inode)
4013  *      stuff();
4014  *      inode->i_size = expr;
4015  *
4016  * is in error because a kswapd-driven write_inode() could occur while
4017  * `stuff()' is running, and the new i_size will be lost.  Plus the inode
4018  * will no longer be on the superblock's dirty inode list.
4019  */
4020 int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
4021 {
4022         int err;
4023
4024         if (current->flags & PF_MEMALLOC)
4025                 return 0;
4026
4027         if (EXT4_SB(inode->i_sb)->s_journal) {
4028                 if (ext4_journal_current_handle()) {
4029                         jbd_debug(1, "called recursively, non-PF_MEMALLOC!\n");
4030                         dump_stack();
4031                         return -EIO;
4032                 }
4033
4034                 if (wbc->sync_mode != WB_SYNC_ALL)
4035                         return 0;
4036
4037                 err = ext4_force_commit(inode->i_sb);
4038         } else {
4039                 struct ext4_iloc iloc;
4040
4041                 err = __ext4_get_inode_loc(inode, &iloc, 0);
4042                 if (err)
4043                         return err;
4044                 if (wbc->sync_mode == WB_SYNC_ALL)
4045                         sync_dirty_buffer(iloc.bh);
4046                 if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
4047                         EXT4_ERROR_INODE_BLOCK(inode, iloc.bh->b_blocknr,
4048                                          "IO error syncing inode");
4049                         err = -EIO;
4050                 }
4051                 brelse(iloc.bh);
4052         }
4053         return err;
4054 }
4055
4056 /*
4057  * ext4_setattr()
4058  *
4059  * Called from notify_change.
4060  *
4061  * We want to trap VFS attempts to truncate the file as soon as
4062  * possible.  In particular, we want to make sure that when the VFS
4063  * shrinks i_size, we put the inode on the orphan list and modify
4064  * i_disksize immediately, so that during the subsequent flushing of
4065  * dirty pages and freeing of disk blocks, we can guarantee that any
4066  * commit will leave the blocks being flushed in an unused state on
4067  * disk.  (On recovery, the inode will get truncated and the blocks will
4068  * be freed, so we have a strong guarantee that no future commit will
4069  * leave these blocks visible to the user.)
4070  *
4071  * Another thing we have to assure is that if we are in ordered mode
4072  * and inode is still attached to the committing transaction, we must
4073  * we start writeout of all the dirty pages which are being truncated.
4074  * This way we are sure that all the data written in the previous
4075  * transaction are already on disk (truncate waits for pages under
4076  * writeback).
4077  *
4078  * Called with inode->i_mutex down.
4079  */
4080 int ext4_setattr(struct dentry *dentry, struct iattr *attr)
4081 {
4082         struct inode *inode = dentry->d_inode;
4083         int error, rc = 0;
4084         int orphan = 0;
4085         const unsigned int ia_valid = attr->ia_valid;
4086
4087         error = inode_change_ok(inode, attr);
4088         if (error)
4089                 return error;
4090
4091         if (is_quota_modification(inode, attr))
4092                 dquot_initialize(inode);
4093         if ((ia_valid & ATTR_UID && !uid_eq(attr->ia_uid, inode->i_uid)) ||
4094             (ia_valid & ATTR_GID && !gid_eq(attr->ia_gid, inode->i_gid))) {
4095                 handle_t *handle;
4096
4097                 /* (user+group)*(old+new) structure, inode write (sb,
4098                  * inode block, ? - but truncate inode update has it) */
4099                 handle = ext4_journal_start(inode, (EXT4_MAXQUOTAS_INIT_BLOCKS(inode->i_sb)+
4100                                         EXT4_MAXQUOTAS_DEL_BLOCKS(inode->i_sb))+3);
4101                 if (IS_ERR(handle)) {
4102                         error = PTR_ERR(handle);
4103                         goto err_out;
4104                 }
4105                 error = dquot_transfer(inode, attr);
4106                 if (error) {
4107                         ext4_journal_stop(handle);
4108                         return error;
4109                 }
4110                 /* Update corresponding info in inode so that everything is in
4111                  * one transaction */
4112                 if (attr->ia_valid & ATTR_UID)
4113                         inode->i_uid = attr->ia_uid;
4114                 if (attr->ia_valid & ATTR_GID)
4115                         inode->i_gid = attr->ia_gid;
4116                 error = ext4_mark_inode_dirty(handle, inode);
4117                 ext4_journal_stop(handle);
4118         }
4119
4120         if (attr->ia_valid & ATTR_SIZE) {
4121                 inode_dio_wait(inode);
4122
4123                 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
4124                         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4125
4126                         if (attr->ia_size > sbi->s_bitmap_maxbytes)
4127                                 return -EFBIG;
4128                 }
4129         }
4130
4131         if (S_ISREG(inode->i_mode) &&
4132             attr->ia_valid & ATTR_SIZE &&
4133             (attr->ia_size < inode->i_size)) {
4134                 handle_t *handle;
4135
4136                 handle = ext4_journal_start(inode, 3);
4137                 if (IS_ERR(handle)) {
4138                         error = PTR_ERR(handle);
4139                         goto err_out;
4140                 }
4141                 if (ext4_handle_valid(handle)) {
4142                         error = ext4_orphan_add(handle, inode);
4143                         orphan = 1;
4144                 }
4145                 EXT4_I(inode)->i_disksize = attr->ia_size;
4146                 rc = ext4_mark_inode_dirty(handle, inode);
4147                 if (!error)
4148                         error = rc;
4149                 ext4_journal_stop(handle);
4150
4151                 if (ext4_should_order_data(inode)) {
4152                         error = ext4_begin_ordered_truncate(inode,
4153                                                             attr->ia_size);
4154                         if (error) {
4155                                 /* Do as much error cleanup as possible */
4156                                 handle = ext4_journal_start(inode, 3);
4157                                 if (IS_ERR(handle)) {
4158                                         ext4_orphan_del(NULL, inode);
4159                                         goto err_out;
4160                                 }
4161                                 ext4_orphan_del(handle, inode);
4162                                 orphan = 0;
4163                                 ext4_journal_stop(handle);
4164                                 goto err_out;
4165                         }
4166                 }
4167         }
4168
4169         if (attr->ia_valid & ATTR_SIZE) {
4170                 if (attr->ia_size != i_size_read(inode))
4171                         truncate_setsize(inode, attr->ia_size);
4172                 ext4_truncate(inode);
4173         }
4174
4175         if (!rc) {
4176                 setattr_copy(inode, attr);
4177                 mark_inode_dirty(inode);
4178         }
4179
4180         /*
4181          * If the call to ext4_truncate failed to get a transaction handle at
4182          * all, we need to clean up the in-core orphan list manually.
4183          */
4184         if (orphan && inode->i_nlink)
4185                 ext4_orphan_del(NULL, inode);
4186
4187         if (!rc && (ia_valid & ATTR_MODE))
4188                 rc = ext4_acl_chmod(inode);
4189
4190 err_out:
4191         ext4_std_error(inode->i_sb, error);
4192         if (!error)
4193                 error = rc;
4194         return error;
4195 }
4196
4197 int ext4_getattr(struct vfsmount *mnt, struct dentry *dentry,
4198                  struct kstat *stat)
4199 {
4200         struct inode *inode;
4201         unsigned long delalloc_blocks;
4202
4203         inode = dentry->d_inode;
4204         generic_fillattr(inode, stat);
4205
4206         /*
4207          * We can't update i_blocks if the block allocation is delayed
4208          * otherwise in the case of system crash before the real block
4209          * allocation is done, we will have i_blocks inconsistent with
4210          * on-disk file blocks.
4211          * We always keep i_blocks updated together with real
4212          * allocation. But to not confuse with user, stat
4213          * will return the blocks that include the delayed allocation
4214          * blocks for this file.
4215          */
4216         delalloc_blocks = EXT4_I(inode)->i_reserved_data_blocks;
4217
4218         stat->blocks += (delalloc_blocks << inode->i_sb->s_blocksize_bits)>>9;
4219         return 0;
4220 }
4221
4222 static int ext4_index_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4223 {
4224         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
4225                 return ext4_ind_trans_blocks(inode, nrblocks, chunk);
4226         return ext4_ext_index_trans_blocks(inode, nrblocks, chunk);
4227 }
4228
4229 /*
4230  * Account for index blocks, block groups bitmaps and block group
4231  * descriptor blocks if modify datablocks and index blocks
4232  * worse case, the indexs blocks spread over different block groups
4233  *
4234  * If datablocks are discontiguous, they are possible to spread over
4235  * different block groups too. If they are contiuguous, with flexbg,
4236  * they could still across block group boundary.
4237  *
4238  * Also account for superblock, inode, quota and xattr blocks
4239  */
4240 static int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
4241 {
4242         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
4243         int gdpblocks;
4244         int idxblocks;
4245         int ret = 0;
4246
4247         /*
4248          * How many index blocks need to touch to modify nrblocks?
4249          * The "Chunk" flag indicating whether the nrblocks is
4250          * physically contiguous on disk
4251          *
4252          * For Direct IO and fallocate, they calls get_block to allocate
4253          * one single extent at a time, so they could set the "Chunk" flag
4254          */
4255         idxblocks = ext4_index_trans_blocks(inode, nrblocks, chunk);
4256
4257         ret = idxblocks;
4258
4259         /*
4260          * Now let's see how many group bitmaps and group descriptors need
4261          * to account
4262          */
4263         groups = idxblocks;
4264         if (chunk)
4265                 groups += 1;
4266         else
4267                 groups += nrblocks;
4268
4269         gdpblocks = groups;
4270         if (groups > ngroups)
4271                 groups = ngroups;
4272         if (groups > EXT4_SB(inode->i_sb)->s_gdb_count)
4273                 gdpblocks = EXT4_SB(inode->i_sb)->s_gdb_count;
4274
4275         /* bitmaps and block group descriptor blocks */
4276         ret += groups + gdpblocks;
4277
4278         /* Blocks for super block, inode, quota and xattr blocks */
4279         ret += EXT4_META_TRANS_BLOCKS(inode->i_sb);
4280
4281         return ret;
4282 }
4283
4284 /*
4285  * Calculate the total number of credits to reserve to fit
4286  * the modification of a single pages into a single transaction,
4287  * which may include multiple chunks of block allocations.
4288  *
4289  * This could be called via ext4_write_begin()
4290  *
4291  * We need to consider the worse case, when
4292  * one new block per extent.
4293  */
4294 int ext4_writepage_trans_blocks(struct inode *inode)
4295 {
4296         int bpp = ext4_journal_blocks_per_page(inode);
4297         int ret;
4298
4299         ret = ext4_meta_trans_blocks(inode, bpp, 0);
4300
4301         /* Account for data blocks for journalled mode */
4302         if (ext4_should_journal_data(inode))
4303                 ret += bpp;
4304         return ret;
4305 }
4306
4307 /*
4308  * Calculate the journal credits for a chunk of data modification.
4309  *
4310  * This is called from DIO, fallocate or whoever calling
4311  * ext4_map_blocks() to map/allocate a chunk of contiguous disk blocks.
4312  *
4313  * journal buffers for data blocks are not included here, as DIO
4314  * and fallocate do no need to journal data buffers.
4315  */
4316 int ext4_chunk_trans_blocks(struct inode *inode, int nrblocks)
4317 {
4318         return ext4_meta_trans_blocks(inode, nrblocks, 1);
4319 }
4320
4321 /*
4322  * The caller must have previously called ext4_reserve_inode_write().
4323  * Give this, we know that the caller already has write access to iloc->bh.
4324  */
4325 int ext4_mark_iloc_dirty(handle_t *handle,
4326                          struct inode *inode, struct ext4_iloc *iloc)
4327 {
4328         int err = 0;
4329
4330         if (IS_I_VERSION(inode))
4331                 inode_inc_iversion(inode);
4332
4333         /* the do_update_inode consumes one bh->b_count */
4334         get_bh(iloc->bh);
4335
4336         /* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
4337         err = ext4_do_update_inode(handle, inode, iloc);
4338         put_bh(iloc->bh);
4339         return err;
4340 }
4341
4342 /*
4343  * On success, We end up with an outstanding reference count against
4344  * iloc->bh.  This _must_ be cleaned up later.
4345  */
4346
4347 int
4348 ext4_reserve_inode_write(handle_t *handle, struct inode *inode,
4349                          struct ext4_iloc *iloc)
4350 {
4351         int err;
4352
4353         err = ext4_get_inode_loc(inode, iloc);
4354         if (!err) {
4355                 BUFFER_TRACE(iloc->bh, "get_write_access");
4356                 err = ext4_journal_get_write_access(handle, iloc->bh);
4357                 if (err) {
4358                         brelse(iloc->bh);
4359                         iloc->bh = NULL;
4360                 }
4361         }
4362         ext4_std_error(inode->i_sb, err);
4363         return err;
4364 }
4365
4366 /*
4367  * Expand an inode by new_extra_isize bytes.
4368  * Returns 0 on success or negative error number on failure.
4369  */
4370 static int ext4_expand_extra_isize(struct inode *inode,
4371                                    unsigned int new_extra_isize,
4372                                    struct ext4_iloc iloc,
4373                                    handle_t *handle)
4374 {
4375         struct ext4_inode *raw_inode;
4376         struct ext4_xattr_ibody_header *header;
4377
4378         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize)
4379                 return 0;
4380
4381         raw_inode = ext4_raw_inode(&iloc);
4382
4383         header = IHDR(inode, raw_inode);
4384
4385         /* No extended attributes present */
4386         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR) ||
4387             header->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC)) {
4388                 memset((void *)raw_inode + EXT4_GOOD_OLD_INODE_SIZE, 0,
4389                         new_extra_isize);
4390                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
4391                 return 0;
4392         }
4393
4394         /* try to expand with EAs present */
4395         return ext4_expand_extra_isize_ea(inode, new_extra_isize,
4396                                           raw_inode, handle);
4397 }
4398
4399 /*
4400  * What we do here is to mark the in-core inode as clean with respect to inode
4401  * dirtiness (it may still be data-dirty).
4402  * This means that the in-core inode may be reaped by prune_icache
4403  * without having to perform any I/O.  This is a very good thing,
4404  * because *any* task may call prune_icache - even ones which
4405  * have a transaction open against a different journal.
4406  *
4407  * Is this cheating?  Not really.  Sure, we haven't written the
4408  * inode out, but prune_icache isn't a user-visible syncing function.
4409  * Whenever the user wants stuff synced (sys_sync, sys_msync, sys_fsync)
4410  * we start and wait on commits.
4411  *
4412  * Is this efficient/effective?  Well, we're being nice to the system
4413  * by cleaning up our inodes proactively so they can be reaped
4414  * without I/O.  But we are potentially leaving up to five seconds'
4415  * worth of inodes floating about which prune_icache wants us to
4416  * write out.  One way to fix that would be to get prune_icache()
4417  * to do a write_super() to free up some memory.  It has the desired
4418  * effect.
4419  */
4420 int ext4_mark_inode_dirty(handle_t *handle, struct inode *inode)
4421 {
4422         struct ext4_iloc iloc;
4423         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
4424         static unsigned int mnt_count;
4425         int err, ret;
4426
4427         might_sleep();
4428         trace_ext4_mark_inode_dirty(inode, _RET_IP_);
4429         err = ext4_reserve_inode_write(handle, inode, &iloc);
4430         if (ext4_handle_valid(handle) &&
4431             EXT4_I(inode)->i_extra_isize < sbi->s_want_extra_isize &&
4432             !ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND)) {
4433                 /*
4434                  * We need extra buffer credits since we may write into EA block
4435                  * with this same handle. If journal_extend fails, then it will
4436                  * only result in a minor loss of functionality for that inode.
4437                  * If this is felt to be critical, then e2fsck should be run to
4438                  * force a large enough s_min_extra_isize.
4439                  */
4440                 if ((jbd2_journal_extend(handle,
4441                              EXT4_DATA_TRANS_BLOCKS(inode->i_sb))) == 0) {
4442                         ret = ext4_expand_extra_isize(inode,
4443                                                       sbi->s_want_extra_isize,
4444                                                       iloc, handle);
4445                         if (ret) {
4446                                 ext4_set_inode_state(inode,
4447                                                      EXT4_STATE_NO_EXPAND);
4448                                 if (mnt_count !=
4449                                         le16_to_cpu(sbi->s_es->s_mnt_count)) {
4450                                         ext4_warning(inode->i_sb,
4451                                         "Unable to expand inode %lu. Delete"
4452                                         " some EAs or run e2fsck.",
4453                                         inode->i_ino);
4454                                         mnt_count =
4455                                           le16_to_cpu(sbi->s_es->s_mnt_count);
4456                                 }
4457                         }
4458                 }
4459         }
4460         if (!err)
4461                 err = ext4_mark_iloc_dirty(handle, inode, &iloc);
4462         return err;
4463 }
4464
4465 /*
4466  * ext4_dirty_inode() is called from __mark_inode_dirty()
4467  *
4468  * We're really interested in the case where a file is being extended.
4469  * i_size has been changed by generic_commit_write() and we thus need
4470  * to include the updated inode in the current transaction.
4471  *
4472  * Also, dquot_alloc_block() will always dirty the inode when blocks
4473  * are allocated to the file.
4474  *
4475  * If the inode is marked synchronous, we don't honour that here - doing
4476  * so would cause a commit on atime updates, which we don't bother doing.
4477  * We handle synchronous inodes at the highest possible level.
4478  */
4479 void ext4_dirty_inode(struct inode *inode, int flags)
4480 {
4481         handle_t *handle;
4482
4483         handle = ext4_journal_start(inode, 2);
4484         if (IS_ERR(handle))
4485                 goto out;
4486
4487         ext4_mark_inode_dirty(handle, inode);
4488
4489         ext4_journal_stop(handle);
4490 out:
4491         return;
4492 }
4493
4494 #if 0
4495 /*
4496  * Bind an inode's backing buffer_head into this transaction, to prevent
4497  * it from being flushed to disk early.  Unlike
4498  * ext4_reserve_inode_write, this leaves behind no bh reference and
4499  * returns no iloc structure, so the caller needs to repeat the iloc
4500  * lookup to mark the inode dirty later.
4501  */
4502 static int ext4_pin_inode(handle_t *handle, struct inode *inode)
4503 {
4504         struct ext4_iloc iloc;
4505
4506         int err = 0;
4507         if (handle) {
4508                 err = ext4_get_inode_loc(inode, &iloc);
4509                 if (!err) {
4510                         BUFFER_TRACE(iloc.bh, "get_write_access");
4511                         err = jbd2_journal_get_write_access(handle, iloc.bh);
4512                         if (!err)
4513                                 err = ext4_handle_dirty_metadata(handle,
4514                                                                  NULL,
4515                                                                  iloc.bh);
4516                         brelse(iloc.bh);
4517                 }
4518         }
4519         ext4_std_error(inode->i_sb, err);
4520         return err;
4521 }
4522 #endif
4523
4524 int ext4_change_inode_journal_flag(struct inode *inode, int val)
4525 {
4526         journal_t *journal;
4527         handle_t *handle;
4528         int err;
4529
4530         /*
4531          * We have to be very careful here: changing a data block's
4532          * journaling status dynamically is dangerous.  If we write a
4533          * data block to the journal, change the status and then delete
4534          * that block, we risk forgetting to revoke the old log record
4535          * from the journal and so a subsequent replay can corrupt data.
4536          * So, first we make sure that the journal is empty and that
4537          * nobody is changing anything.
4538          */
4539
4540         journal = EXT4_JOURNAL(inode);
4541         if (!journal)
4542                 return 0;
4543         if (is_journal_aborted(journal))
4544                 return -EROFS;
4545         /* We have to allocate physical blocks for delalloc blocks
4546          * before flushing journal. otherwise delalloc blocks can not
4547          * be allocated any more. even more truncate on delalloc blocks
4548          * could trigger BUG by flushing delalloc blocks in journal.
4549          * There is no delalloc block in non-journal data mode.
4550          */
4551         if (val && test_opt(inode->i_sb, DELALLOC)) {
4552                 err = ext4_alloc_da_blocks(inode);
4553                 if (err < 0)
4554                         return err;
4555         }
4556
4557         jbd2_journal_lock_updates(journal);
4558
4559         /*
4560          * OK, there are no updates running now, and all cached data is
4561          * synced to disk.  We are now in a completely consistent state
4562          * which doesn't have anything in the journal, and we know that
4563          * no filesystem updates are running, so it is safe to modify
4564          * the inode's in-core data-journaling state flag now.
4565          */
4566
4567         if (val)
4568                 ext4_set_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4569         else {
4570                 jbd2_journal_flush(journal);
4571                 ext4_clear_inode_flag(inode, EXT4_INODE_JOURNAL_DATA);
4572         }
4573         ext4_set_aops(inode);
4574
4575         jbd2_journal_unlock_updates(journal);
4576
4577         /* Finally we can mark the inode as dirty. */
4578
4579         handle = ext4_journal_start(inode, 1);
4580         if (IS_ERR(handle))
4581                 return PTR_ERR(handle);
4582
4583         err = ext4_mark_inode_dirty(handle, inode);
4584         ext4_handle_sync(handle);
4585         ext4_journal_stop(handle);
4586         ext4_std_error(inode->i_sb, err);
4587
4588         return err;
4589 }
4590
4591 static int ext4_bh_unmapped(handle_t *handle, struct buffer_head *bh)
4592 {
4593         return !buffer_mapped(bh);
4594 }
4595
4596 int ext4_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
4597 {
4598         struct page *page = vmf->page;
4599         loff_t size;
4600         unsigned long len;
4601         int ret;
4602         struct file *file = vma->vm_file;
4603         struct inode *inode = file->f_path.dentry->d_inode;
4604         struct address_space *mapping = inode->i_mapping;
4605         handle_t *handle;
4606         get_block_t *get_block;
4607         int retries = 0;
4608
4609         /*
4610          * This check is racy but catches the common case. We rely on
4611          * __block_page_mkwrite() to do a reliable check.
4612          */
4613         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
4614         /* Delalloc case is easy... */
4615         if (test_opt(inode->i_sb, DELALLOC) &&
4616             !ext4_should_journal_data(inode) &&
4617             !ext4_nonda_switch(inode->i_sb)) {
4618                 do {
4619                         ret = __block_page_mkwrite(vma, vmf,
4620                                                    ext4_da_get_block_prep);
4621                 } while (ret == -ENOSPC &&
4622                        ext4_should_retry_alloc(inode->i_sb, &retries));
4623                 goto out_ret;
4624         }
4625
4626         lock_page(page);
4627         size = i_size_read(inode);
4628         /* Page got truncated from under us? */
4629         if (page->mapping != mapping || page_offset(page) > size) {
4630                 unlock_page(page);
4631                 ret = VM_FAULT_NOPAGE;
4632                 goto out;
4633         }
4634
4635         if (page->index == size >> PAGE_CACHE_SHIFT)
4636                 len = size & ~PAGE_CACHE_MASK;
4637         else
4638                 len = PAGE_CACHE_SIZE;
4639         /*
4640          * Return if we have all the buffers mapped. This avoids the need to do
4641          * journal_start/journal_stop which can block and take a long time
4642          */
4643         if (page_has_buffers(page)) {
4644                 if (!walk_page_buffers(NULL, page_buffers(page), 0, len, NULL,
4645                                         ext4_bh_unmapped)) {
4646                         /* Wait so that we don't change page under IO */
4647                         wait_on_page_writeback(page);
4648                         ret = VM_FAULT_LOCKED;
4649                         goto out;
4650                 }
4651         }
4652         unlock_page(page);
4653         /* OK, we need to fill the hole... */
4654         if (ext4_should_dioread_nolock(inode))
4655                 get_block = ext4_get_block_write;
4656         else
4657                 get_block = ext4_get_block;
4658 retry_alloc:
4659         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
4660         if (IS_ERR(handle)) {
4661                 ret = VM_FAULT_SIGBUS;
4662                 goto out;
4663         }
4664         ret = __block_page_mkwrite(vma, vmf, get_block);
4665         if (!ret && ext4_should_journal_data(inode)) {
4666                 if (walk_page_buffers(handle, page_buffers(page), 0,
4667                           PAGE_CACHE_SIZE, NULL, do_journal_get_write_access)) {
4668                         unlock_page(page);
4669                         ret = VM_FAULT_SIGBUS;
4670                         ext4_journal_stop(handle);
4671                         goto out;
4672                 }
4673                 ext4_set_inode_state(inode, EXT4_STATE_JDATA);
4674         }
4675         ext4_journal_stop(handle);
4676         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
4677                 goto retry_alloc;
4678 out_ret:
4679         ret = block_page_mkwrite_return(ret);
4680 out:
4681         return ret;
4682 }