Merge branch 'dev' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4
[pandora-kernel.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  64-bit file support on 64-bit platforms by Jakub Jelinek
16  *      (jj@sunsite.ms.mff.cuni.cz)
17  *
18  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
19  */
20
21 #include <linux/module.h>
22 #include <linux/fs.h>
23 #include <linux/time.h>
24 #include <linux/jbd2.h>
25 #include <linux/highuid.h>
26 #include <linux/pagemap.h>
27 #include <linux/quotaops.h>
28 #include <linux/string.h>
29 #include <linux/buffer_head.h>
30 #include <linux/writeback.h>
31 #include <linux/pagevec.h>
32 #include <linux/mpage.h>
33 #include <linux/namei.h>
34 #include <linux/uio.h>
35 #include <linux/bio.h>
36 #include <linux/workqueue.h>
37 #include <linux/kernel.h>
38 #include <linux/printk.h>
39 #include <linux/slab.h>
40 #include <linux/ratelimit.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "truncate.h"
46
47 #include <trace/events/ext4.h>
48
49 #define MPAGE_DA_EXTENT_TAIL 0x01
50
51 static inline int ext4_begin_ordered_truncate(struct inode *inode,
52                                               loff_t new_size)
53 {
54         trace_ext4_begin_ordered_truncate(inode, new_size);
55         /*
56          * If jinode is zero, then we never opened the file for
57          * writing, so there's no need to call
58          * jbd2_journal_begin_ordered_truncate() since there's no
59          * outstanding writes we need to flush.
60          */
61         if (!EXT4_I(inode)->jinode)
62                 return 0;
63         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
64                                                    EXT4_I(inode)->jinode,
65                                                    new_size);
66 }
67
68 static void ext4_invalidatepage(struct page *page, unsigned long offset);
69 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
70                                    struct buffer_head *bh_result, int create);
71 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
72 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
73 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
74 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
75
76 /*
77  * Test whether an inode is a fast symlink.
78  */
79 static int ext4_inode_is_fast_symlink(struct inode *inode)
80 {
81         int ea_blocks = EXT4_I(inode)->i_file_acl ?
82                 (inode->i_sb->s_blocksize >> 9) : 0;
83
84         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
85 }
86
87 /*
88  * Restart the transaction associated with *handle.  This does a commit,
89  * so before we call here everything must be consistently dirtied against
90  * this transaction.
91  */
92 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
93                                  int nblocks)
94 {
95         int ret;
96
97         /*
98          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
99          * moment, get_block can be called only for blocks inside i_size since
100          * page cache has been already dropped and writes are blocked by
101          * i_mutex. So we can safely drop the i_data_sem here.
102          */
103         BUG_ON(EXT4_JOURNAL(inode) == NULL);
104         jbd_debug(2, "restarting handle %p\n", handle);
105         up_write(&EXT4_I(inode)->i_data_sem);
106         ret = ext4_journal_restart(handle, nblocks);
107         down_write(&EXT4_I(inode)->i_data_sem);
108         ext4_discard_preallocations(inode);
109
110         return ret;
111 }
112
113 /*
114  * Called at the last iput() if i_nlink is zero.
115  */
116 void ext4_evict_inode(struct inode *inode)
117 {
118         handle_t *handle;
119         int err;
120
121         trace_ext4_evict_inode(inode);
122
123         ext4_ioend_wait(inode);
124
125         if (inode->i_nlink) {
126                 /*
127                  * When journalling data dirty buffers are tracked only in the
128                  * journal. So although mm thinks everything is clean and
129                  * ready for reaping the inode might still have some pages to
130                  * write in the running transaction or waiting to be
131                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
132                  * (via truncate_inode_pages()) to discard these buffers can
133                  * cause data loss. Also even if we did not discard these
134                  * buffers, we would have no way to find them after the inode
135                  * is reaped and thus user could see stale data if he tries to
136                  * read them before the transaction is checkpointed. So be
137                  * careful and force everything to disk here... We use
138                  * ei->i_datasync_tid to store the newest transaction
139                  * containing inode's data.
140                  *
141                  * Note that directories do not have this problem because they
142                  * don't use page cache.
143                  */
144                 if (ext4_should_journal_data(inode) &&
145                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
146                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
147                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
148
149                         jbd2_log_start_commit(journal, commit_tid);
150                         jbd2_log_wait_commit(journal, commit_tid);
151                         filemap_write_and_wait(&inode->i_data);
152                 }
153                 truncate_inode_pages(&inode->i_data, 0);
154                 goto no_delete;
155         }
156
157         if (!is_bad_inode(inode))
158                 dquot_initialize(inode);
159
160         if (ext4_should_order_data(inode))
161                 ext4_begin_ordered_truncate(inode, 0);
162         truncate_inode_pages(&inode->i_data, 0);
163
164         if (is_bad_inode(inode))
165                 goto no_delete;
166
167         handle = ext4_journal_start(inode, ext4_blocks_for_truncate(inode)+3);
168         if (IS_ERR(handle)) {
169                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
170                 /*
171                  * If we're going to skip the normal cleanup, we still need to
172                  * make sure that the in-core orphan linked list is properly
173                  * cleaned up.
174                  */
175                 ext4_orphan_del(NULL, inode);
176                 goto no_delete;
177         }
178
179         if (IS_SYNC(inode))
180                 ext4_handle_sync(handle);
181         inode->i_size = 0;
182         err = ext4_mark_inode_dirty(handle, inode);
183         if (err) {
184                 ext4_warning(inode->i_sb,
185                              "couldn't mark inode dirty (err %d)", err);
186                 goto stop_handle;
187         }
188         if (inode->i_blocks)
189                 ext4_truncate(inode);
190
191         /*
192          * ext4_ext_truncate() doesn't reserve any slop when it
193          * restarts journal transactions; therefore there may not be
194          * enough credits left in the handle to remove the inode from
195          * the orphan list and set the dtime field.
196          */
197         if (!ext4_handle_has_enough_credits(handle, 3)) {
198                 err = ext4_journal_extend(handle, 3);
199                 if (err > 0)
200                         err = ext4_journal_restart(handle, 3);
201                 if (err != 0) {
202                         ext4_warning(inode->i_sb,
203                                      "couldn't extend journal (err %d)", err);
204                 stop_handle:
205                         ext4_journal_stop(handle);
206                         ext4_orphan_del(NULL, inode);
207                         goto no_delete;
208                 }
209         }
210
211         /*
212          * Kill off the orphan record which ext4_truncate created.
213          * AKPM: I think this can be inside the above `if'.
214          * Note that ext4_orphan_del() has to be able to cope with the
215          * deletion of a non-existent orphan - this is because we don't
216          * know if ext4_truncate() actually created an orphan record.
217          * (Well, we could do this if we need to, but heck - it works)
218          */
219         ext4_orphan_del(handle, inode);
220         EXT4_I(inode)->i_dtime  = get_seconds();
221
222         /*
223          * One subtle ordering requirement: if anything has gone wrong
224          * (transaction abort, IO errors, whatever), then we can still
225          * do these next steps (the fs will already have been marked as
226          * having errors), but we can't free the inode if the mark_dirty
227          * fails.
228          */
229         if (ext4_mark_inode_dirty(handle, inode))
230                 /* If that failed, just do the required in-core inode clear. */
231                 ext4_clear_inode(inode);
232         else
233                 ext4_free_inode(handle, inode);
234         ext4_journal_stop(handle);
235         return;
236 no_delete:
237         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
238 }
239
240 #ifdef CONFIG_QUOTA
241 qsize_t *ext4_get_reserved_space(struct inode *inode)
242 {
243         return &EXT4_I(inode)->i_reserved_quota;
244 }
245 #endif
246
247 /*
248  * Calculate the number of metadata blocks need to reserve
249  * to allocate a block located at @lblock
250  */
251 static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
252 {
253         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
254                 return ext4_ext_calc_metadata_amount(inode, lblock);
255
256         return ext4_ind_calc_metadata_amount(inode, lblock);
257 }
258
259 /*
260  * Called with i_data_sem down, which is important since we can call
261  * ext4_discard_preallocations() from here.
262  */
263 void ext4_da_update_reserve_space(struct inode *inode,
264                                         int used, int quota_claim)
265 {
266         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
267         struct ext4_inode_info *ei = EXT4_I(inode);
268
269         spin_lock(&ei->i_block_reservation_lock);
270         trace_ext4_da_update_reserve_space(inode, used, quota_claim);
271         if (unlikely(used > ei->i_reserved_data_blocks)) {
272                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
273                          "with only %d reserved data blocks\n",
274                          __func__, inode->i_ino, used,
275                          ei->i_reserved_data_blocks);
276                 WARN_ON(1);
277                 used = ei->i_reserved_data_blocks;
278         }
279
280         /* Update per-inode reservations */
281         ei->i_reserved_data_blocks -= used;
282         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
283         percpu_counter_sub(&sbi->s_dirtyclusters_counter,
284                            used + ei->i_allocated_meta_blocks);
285         ei->i_allocated_meta_blocks = 0;
286
287         if (ei->i_reserved_data_blocks == 0) {
288                 /*
289                  * We can release all of the reserved metadata blocks
290                  * only when we have written all of the delayed
291                  * allocation blocks.
292                  */
293                 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
294                                    ei->i_reserved_meta_blocks);
295                 ei->i_reserved_meta_blocks = 0;
296                 ei->i_da_metadata_calc_len = 0;
297         }
298         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
299
300         /* Update quota subsystem for data blocks */
301         if (quota_claim)
302                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
303         else {
304                 /*
305                  * We did fallocate with an offset that is already delayed
306                  * allocated. So on delayed allocated writeback we should
307                  * not re-claim the quota for fallocated blocks.
308                  */
309                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
310         }
311
312         /*
313          * If we have done all the pending block allocations and if
314          * there aren't any writers on the inode, we can discard the
315          * inode's preallocations.
316          */
317         if ((ei->i_reserved_data_blocks == 0) &&
318             (atomic_read(&inode->i_writecount) == 0))
319                 ext4_discard_preallocations(inode);
320 }
321
322 static int __check_block_validity(struct inode *inode, const char *func,
323                                 unsigned int line,
324                                 struct ext4_map_blocks *map)
325 {
326         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
327                                    map->m_len)) {
328                 ext4_error_inode(inode, func, line, map->m_pblk,
329                                  "lblock %lu mapped to illegal pblock "
330                                  "(length %d)", (unsigned long) map->m_lblk,
331                                  map->m_len);
332                 return -EIO;
333         }
334         return 0;
335 }
336
337 #define check_block_validity(inode, map)        \
338         __check_block_validity((inode), __func__, __LINE__, (map))
339
340 /*
341  * Return the number of contiguous dirty pages in a given inode
342  * starting at page frame idx.
343  */
344 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
345                                     unsigned int max_pages)
346 {
347         struct address_space *mapping = inode->i_mapping;
348         pgoff_t index;
349         struct pagevec pvec;
350         pgoff_t num = 0;
351         int i, nr_pages, done = 0;
352
353         if (max_pages == 0)
354                 return 0;
355         pagevec_init(&pvec, 0);
356         while (!done) {
357                 index = idx;
358                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
359                                               PAGECACHE_TAG_DIRTY,
360                                               (pgoff_t)PAGEVEC_SIZE);
361                 if (nr_pages == 0)
362                         break;
363                 for (i = 0; i < nr_pages; i++) {
364                         struct page *page = pvec.pages[i];
365                         struct buffer_head *bh, *head;
366
367                         lock_page(page);
368                         if (unlikely(page->mapping != mapping) ||
369                             !PageDirty(page) ||
370                             PageWriteback(page) ||
371                             page->index != idx) {
372                                 done = 1;
373                                 unlock_page(page);
374                                 break;
375                         }
376                         if (page_has_buffers(page)) {
377                                 bh = head = page_buffers(page);
378                                 do {
379                                         if (!buffer_delay(bh) &&
380                                             !buffer_unwritten(bh))
381                                                 done = 1;
382                                         bh = bh->b_this_page;
383                                 } while (!done && (bh != head));
384                         }
385                         unlock_page(page);
386                         if (done)
387                                 break;
388                         idx++;
389                         num++;
390                         if (num >= max_pages) {
391                                 done = 1;
392                                 break;
393                         }
394                 }
395                 pagevec_release(&pvec);
396         }
397         return num;
398 }
399
400 /*
401  * Sets the BH_Da_Mapped bit on the buffer heads corresponding to the given map.
402  */
403 static void set_buffers_da_mapped(struct inode *inode,
404                                    struct ext4_map_blocks *map)
405 {
406         struct address_space *mapping = inode->i_mapping;
407         struct pagevec pvec;
408         int i, nr_pages;
409         pgoff_t index, end;
410
411         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
412         end = (map->m_lblk + map->m_len - 1) >>
413                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
414
415         pagevec_init(&pvec, 0);
416         while (index <= end) {
417                 nr_pages = pagevec_lookup(&pvec, mapping, index,
418                                           min(end - index + 1,
419                                               (pgoff_t)PAGEVEC_SIZE));
420                 if (nr_pages == 0)
421                         break;
422                 for (i = 0; i < nr_pages; i++) {
423                         struct page *page = pvec.pages[i];
424                         struct buffer_head *bh, *head;
425
426                         if (unlikely(page->mapping != mapping) ||
427                             !PageDirty(page))
428                                 break;
429
430                         if (page_has_buffers(page)) {
431                                 bh = head = page_buffers(page);
432                                 do {
433                                         set_buffer_da_mapped(bh);
434                                         bh = bh->b_this_page;
435                                 } while (bh != head);
436                         }
437                         index++;
438                 }
439                 pagevec_release(&pvec);
440         }
441 }
442
443 /*
444  * The ext4_map_blocks() function tries to look up the requested blocks,
445  * and returns if the blocks are already mapped.
446  *
447  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
448  * and store the allocated blocks in the result buffer head and mark it
449  * mapped.
450  *
451  * If file type is extents based, it will call ext4_ext_map_blocks(),
452  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
453  * based files
454  *
455  * On success, it returns the number of blocks being mapped or allocate.
456  * if create==0 and the blocks are pre-allocated and uninitialized block,
457  * the result buffer head is unmapped. If the create ==1, it will make sure
458  * the buffer head is mapped.
459  *
460  * It returns 0 if plain look up failed (blocks have not been allocated), in
461  * that case, buffer head is unmapped
462  *
463  * It returns the error in case of allocation failure.
464  */
465 int ext4_map_blocks(handle_t *handle, struct inode *inode,
466                     struct ext4_map_blocks *map, int flags)
467 {
468         int retval;
469
470         map->m_flags = 0;
471         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
472                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
473                   (unsigned long) map->m_lblk);
474         /*
475          * Try to see if we can get the block without requesting a new
476          * file system block.
477          */
478         down_read((&EXT4_I(inode)->i_data_sem));
479         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
480                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
481                                              EXT4_GET_BLOCKS_KEEP_SIZE);
482         } else {
483                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
484                                              EXT4_GET_BLOCKS_KEEP_SIZE);
485         }
486         up_read((&EXT4_I(inode)->i_data_sem));
487
488         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
489                 int ret = check_block_validity(inode, map);
490                 if (ret != 0)
491                         return ret;
492         }
493
494         /* If it is only a block(s) look up */
495         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
496                 return retval;
497
498         /*
499          * Returns if the blocks have already allocated
500          *
501          * Note that if blocks have been preallocated
502          * ext4_ext_get_block() returns the create = 0
503          * with buffer head unmapped.
504          */
505         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
506                 return retval;
507
508         /*
509          * When we call get_blocks without the create flag, the
510          * BH_Unwritten flag could have gotten set if the blocks
511          * requested were part of a uninitialized extent.  We need to
512          * clear this flag now that we are committed to convert all or
513          * part of the uninitialized extent to be an initialized
514          * extent.  This is because we need to avoid the combination
515          * of BH_Unwritten and BH_Mapped flags being simultaneously
516          * set on the buffer_head.
517          */
518         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
519
520         /*
521          * New blocks allocate and/or writing to uninitialized extent
522          * will possibly result in updating i_data, so we take
523          * the write lock of i_data_sem, and call get_blocks()
524          * with create == 1 flag.
525          */
526         down_write((&EXT4_I(inode)->i_data_sem));
527
528         /*
529          * if the caller is from delayed allocation writeout path
530          * we have already reserved fs blocks for allocation
531          * let the underlying get_block() function know to
532          * avoid double accounting
533          */
534         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
535                 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
536         /*
537          * We need to check for EXT4 here because migrate
538          * could have changed the inode type in between
539          */
540         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
541                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
542         } else {
543                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
544
545                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
546                         /*
547                          * We allocated new blocks which will result in
548                          * i_data's format changing.  Force the migrate
549                          * to fail by clearing migrate flags
550                          */
551                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
552                 }
553
554                 /*
555                  * Update reserved blocks/metadata blocks after successful
556                  * block allocation which had been deferred till now. We don't
557                  * support fallocate for non extent files. So we can update
558                  * reserve space here.
559                  */
560                 if ((retval > 0) &&
561                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
562                         ext4_da_update_reserve_space(inode, retval, 1);
563         }
564         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
565                 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
566
567                 /* If we have successfully mapped the delayed allocated blocks,
568                  * set the BH_Da_Mapped bit on them. Its important to do this
569                  * under the protection of i_data_sem.
570                  */
571                 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
572                         set_buffers_da_mapped(inode, map);
573         }
574
575         up_write((&EXT4_I(inode)->i_data_sem));
576         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
577                 int ret = check_block_validity(inode, map);
578                 if (ret != 0)
579                         return ret;
580         }
581         return retval;
582 }
583
584 /* Maximum number of blocks we map for direct IO at once. */
585 #define DIO_MAX_BLOCKS 4096
586
587 static int _ext4_get_block(struct inode *inode, sector_t iblock,
588                            struct buffer_head *bh, int flags)
589 {
590         handle_t *handle = ext4_journal_current_handle();
591         struct ext4_map_blocks map;
592         int ret = 0, started = 0;
593         int dio_credits;
594
595         map.m_lblk = iblock;
596         map.m_len = bh->b_size >> inode->i_blkbits;
597
598         if (flags && !handle) {
599                 /* Direct IO write... */
600                 if (map.m_len > DIO_MAX_BLOCKS)
601                         map.m_len = DIO_MAX_BLOCKS;
602                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
603                 handle = ext4_journal_start(inode, dio_credits);
604                 if (IS_ERR(handle)) {
605                         ret = PTR_ERR(handle);
606                         return ret;
607                 }
608                 started = 1;
609         }
610
611         ret = ext4_map_blocks(handle, inode, &map, flags);
612         if (ret > 0) {
613                 map_bh(bh, inode->i_sb, map.m_pblk);
614                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
615                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
616                 ret = 0;
617         }
618         if (started)
619                 ext4_journal_stop(handle);
620         return ret;
621 }
622
623 int ext4_get_block(struct inode *inode, sector_t iblock,
624                    struct buffer_head *bh, int create)
625 {
626         return _ext4_get_block(inode, iblock, bh,
627                                create ? EXT4_GET_BLOCKS_CREATE : 0);
628 }
629
630 /*
631  * `handle' can be NULL if create is zero
632  */
633 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
634                                 ext4_lblk_t block, int create, int *errp)
635 {
636         struct ext4_map_blocks map;
637         struct buffer_head *bh;
638         int fatal = 0, err;
639
640         J_ASSERT(handle != NULL || create == 0);
641
642         map.m_lblk = block;
643         map.m_len = 1;
644         err = ext4_map_blocks(handle, inode, &map,
645                               create ? EXT4_GET_BLOCKS_CREATE : 0);
646
647         if (err < 0)
648                 *errp = err;
649         if (err <= 0)
650                 return NULL;
651         *errp = 0;
652
653         bh = sb_getblk(inode->i_sb, map.m_pblk);
654         if (!bh) {
655                 *errp = -EIO;
656                 return NULL;
657         }
658         if (map.m_flags & EXT4_MAP_NEW) {
659                 J_ASSERT(create != 0);
660                 J_ASSERT(handle != NULL);
661
662                 /*
663                  * Now that we do not always journal data, we should
664                  * keep in mind whether this should always journal the
665                  * new buffer as metadata.  For now, regular file
666                  * writes use ext4_get_block instead, so it's not a
667                  * problem.
668                  */
669                 lock_buffer(bh);
670                 BUFFER_TRACE(bh, "call get_create_access");
671                 fatal = ext4_journal_get_create_access(handle, bh);
672                 if (!fatal && !buffer_uptodate(bh)) {
673                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
674                         set_buffer_uptodate(bh);
675                 }
676                 unlock_buffer(bh);
677                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
678                 err = ext4_handle_dirty_metadata(handle, inode, bh);
679                 if (!fatal)
680                         fatal = err;
681         } else {
682                 BUFFER_TRACE(bh, "not a new buffer");
683         }
684         if (fatal) {
685                 *errp = fatal;
686                 brelse(bh);
687                 bh = NULL;
688         }
689         return bh;
690 }
691
692 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
693                                ext4_lblk_t block, int create, int *err)
694 {
695         struct buffer_head *bh;
696
697         bh = ext4_getblk(handle, inode, block, create, err);
698         if (!bh)
699                 return bh;
700         if (buffer_uptodate(bh))
701                 return bh;
702         ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
703         wait_on_buffer(bh);
704         if (buffer_uptodate(bh))
705                 return bh;
706         put_bh(bh);
707         *err = -EIO;
708         return NULL;
709 }
710
711 static int walk_page_buffers(handle_t *handle,
712                              struct buffer_head *head,
713                              unsigned from,
714                              unsigned to,
715                              int *partial,
716                              int (*fn)(handle_t *handle,
717                                        struct buffer_head *bh))
718 {
719         struct buffer_head *bh;
720         unsigned block_start, block_end;
721         unsigned blocksize = head->b_size;
722         int err, ret = 0;
723         struct buffer_head *next;
724
725         for (bh = head, block_start = 0;
726              ret == 0 && (bh != head || !block_start);
727              block_start = block_end, bh = next) {
728                 next = bh->b_this_page;
729                 block_end = block_start + blocksize;
730                 if (block_end <= from || block_start >= to) {
731                         if (partial && !buffer_uptodate(bh))
732                                 *partial = 1;
733                         continue;
734                 }
735                 err = (*fn)(handle, bh);
736                 if (!ret)
737                         ret = err;
738         }
739         return ret;
740 }
741
742 /*
743  * To preserve ordering, it is essential that the hole instantiation and
744  * the data write be encapsulated in a single transaction.  We cannot
745  * close off a transaction and start a new one between the ext4_get_block()
746  * and the commit_write().  So doing the jbd2_journal_start at the start of
747  * prepare_write() is the right place.
748  *
749  * Also, this function can nest inside ext4_writepage() ->
750  * block_write_full_page(). In that case, we *know* that ext4_writepage()
751  * has generated enough buffer credits to do the whole page.  So we won't
752  * block on the journal in that case, which is good, because the caller may
753  * be PF_MEMALLOC.
754  *
755  * By accident, ext4 can be reentered when a transaction is open via
756  * quota file writes.  If we were to commit the transaction while thus
757  * reentered, there can be a deadlock - we would be holding a quota
758  * lock, and the commit would never complete if another thread had a
759  * transaction open and was blocking on the quota lock - a ranking
760  * violation.
761  *
762  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
763  * will _not_ run commit under these circumstances because handle->h_ref
764  * is elevated.  We'll still have enough credits for the tiny quotafile
765  * write.
766  */
767 static int do_journal_get_write_access(handle_t *handle,
768                                        struct buffer_head *bh)
769 {
770         int dirty = buffer_dirty(bh);
771         int ret;
772
773         if (!buffer_mapped(bh) || buffer_freed(bh))
774                 return 0;
775         /*
776          * __block_write_begin() could have dirtied some buffers. Clean
777          * the dirty bit as jbd2_journal_get_write_access() could complain
778          * otherwise about fs integrity issues. Setting of the dirty bit
779          * by __block_write_begin() isn't a real problem here as we clear
780          * the bit before releasing a page lock and thus writeback cannot
781          * ever write the buffer.
782          */
783         if (dirty)
784                 clear_buffer_dirty(bh);
785         ret = ext4_journal_get_write_access(handle, bh);
786         if (!ret && dirty)
787                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
788         return ret;
789 }
790
791 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
792                    struct buffer_head *bh_result, int create);
793 static int ext4_write_begin(struct file *file, struct address_space *mapping,
794                             loff_t pos, unsigned len, unsigned flags,
795                             struct page **pagep, void **fsdata)
796 {
797         struct inode *inode = mapping->host;
798         int ret, needed_blocks;
799         handle_t *handle;
800         int retries = 0;
801         struct page *page;
802         pgoff_t index;
803         unsigned from, to;
804
805         trace_ext4_write_begin(inode, pos, len, flags);
806         /*
807          * Reserve one block more for addition to orphan list in case
808          * we allocate blocks but write fails for some reason
809          */
810         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
811         index = pos >> PAGE_CACHE_SHIFT;
812         from = pos & (PAGE_CACHE_SIZE - 1);
813         to = from + len;
814
815 retry:
816         handle = ext4_journal_start(inode, needed_blocks);
817         if (IS_ERR(handle)) {
818                 ret = PTR_ERR(handle);
819                 goto out;
820         }
821
822         /* We cannot recurse into the filesystem as the transaction is already
823          * started */
824         flags |= AOP_FLAG_NOFS;
825
826         page = grab_cache_page_write_begin(mapping, index, flags);
827         if (!page) {
828                 ext4_journal_stop(handle);
829                 ret = -ENOMEM;
830                 goto out;
831         }
832         *pagep = page;
833
834         if (ext4_should_dioread_nolock(inode))
835                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
836         else
837                 ret = __block_write_begin(page, pos, len, ext4_get_block);
838
839         if (!ret && ext4_should_journal_data(inode)) {
840                 ret = walk_page_buffers(handle, page_buffers(page),
841                                 from, to, NULL, do_journal_get_write_access);
842         }
843
844         if (ret) {
845                 unlock_page(page);
846                 page_cache_release(page);
847                 /*
848                  * __block_write_begin may have instantiated a few blocks
849                  * outside i_size.  Trim these off again. Don't need
850                  * i_size_read because we hold i_mutex.
851                  *
852                  * Add inode to orphan list in case we crash before
853                  * truncate finishes
854                  */
855                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
856                         ext4_orphan_add(handle, inode);
857
858                 ext4_journal_stop(handle);
859                 if (pos + len > inode->i_size) {
860                         ext4_truncate_failed_write(inode);
861                         /*
862                          * If truncate failed early the inode might
863                          * still be on the orphan list; we need to
864                          * make sure the inode is removed from the
865                          * orphan list in that case.
866                          */
867                         if (inode->i_nlink)
868                                 ext4_orphan_del(NULL, inode);
869                 }
870         }
871
872         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
873                 goto retry;
874 out:
875         return ret;
876 }
877
878 /* For write_end() in data=journal mode */
879 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
880 {
881         if (!buffer_mapped(bh) || buffer_freed(bh))
882                 return 0;
883         set_buffer_uptodate(bh);
884         return ext4_handle_dirty_metadata(handle, NULL, bh);
885 }
886
887 static int ext4_generic_write_end(struct file *file,
888                                   struct address_space *mapping,
889                                   loff_t pos, unsigned len, unsigned copied,
890                                   struct page *page, void *fsdata)
891 {
892         int i_size_changed = 0;
893         struct inode *inode = mapping->host;
894         handle_t *handle = ext4_journal_current_handle();
895
896         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
897
898         /*
899          * No need to use i_size_read() here, the i_size
900          * cannot change under us because we hold i_mutex.
901          *
902          * But it's important to update i_size while still holding page lock:
903          * page writeout could otherwise come in and zero beyond i_size.
904          */
905         if (pos + copied > inode->i_size) {
906                 i_size_write(inode, pos + copied);
907                 i_size_changed = 1;
908         }
909
910         if (pos + copied >  EXT4_I(inode)->i_disksize) {
911                 /* We need to mark inode dirty even if
912                  * new_i_size is less that inode->i_size
913                  * bu greater than i_disksize.(hint delalloc)
914                  */
915                 ext4_update_i_disksize(inode, (pos + copied));
916                 i_size_changed = 1;
917         }
918         unlock_page(page);
919         page_cache_release(page);
920
921         /*
922          * Don't mark the inode dirty under page lock. First, it unnecessarily
923          * makes the holding time of page lock longer. Second, it forces lock
924          * ordering of page lock and transaction start for journaling
925          * filesystems.
926          */
927         if (i_size_changed)
928                 ext4_mark_inode_dirty(handle, inode);
929
930         return copied;
931 }
932
933 /*
934  * We need to pick up the new inode size which generic_commit_write gave us
935  * `file' can be NULL - eg, when called from page_symlink().
936  *
937  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
938  * buffers are managed internally.
939  */
940 static int ext4_ordered_write_end(struct file *file,
941                                   struct address_space *mapping,
942                                   loff_t pos, unsigned len, unsigned copied,
943                                   struct page *page, void *fsdata)
944 {
945         handle_t *handle = ext4_journal_current_handle();
946         struct inode *inode = mapping->host;
947         int ret = 0, ret2;
948
949         trace_ext4_ordered_write_end(inode, pos, len, copied);
950         ret = ext4_jbd2_file_inode(handle, inode);
951
952         if (ret == 0) {
953                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
954                                                         page, fsdata);
955                 copied = ret2;
956                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
957                         /* if we have allocated more blocks and copied
958                          * less. We will have blocks allocated outside
959                          * inode->i_size. So truncate them
960                          */
961                         ext4_orphan_add(handle, inode);
962                 if (ret2 < 0)
963                         ret = ret2;
964         } else {
965                 unlock_page(page);
966                 page_cache_release(page);
967         }
968
969         ret2 = ext4_journal_stop(handle);
970         if (!ret)
971                 ret = ret2;
972
973         if (pos + len > inode->i_size) {
974                 ext4_truncate_failed_write(inode);
975                 /*
976                  * If truncate failed early the inode might still be
977                  * on the orphan list; we need to make sure the inode
978                  * is removed from the orphan list in that case.
979                  */
980                 if (inode->i_nlink)
981                         ext4_orphan_del(NULL, inode);
982         }
983
984
985         return ret ? ret : copied;
986 }
987
988 static int ext4_writeback_write_end(struct file *file,
989                                     struct address_space *mapping,
990                                     loff_t pos, unsigned len, unsigned copied,
991                                     struct page *page, void *fsdata)
992 {
993         handle_t *handle = ext4_journal_current_handle();
994         struct inode *inode = mapping->host;
995         int ret = 0, ret2;
996
997         trace_ext4_writeback_write_end(inode, pos, len, copied);
998         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
999                                                         page, fsdata);
1000         copied = ret2;
1001         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1002                 /* if we have allocated more blocks and copied
1003                  * less. We will have blocks allocated outside
1004                  * inode->i_size. So truncate them
1005                  */
1006                 ext4_orphan_add(handle, inode);
1007
1008         if (ret2 < 0)
1009                 ret = ret2;
1010
1011         ret2 = ext4_journal_stop(handle);
1012         if (!ret)
1013                 ret = ret2;
1014
1015         if (pos + len > inode->i_size) {
1016                 ext4_truncate_failed_write(inode);
1017                 /*
1018                  * If truncate failed early the inode might still be
1019                  * on the orphan list; we need to make sure the inode
1020                  * is removed from the orphan list in that case.
1021                  */
1022                 if (inode->i_nlink)
1023                         ext4_orphan_del(NULL, inode);
1024         }
1025
1026         return ret ? ret : copied;
1027 }
1028
1029 static int ext4_journalled_write_end(struct file *file,
1030                                      struct address_space *mapping,
1031                                      loff_t pos, unsigned len, unsigned copied,
1032                                      struct page *page, void *fsdata)
1033 {
1034         handle_t *handle = ext4_journal_current_handle();
1035         struct inode *inode = mapping->host;
1036         int ret = 0, ret2;
1037         int partial = 0;
1038         unsigned from, to;
1039         loff_t new_i_size;
1040
1041         trace_ext4_journalled_write_end(inode, pos, len, copied);
1042         from = pos & (PAGE_CACHE_SIZE - 1);
1043         to = from + len;
1044
1045         BUG_ON(!ext4_handle_valid(handle));
1046
1047         if (copied < len) {
1048                 if (!PageUptodate(page))
1049                         copied = 0;
1050                 page_zero_new_buffers(page, from+copied, to);
1051         }
1052
1053         ret = walk_page_buffers(handle, page_buffers(page), from,
1054                                 to, &partial, write_end_fn);
1055         if (!partial)
1056                 SetPageUptodate(page);
1057         new_i_size = pos + copied;
1058         if (new_i_size > inode->i_size)
1059                 i_size_write(inode, pos+copied);
1060         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1061         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1062         if (new_i_size > EXT4_I(inode)->i_disksize) {
1063                 ext4_update_i_disksize(inode, new_i_size);
1064                 ret2 = ext4_mark_inode_dirty(handle, inode);
1065                 if (!ret)
1066                         ret = ret2;
1067         }
1068
1069         unlock_page(page);
1070         page_cache_release(page);
1071         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1072                 /* if we have allocated more blocks and copied
1073                  * less. We will have blocks allocated outside
1074                  * inode->i_size. So truncate them
1075                  */
1076                 ext4_orphan_add(handle, inode);
1077
1078         ret2 = ext4_journal_stop(handle);
1079         if (!ret)
1080                 ret = ret2;
1081         if (pos + len > inode->i_size) {
1082                 ext4_truncate_failed_write(inode);
1083                 /*
1084                  * If truncate failed early the inode might still be
1085                  * on the orphan list; we need to make sure the inode
1086                  * is removed from the orphan list in that case.
1087                  */
1088                 if (inode->i_nlink)
1089                         ext4_orphan_del(NULL, inode);
1090         }
1091
1092         return ret ? ret : copied;
1093 }
1094
1095 /*
1096  * Reserve a single cluster located at lblock
1097  */
1098 static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1099 {
1100         int retries = 0;
1101         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1102         struct ext4_inode_info *ei = EXT4_I(inode);
1103         unsigned int md_needed;
1104         int ret;
1105
1106         /*
1107          * recalculate the amount of metadata blocks to reserve
1108          * in order to allocate nrblocks
1109          * worse case is one extent per block
1110          */
1111 repeat:
1112         spin_lock(&ei->i_block_reservation_lock);
1113         md_needed = EXT4_NUM_B2C(sbi,
1114                                  ext4_calc_metadata_amount(inode, lblock));
1115         trace_ext4_da_reserve_space(inode, md_needed);
1116         spin_unlock(&ei->i_block_reservation_lock);
1117
1118         /*
1119          * We will charge metadata quota at writeout time; this saves
1120          * us from metadata over-estimation, though we may go over by
1121          * a small amount in the end.  Here we just reserve for data.
1122          */
1123         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1124         if (ret)
1125                 return ret;
1126         /*
1127          * We do still charge estimated metadata to the sb though;
1128          * we cannot afford to run out of free blocks.
1129          */
1130         if (ext4_claim_free_clusters(sbi, md_needed + 1, 0)) {
1131                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1132                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1133                         yield();
1134                         goto repeat;
1135                 }
1136                 return -ENOSPC;
1137         }
1138         spin_lock(&ei->i_block_reservation_lock);
1139         ei->i_reserved_data_blocks++;
1140         ei->i_reserved_meta_blocks += md_needed;
1141         spin_unlock(&ei->i_block_reservation_lock);
1142
1143         return 0;       /* success */
1144 }
1145
1146 static void ext4_da_release_space(struct inode *inode, int to_free)
1147 {
1148         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1149         struct ext4_inode_info *ei = EXT4_I(inode);
1150
1151         if (!to_free)
1152                 return;         /* Nothing to release, exit */
1153
1154         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1155
1156         trace_ext4_da_release_space(inode, to_free);
1157         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1158                 /*
1159                  * if there aren't enough reserved blocks, then the
1160                  * counter is messed up somewhere.  Since this
1161                  * function is called from invalidate page, it's
1162                  * harmless to return without any action.
1163                  */
1164                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1165                          "ino %lu, to_free %d with only %d reserved "
1166                          "data blocks\n", inode->i_ino, to_free,
1167                          ei->i_reserved_data_blocks);
1168                 WARN_ON(1);
1169                 to_free = ei->i_reserved_data_blocks;
1170         }
1171         ei->i_reserved_data_blocks -= to_free;
1172
1173         if (ei->i_reserved_data_blocks == 0) {
1174                 /*
1175                  * We can release all of the reserved metadata blocks
1176                  * only when we have written all of the delayed
1177                  * allocation blocks.
1178                  * Note that in case of bigalloc, i_reserved_meta_blocks,
1179                  * i_reserved_data_blocks, etc. refer to number of clusters.
1180                  */
1181                 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
1182                                    ei->i_reserved_meta_blocks);
1183                 ei->i_reserved_meta_blocks = 0;
1184                 ei->i_da_metadata_calc_len = 0;
1185         }
1186
1187         /* update fs dirty data blocks counter */
1188         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1189
1190         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1191
1192         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1193 }
1194
1195 static void ext4_da_page_release_reservation(struct page *page,
1196                                              unsigned long offset)
1197 {
1198         int to_release = 0;
1199         struct buffer_head *head, *bh;
1200         unsigned int curr_off = 0;
1201         struct inode *inode = page->mapping->host;
1202         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1203         int num_clusters;
1204
1205         head = page_buffers(page);
1206         bh = head;
1207         do {
1208                 unsigned int next_off = curr_off + bh->b_size;
1209
1210                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1211                         to_release++;
1212                         clear_buffer_delay(bh);
1213                         clear_buffer_da_mapped(bh);
1214                 }
1215                 curr_off = next_off;
1216         } while ((bh = bh->b_this_page) != head);
1217
1218         /* If we have released all the blocks belonging to a cluster, then we
1219          * need to release the reserved space for that cluster. */
1220         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1221         while (num_clusters > 0) {
1222                 ext4_fsblk_t lblk;
1223                 lblk = (page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits)) +
1224                         ((num_clusters - 1) << sbi->s_cluster_bits);
1225                 if (sbi->s_cluster_ratio == 1 ||
1226                     !ext4_find_delalloc_cluster(inode, lblk, 1))
1227                         ext4_da_release_space(inode, 1);
1228
1229                 num_clusters--;
1230         }
1231 }
1232
1233 /*
1234  * Delayed allocation stuff
1235  */
1236
1237 /*
1238  * mpage_da_submit_io - walks through extent of pages and try to write
1239  * them with writepage() call back
1240  *
1241  * @mpd->inode: inode
1242  * @mpd->first_page: first page of the extent
1243  * @mpd->next_page: page after the last page of the extent
1244  *
1245  * By the time mpage_da_submit_io() is called we expect all blocks
1246  * to be allocated. this may be wrong if allocation failed.
1247  *
1248  * As pages are already locked by write_cache_pages(), we can't use it
1249  */
1250 static int mpage_da_submit_io(struct mpage_da_data *mpd,
1251                               struct ext4_map_blocks *map)
1252 {
1253         struct pagevec pvec;
1254         unsigned long index, end;
1255         int ret = 0, err, nr_pages, i;
1256         struct inode *inode = mpd->inode;
1257         struct address_space *mapping = inode->i_mapping;
1258         loff_t size = i_size_read(inode);
1259         unsigned int len, block_start;
1260         struct buffer_head *bh, *page_bufs = NULL;
1261         int journal_data = ext4_should_journal_data(inode);
1262         sector_t pblock = 0, cur_logical = 0;
1263         struct ext4_io_submit io_submit;
1264
1265         BUG_ON(mpd->next_page <= mpd->first_page);
1266         memset(&io_submit, 0, sizeof(io_submit));
1267         /*
1268          * We need to start from the first_page to the next_page - 1
1269          * to make sure we also write the mapped dirty buffer_heads.
1270          * If we look at mpd->b_blocknr we would only be looking
1271          * at the currently mapped buffer_heads.
1272          */
1273         index = mpd->first_page;
1274         end = mpd->next_page - 1;
1275
1276         pagevec_init(&pvec, 0);
1277         while (index <= end) {
1278                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1279                 if (nr_pages == 0)
1280                         break;
1281                 for (i = 0; i < nr_pages; i++) {
1282                         int commit_write = 0, skip_page = 0;
1283                         struct page *page = pvec.pages[i];
1284
1285                         index = page->index;
1286                         if (index > end)
1287                                 break;
1288
1289                         if (index == size >> PAGE_CACHE_SHIFT)
1290                                 len = size & ~PAGE_CACHE_MASK;
1291                         else
1292                                 len = PAGE_CACHE_SIZE;
1293                         if (map) {
1294                                 cur_logical = index << (PAGE_CACHE_SHIFT -
1295                                                         inode->i_blkbits);
1296                                 pblock = map->m_pblk + (cur_logical -
1297                                                         map->m_lblk);
1298                         }
1299                         index++;
1300
1301                         BUG_ON(!PageLocked(page));
1302                         BUG_ON(PageWriteback(page));
1303
1304                         /*
1305                          * If the page does not have buffers (for
1306                          * whatever reason), try to create them using
1307                          * __block_write_begin.  If this fails,
1308                          * skip the page and move on.
1309                          */
1310                         if (!page_has_buffers(page)) {
1311                                 if (__block_write_begin(page, 0, len,
1312                                                 noalloc_get_block_write)) {
1313                                 skip_page:
1314                                         unlock_page(page);
1315                                         continue;
1316                                 }
1317                                 commit_write = 1;
1318                         }
1319
1320                         bh = page_bufs = page_buffers(page);
1321                         block_start = 0;
1322                         do {
1323                                 if (!bh)
1324                                         goto skip_page;
1325                                 if (map && (cur_logical >= map->m_lblk) &&
1326                                     (cur_logical <= (map->m_lblk +
1327                                                      (map->m_len - 1)))) {
1328                                         if (buffer_delay(bh)) {
1329                                                 clear_buffer_delay(bh);
1330                                                 bh->b_blocknr = pblock;
1331                                         }
1332                                         if (buffer_da_mapped(bh))
1333                                                 clear_buffer_da_mapped(bh);
1334                                         if (buffer_unwritten(bh) ||
1335                                             buffer_mapped(bh))
1336                                                 BUG_ON(bh->b_blocknr != pblock);
1337                                         if (map->m_flags & EXT4_MAP_UNINIT)
1338                                                 set_buffer_uninit(bh);
1339                                         clear_buffer_unwritten(bh);
1340                                 }
1341
1342                                 /* skip page if block allocation undone */
1343                                 if (buffer_delay(bh) || buffer_unwritten(bh))
1344                                         skip_page = 1;
1345                                 bh = bh->b_this_page;
1346                                 block_start += bh->b_size;
1347                                 cur_logical++;
1348                                 pblock++;
1349                         } while (bh != page_bufs);
1350
1351                         if (skip_page)
1352                                 goto skip_page;
1353
1354                         if (commit_write)
1355                                 /* mark the buffer_heads as dirty & uptodate */
1356                                 block_commit_write(page, 0, len);
1357
1358                         clear_page_dirty_for_io(page);
1359                         /*
1360                          * Delalloc doesn't support data journalling,
1361                          * but eventually maybe we'll lift this
1362                          * restriction.
1363                          */
1364                         if (unlikely(journal_data && PageChecked(page)))
1365                                 err = __ext4_journalled_writepage(page, len);
1366                         else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
1367                                 err = ext4_bio_write_page(&io_submit, page,
1368                                                           len, mpd->wbc);
1369                         else if (buffer_uninit(page_bufs)) {
1370                                 ext4_set_bh_endio(page_bufs, inode);
1371                                 err = block_write_full_page_endio(page,
1372                                         noalloc_get_block_write,
1373                                         mpd->wbc, ext4_end_io_buffer_write);
1374                         } else
1375                                 err = block_write_full_page(page,
1376                                         noalloc_get_block_write, mpd->wbc);
1377
1378                         if (!err)
1379                                 mpd->pages_written++;
1380                         /*
1381                          * In error case, we have to continue because
1382                          * remaining pages are still locked
1383                          */
1384                         if (ret == 0)
1385                                 ret = err;
1386                 }
1387                 pagevec_release(&pvec);
1388         }
1389         ext4_io_submit(&io_submit);
1390         return ret;
1391 }
1392
1393 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd)
1394 {
1395         int nr_pages, i;
1396         pgoff_t index, end;
1397         struct pagevec pvec;
1398         struct inode *inode = mpd->inode;
1399         struct address_space *mapping = inode->i_mapping;
1400
1401         index = mpd->first_page;
1402         end   = mpd->next_page - 1;
1403         while (index <= end) {
1404                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1405                 if (nr_pages == 0)
1406                         break;
1407                 for (i = 0; i < nr_pages; i++) {
1408                         struct page *page = pvec.pages[i];
1409                         if (page->index > end)
1410                                 break;
1411                         BUG_ON(!PageLocked(page));
1412                         BUG_ON(PageWriteback(page));
1413                         block_invalidatepage(page, 0);
1414                         ClearPageUptodate(page);
1415                         unlock_page(page);
1416                 }
1417                 index = pvec.pages[nr_pages - 1]->index + 1;
1418                 pagevec_release(&pvec);
1419         }
1420         return;
1421 }
1422
1423 static void ext4_print_free_blocks(struct inode *inode)
1424 {
1425         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1426         printk(KERN_CRIT "Total free blocks count %lld\n",
1427                EXT4_C2B(EXT4_SB(inode->i_sb),
1428                         ext4_count_free_clusters(inode->i_sb)));
1429         printk(KERN_CRIT "Free/Dirty block details\n");
1430         printk(KERN_CRIT "free_blocks=%lld\n",
1431                (long long) EXT4_C2B(EXT4_SB(inode->i_sb),
1432                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1433         printk(KERN_CRIT "dirty_blocks=%lld\n",
1434                (long long) EXT4_C2B(EXT4_SB(inode->i_sb),
1435                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1436         printk(KERN_CRIT "Block reservation details\n");
1437         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
1438                EXT4_I(inode)->i_reserved_data_blocks);
1439         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
1440                EXT4_I(inode)->i_reserved_meta_blocks);
1441         return;
1442 }
1443
1444 /*
1445  * mpage_da_map_and_submit - go through given space, map them
1446  *       if necessary, and then submit them for I/O
1447  *
1448  * @mpd - bh describing space
1449  *
1450  * The function skips space we know is already mapped to disk blocks.
1451  *
1452  */
1453 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
1454 {
1455         int err, blks, get_blocks_flags;
1456         struct ext4_map_blocks map, *mapp = NULL;
1457         sector_t next = mpd->b_blocknr;
1458         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
1459         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
1460         handle_t *handle = NULL;
1461
1462         /*
1463          * If the blocks are mapped already, or we couldn't accumulate
1464          * any blocks, then proceed immediately to the submission stage.
1465          */
1466         if ((mpd->b_size == 0) ||
1467             ((mpd->b_state  & (1 << BH_Mapped)) &&
1468              !(mpd->b_state & (1 << BH_Delay)) &&
1469              !(mpd->b_state & (1 << BH_Unwritten))))
1470                 goto submit_io;
1471
1472         handle = ext4_journal_current_handle();
1473         BUG_ON(!handle);
1474
1475         /*
1476          * Call ext4_map_blocks() to allocate any delayed allocation
1477          * blocks, or to convert an uninitialized extent to be
1478          * initialized (in the case where we have written into
1479          * one or more preallocated blocks).
1480          *
1481          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
1482          * indicate that we are on the delayed allocation path.  This
1483          * affects functions in many different parts of the allocation
1484          * call path.  This flag exists primarily because we don't
1485          * want to change *many* call functions, so ext4_map_blocks()
1486          * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
1487          * inode's allocation semaphore is taken.
1488          *
1489          * If the blocks in questions were delalloc blocks, set
1490          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
1491          * variables are updated after the blocks have been allocated.
1492          */
1493         map.m_lblk = next;
1494         map.m_len = max_blocks;
1495         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
1496         if (ext4_should_dioread_nolock(mpd->inode))
1497                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
1498         if (mpd->b_state & (1 << BH_Delay))
1499                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
1500
1501         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
1502         if (blks < 0) {
1503                 struct super_block *sb = mpd->inode->i_sb;
1504
1505                 err = blks;
1506                 /*
1507                  * If get block returns EAGAIN or ENOSPC and there
1508                  * appears to be free blocks we will just let
1509                  * mpage_da_submit_io() unlock all of the pages.
1510                  */
1511                 if (err == -EAGAIN)
1512                         goto submit_io;
1513
1514                 if (err == -ENOSPC && ext4_count_free_clusters(sb)) {
1515                         mpd->retval = err;
1516                         goto submit_io;
1517                 }
1518
1519                 /*
1520                  * get block failure will cause us to loop in
1521                  * writepages, because a_ops->writepage won't be able
1522                  * to make progress. The page will be redirtied by
1523                  * writepage and writepages will again try to write
1524                  * the same.
1525                  */
1526                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
1527                         ext4_msg(sb, KERN_CRIT,
1528                                  "delayed block allocation failed for inode %lu "
1529                                  "at logical offset %llu with max blocks %zd "
1530                                  "with error %d", mpd->inode->i_ino,
1531                                  (unsigned long long) next,
1532                                  mpd->b_size >> mpd->inode->i_blkbits, err);
1533                         ext4_msg(sb, KERN_CRIT,
1534                                 "This should not happen!! Data will be lost\n");
1535                         if (err == -ENOSPC)
1536                                 ext4_print_free_blocks(mpd->inode);
1537                 }
1538                 /* invalidate all the pages */
1539                 ext4_da_block_invalidatepages(mpd);
1540
1541                 /* Mark this page range as having been completed */
1542                 mpd->io_done = 1;
1543                 return;
1544         }
1545         BUG_ON(blks == 0);
1546
1547         mapp = &map;
1548         if (map.m_flags & EXT4_MAP_NEW) {
1549                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
1550                 int i;
1551
1552                 for (i = 0; i < map.m_len; i++)
1553                         unmap_underlying_metadata(bdev, map.m_pblk + i);
1554
1555                 if (ext4_should_order_data(mpd->inode)) {
1556                         err = ext4_jbd2_file_inode(handle, mpd->inode);
1557                         if (err) {
1558                                 /* Only if the journal is aborted */
1559                                 mpd->retval = err;
1560                                 goto submit_io;
1561                         }
1562                 }
1563         }
1564
1565         /*
1566          * Update on-disk size along with block allocation.
1567          */
1568         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
1569         if (disksize > i_size_read(mpd->inode))
1570                 disksize = i_size_read(mpd->inode);
1571         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
1572                 ext4_update_i_disksize(mpd->inode, disksize);
1573                 err = ext4_mark_inode_dirty(handle, mpd->inode);
1574                 if (err)
1575                         ext4_error(mpd->inode->i_sb,
1576                                    "Failed to mark inode %lu dirty",
1577                                    mpd->inode->i_ino);
1578         }
1579
1580 submit_io:
1581         mpage_da_submit_io(mpd, mapp);
1582         mpd->io_done = 1;
1583 }
1584
1585 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1586                 (1 << BH_Delay) | (1 << BH_Unwritten))
1587
1588 /*
1589  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1590  *
1591  * @mpd->lbh - extent of blocks
1592  * @logical - logical number of the block in the file
1593  * @bh - bh of the block (used to access block's state)
1594  *
1595  * the function is used to collect contig. blocks in same state
1596  */
1597 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1598                                    sector_t logical, size_t b_size,
1599                                    unsigned long b_state)
1600 {
1601         sector_t next;
1602         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
1603
1604         /*
1605          * XXX Don't go larger than mballoc is willing to allocate
1606          * This is a stopgap solution.  We eventually need to fold
1607          * mpage_da_submit_io() into this function and then call
1608          * ext4_map_blocks() multiple times in a loop
1609          */
1610         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
1611                 goto flush_it;
1612
1613         /* check if thereserved journal credits might overflow */
1614         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
1615                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1616                         /*
1617                          * With non-extent format we are limited by the journal
1618                          * credit available.  Total credit needed to insert
1619                          * nrblocks contiguous blocks is dependent on the
1620                          * nrblocks.  So limit nrblocks.
1621                          */
1622                         goto flush_it;
1623                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1624                                 EXT4_MAX_TRANS_DATA) {
1625                         /*
1626                          * Adding the new buffer_head would make it cross the
1627                          * allowed limit for which we have journal credit
1628                          * reserved. So limit the new bh->b_size
1629                          */
1630                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1631                                                 mpd->inode->i_blkbits;
1632                         /* we will do mpage_da_submit_io in the next loop */
1633                 }
1634         }
1635         /*
1636          * First block in the extent
1637          */
1638         if (mpd->b_size == 0) {
1639                 mpd->b_blocknr = logical;
1640                 mpd->b_size = b_size;
1641                 mpd->b_state = b_state & BH_FLAGS;
1642                 return;
1643         }
1644
1645         next = mpd->b_blocknr + nrblocks;
1646         /*
1647          * Can we merge the block to our big extent?
1648          */
1649         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
1650                 mpd->b_size += b_size;
1651                 return;
1652         }
1653
1654 flush_it:
1655         /*
1656          * We couldn't merge the block to our extent, so we
1657          * need to flush current  extent and start new one
1658          */
1659         mpage_da_map_and_submit(mpd);
1660         return;
1661 }
1662
1663 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1664 {
1665         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1666 }
1667
1668 /*
1669  * This function is grabs code from the very beginning of
1670  * ext4_map_blocks, but assumes that the caller is from delayed write
1671  * time. This function looks up the requested blocks and sets the
1672  * buffer delay bit under the protection of i_data_sem.
1673  */
1674 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1675                               struct ext4_map_blocks *map,
1676                               struct buffer_head *bh)
1677 {
1678         int retval;
1679         sector_t invalid_block = ~((sector_t) 0xffff);
1680
1681         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1682                 invalid_block = ~0;
1683
1684         map->m_flags = 0;
1685         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1686                   "logical block %lu\n", inode->i_ino, map->m_len,
1687                   (unsigned long) map->m_lblk);
1688         /*
1689          * Try to see if we can get the block without requesting a new
1690          * file system block.
1691          */
1692         down_read((&EXT4_I(inode)->i_data_sem));
1693         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1694                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1695         else
1696                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1697
1698         if (retval == 0) {
1699                 /*
1700                  * XXX: __block_prepare_write() unmaps passed block,
1701                  * is it OK?
1702                  */
1703                 /* If the block was allocated from previously allocated cluster,
1704                  * then we dont need to reserve it again. */
1705                 if (!(map->m_flags & EXT4_MAP_FROM_CLUSTER)) {
1706                         retval = ext4_da_reserve_space(inode, iblock);
1707                         if (retval)
1708                                 /* not enough space to reserve */
1709                                 goto out_unlock;
1710                 }
1711
1712                 /* Clear EXT4_MAP_FROM_CLUSTER flag since its purpose is served
1713                  * and it should not appear on the bh->b_state.
1714                  */
1715                 map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
1716
1717                 map_bh(bh, inode->i_sb, invalid_block);
1718                 set_buffer_new(bh);
1719                 set_buffer_delay(bh);
1720         }
1721
1722 out_unlock:
1723         up_read((&EXT4_I(inode)->i_data_sem));
1724
1725         return retval;
1726 }
1727
1728 /*
1729  * This is a special get_blocks_t callback which is used by
1730  * ext4_da_write_begin().  It will either return mapped block or
1731  * reserve space for a single block.
1732  *
1733  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1734  * We also have b_blocknr = -1 and b_bdev initialized properly
1735  *
1736  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1737  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1738  * initialized properly.
1739  */
1740 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1741                                   struct buffer_head *bh, int create)
1742 {
1743         struct ext4_map_blocks map;
1744         int ret = 0;
1745
1746         BUG_ON(create == 0);
1747         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1748
1749         map.m_lblk = iblock;
1750         map.m_len = 1;
1751
1752         /*
1753          * first, we need to know whether the block is allocated already
1754          * preallocated blocks are unmapped but should treated
1755          * the same as allocated blocks.
1756          */
1757         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1758         if (ret <= 0)
1759                 return ret;
1760
1761         map_bh(bh, inode->i_sb, map.m_pblk);
1762         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1763
1764         if (buffer_unwritten(bh)) {
1765                 /* A delayed write to unwritten bh should be marked
1766                  * new and mapped.  Mapped ensures that we don't do
1767                  * get_block multiple times when we write to the same
1768                  * offset and new ensures that we do proper zero out
1769                  * for partial write.
1770                  */
1771                 set_buffer_new(bh);
1772                 set_buffer_mapped(bh);
1773         }
1774         return 0;
1775 }
1776
1777 /*
1778  * This function is used as a standard get_block_t calback function
1779  * when there is no desire to allocate any blocks.  It is used as a
1780  * callback function for block_write_begin() and block_write_full_page().
1781  * These functions should only try to map a single block at a time.
1782  *
1783  * Since this function doesn't do block allocations even if the caller
1784  * requests it by passing in create=1, it is critically important that
1785  * any caller checks to make sure that any buffer heads are returned
1786  * by this function are either all already mapped or marked for
1787  * delayed allocation before calling  block_write_full_page().  Otherwise,
1788  * b_blocknr could be left unitialized, and the page write functions will
1789  * be taken by surprise.
1790  */
1791 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
1792                                    struct buffer_head *bh_result, int create)
1793 {
1794         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
1795         return _ext4_get_block(inode, iblock, bh_result, 0);
1796 }
1797
1798 static int bget_one(handle_t *handle, struct buffer_head *bh)
1799 {
1800         get_bh(bh);
1801         return 0;
1802 }
1803
1804 static int bput_one(handle_t *handle, struct buffer_head *bh)
1805 {
1806         put_bh(bh);
1807         return 0;
1808 }
1809
1810 static int __ext4_journalled_writepage(struct page *page,
1811                                        unsigned int len)
1812 {
1813         struct address_space *mapping = page->mapping;
1814         struct inode *inode = mapping->host;
1815         struct buffer_head *page_bufs;
1816         handle_t *handle = NULL;
1817         int ret = 0;
1818         int err;
1819
1820         ClearPageChecked(page);
1821         page_bufs = page_buffers(page);
1822         BUG_ON(!page_bufs);
1823         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
1824         /* As soon as we unlock the page, it can go away, but we have
1825          * references to buffers so we are safe */
1826         unlock_page(page);
1827
1828         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1829         if (IS_ERR(handle)) {
1830                 ret = PTR_ERR(handle);
1831                 goto out;
1832         }
1833
1834         BUG_ON(!ext4_handle_valid(handle));
1835
1836         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1837                                 do_journal_get_write_access);
1838
1839         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1840                                 write_end_fn);
1841         if (ret == 0)
1842                 ret = err;
1843         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1844         err = ext4_journal_stop(handle);
1845         if (!ret)
1846                 ret = err;
1847
1848         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
1849         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1850 out:
1851         return ret;
1852 }
1853
1854 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
1855 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
1856
1857 /*
1858  * Note that we don't need to start a transaction unless we're journaling data
1859  * because we should have holes filled from ext4_page_mkwrite(). We even don't
1860  * need to file the inode to the transaction's list in ordered mode because if
1861  * we are writing back data added by write(), the inode is already there and if
1862  * we are writing back data modified via mmap(), no one guarantees in which
1863  * transaction the data will hit the disk. In case we are journaling data, we
1864  * cannot start transaction directly because transaction start ranks above page
1865  * lock so we have to do some magic.
1866  *
1867  * This function can get called via...
1868  *   - ext4_da_writepages after taking page lock (have journal handle)
1869  *   - journal_submit_inode_data_buffers (no journal handle)
1870  *   - shrink_page_list via pdflush (no journal handle)
1871  *   - grab_page_cache when doing write_begin (have journal handle)
1872  *
1873  * We don't do any block allocation in this function. If we have page with
1874  * multiple blocks we need to write those buffer_heads that are mapped. This
1875  * is important for mmaped based write. So if we do with blocksize 1K
1876  * truncate(f, 1024);
1877  * a = mmap(f, 0, 4096);
1878  * a[0] = 'a';
1879  * truncate(f, 4096);
1880  * we have in the page first buffer_head mapped via page_mkwrite call back
1881  * but other bufer_heads would be unmapped but dirty(dirty done via the
1882  * do_wp_page). So writepage should write the first block. If we modify
1883  * the mmap area beyond 1024 we will again get a page_fault and the
1884  * page_mkwrite callback will do the block allocation and mark the
1885  * buffer_heads mapped.
1886  *
1887  * We redirty the page if we have any buffer_heads that is either delay or
1888  * unwritten in the page.
1889  *
1890  * We can get recursively called as show below.
1891  *
1892  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1893  *              ext4_writepage()
1894  *
1895  * But since we don't do any block allocation we should not deadlock.
1896  * Page also have the dirty flag cleared so we don't get recurive page_lock.
1897  */
1898 static int ext4_writepage(struct page *page,
1899                           struct writeback_control *wbc)
1900 {
1901         int ret = 0, commit_write = 0;
1902         loff_t size;
1903         unsigned int len;
1904         struct buffer_head *page_bufs = NULL;
1905         struct inode *inode = page->mapping->host;
1906
1907         trace_ext4_writepage(page);
1908         size = i_size_read(inode);
1909         if (page->index == size >> PAGE_CACHE_SHIFT)
1910                 len = size & ~PAGE_CACHE_MASK;
1911         else
1912                 len = PAGE_CACHE_SIZE;
1913
1914         /*
1915          * If the page does not have buffers (for whatever reason),
1916          * try to create them using __block_write_begin.  If this
1917          * fails, redirty the page and move on.
1918          */
1919         if (!page_has_buffers(page)) {
1920                 if (__block_write_begin(page, 0, len,
1921                                         noalloc_get_block_write)) {
1922                 redirty_page:
1923                         redirty_page_for_writepage(wbc, page);
1924                         unlock_page(page);
1925                         return 0;
1926                 }
1927                 commit_write = 1;
1928         }
1929         page_bufs = page_buffers(page);
1930         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
1931                               ext4_bh_delay_or_unwritten)) {
1932                 /*
1933                  * We don't want to do block allocation, so redirty
1934                  * the page and return.  We may reach here when we do
1935                  * a journal commit via journal_submit_inode_data_buffers.
1936                  * We can also reach here via shrink_page_list but it
1937                  * should never be for direct reclaim so warn if that
1938                  * happens
1939                  */
1940                 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
1941                                                                 PF_MEMALLOC);
1942                 goto redirty_page;
1943         }
1944         if (commit_write)
1945                 /* now mark the buffer_heads as dirty and uptodate */
1946                 block_commit_write(page, 0, len);
1947
1948         if (PageChecked(page) && ext4_should_journal_data(inode))
1949                 /*
1950                  * It's mmapped pagecache.  Add buffers and journal it.  There
1951                  * doesn't seem much point in redirtying the page here.
1952                  */
1953                 return __ext4_journalled_writepage(page, len);
1954
1955         if (buffer_uninit(page_bufs)) {
1956                 ext4_set_bh_endio(page_bufs, inode);
1957                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
1958                                             wbc, ext4_end_io_buffer_write);
1959         } else
1960                 ret = block_write_full_page(page, noalloc_get_block_write,
1961                                             wbc);
1962
1963         return ret;
1964 }
1965
1966 /*
1967  * This is called via ext4_da_writepages() to
1968  * calculate the total number of credits to reserve to fit
1969  * a single extent allocation into a single transaction,
1970  * ext4_da_writpeages() will loop calling this before
1971  * the block allocation.
1972  */
1973
1974 static int ext4_da_writepages_trans_blocks(struct inode *inode)
1975 {
1976         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
1977
1978         /*
1979          * With non-extent format the journal credit needed to
1980          * insert nrblocks contiguous block is dependent on
1981          * number of contiguous block. So we will limit
1982          * number of contiguous block to a sane value
1983          */
1984         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
1985             (max_blocks > EXT4_MAX_TRANS_DATA))
1986                 max_blocks = EXT4_MAX_TRANS_DATA;
1987
1988         return ext4_chunk_trans_blocks(inode, max_blocks);
1989 }
1990
1991 /*
1992  * write_cache_pages_da - walk the list of dirty pages of the given
1993  * address space and accumulate pages that need writing, and call
1994  * mpage_da_map_and_submit to map a single contiguous memory region
1995  * and then write them.
1996  */
1997 static int write_cache_pages_da(struct address_space *mapping,
1998                                 struct writeback_control *wbc,
1999                                 struct mpage_da_data *mpd,
2000                                 pgoff_t *done_index)
2001 {
2002         struct buffer_head      *bh, *head;
2003         struct inode            *inode = mapping->host;
2004         struct pagevec          pvec;
2005         unsigned int            nr_pages;
2006         sector_t                logical;
2007         pgoff_t                 index, end;
2008         long                    nr_to_write = wbc->nr_to_write;
2009         int                     i, tag, ret = 0;
2010
2011         memset(mpd, 0, sizeof(struct mpage_da_data));
2012         mpd->wbc = wbc;
2013         mpd->inode = inode;
2014         pagevec_init(&pvec, 0);
2015         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2016         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2017
2018         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2019                 tag = PAGECACHE_TAG_TOWRITE;
2020         else
2021                 tag = PAGECACHE_TAG_DIRTY;
2022
2023         *done_index = index;
2024         while (index <= end) {
2025                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2026                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2027                 if (nr_pages == 0)
2028                         return 0;
2029
2030                 for (i = 0; i < nr_pages; i++) {
2031                         struct page *page = pvec.pages[i];
2032
2033                         /*
2034                          * At this point, the page may be truncated or
2035                          * invalidated (changing page->mapping to NULL), or
2036                          * even swizzled back from swapper_space to tmpfs file
2037                          * mapping. However, page->index will not change
2038                          * because we have a reference on the page.
2039                          */
2040                         if (page->index > end)
2041                                 goto out;
2042
2043                         *done_index = page->index + 1;
2044
2045                         /*
2046                          * If we can't merge this page, and we have
2047                          * accumulated an contiguous region, write it
2048                          */
2049                         if ((mpd->next_page != page->index) &&
2050                             (mpd->next_page != mpd->first_page)) {
2051                                 mpage_da_map_and_submit(mpd);
2052                                 goto ret_extent_tail;
2053                         }
2054
2055                         lock_page(page);
2056
2057                         /*
2058                          * If the page is no longer dirty, or its
2059                          * mapping no longer corresponds to inode we
2060                          * are writing (which means it has been
2061                          * truncated or invalidated), or the page is
2062                          * already under writeback and we are not
2063                          * doing a data integrity writeback, skip the page
2064                          */
2065                         if (!PageDirty(page) ||
2066                             (PageWriteback(page) &&
2067                              (wbc->sync_mode == WB_SYNC_NONE)) ||
2068                             unlikely(page->mapping != mapping)) {
2069                                 unlock_page(page);
2070                                 continue;
2071                         }
2072
2073                         wait_on_page_writeback(page);
2074                         BUG_ON(PageWriteback(page));
2075
2076                         if (mpd->next_page != page->index)
2077                                 mpd->first_page = page->index;
2078                         mpd->next_page = page->index + 1;
2079                         logical = (sector_t) page->index <<
2080                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2081
2082                         if (!page_has_buffers(page)) {
2083                                 mpage_add_bh_to_extent(mpd, logical,
2084                                                        PAGE_CACHE_SIZE,
2085                                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2086                                 if (mpd->io_done)
2087                                         goto ret_extent_tail;
2088                         } else {
2089                                 /*
2090                                  * Page with regular buffer heads,
2091                                  * just add all dirty ones
2092                                  */
2093                                 head = page_buffers(page);
2094                                 bh = head;
2095                                 do {
2096                                         BUG_ON(buffer_locked(bh));
2097                                         /*
2098                                          * We need to try to allocate
2099                                          * unmapped blocks in the same page.
2100                                          * Otherwise we won't make progress
2101                                          * with the page in ext4_writepage
2102                                          */
2103                                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2104                                                 mpage_add_bh_to_extent(mpd, logical,
2105                                                                        bh->b_size,
2106                                                                        bh->b_state);
2107                                                 if (mpd->io_done)
2108                                                         goto ret_extent_tail;
2109                                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2110                                                 /*
2111                                                  * mapped dirty buffer. We need
2112                                                  * to update the b_state
2113                                                  * because we look at b_state
2114                                                  * in mpage_da_map_blocks.  We
2115                                                  * don't update b_size because
2116                                                  * if we find an unmapped
2117                                                  * buffer_head later we need to
2118                                                  * use the b_state flag of that
2119                                                  * buffer_head.
2120                                                  */
2121                                                 if (mpd->b_size == 0)
2122                                                         mpd->b_state = bh->b_state & BH_FLAGS;
2123                                         }
2124                                         logical++;
2125                                 } while ((bh = bh->b_this_page) != head);
2126                         }
2127
2128                         if (nr_to_write > 0) {
2129                                 nr_to_write--;
2130                                 if (nr_to_write == 0 &&
2131                                     wbc->sync_mode == WB_SYNC_NONE)
2132                                         /*
2133                                          * We stop writing back only if we are
2134                                          * not doing integrity sync. In case of
2135                                          * integrity sync we have to keep going
2136                                          * because someone may be concurrently
2137                                          * dirtying pages, and we might have
2138                                          * synced a lot of newly appeared dirty
2139                                          * pages, but have not synced all of the
2140                                          * old dirty pages.
2141                                          */
2142                                         goto out;
2143                         }
2144                 }
2145                 pagevec_release(&pvec);
2146                 cond_resched();
2147         }
2148         return 0;
2149 ret_extent_tail:
2150         ret = MPAGE_DA_EXTENT_TAIL;
2151 out:
2152         pagevec_release(&pvec);
2153         cond_resched();
2154         return ret;
2155 }
2156
2157
2158 static int ext4_da_writepages(struct address_space *mapping,
2159                               struct writeback_control *wbc)
2160 {
2161         pgoff_t index;
2162         int range_whole = 0;
2163         handle_t *handle = NULL;
2164         struct mpage_da_data mpd;
2165         struct inode *inode = mapping->host;
2166         int pages_written = 0;
2167         unsigned int max_pages;
2168         int range_cyclic, cycled = 1, io_done = 0;
2169         int needed_blocks, ret = 0;
2170         long desired_nr_to_write, nr_to_writebump = 0;
2171         loff_t range_start = wbc->range_start;
2172         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2173         pgoff_t done_index = 0;
2174         pgoff_t end;
2175         struct blk_plug plug;
2176
2177         trace_ext4_da_writepages(inode, wbc);
2178
2179         /*
2180          * No pages to write? This is mainly a kludge to avoid starting
2181          * a transaction for special inodes like journal inode on last iput()
2182          * because that could violate lock ordering on umount
2183          */
2184         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2185                 return 0;
2186
2187         /*
2188          * If the filesystem has aborted, it is read-only, so return
2189          * right away instead of dumping stack traces later on that
2190          * will obscure the real source of the problem.  We test
2191          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2192          * the latter could be true if the filesystem is mounted
2193          * read-only, and in that case, ext4_da_writepages should
2194          * *never* be called, so if that ever happens, we would want
2195          * the stack trace.
2196          */
2197         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2198                 return -EROFS;
2199
2200         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2201                 range_whole = 1;
2202
2203         range_cyclic = wbc->range_cyclic;
2204         if (wbc->range_cyclic) {
2205                 index = mapping->writeback_index;
2206                 if (index)
2207                         cycled = 0;
2208                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2209                 wbc->range_end  = LLONG_MAX;
2210                 wbc->range_cyclic = 0;
2211                 end = -1;
2212         } else {
2213                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2214                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2215         }
2216
2217         /*
2218          * This works around two forms of stupidity.  The first is in
2219          * the writeback code, which caps the maximum number of pages
2220          * written to be 1024 pages.  This is wrong on multiple
2221          * levels; different architectues have a different page size,
2222          * which changes the maximum amount of data which gets
2223          * written.  Secondly, 4 megabytes is way too small.  XFS
2224          * forces this value to be 16 megabytes by multiplying
2225          * nr_to_write parameter by four, and then relies on its
2226          * allocator to allocate larger extents to make them
2227          * contiguous.  Unfortunately this brings us to the second
2228          * stupidity, which is that ext4's mballoc code only allocates
2229          * at most 2048 blocks.  So we force contiguous writes up to
2230          * the number of dirty blocks in the inode, or
2231          * sbi->max_writeback_mb_bump whichever is smaller.
2232          */
2233         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2234         if (!range_cyclic && range_whole) {
2235                 if (wbc->nr_to_write == LONG_MAX)
2236                         desired_nr_to_write = wbc->nr_to_write;
2237                 else
2238                         desired_nr_to_write = wbc->nr_to_write * 8;
2239         } else
2240                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2241                                                            max_pages);
2242         if (desired_nr_to_write > max_pages)
2243                 desired_nr_to_write = max_pages;
2244
2245         if (wbc->nr_to_write < desired_nr_to_write) {
2246                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2247                 wbc->nr_to_write = desired_nr_to_write;
2248         }
2249
2250 retry:
2251         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2252                 tag_pages_for_writeback(mapping, index, end);
2253
2254         blk_start_plug(&plug);
2255         while (!ret && wbc->nr_to_write > 0) {
2256
2257                 /*
2258                  * we  insert one extent at a time. So we need
2259                  * credit needed for single extent allocation.
2260                  * journalled mode is currently not supported
2261                  * by delalloc
2262                  */
2263                 BUG_ON(ext4_should_journal_data(inode));
2264                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2265
2266                 /* start a new transaction*/
2267                 handle = ext4_journal_start(inode, needed_blocks);
2268                 if (IS_ERR(handle)) {
2269                         ret = PTR_ERR(handle);
2270                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2271                                "%ld pages, ino %lu; err %d", __func__,
2272                                 wbc->nr_to_write, inode->i_ino, ret);
2273                         blk_finish_plug(&plug);
2274                         goto out_writepages;
2275                 }
2276
2277                 /*
2278                  * Now call write_cache_pages_da() to find the next
2279                  * contiguous region of logical blocks that need
2280                  * blocks to be allocated by ext4 and submit them.
2281                  */
2282                 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
2283                 /*
2284                  * If we have a contiguous extent of pages and we
2285                  * haven't done the I/O yet, map the blocks and submit
2286                  * them for I/O.
2287                  */
2288                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2289                         mpage_da_map_and_submit(&mpd);
2290                         ret = MPAGE_DA_EXTENT_TAIL;
2291                 }
2292                 trace_ext4_da_write_pages(inode, &mpd);
2293                 wbc->nr_to_write -= mpd.pages_written;
2294
2295                 ext4_journal_stop(handle);
2296
2297                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2298                         /* commit the transaction which would
2299                          * free blocks released in the transaction
2300                          * and try again
2301                          */
2302                         jbd2_journal_force_commit_nested(sbi->s_journal);
2303                         ret = 0;
2304                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2305                         /*
2306                          * Got one extent now try with rest of the pages.
2307                          * If mpd.retval is set -EIO, journal is aborted.
2308                          * So we don't need to write any more.
2309                          */
2310                         pages_written += mpd.pages_written;
2311                         ret = mpd.retval;
2312                         io_done = 1;
2313                 } else if (wbc->nr_to_write)
2314                         /*
2315                          * There is no more writeout needed
2316                          * or we requested for a noblocking writeout
2317                          * and we found the device congested
2318                          */
2319                         break;
2320         }
2321         blk_finish_plug(&plug);
2322         if (!io_done && !cycled) {
2323                 cycled = 1;
2324                 index = 0;
2325                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2326                 wbc->range_end  = mapping->writeback_index - 1;
2327                 goto retry;
2328         }
2329
2330         /* Update index */
2331         wbc->range_cyclic = range_cyclic;
2332         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2333                 /*
2334                  * set the writeback_index so that range_cyclic
2335                  * mode will write it back later
2336                  */
2337                 mapping->writeback_index = done_index;
2338
2339 out_writepages:
2340         wbc->nr_to_write -= nr_to_writebump;
2341         wbc->range_start = range_start;
2342         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2343         return ret;
2344 }
2345
2346 #define FALL_BACK_TO_NONDELALLOC 1
2347 static int ext4_nonda_switch(struct super_block *sb)
2348 {
2349         s64 free_blocks, dirty_blocks;
2350         struct ext4_sb_info *sbi = EXT4_SB(sb);
2351
2352         /*
2353          * switch to non delalloc mode if we are running low
2354          * on free block. The free block accounting via percpu
2355          * counters can get slightly wrong with percpu_counter_batch getting
2356          * accumulated on each CPU without updating global counters
2357          * Delalloc need an accurate free block accounting. So switch
2358          * to non delalloc when we are near to error range.
2359          */
2360         free_blocks  = EXT4_C2B(sbi,
2361                 percpu_counter_read_positive(&sbi->s_freeclusters_counter));
2362         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2363         if (2 * free_blocks < 3 * dirty_blocks ||
2364                 free_blocks < (dirty_blocks + EXT4_FREECLUSTERS_WATERMARK)) {
2365                 /*
2366                  * free block count is less than 150% of dirty blocks
2367                  * or free blocks is less than watermark
2368                  */
2369                 return 1;
2370         }
2371         /*
2372          * Even if we don't switch but are nearing capacity,
2373          * start pushing delalloc when 1/2 of free blocks are dirty.
2374          */
2375         if (free_blocks < 2 * dirty_blocks)
2376                 writeback_inodes_sb_if_idle(sb, WB_REASON_FS_FREE_SPACE);
2377
2378         return 0;
2379 }
2380
2381 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2382                                loff_t pos, unsigned len, unsigned flags,
2383                                struct page **pagep, void **fsdata)
2384 {
2385         int ret, retries = 0;
2386         struct page *page;
2387         pgoff_t index;
2388         struct inode *inode = mapping->host;
2389         handle_t *handle;
2390         loff_t page_len;
2391
2392         index = pos >> PAGE_CACHE_SHIFT;
2393
2394         if (ext4_nonda_switch(inode->i_sb)) {
2395                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2396                 return ext4_write_begin(file, mapping, pos,
2397                                         len, flags, pagep, fsdata);
2398         }
2399         *fsdata = (void *)0;
2400         trace_ext4_da_write_begin(inode, pos, len, flags);
2401 retry:
2402         /*
2403          * With delayed allocation, we don't log the i_disksize update
2404          * if there is delayed block allocation. But we still need
2405          * to journalling the i_disksize update if writes to the end
2406          * of file which has an already mapped buffer.
2407          */
2408         handle = ext4_journal_start(inode, 1);
2409         if (IS_ERR(handle)) {
2410                 ret = PTR_ERR(handle);
2411                 goto out;
2412         }
2413         /* We cannot recurse into the filesystem as the transaction is already
2414          * started */
2415         flags |= AOP_FLAG_NOFS;
2416
2417         page = grab_cache_page_write_begin(mapping, index, flags);
2418         if (!page) {
2419                 ext4_journal_stop(handle);
2420                 ret = -ENOMEM;
2421                 goto out;
2422         }
2423         *pagep = page;
2424
2425         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2426         if (ret < 0) {
2427                 unlock_page(page);
2428                 ext4_journal_stop(handle);
2429                 page_cache_release(page);
2430                 /*
2431                  * block_write_begin may have instantiated a few blocks
2432                  * outside i_size.  Trim these off again. Don't need
2433                  * i_size_read because we hold i_mutex.
2434                  */
2435                 if (pos + len > inode->i_size)
2436                         ext4_truncate_failed_write(inode);
2437         } else {
2438                 page_len = pos & (PAGE_CACHE_SIZE - 1);
2439                 if (page_len > 0) {
2440                         ret = ext4_discard_partial_page_buffers_no_lock(handle,
2441                                 inode, page, pos - page_len, page_len,
2442                                 EXT4_DISCARD_PARTIAL_PG_ZERO_UNMAPPED);
2443                 }
2444         }
2445
2446         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2447                 goto retry;
2448 out:
2449         return ret;
2450 }
2451
2452 /*
2453  * Check if we should update i_disksize
2454  * when write to the end of file but not require block allocation
2455  */
2456 static int ext4_da_should_update_i_disksize(struct page *page,
2457                                             unsigned long offset)
2458 {
2459         struct buffer_head *bh;
2460         struct inode *inode = page->mapping->host;
2461         unsigned int idx;
2462         int i;
2463
2464         bh = page_buffers(page);
2465         idx = offset >> inode->i_blkbits;
2466
2467         for (i = 0; i < idx; i++)
2468                 bh = bh->b_this_page;
2469
2470         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2471                 return 0;
2472         return 1;
2473 }
2474
2475 static int ext4_da_write_end(struct file *file,
2476                              struct address_space *mapping,
2477                              loff_t pos, unsigned len, unsigned copied,
2478                              struct page *page, void *fsdata)
2479 {
2480         struct inode *inode = mapping->host;
2481         int ret = 0, ret2;
2482         handle_t *handle = ext4_journal_current_handle();
2483         loff_t new_i_size;
2484         unsigned long start, end;
2485         int write_mode = (int)(unsigned long)fsdata;
2486         loff_t page_len;
2487
2488         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2489                 if (ext4_should_order_data(inode)) {
2490                         return ext4_ordered_write_end(file, mapping, pos,
2491                                         len, copied, page, fsdata);
2492                 } else if (ext4_should_writeback_data(inode)) {
2493                         return ext4_writeback_write_end(file, mapping, pos,
2494                                         len, copied, page, fsdata);
2495                 } else {
2496                         BUG();
2497                 }
2498         }
2499
2500         trace_ext4_da_write_end(inode, pos, len, copied);
2501         start = pos & (PAGE_CACHE_SIZE - 1);
2502         end = start + copied - 1;
2503
2504         /*
2505          * generic_write_end() will run mark_inode_dirty() if i_size
2506          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2507          * into that.
2508          */
2509
2510         new_i_size = pos + copied;
2511         if (new_i_size > EXT4_I(inode)->i_disksize) {
2512                 if (ext4_da_should_update_i_disksize(page, end)) {
2513                         down_write(&EXT4_I(inode)->i_data_sem);
2514                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2515                                 /*
2516                                  * Updating i_disksize when extending file
2517                                  * without needing block allocation
2518                                  */
2519                                 if (ext4_should_order_data(inode))
2520                                         ret = ext4_jbd2_file_inode(handle,
2521                                                                    inode);
2522
2523                                 EXT4_I(inode)->i_disksize = new_i_size;
2524                         }
2525                         up_write(&EXT4_I(inode)->i_data_sem);
2526                         /* We need to mark inode dirty even if
2527                          * new_i_size is less that inode->i_size
2528                          * bu greater than i_disksize.(hint delalloc)
2529                          */
2530                         ext4_mark_inode_dirty(handle, inode);
2531                 }
2532         }
2533         ret2 = generic_write_end(file, mapping, pos, len, copied,
2534                                                         page, fsdata);
2535
2536         page_len = PAGE_CACHE_SIZE -
2537                         ((pos + copied - 1) & (PAGE_CACHE_SIZE - 1));
2538
2539         if (page_len > 0) {
2540                 ret = ext4_discard_partial_page_buffers_no_lock(handle,
2541                         inode, page, pos + copied - 1, page_len,
2542                         EXT4_DISCARD_PARTIAL_PG_ZERO_UNMAPPED);
2543         }
2544
2545         copied = ret2;
2546         if (ret2 < 0)
2547                 ret = ret2;
2548         ret2 = ext4_journal_stop(handle);
2549         if (!ret)
2550                 ret = ret2;
2551
2552         return ret ? ret : copied;
2553 }
2554
2555 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2556 {
2557         /*
2558          * Drop reserved blocks
2559          */
2560         BUG_ON(!PageLocked(page));
2561         if (!page_has_buffers(page))
2562                 goto out;
2563
2564         ext4_da_page_release_reservation(page, offset);
2565
2566 out:
2567         ext4_invalidatepage(page, offset);
2568
2569         return;
2570 }
2571
2572 /*
2573  * Force all delayed allocation blocks to be allocated for a given inode.
2574  */
2575 int ext4_alloc_da_blocks(struct inode *inode)
2576 {
2577         trace_ext4_alloc_da_blocks(inode);
2578
2579         if (!EXT4_I(inode)->i_reserved_data_blocks &&
2580             !EXT4_I(inode)->i_reserved_meta_blocks)
2581                 return 0;
2582
2583         /*
2584          * We do something simple for now.  The filemap_flush() will
2585          * also start triggering a write of the data blocks, which is
2586          * not strictly speaking necessary (and for users of
2587          * laptop_mode, not even desirable).  However, to do otherwise
2588          * would require replicating code paths in:
2589          *
2590          * ext4_da_writepages() ->
2591          *    write_cache_pages() ---> (via passed in callback function)
2592          *        __mpage_da_writepage() -->
2593          *           mpage_add_bh_to_extent()
2594          *           mpage_da_map_blocks()
2595          *
2596          * The problem is that write_cache_pages(), located in
2597          * mm/page-writeback.c, marks pages clean in preparation for
2598          * doing I/O, which is not desirable if we're not planning on
2599          * doing I/O at all.
2600          *
2601          * We could call write_cache_pages(), and then redirty all of
2602          * the pages by calling redirty_page_for_writepage() but that
2603          * would be ugly in the extreme.  So instead we would need to
2604          * replicate parts of the code in the above functions,
2605          * simplifying them because we wouldn't actually intend to
2606          * write out the pages, but rather only collect contiguous
2607          * logical block extents, call the multi-block allocator, and
2608          * then update the buffer heads with the block allocations.
2609          *
2610          * For now, though, we'll cheat by calling filemap_flush(),
2611          * which will map the blocks, and start the I/O, but not
2612          * actually wait for the I/O to complete.
2613          */
2614         return filemap_flush(inode->i_mapping);
2615 }
2616
2617 /*
2618  * bmap() is special.  It gets used by applications such as lilo and by
2619  * the swapper to find the on-disk block of a specific piece of data.
2620  *
2621  * Naturally, this is dangerous if the block concerned is still in the
2622  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2623  * filesystem and enables swap, then they may get a nasty shock when the
2624  * data getting swapped to that swapfile suddenly gets overwritten by
2625  * the original zero's written out previously to the journal and
2626  * awaiting writeback in the kernel's buffer cache.
2627  *
2628  * So, if we see any bmap calls here on a modified, data-journaled file,
2629  * take extra steps to flush any blocks which might be in the cache.
2630  */
2631 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2632 {
2633         struct inode *inode = mapping->host;
2634         journal_t *journal;
2635         int err;
2636
2637         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2638                         test_opt(inode->i_sb, DELALLOC)) {
2639                 /*
2640                  * With delalloc we want to sync the file
2641                  * so that we can make sure we allocate
2642                  * blocks for file
2643                  */
2644                 filemap_write_and_wait(mapping);
2645         }
2646
2647         if (EXT4_JOURNAL(inode) &&
2648             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2649                 /*
2650                  * This is a REALLY heavyweight approach, but the use of
2651                  * bmap on dirty files is expected to be extremely rare:
2652                  * only if we run lilo or swapon on a freshly made file
2653                  * do we expect this to happen.
2654                  *
2655                  * (bmap requires CAP_SYS_RAWIO so this does not
2656                  * represent an unprivileged user DOS attack --- we'd be
2657                  * in trouble if mortal users could trigger this path at
2658                  * will.)
2659                  *
2660                  * NB. EXT4_STATE_JDATA is not set on files other than
2661                  * regular files.  If somebody wants to bmap a directory
2662                  * or symlink and gets confused because the buffer
2663                  * hasn't yet been flushed to disk, they deserve
2664                  * everything they get.
2665                  */
2666
2667                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2668                 journal = EXT4_JOURNAL(inode);
2669                 jbd2_journal_lock_updates(journal);
2670                 err = jbd2_journal_flush(journal);
2671                 jbd2_journal_unlock_updates(journal);
2672
2673                 if (err)
2674                         return 0;
2675         }
2676
2677         return generic_block_bmap(mapping, block, ext4_get_block);
2678 }
2679
2680 static int ext4_readpage(struct file *file, struct page *page)
2681 {
2682         trace_ext4_readpage(page);
2683         return mpage_readpage(page, ext4_get_block);
2684 }
2685
2686 static int
2687 ext4_readpages(struct file *file, struct address_space *mapping,
2688                 struct list_head *pages, unsigned nr_pages)
2689 {
2690         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2691 }
2692
2693 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
2694 {
2695         struct buffer_head *head, *bh;
2696         unsigned int curr_off = 0;
2697
2698         if (!page_has_buffers(page))
2699                 return;
2700         head = bh = page_buffers(page);
2701         do {
2702                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
2703                                         && bh->b_private) {
2704                         ext4_free_io_end(bh->b_private);
2705                         bh->b_private = NULL;
2706                         bh->b_end_io = NULL;
2707                 }
2708                 curr_off = curr_off + bh->b_size;
2709                 bh = bh->b_this_page;
2710         } while (bh != head);
2711 }
2712
2713 static void ext4_invalidatepage(struct page *page, unsigned long offset)
2714 {
2715         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2716
2717         trace_ext4_invalidatepage(page, offset);
2718
2719         /*
2720          * free any io_end structure allocated for buffers to be discarded
2721          */
2722         if (ext4_should_dioread_nolock(page->mapping->host))
2723                 ext4_invalidatepage_free_endio(page, offset);
2724         /*
2725          * If it's a full truncate we just forget about the pending dirtying
2726          */
2727         if (offset == 0)
2728                 ClearPageChecked(page);
2729
2730         if (journal)
2731                 jbd2_journal_invalidatepage(journal, page, offset);
2732         else
2733                 block_invalidatepage(page, offset);
2734 }
2735
2736 static int ext4_releasepage(struct page *page, gfp_t wait)
2737 {
2738         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2739
2740         trace_ext4_releasepage(page);
2741
2742         WARN_ON(PageChecked(page));
2743         if (!page_has_buffers(page))
2744                 return 0;
2745         if (journal)
2746                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
2747         else
2748                 return try_to_free_buffers(page);
2749 }
2750
2751 /*
2752  * ext4_get_block used when preparing for a DIO write or buffer write.
2753  * We allocate an uinitialized extent if blocks haven't been allocated.
2754  * The extent will be converted to initialized after the IO is complete.
2755  */
2756 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
2757                    struct buffer_head *bh_result, int create)
2758 {
2759         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
2760                    inode->i_ino, create);
2761         return _ext4_get_block(inode, iblock, bh_result,
2762                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
2763 }
2764
2765 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
2766                             ssize_t size, void *private, int ret,
2767                             bool is_async)
2768 {
2769         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
2770         ext4_io_end_t *io_end = iocb->private;
2771         struct workqueue_struct *wq;
2772         unsigned long flags;
2773         struct ext4_inode_info *ei;
2774
2775         /* if not async direct IO or dio with 0 bytes write, just return */
2776         if (!io_end || !size)
2777                 goto out;
2778
2779         ext_debug("ext4_end_io_dio(): io_end 0x%p"
2780                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
2781                   iocb->private, io_end->inode->i_ino, iocb, offset,
2782                   size);
2783
2784         /* if not aio dio with unwritten extents, just free io and return */
2785         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2786                 ext4_free_io_end(io_end);
2787                 iocb->private = NULL;
2788 out:
2789                 if (is_async)
2790                         aio_complete(iocb, ret, 0);
2791                 inode_dio_done(inode);
2792                 return;
2793         }
2794
2795         io_end->offset = offset;
2796         io_end->size = size;
2797         if (is_async) {
2798                 io_end->iocb = iocb;
2799                 io_end->result = ret;
2800         }
2801         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
2802
2803         /* Add the io_end to per-inode completed aio dio list*/
2804         ei = EXT4_I(io_end->inode);
2805         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
2806         list_add_tail(&io_end->list, &ei->i_completed_io_list);
2807         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
2808
2809         /* queue the work to convert unwritten extents to written */
2810         queue_work(wq, &io_end->work);
2811         iocb->private = NULL;
2812
2813         /* XXX: probably should move into the real I/O completion handler */
2814         inode_dio_done(inode);
2815 }
2816
2817 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
2818 {
2819         ext4_io_end_t *io_end = bh->b_private;
2820         struct workqueue_struct *wq;
2821         struct inode *inode;
2822         unsigned long flags;
2823
2824         if (!test_clear_buffer_uninit(bh) || !io_end)
2825                 goto out;
2826
2827         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
2828                 printk("sb umounted, discard end_io request for inode %lu\n",
2829                         io_end->inode->i_ino);
2830                 ext4_free_io_end(io_end);
2831                 goto out;
2832         }
2833
2834         /*
2835          * It may be over-defensive here to check EXT4_IO_END_UNWRITTEN now,
2836          * but being more careful is always safe for the future change.
2837          */
2838         inode = io_end->inode;
2839         ext4_set_io_unwritten_flag(inode, io_end);
2840
2841         /* Add the io_end to per-inode completed io list*/
2842         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
2843         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
2844         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
2845
2846         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
2847         /* queue the work to convert unwritten extents to written */
2848         queue_work(wq, &io_end->work);
2849 out:
2850         bh->b_private = NULL;
2851         bh->b_end_io = NULL;
2852         clear_buffer_uninit(bh);
2853         end_buffer_async_write(bh, uptodate);
2854 }
2855
2856 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
2857 {
2858         ext4_io_end_t *io_end;
2859         struct page *page = bh->b_page;
2860         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
2861         size_t size = bh->b_size;
2862
2863 retry:
2864         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
2865         if (!io_end) {
2866                 pr_warn_ratelimited("%s: allocation fail\n", __func__);
2867                 schedule();
2868                 goto retry;
2869         }
2870         io_end->offset = offset;
2871         io_end->size = size;
2872         /*
2873          * We need to hold a reference to the page to make sure it
2874          * doesn't get evicted before ext4_end_io_work() has a chance
2875          * to convert the extent from written to unwritten.
2876          */
2877         io_end->page = page;
2878         get_page(io_end->page);
2879
2880         bh->b_private = io_end;
2881         bh->b_end_io = ext4_end_io_buffer_write;
2882         return 0;
2883 }
2884
2885 /*
2886  * For ext4 extent files, ext4 will do direct-io write to holes,
2887  * preallocated extents, and those write extend the file, no need to
2888  * fall back to buffered IO.
2889  *
2890  * For holes, we fallocate those blocks, mark them as uninitialized
2891  * If those blocks were preallocated, we mark sure they are splited, but
2892  * still keep the range to write as uninitialized.
2893  *
2894  * The unwrritten extents will be converted to written when DIO is completed.
2895  * For async direct IO, since the IO may still pending when return, we
2896  * set up an end_io call back function, which will do the conversion
2897  * when async direct IO completed.
2898  *
2899  * If the O_DIRECT write will extend the file then add this inode to the
2900  * orphan list.  So recovery will truncate it back to the original size
2901  * if the machine crashes during the write.
2902  *
2903  */
2904 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
2905                               const struct iovec *iov, loff_t offset,
2906                               unsigned long nr_segs)
2907 {
2908         struct file *file = iocb->ki_filp;
2909         struct inode *inode = file->f_mapping->host;
2910         ssize_t ret;
2911         size_t count = iov_length(iov, nr_segs);
2912
2913         loff_t final_size = offset + count;
2914         if (rw == WRITE && final_size <= inode->i_size) {
2915                 /*
2916                  * We could direct write to holes and fallocate.
2917                  *
2918                  * Allocated blocks to fill the hole are marked as uninitialized
2919                  * to prevent parallel buffered read to expose the stale data
2920                  * before DIO complete the data IO.
2921                  *
2922                  * As to previously fallocated extents, ext4 get_block
2923                  * will just simply mark the buffer mapped but still
2924                  * keep the extents uninitialized.
2925                  *
2926                  * for non AIO case, we will convert those unwritten extents
2927                  * to written after return back from blockdev_direct_IO.
2928                  *
2929                  * for async DIO, the conversion needs to be defered when
2930                  * the IO is completed. The ext4 end_io callback function
2931                  * will be called to take care of the conversion work.
2932                  * Here for async case, we allocate an io_end structure to
2933                  * hook to the iocb.
2934                  */
2935                 iocb->private = NULL;
2936                 EXT4_I(inode)->cur_aio_dio = NULL;
2937                 if (!is_sync_kiocb(iocb)) {
2938                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
2939                         if (!iocb->private)
2940                                 return -ENOMEM;
2941                         /*
2942                          * we save the io structure for current async
2943                          * direct IO, so that later ext4_map_blocks()
2944                          * could flag the io structure whether there
2945                          * is a unwritten extents needs to be converted
2946                          * when IO is completed.
2947                          */
2948                         EXT4_I(inode)->cur_aio_dio = iocb->private;
2949                 }
2950
2951                 ret = __blockdev_direct_IO(rw, iocb, inode,
2952                                          inode->i_sb->s_bdev, iov,
2953                                          offset, nr_segs,
2954                                          ext4_get_block_write,
2955                                          ext4_end_io_dio,
2956                                          NULL,
2957                                          DIO_LOCKING | DIO_SKIP_HOLES);
2958                 if (iocb->private)
2959                         EXT4_I(inode)->cur_aio_dio = NULL;
2960                 /*
2961                  * The io_end structure takes a reference to the inode,
2962                  * that structure needs to be destroyed and the
2963                  * reference to the inode need to be dropped, when IO is
2964                  * complete, even with 0 byte write, or failed.
2965                  *
2966                  * In the successful AIO DIO case, the io_end structure will be
2967                  * desctroyed and the reference to the inode will be dropped
2968                  * after the end_io call back function is called.
2969                  *
2970                  * In the case there is 0 byte write, or error case, since
2971                  * VFS direct IO won't invoke the end_io call back function,
2972                  * we need to free the end_io structure here.
2973                  */
2974                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
2975                         ext4_free_io_end(iocb->private);
2976                         iocb->private = NULL;
2977                 } else if (ret > 0 && ext4_test_inode_state(inode,
2978                                                 EXT4_STATE_DIO_UNWRITTEN)) {
2979                         int err;
2980                         /*
2981                          * for non AIO case, since the IO is already
2982                          * completed, we could do the conversion right here
2983                          */
2984                         err = ext4_convert_unwritten_extents(inode,
2985                                                              offset, ret);
2986                         if (err < 0)
2987                                 ret = err;
2988                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
2989                 }
2990                 return ret;
2991         }
2992
2993         /* for write the the end of file case, we fall back to old way */
2994         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2995 }
2996
2997 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
2998                               const struct iovec *iov, loff_t offset,
2999                               unsigned long nr_segs)
3000 {
3001         struct file *file = iocb->ki_filp;
3002         struct inode *inode = file->f_mapping->host;
3003         ssize_t ret;
3004
3005         /*
3006          * If we are doing data journalling we don't support O_DIRECT
3007          */
3008         if (ext4_should_journal_data(inode))
3009                 return 0;
3010
3011         trace_ext4_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
3012         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
3013                 ret = ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
3014         else
3015                 ret = ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
3016         trace_ext4_direct_IO_exit(inode, offset,
3017                                 iov_length(iov, nr_segs), rw, ret);
3018         return ret;
3019 }
3020
3021 /*
3022  * Pages can be marked dirty completely asynchronously from ext4's journalling
3023  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
3024  * much here because ->set_page_dirty is called under VFS locks.  The page is
3025  * not necessarily locked.
3026  *
3027  * We cannot just dirty the page and leave attached buffers clean, because the
3028  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
3029  * or jbddirty because all the journalling code will explode.
3030  *
3031  * So what we do is to mark the page "pending dirty" and next time writepage
3032  * is called, propagate that into the buffers appropriately.
3033  */
3034 static int ext4_journalled_set_page_dirty(struct page *page)
3035 {
3036         SetPageChecked(page);
3037         return __set_page_dirty_nobuffers(page);
3038 }
3039
3040 static const struct address_space_operations ext4_ordered_aops = {
3041         .readpage               = ext4_readpage,
3042         .readpages              = ext4_readpages,
3043         .writepage              = ext4_writepage,
3044         .write_begin            = ext4_write_begin,
3045         .write_end              = ext4_ordered_write_end,
3046         .bmap                   = ext4_bmap,
3047         .invalidatepage         = ext4_invalidatepage,
3048         .releasepage            = ext4_releasepage,
3049         .direct_IO              = ext4_direct_IO,
3050         .migratepage            = buffer_migrate_page,
3051         .is_partially_uptodate  = block_is_partially_uptodate,
3052         .error_remove_page      = generic_error_remove_page,
3053 };