ext4: fix data corruption for mmap writes
[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 #include <linux/bitops.h>
42
43 #include "ext4_jbd2.h"
44 #include "xattr.h"
45 #include "acl.h"
46 #include "truncate.h"
47
48 #include <trace/events/ext4.h>
49
50 #define MPAGE_DA_EXTENT_TAIL 0x01
51
52 static inline int ext4_begin_ordered_truncate(struct inode *inode,
53                                               loff_t new_size)
54 {
55         trace_ext4_begin_ordered_truncate(inode, new_size);
56         /*
57          * If jinode is zero, then we never opened the file for
58          * writing, so there's no need to call
59          * jbd2_journal_begin_ordered_truncate() since there's no
60          * outstanding writes we need to flush.
61          */
62         if (!EXT4_I(inode)->jinode)
63                 return 0;
64         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
65                                                    EXT4_I(inode)->jinode,
66                                                    new_size);
67 }
68
69 static void ext4_invalidatepage(struct page *page, unsigned long offset);
70 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
71                                    struct buffer_head *bh_result, int create);
72 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
73 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
74 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
75 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
76
77 /*
78  * Test whether an inode is a fast symlink.
79  */
80 static int ext4_inode_is_fast_symlink(struct inode *inode)
81 {
82         int ea_blocks = EXT4_I(inode)->i_file_acl ?
83                 (inode->i_sb->s_blocksize >> 9) : 0;
84
85         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
86 }
87
88 /*
89  * Restart the transaction associated with *handle.  This does a commit,
90  * so before we call here everything must be consistently dirtied against
91  * this transaction.
92  */
93 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
94                                  int nblocks)
95 {
96         int ret;
97
98         /*
99          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
100          * moment, get_block can be called only for blocks inside i_size since
101          * page cache has been already dropped and writes are blocked by
102          * i_mutex. So we can safely drop the i_data_sem here.
103          */
104         BUG_ON(EXT4_JOURNAL(inode) == NULL);
105         jbd_debug(2, "restarting handle %p\n", handle);
106         up_write(&EXT4_I(inode)->i_data_sem);
107         ret = ext4_journal_restart(handle, nblocks);
108         down_write(&EXT4_I(inode)->i_data_sem);
109         ext4_discard_preallocations(inode);
110
111         return ret;
112 }
113
114 /*
115  * Called at the last iput() if i_nlink is zero.
116  */
117 void ext4_evict_inode(struct inode *inode)
118 {
119         handle_t *handle;
120         int err;
121
122         trace_ext4_evict_inode(inode);
123
124         ext4_ioend_wait(inode);
125
126         if (inode->i_nlink) {
127                 /*
128                  * When journalling data dirty buffers are tracked only in the
129                  * journal. So although mm thinks everything is clean and
130                  * ready for reaping the inode might still have some pages to
131                  * write in the running transaction or waiting to be
132                  * checkpointed. Thus calling jbd2_journal_invalidatepage()
133                  * (via truncate_inode_pages()) to discard these buffers can
134                  * cause data loss. Also even if we did not discard these
135                  * buffers, we would have no way to find them after the inode
136                  * is reaped and thus user could see stale data if he tries to
137                  * read them before the transaction is checkpointed. So be
138                  * careful and force everything to disk here... We use
139                  * ei->i_datasync_tid to store the newest transaction
140                  * containing inode's data.
141                  *
142                  * Note that directories do not have this problem because they
143                  * don't use page cache.
144                  */
145                 if (inode->i_ino != EXT4_JOURNAL_INO &&
146                     ext4_should_journal_data(inode) &&
147                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
148                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
149                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
150
151                         jbd2_complete_transaction(journal, commit_tid);
152                         filemap_write_and_wait(&inode->i_data);
153                 }
154                 truncate_inode_pages(&inode->i_data, 0);
155                 goto no_delete;
156         }
157
158         if (is_bad_inode(inode))
159                 goto no_delete;
160         dquot_initialize(inode);
161
162         if (ext4_should_order_data(inode))
163                 ext4_begin_ordered_truncate(inode, 0);
164         truncate_inode_pages(&inode->i_data, 0);
165
166
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         if (unlikely(ei->i_allocated_meta_blocks > ei->i_reserved_meta_blocks)) {
281                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, allocated %d "
282                          "with only %d reserved metadata blocks\n", __func__,
283                          inode->i_ino, ei->i_allocated_meta_blocks,
284                          ei->i_reserved_meta_blocks);
285                 WARN_ON(1);
286                 ei->i_allocated_meta_blocks = ei->i_reserved_meta_blocks;
287         }
288
289         /* Update per-inode reservations */
290         ei->i_reserved_data_blocks -= used;
291         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
292         percpu_counter_sub(&sbi->s_dirtyclusters_counter,
293                            used + ei->i_allocated_meta_blocks);
294         ei->i_allocated_meta_blocks = 0;
295
296         if (ei->i_reserved_data_blocks == 0) {
297                 /*
298                  * We can release all of the reserved metadata blocks
299                  * only when we have written all of the delayed
300                  * allocation blocks.
301                  */
302                 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
303                                    ei->i_reserved_meta_blocks);
304                 ei->i_reserved_meta_blocks = 0;
305                 ei->i_da_metadata_calc_len = 0;
306         }
307         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
308
309         /* Update quota subsystem for data blocks */
310         if (quota_claim)
311                 dquot_claim_block(inode, EXT4_C2B(sbi, used));
312         else {
313                 /*
314                  * We did fallocate with an offset that is already delayed
315                  * allocated. So on delayed allocated writeback we should
316                  * not re-claim the quota for fallocated blocks.
317                  */
318                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, used));
319         }
320
321         /*
322          * If we have done all the pending block allocations and if
323          * there aren't any writers on the inode, we can discard the
324          * inode's preallocations.
325          */
326         if ((ei->i_reserved_data_blocks == 0) &&
327             (atomic_read(&inode->i_writecount) == 0))
328                 ext4_discard_preallocations(inode);
329 }
330
331 static int __check_block_validity(struct inode *inode, const char *func,
332                                 unsigned int line,
333                                 struct ext4_map_blocks *map)
334 {
335         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
336                                    map->m_len)) {
337                 ext4_error_inode(inode, func, line, map->m_pblk,
338                                  "lblock %lu mapped to illegal pblock "
339                                  "(length %d)", (unsigned long) map->m_lblk,
340                                  map->m_len);
341                 return -EIO;
342         }
343         return 0;
344 }
345
346 #define check_block_validity(inode, map)        \
347         __check_block_validity((inode), __func__, __LINE__, (map))
348
349 /*
350  * Return the number of contiguous dirty pages in a given inode
351  * starting at page frame idx.
352  */
353 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
354                                     unsigned int max_pages)
355 {
356         struct address_space *mapping = inode->i_mapping;
357         pgoff_t index;
358         struct pagevec pvec;
359         pgoff_t num = 0;
360         int i, nr_pages, done = 0;
361
362         if (max_pages == 0)
363                 return 0;
364         pagevec_init(&pvec, 0);
365         while (!done) {
366                 index = idx;
367                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
368                                               PAGECACHE_TAG_DIRTY,
369                                               (pgoff_t)PAGEVEC_SIZE);
370                 if (nr_pages == 0)
371                         break;
372                 for (i = 0; i < nr_pages; i++) {
373                         struct page *page = pvec.pages[i];
374                         struct buffer_head *bh, *head;
375
376                         lock_page(page);
377                         if (unlikely(page->mapping != mapping) ||
378                             !PageDirty(page) ||
379                             PageWriteback(page) ||
380                             page->index != idx) {
381                                 done = 1;
382                                 unlock_page(page);
383                                 break;
384                         }
385                         if (page_has_buffers(page)) {
386                                 bh = head = page_buffers(page);
387                                 do {
388                                         if (!buffer_delay(bh) &&
389                                             !buffer_unwritten(bh))
390                                                 done = 1;
391                                         bh = bh->b_this_page;
392                                 } while (!done && (bh != head));
393                         }
394                         unlock_page(page);
395                         if (done)
396                                 break;
397                         idx++;
398                         num++;
399                         if (num >= max_pages) {
400                                 done = 1;
401                                 break;
402                         }
403                 }
404                 pagevec_release(&pvec);
405         }
406         return num;
407 }
408
409 /*
410  * Sets the BH_Da_Mapped bit on the buffer heads corresponding to the given map.
411  */
412 static void set_buffers_da_mapped(struct inode *inode,
413                                    struct ext4_map_blocks *map)
414 {
415         struct address_space *mapping = inode->i_mapping;
416         struct pagevec pvec;
417         int i, nr_pages;
418         pgoff_t index, end;
419
420         index = map->m_lblk >> (PAGE_CACHE_SHIFT - inode->i_blkbits);
421         end = (map->m_lblk + map->m_len - 1) >>
422                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
423
424         pagevec_init(&pvec, 0);
425         while (index <= end) {
426                 nr_pages = pagevec_lookup(&pvec, mapping, index,
427                                           min(end - index + 1,
428                                               (pgoff_t)PAGEVEC_SIZE));
429                 if (nr_pages == 0)
430                         break;
431                 for (i = 0; i < nr_pages; i++) {
432                         struct page *page = pvec.pages[i];
433                         struct buffer_head *bh, *head;
434
435                         if (unlikely(page->mapping != mapping) ||
436                             !PageDirty(page))
437                                 break;
438
439                         if (page_has_buffers(page)) {
440                                 bh = head = page_buffers(page);
441                                 do {
442                                         set_buffer_da_mapped(bh);
443                                         bh = bh->b_this_page;
444                                 } while (bh != head);
445                         }
446                         index++;
447                 }
448                 pagevec_release(&pvec);
449         }
450 }
451
452 /*
453  * The ext4_map_blocks() function tries to look up the requested blocks,
454  * and returns if the blocks are already mapped.
455  *
456  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
457  * and store the allocated blocks in the result buffer head and mark it
458  * mapped.
459  *
460  * If file type is extents based, it will call ext4_ext_map_blocks(),
461  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
462  * based files
463  *
464  * On success, it returns the number of blocks being mapped or allocate.
465  * if create==0 and the blocks are pre-allocated and uninitialized block,
466  * the result buffer head is unmapped. If the create ==1, it will make sure
467  * the buffer head is mapped.
468  *
469  * It returns 0 if plain look up failed (blocks have not been allocated), in
470  * that case, buffer head is unmapped
471  *
472  * It returns the error in case of allocation failure.
473  */
474 int ext4_map_blocks(handle_t *handle, struct inode *inode,
475                     struct ext4_map_blocks *map, int flags)
476 {
477         int retval;
478
479         map->m_flags = 0;
480         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
481                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
482                   (unsigned long) map->m_lblk);
483
484         /* We can handle the block number less than EXT_MAX_BLOCKS */
485         if (unlikely(map->m_lblk >= EXT_MAX_BLOCKS))
486                 return -EIO;
487
488         /*
489          * Try to see if we can get the block without requesting a new
490          * file system block.
491          */
492         down_read((&EXT4_I(inode)->i_data_sem));
493         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
494                 retval = ext4_ext_map_blocks(handle, inode, map, flags &
495                                              EXT4_GET_BLOCKS_KEEP_SIZE);
496         } else {
497                 retval = ext4_ind_map_blocks(handle, inode, map, flags &
498                                              EXT4_GET_BLOCKS_KEEP_SIZE);
499         }
500         up_read((&EXT4_I(inode)->i_data_sem));
501
502         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
503                 int ret = check_block_validity(inode, map);
504                 if (ret != 0)
505                         return ret;
506         }
507
508         /* If it is only a block(s) look up */
509         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
510                 return retval;
511
512         /*
513          * Returns if the blocks have already allocated
514          *
515          * Note that if blocks have been preallocated
516          * ext4_ext_get_block() returns the create = 0
517          * with buffer head unmapped.
518          */
519         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
520                 return retval;
521
522         /*
523          * When we call get_blocks without the create flag, the
524          * BH_Unwritten flag could have gotten set if the blocks
525          * requested were part of a uninitialized extent.  We need to
526          * clear this flag now that we are committed to convert all or
527          * part of the uninitialized extent to be an initialized
528          * extent.  This is because we need to avoid the combination
529          * of BH_Unwritten and BH_Mapped flags being simultaneously
530          * set on the buffer_head.
531          */
532         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
533
534         /*
535          * New blocks allocate and/or writing to uninitialized extent
536          * will possibly result in updating i_data, so we take
537          * the write lock of i_data_sem, and call get_blocks()
538          * with create == 1 flag.
539          */
540         down_write((&EXT4_I(inode)->i_data_sem));
541
542         /*
543          * if the caller is from delayed allocation writeout path
544          * we have already reserved fs blocks for allocation
545          * let the underlying get_block() function know to
546          * avoid double accounting
547          */
548         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
549                 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
550         /*
551          * We need to check for EXT4 here because migrate
552          * could have changed the inode type in between
553          */
554         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
555                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
556         } else {
557                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
558
559                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
560                         /*
561                          * We allocated new blocks which will result in
562                          * i_data's format changing.  Force the migrate
563                          * to fail by clearing migrate flags
564                          */
565                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
566                 }
567
568                 /*
569                  * Update reserved blocks/metadata blocks after successful
570                  * block allocation which had been deferred till now. We don't
571                  * support fallocate for non extent files. So we can update
572                  * reserve space here.
573                  */
574                 if ((retval > 0) &&
575                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
576                         ext4_da_update_reserve_space(inode, retval, 1);
577         }
578         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE) {
579                 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
580
581                 /* If we have successfully mapped the delayed allocated blocks,
582                  * set the BH_Da_Mapped bit on them. Its important to do this
583                  * under the protection of i_data_sem.
584                  */
585                 if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
586                         set_buffers_da_mapped(inode, map);
587         }
588
589         up_write((&EXT4_I(inode)->i_data_sem));
590         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
591                 int ret = check_block_validity(inode, map);
592                 if (ret != 0)
593                         return ret;
594         }
595         return retval;
596 }
597
598 /*
599  * Update EXT4_MAP_FLAGS in bh->b_state. For buffer heads attached to pages
600  * we have to be careful as someone else may be manipulating b_state as well.
601  */
602 static void ext4_update_bh_state(struct buffer_head *bh, unsigned long flags)
603 {
604         unsigned long old_state;
605         unsigned long new_state;
606
607         flags &= EXT4_MAP_FLAGS;
608
609         /* Dummy buffer_head? Set non-atomically. */
610         if (!bh->b_page) {
611                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | flags;
612                 return;
613         }
614         /*
615          * Someone else may be modifying b_state. Be careful! This is ugly but
616          * once we get rid of using bh as a container for mapping information
617          * to pass to / from get_block functions, this can go away.
618          */
619         do {
620                 old_state = ACCESS_ONCE(bh->b_state);
621                 new_state = (old_state & ~EXT4_MAP_FLAGS) | flags;
622         } while (unlikely(
623                  cmpxchg(&bh->b_state, old_state, new_state) != old_state));
624 }
625
626 /* Maximum number of blocks we map for direct IO at once. */
627 #define DIO_MAX_BLOCKS 4096
628
629 static int _ext4_get_block(struct inode *inode, sector_t iblock,
630                            struct buffer_head *bh, int flags)
631 {
632         handle_t *handle = ext4_journal_current_handle();
633         struct ext4_map_blocks map;
634         int ret = 0, started = 0;
635         int dio_credits;
636
637         map.m_lblk = iblock;
638         map.m_len = bh->b_size >> inode->i_blkbits;
639
640         if (flags && !handle) {
641                 /* Direct IO write... */
642                 if (map.m_len > DIO_MAX_BLOCKS)
643                         map.m_len = DIO_MAX_BLOCKS;
644                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
645                 handle = ext4_journal_start(inode, dio_credits);
646                 if (IS_ERR(handle)) {
647                         ret = PTR_ERR(handle);
648                         return ret;
649                 }
650                 started = 1;
651         }
652
653         ret = ext4_map_blocks(handle, inode, &map, flags);
654         if (ret > 0) {
655                 map_bh(bh, inode->i_sb, map.m_pblk);
656                 ext4_update_bh_state(bh, map.m_flags);
657                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
658                 ret = 0;
659         }
660         if (started)
661                 ext4_journal_stop(handle);
662         return ret;
663 }
664
665 int ext4_get_block(struct inode *inode, sector_t iblock,
666                    struct buffer_head *bh, int create)
667 {
668         return _ext4_get_block(inode, iblock, bh,
669                                create ? EXT4_GET_BLOCKS_CREATE : 0);
670 }
671
672 /*
673  * `handle' can be NULL if create is zero
674  */
675 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
676                                 ext4_lblk_t block, int create, int *errp)
677 {
678         struct ext4_map_blocks map;
679         struct buffer_head *bh;
680         int fatal = 0, err;
681
682         J_ASSERT(handle != NULL || create == 0);
683
684         map.m_lblk = block;
685         map.m_len = 1;
686         err = ext4_map_blocks(handle, inode, &map,
687                               create ? EXT4_GET_BLOCKS_CREATE : 0);
688
689         if (err < 0)
690                 *errp = err;
691         if (err <= 0)
692                 return NULL;
693         *errp = 0;
694
695         bh = sb_getblk(inode->i_sb, map.m_pblk);
696         if (!bh) {
697                 *errp = -ENOMEM;
698                 return NULL;
699         }
700         if (map.m_flags & EXT4_MAP_NEW) {
701                 J_ASSERT(create != 0);
702                 J_ASSERT(handle != NULL);
703
704                 /*
705                  * Now that we do not always journal data, we should
706                  * keep in mind whether this should always journal the
707                  * new buffer as metadata.  For now, regular file
708                  * writes use ext4_get_block instead, so it's not a
709                  * problem.
710                  */
711                 lock_buffer(bh);
712                 BUFFER_TRACE(bh, "call get_create_access");
713                 fatal = ext4_journal_get_create_access(handle, bh);
714                 if (!fatal && !buffer_uptodate(bh)) {
715                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
716                         set_buffer_uptodate(bh);
717                 }
718                 unlock_buffer(bh);
719                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
720                 err = ext4_handle_dirty_metadata(handle, inode, bh);
721                 if (!fatal)
722                         fatal = err;
723         } else {
724                 BUFFER_TRACE(bh, "not a new buffer");
725         }
726         if (fatal) {
727                 *errp = fatal;
728                 brelse(bh);
729                 bh = NULL;
730         }
731         return bh;
732 }
733
734 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
735                                ext4_lblk_t block, int create, int *err)
736 {
737         struct buffer_head *bh;
738
739         bh = ext4_getblk(handle, inode, block, create, err);
740         if (!bh)
741                 return bh;
742         if (buffer_uptodate(bh))
743                 return bh;
744         ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
745         wait_on_buffer(bh);
746         if (buffer_uptodate(bh))
747                 return bh;
748         put_bh(bh);
749         *err = -EIO;
750         return NULL;
751 }
752
753 static int walk_page_buffers(handle_t *handle,
754                              struct buffer_head *head,
755                              unsigned from,
756                              unsigned to,
757                              int *partial,
758                              int (*fn)(handle_t *handle,
759                                        struct buffer_head *bh))
760 {
761         struct buffer_head *bh;
762         unsigned block_start, block_end;
763         unsigned blocksize = head->b_size;
764         int err, ret = 0;
765         struct buffer_head *next;
766
767         for (bh = head, block_start = 0;
768              ret == 0 && (bh != head || !block_start);
769              block_start = block_end, bh = next) {
770                 next = bh->b_this_page;
771                 block_end = block_start + blocksize;
772                 if (block_end <= from || block_start >= to) {
773                         if (partial && !buffer_uptodate(bh))
774                                 *partial = 1;
775                         continue;
776                 }
777                 err = (*fn)(handle, bh);
778                 if (!ret)
779                         ret = err;
780         }
781         return ret;
782 }
783
784 /*
785  * To preserve ordering, it is essential that the hole instantiation and
786  * the data write be encapsulated in a single transaction.  We cannot
787  * close off a transaction and start a new one between the ext4_get_block()
788  * and the commit_write().  So doing the jbd2_journal_start at the start of
789  * prepare_write() is the right place.
790  *
791  * Also, this function can nest inside ext4_writepage() ->
792  * block_write_full_page(). In that case, we *know* that ext4_writepage()
793  * has generated enough buffer credits to do the whole page.  So we won't
794  * block on the journal in that case, which is good, because the caller may
795  * be PF_MEMALLOC.
796  *
797  * By accident, ext4 can be reentered when a transaction is open via
798  * quota file writes.  If we were to commit the transaction while thus
799  * reentered, there can be a deadlock - we would be holding a quota
800  * lock, and the commit would never complete if another thread had a
801  * transaction open and was blocking on the quota lock - a ranking
802  * violation.
803  *
804  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
805  * will _not_ run commit under these circumstances because handle->h_ref
806  * is elevated.  We'll still have enough credits for the tiny quotafile
807  * write.
808  */
809 static int do_journal_get_write_access(handle_t *handle,
810                                        struct buffer_head *bh)
811 {
812         int dirty = buffer_dirty(bh);
813         int ret;
814
815         if (!buffer_mapped(bh) || buffer_freed(bh))
816                 return 0;
817         /*
818          * __block_write_begin() could have dirtied some buffers. Clean
819          * the dirty bit as jbd2_journal_get_write_access() could complain
820          * otherwise about fs integrity issues. Setting of the dirty bit
821          * by __block_write_begin() isn't a real problem here as we clear
822          * the bit before releasing a page lock and thus writeback cannot
823          * ever write the buffer.
824          */
825         if (dirty)
826                 clear_buffer_dirty(bh);
827         ret = ext4_journal_get_write_access(handle, bh);
828         if (!ret && dirty)
829                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
830         return ret;
831 }
832
833 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
834                    struct buffer_head *bh_result, int create);
835 static int ext4_write_begin(struct file *file, struct address_space *mapping,
836                             loff_t pos, unsigned len, unsigned flags,
837                             struct page **pagep, void **fsdata)
838 {
839         struct inode *inode = mapping->host;
840         int ret, needed_blocks;
841         handle_t *handle;
842         int retries = 0;
843         struct page *page;
844         pgoff_t index;
845         unsigned from, to;
846
847         trace_ext4_write_begin(inode, pos, len, flags);
848         /*
849          * Reserve one block more for addition to orphan list in case
850          * we allocate blocks but write fails for some reason
851          */
852         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
853         index = pos >> PAGE_CACHE_SHIFT;
854         from = pos & (PAGE_CACHE_SIZE - 1);
855         to = from + len;
856
857 retry:
858         handle = ext4_journal_start(inode, needed_blocks);
859         if (IS_ERR(handle)) {
860                 ret = PTR_ERR(handle);
861                 goto out;
862         }
863
864         /* We cannot recurse into the filesystem as the transaction is already
865          * started */
866         flags |= AOP_FLAG_NOFS;
867
868         page = grab_cache_page_write_begin(mapping, index, flags);
869         if (!page) {
870                 ext4_journal_stop(handle);
871                 ret = -ENOMEM;
872                 goto out;
873         }
874         *pagep = page;
875
876         if (ext4_should_dioread_nolock(inode))
877                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
878         else
879                 ret = __block_write_begin(page, pos, len, ext4_get_block);
880
881         if (!ret && ext4_should_journal_data(inode)) {
882                 ret = walk_page_buffers(handle, page_buffers(page),
883                                 from, to, NULL, do_journal_get_write_access);
884         }
885
886         if (ret) {
887                 unlock_page(page);
888                 page_cache_release(page);
889                 /*
890                  * __block_write_begin may have instantiated a few blocks
891                  * outside i_size.  Trim these off again. Don't need
892                  * i_size_read because we hold i_mutex.
893                  *
894                  * Add inode to orphan list in case we crash before
895                  * truncate finishes
896                  */
897                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
898                         ext4_orphan_add(handle, inode);
899
900                 ext4_journal_stop(handle);
901                 if (pos + len > inode->i_size) {
902                         ext4_truncate_failed_write(inode);
903                         /*
904                          * If truncate failed early the inode might
905                          * still be on the orphan list; we need to
906                          * make sure the inode is removed from the
907                          * orphan list in that case.
908                          */
909                         if (inode->i_nlink)
910                                 ext4_orphan_del(NULL, inode);
911                 }
912         }
913
914         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
915                 goto retry;
916 out:
917         return ret;
918 }
919
920 /* For write_end() in data=journal mode */
921 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
922 {
923         if (!buffer_mapped(bh) || buffer_freed(bh))
924                 return 0;
925         set_buffer_uptodate(bh);
926         return ext4_handle_dirty_metadata(handle, NULL, bh);
927 }
928
929 static int ext4_generic_write_end(struct file *file,
930                                   struct address_space *mapping,
931                                   loff_t pos, unsigned len, unsigned copied,
932                                   struct page *page, void *fsdata)
933 {
934         int i_size_changed = 0;
935         struct inode *inode = mapping->host;
936         handle_t *handle = ext4_journal_current_handle();
937
938         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
939
940         /*
941          * No need to use i_size_read() here, the i_size
942          * cannot change under us because we hold i_mutex.
943          *
944          * But it's important to update i_size while still holding page lock:
945          * page writeout could otherwise come in and zero beyond i_size.
946          */
947         if (pos + copied > inode->i_size) {
948                 i_size_write(inode, pos + copied);
949                 i_size_changed = 1;
950         }
951
952         if (pos + copied >  EXT4_I(inode)->i_disksize) {
953                 /* We need to mark inode dirty even if
954                  * new_i_size is less that inode->i_size
955                  * bu greater than i_disksize.(hint delalloc)
956                  */
957                 ext4_update_i_disksize(inode, (pos + copied));
958                 i_size_changed = 1;
959         }
960         unlock_page(page);
961         page_cache_release(page);
962
963         /*
964          * Don't mark the inode dirty under page lock. First, it unnecessarily
965          * makes the holding time of page lock longer. Second, it forces lock
966          * ordering of page lock and transaction start for journaling
967          * filesystems.
968          */
969         if (i_size_changed)
970                 ext4_mark_inode_dirty(handle, inode);
971
972         return copied;
973 }
974
975 /*
976  * We need to pick up the new inode size which generic_commit_write gave us
977  * `file' can be NULL - eg, when called from page_symlink().
978  *
979  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
980  * buffers are managed internally.
981  */
982 static int ext4_ordered_write_end(struct file *file,
983                                   struct address_space *mapping,
984                                   loff_t pos, unsigned len, unsigned copied,
985                                   struct page *page, void *fsdata)
986 {
987         handle_t *handle = ext4_journal_current_handle();
988         struct inode *inode = mapping->host;
989         int ret = 0, ret2;
990
991         trace_ext4_ordered_write_end(inode, pos, len, copied);
992         ret = ext4_jbd2_file_inode(handle, inode);
993
994         if (ret == 0) {
995                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
996                                                         page, fsdata);
997                 copied = ret2;
998                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
999                         /* if we have allocated more blocks and copied
1000                          * less. We will have blocks allocated outside
1001                          * inode->i_size. So truncate them
1002                          */
1003                         ext4_orphan_add(handle, inode);
1004                 if (ret2 < 0)
1005                         ret = ret2;
1006         } else {
1007                 unlock_page(page);
1008                 page_cache_release(page);
1009         }
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
1027         return ret ? ret : copied;
1028 }
1029
1030 static int ext4_writeback_write_end(struct file *file,
1031                                     struct address_space *mapping,
1032                                     loff_t pos, unsigned len, unsigned copied,
1033                                     struct page *page, void *fsdata)
1034 {
1035         handle_t *handle = ext4_journal_current_handle();
1036         struct inode *inode = mapping->host;
1037         int ret = 0, ret2;
1038
1039         trace_ext4_writeback_write_end(inode, pos, len, copied);
1040         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
1041                                                         page, fsdata);
1042         copied = ret2;
1043         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1044                 /* if we have allocated more blocks and copied
1045                  * less. We will have blocks allocated outside
1046                  * inode->i_size. So truncate them
1047                  */
1048                 ext4_orphan_add(handle, inode);
1049
1050         if (ret2 < 0)
1051                 ret = ret2;
1052
1053         ret2 = ext4_journal_stop(handle);
1054         if (!ret)
1055                 ret = ret2;
1056
1057         if (pos + len > inode->i_size) {
1058                 ext4_truncate_failed_write(inode);
1059                 /*
1060                  * If truncate failed early the inode might still be
1061                  * on the orphan list; we need to make sure the inode
1062                  * is removed from the orphan list in that case.
1063                  */
1064                 if (inode->i_nlink)
1065                         ext4_orphan_del(NULL, inode);
1066         }
1067
1068         return ret ? ret : copied;
1069 }
1070
1071 /*
1072  * This is a private version of page_zero_new_buffers() which doesn't
1073  * set the buffer to be dirty, since in data=journalled mode we need
1074  * to call ext4_handle_dirty_metadata() instead.
1075  */
1076 static void ext4_journalled_zero_new_buffers(handle_t *handle,
1077                                             struct page *page,
1078                                             unsigned from, unsigned to)
1079 {
1080         unsigned int block_start = 0, block_end;
1081         struct buffer_head *head, *bh;
1082
1083         bh = head = page_buffers(page);
1084         do {
1085                 block_end = block_start + bh->b_size;
1086                 if (buffer_new(bh)) {
1087                         if (block_end > from && block_start < to) {
1088                                 if (!PageUptodate(page)) {
1089                                         unsigned start, size;
1090
1091                                         start = max(from, block_start);
1092                                         size = min(to, block_end) - start;
1093
1094                                         zero_user(page, start, size);
1095                                         write_end_fn(handle, bh);
1096                                 }
1097                                 clear_buffer_new(bh);
1098                         }
1099                 }
1100                 block_start = block_end;
1101                 bh = bh->b_this_page;
1102         } while (bh != head);
1103 }
1104
1105 static int ext4_journalled_write_end(struct file *file,
1106                                      struct address_space *mapping,
1107                                      loff_t pos, unsigned len, unsigned copied,
1108                                      struct page *page, void *fsdata)
1109 {
1110         handle_t *handle = ext4_journal_current_handle();
1111         struct inode *inode = mapping->host;
1112         int ret = 0, ret2;
1113         int partial = 0;
1114         unsigned from, to;
1115         loff_t new_i_size;
1116
1117         trace_ext4_journalled_write_end(inode, pos, len, copied);
1118         from = pos & (PAGE_CACHE_SIZE - 1);
1119         to = from + len;
1120
1121         BUG_ON(!ext4_handle_valid(handle));
1122
1123         if (unlikely(copied < len) && !PageUptodate(page)) {
1124                 copied = 0;
1125                 ext4_journalled_zero_new_buffers(handle, page, from, to);
1126         } else {
1127                 if (unlikely(copied < len))
1128                         ext4_journalled_zero_new_buffers(handle, page,
1129                                                          from + copied, to);
1130                 ret = walk_page_buffers(handle, page_buffers(page), from,
1131                                         from + copied, &partial,
1132                                         write_end_fn);
1133                 if (!partial)
1134                         SetPageUptodate(page);
1135         }
1136         new_i_size = pos + copied;
1137         if (new_i_size > inode->i_size)
1138                 i_size_write(inode, pos+copied);
1139         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1140         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1141         if (new_i_size > EXT4_I(inode)->i_disksize) {
1142                 ext4_update_i_disksize(inode, new_i_size);
1143                 ret2 = ext4_mark_inode_dirty(handle, inode);
1144                 if (!ret)
1145                         ret = ret2;
1146         }
1147
1148         unlock_page(page);
1149         page_cache_release(page);
1150         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1151                 /* if we have allocated more blocks and copied
1152                  * less. We will have blocks allocated outside
1153                  * inode->i_size. So truncate them
1154                  */
1155                 ext4_orphan_add(handle, inode);
1156
1157         ret2 = ext4_journal_stop(handle);
1158         if (!ret)
1159                 ret = ret2;
1160         if (pos + len > inode->i_size) {
1161                 ext4_truncate_failed_write(inode);
1162                 /*
1163                  * If truncate failed early the inode might still be
1164                  * on the orphan list; we need to make sure the inode
1165                  * is removed from the orphan list in that case.
1166                  */
1167                 if (inode->i_nlink)
1168                         ext4_orphan_del(NULL, inode);
1169         }
1170
1171         return ret ? ret : copied;
1172 }
1173
1174 /*
1175  * Reserve a single cluster located at lblock
1176  */
1177 static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1178 {
1179         int retries = 0;
1180         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1181         struct ext4_inode_info *ei = EXT4_I(inode);
1182         unsigned int md_needed;
1183         int ret;
1184         ext4_lblk_t save_last_lblock;
1185         int save_len;
1186
1187         /*
1188          * We will charge metadata quota at writeout time; this saves
1189          * us from metadata over-estimation, though we may go over by
1190          * a small amount in the end.  Here we just reserve for data.
1191          */
1192         ret = dquot_reserve_block(inode, EXT4_C2B(sbi, 1));
1193         if (ret)
1194                 return ret;
1195
1196         /*
1197          * recalculate the amount of metadata blocks to reserve
1198          * in order to allocate nrblocks
1199          * worse case is one extent per block
1200          */
1201 repeat:
1202         spin_lock(&ei->i_block_reservation_lock);
1203         /*
1204          * ext4_calc_metadata_amount() has side effects, which we have
1205          * to be prepared undo if we fail to claim space.
1206          */
1207         save_len = ei->i_da_metadata_calc_len;
1208         save_last_lblock = ei->i_da_metadata_calc_last_lblock;
1209         md_needed = EXT4_NUM_B2C(sbi,
1210                                  ext4_calc_metadata_amount(inode, lblock));
1211         trace_ext4_da_reserve_space(inode, md_needed);
1212
1213         /*
1214          * We do still charge estimated metadata to the sb though;
1215          * we cannot afford to run out of free blocks.
1216          */
1217         if (ext4_claim_free_clusters(sbi, md_needed + 1, 0)) {
1218                 ei->i_da_metadata_calc_len = save_len;
1219                 ei->i_da_metadata_calc_last_lblock = save_last_lblock;
1220                 spin_unlock(&ei->i_block_reservation_lock);
1221                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1222                         yield();
1223                         goto repeat;
1224                 }
1225                 dquot_release_reservation_block(inode, EXT4_C2B(sbi, 1));
1226                 return -ENOSPC;
1227         }
1228         ei->i_reserved_data_blocks++;
1229         ei->i_reserved_meta_blocks += md_needed;
1230         spin_unlock(&ei->i_block_reservation_lock);
1231
1232         return 0;       /* success */
1233 }
1234
1235 static void ext4_da_release_space(struct inode *inode, int to_free)
1236 {
1237         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1238         struct ext4_inode_info *ei = EXT4_I(inode);
1239
1240         if (!to_free)
1241                 return;         /* Nothing to release, exit */
1242
1243         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1244
1245         trace_ext4_da_release_space(inode, to_free);
1246         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1247                 /*
1248                  * if there aren't enough reserved blocks, then the
1249                  * counter is messed up somewhere.  Since this
1250                  * function is called from invalidate page, it's
1251                  * harmless to return without any action.
1252                  */
1253                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1254                          "ino %lu, to_free %d with only %d reserved "
1255                          "data blocks\n", inode->i_ino, to_free,
1256                          ei->i_reserved_data_blocks);
1257                 WARN_ON(1);
1258                 to_free = ei->i_reserved_data_blocks;
1259         }
1260         ei->i_reserved_data_blocks -= to_free;
1261
1262         if (ei->i_reserved_data_blocks == 0) {
1263                 /*
1264                  * We can release all of the reserved metadata blocks
1265                  * only when we have written all of the delayed
1266                  * allocation blocks.
1267                  * Note that in case of bigalloc, i_reserved_meta_blocks,
1268                  * i_reserved_data_blocks, etc. refer to number of clusters.
1269                  */
1270                 percpu_counter_sub(&sbi->s_dirtyclusters_counter,
1271                                    ei->i_reserved_meta_blocks);
1272                 ei->i_reserved_meta_blocks = 0;
1273                 ei->i_da_metadata_calc_len = 0;
1274         }
1275
1276         /* update fs dirty data blocks counter */
1277         percpu_counter_sub(&sbi->s_dirtyclusters_counter, to_free);
1278
1279         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1280
1281         dquot_release_reservation_block(inode, EXT4_C2B(sbi, to_free));
1282 }
1283
1284 static void ext4_da_page_release_reservation(struct page *page,
1285                                              unsigned long offset)
1286 {
1287         int to_release = 0;
1288         struct buffer_head *head, *bh;
1289         unsigned int curr_off = 0;
1290         struct inode *inode = page->mapping->host;
1291         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1292         int num_clusters;
1293
1294         head = page_buffers(page);
1295         bh = head;
1296         do {
1297                 unsigned int next_off = curr_off + bh->b_size;
1298
1299                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1300                         to_release++;
1301                         clear_buffer_delay(bh);
1302                         clear_buffer_da_mapped(bh);
1303                 }
1304                 curr_off = next_off;
1305         } while ((bh = bh->b_this_page) != head);
1306
1307         /* If we have released all the blocks belonging to a cluster, then we
1308          * need to release the reserved space for that cluster. */
1309         num_clusters = EXT4_NUM_B2C(sbi, to_release);
1310         while (num_clusters > 0) {
1311                 ext4_fsblk_t lblk;
1312                 lblk = (page->index << (PAGE_CACHE_SHIFT - inode->i_blkbits)) +
1313                         ((num_clusters - 1) << sbi->s_cluster_bits);
1314                 if (sbi->s_cluster_ratio == 1 ||
1315                     !ext4_find_delalloc_cluster(inode, lblk, 1))
1316                         ext4_da_release_space(inode, 1);
1317
1318                 num_clusters--;
1319         }
1320 }
1321
1322 /*
1323  * Delayed allocation stuff
1324  */
1325
1326 /*
1327  * mpage_da_submit_io - walks through extent of pages and try to write
1328  * them with writepage() call back
1329  *
1330  * @mpd->inode: inode
1331  * @mpd->first_page: first page of the extent
1332  * @mpd->next_page: page after the last page of the extent
1333  *
1334  * By the time mpage_da_submit_io() is called we expect all blocks
1335  * to be allocated. this may be wrong if allocation failed.
1336  *
1337  * As pages are already locked by write_cache_pages(), we can't use it
1338  */
1339 static int mpage_da_submit_io(struct mpage_da_data *mpd,
1340                               struct ext4_map_blocks *map)
1341 {
1342         struct pagevec pvec;
1343         unsigned long index, end;
1344         int ret = 0, err, nr_pages, i;
1345         struct inode *inode = mpd->inode;
1346         struct address_space *mapping = inode->i_mapping;
1347         unsigned int len, block_start;
1348         struct buffer_head *bh, *page_bufs = NULL;
1349         int journal_data = ext4_should_journal_data(inode);
1350         sector_t pblock = 0, cur_logical = 0;
1351         struct ext4_io_submit io_submit;
1352
1353         BUG_ON(mpd->next_page <= mpd->first_page);
1354         memset(&io_submit, 0, sizeof(io_submit));
1355         /*
1356          * We need to start from the first_page to the next_page - 1
1357          * to make sure we also write the mapped dirty buffer_heads.
1358          * If we look at mpd->b_blocknr we would only be looking
1359          * at the currently mapped buffer_heads.
1360          */
1361         index = mpd->first_page;
1362         end = mpd->next_page - 1;
1363
1364         pagevec_init(&pvec, 0);
1365         while (index <= end) {
1366                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1367                 if (nr_pages == 0)
1368                         break;
1369                 for (i = 0; i < nr_pages; i++) {
1370                         int commit_write = 0, skip_page = 0;
1371                         struct page *page = pvec.pages[i];
1372                         loff_t size = i_size_read(inode);
1373
1374                         index = page->index;
1375                         if (index > end)
1376                                 break;
1377
1378                         if (index == size >> PAGE_CACHE_SHIFT)
1379                                 len = size & ~PAGE_CACHE_MASK;
1380                         else
1381                                 len = PAGE_CACHE_SIZE;
1382                         if (map) {
1383                                 cur_logical = index << (PAGE_CACHE_SHIFT -
1384                                                         inode->i_blkbits);
1385                                 pblock = map->m_pblk + (cur_logical -
1386                                                         map->m_lblk);
1387                         }
1388                         index++;
1389
1390                         BUG_ON(!PageLocked(page));
1391                         BUG_ON(PageWriteback(page));
1392
1393                         /*
1394                          * If the page does not have buffers (for
1395                          * whatever reason), try to create them using
1396                          * __block_write_begin.  If this fails,
1397                          * skip the page and move on.
1398                          */
1399                         if (!page_has_buffers(page)) {
1400                                 if (__block_write_begin(page, 0, len,
1401                                                 noalloc_get_block_write)) {
1402                                 skip_page:
1403                                         unlock_page(page);
1404                                         continue;
1405                                 }
1406                                 commit_write = 1;
1407                         }
1408
1409                         bh = page_bufs = page_buffers(page);
1410                         block_start = 0;
1411                         do {
1412                                 if (!bh)
1413                                         goto skip_page;
1414                                 if (map && (cur_logical >= map->m_lblk) &&
1415                                     (cur_logical <= (map->m_lblk +
1416                                                      (map->m_len - 1)))) {
1417                                         if (buffer_delay(bh)) {
1418                                                 clear_buffer_delay(bh);
1419                                                 bh->b_blocknr = pblock;
1420                                         }
1421                                         if (buffer_da_mapped(bh))
1422                                                 clear_buffer_da_mapped(bh);
1423                                         if (buffer_unwritten(bh) ||
1424                                             buffer_mapped(bh))
1425                                                 BUG_ON(bh->b_blocknr != pblock);
1426                                         if (map->m_flags & EXT4_MAP_UNINIT)
1427                                                 set_buffer_uninit(bh);
1428                                         clear_buffer_unwritten(bh);
1429                                 }
1430
1431                                 /*
1432                                  * skip page if block allocation undone and
1433                                  * block is dirty
1434                                  */
1435                                 if (ext4_bh_delay_or_unwritten(NULL, bh))
1436                                         skip_page = 1;
1437                                 bh = bh->b_this_page;
1438                                 block_start += bh->b_size;
1439                                 cur_logical++;
1440                                 pblock++;
1441                         } while (bh != page_bufs);
1442
1443                         if (skip_page)
1444                                 goto skip_page;
1445
1446                         clear_page_dirty_for_io(page);
1447                         /*
1448                          * We have to be very careful here!  Nothing protects
1449                          * writeback path against i_size changes and the page
1450                          * can be writeably mapped into page tables. So an
1451                          * application can be growing i_size and writing data
1452                          * through mmap while writeback runs.
1453                          * clear_page_dirty_for_io() write-protects our page in
1454                          * page tables and the page cannot get written to again
1455                          * until we release page lock. So only after
1456                          * clear_page_dirty_for_io() we are safe to sample
1457                          * i_size for ext4_bio_write_page() to zero-out tail of
1458                          * the written page. We rely on the barrier provided by
1459                          * TestClearPageDirty in clear_page_dirty_for_io() to
1460                          * make sure i_size is really sampled only after page
1461                          * tables are updated.
1462                          */
1463                         if (size != i_size_read(inode)) {
1464                                 set_page_dirty(page);
1465                                 goto skip_page;
1466                         }
1467
1468                         if (commit_write)
1469                                 /* mark the buffer_heads as dirty & uptodate */
1470                                 block_commit_write(page, 0, len);
1471                         /*
1472                          * Delalloc doesn't support data journalling,
1473                          * but eventually maybe we'll lift this
1474                          * restriction.
1475                          */
1476                         if (unlikely(journal_data && PageChecked(page)))
1477                                 err = __ext4_journalled_writepage(page, len);
1478                         else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
1479                                 err = ext4_bio_write_page(&io_submit, page,
1480                                                           len, mpd->wbc);
1481                         else if (buffer_uninit(page_bufs)) {
1482                                 ext4_set_bh_endio(page_bufs, inode);
1483                                 err = block_write_full_page_endio(page,
1484                                         noalloc_get_block_write,
1485                                         mpd->wbc, ext4_end_io_buffer_write);
1486                         } else
1487                                 err = block_write_full_page(page,
1488                                         noalloc_get_block_write, mpd->wbc);
1489
1490                         if (!err)
1491                                 mpd->pages_written++;
1492                         /*
1493                          * In error case, we have to continue because
1494                          * remaining pages are still locked
1495                          */
1496                         if (ret == 0)
1497                                 ret = err;
1498                 }
1499                 pagevec_release(&pvec);
1500         }
1501         ext4_io_submit(&io_submit);
1502         return ret;
1503 }
1504
1505 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd)
1506 {
1507         int nr_pages, i;
1508         pgoff_t index, end;
1509         struct pagevec pvec;
1510         struct inode *inode = mpd->inode;
1511         struct address_space *mapping = inode->i_mapping;
1512
1513         index = mpd->first_page;
1514         end   = mpd->next_page - 1;
1515         pagevec_init(&pvec, 0);
1516         while (index <= end) {
1517                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1518                 if (nr_pages == 0)
1519                         break;
1520                 for (i = 0; i < nr_pages; i++) {
1521                         struct page *page = pvec.pages[i];
1522                         if (page->index > end)
1523                                 break;
1524                         BUG_ON(!PageLocked(page));
1525                         BUG_ON(PageWriteback(page));
1526                         block_invalidatepage(page, 0);
1527                         ClearPageUptodate(page);
1528                         unlock_page(page);
1529                 }
1530                 index = pvec.pages[nr_pages - 1]->index + 1;
1531                 pagevec_release(&pvec);
1532         }
1533         return;
1534 }
1535
1536 static void ext4_print_free_blocks(struct inode *inode)
1537 {
1538         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1539         printk(KERN_CRIT "Total free blocks count %lld\n",
1540                EXT4_C2B(EXT4_SB(inode->i_sb),
1541                         ext4_count_free_clusters(inode->i_sb)));
1542         printk(KERN_CRIT "Free/Dirty block details\n");
1543         printk(KERN_CRIT "free_blocks=%lld\n",
1544                (long long) EXT4_C2B(EXT4_SB(inode->i_sb),
1545                 percpu_counter_sum(&sbi->s_freeclusters_counter)));
1546         printk(KERN_CRIT "dirty_blocks=%lld\n",
1547                (long long) EXT4_C2B(EXT4_SB(inode->i_sb),
1548                 percpu_counter_sum(&sbi->s_dirtyclusters_counter)));
1549         printk(KERN_CRIT "Block reservation details\n");
1550         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
1551                EXT4_I(inode)->i_reserved_data_blocks);
1552         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
1553                EXT4_I(inode)->i_reserved_meta_blocks);
1554         return;
1555 }
1556
1557 /*
1558  * mpage_da_map_and_submit - go through given space, map them
1559  *       if necessary, and then submit them for I/O
1560  *
1561  * @mpd - bh describing space
1562  *
1563  * The function skips space we know is already mapped to disk blocks.
1564  *
1565  */
1566 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
1567 {
1568         int err, blks, get_blocks_flags;
1569         struct ext4_map_blocks map, *mapp = NULL;
1570         sector_t next = mpd->b_blocknr;
1571         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
1572         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
1573         handle_t *handle = NULL;
1574
1575         /*
1576          * If the blocks are mapped already, or we couldn't accumulate
1577          * any blocks, then proceed immediately to the submission stage.
1578          */
1579         if ((mpd->b_size == 0) ||
1580             ((mpd->b_state  & (1 << BH_Mapped)) &&
1581              !(mpd->b_state & (1 << BH_Delay)) &&
1582              !(mpd->b_state & (1 << BH_Unwritten))))
1583                 goto submit_io;
1584
1585         handle = ext4_journal_current_handle();
1586         BUG_ON(!handle);
1587
1588         /*
1589          * Call ext4_map_blocks() to allocate any delayed allocation
1590          * blocks, or to convert an uninitialized extent to be
1591          * initialized (in the case where we have written into
1592          * one or more preallocated blocks).
1593          *
1594          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
1595          * indicate that we are on the delayed allocation path.  This
1596          * affects functions in many different parts of the allocation
1597          * call path.  This flag exists primarily because we don't
1598          * want to change *many* call functions, so ext4_map_blocks()
1599          * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
1600          * inode's allocation semaphore is taken.
1601          *
1602          * If the blocks in questions were delalloc blocks, set
1603          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
1604          * variables are updated after the blocks have been allocated.
1605          */
1606         map.m_lblk = next;
1607         map.m_len = max_blocks;
1608         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
1609         if (ext4_should_dioread_nolock(mpd->inode))
1610                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
1611         if (mpd->b_state & (1 << BH_Delay))
1612                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
1613
1614         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
1615         if (blks < 0) {
1616                 struct super_block *sb = mpd->inode->i_sb;
1617
1618                 err = blks;
1619                 /*
1620                  * If get block returns EAGAIN or ENOSPC and there
1621                  * appears to be free blocks we will just let
1622                  * mpage_da_submit_io() unlock all of the pages.
1623                  */
1624                 if (err == -EAGAIN)
1625                         goto submit_io;
1626
1627                 if (err == -ENOSPC && ext4_count_free_clusters(sb)) {
1628                         mpd->retval = err;
1629                         goto submit_io;
1630                 }
1631
1632                 /*
1633                  * get block failure will cause us to loop in
1634                  * writepages, because a_ops->writepage won't be able
1635                  * to make progress. The page will be redirtied by
1636                  * writepage and writepages will again try to write
1637                  * the same.
1638                  */
1639                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
1640                         ext4_msg(sb, KERN_CRIT,
1641                                  "delayed block allocation failed for inode %lu "
1642                                  "at logical offset %llu with max blocks %zd "
1643                                  "with error %d", mpd->inode->i_ino,
1644                                  (unsigned long long) next,
1645                                  mpd->b_size >> mpd->inode->i_blkbits, err);
1646                         ext4_msg(sb, KERN_CRIT,
1647                                 "This should not happen!! Data will be lost\n");
1648                         if (err == -ENOSPC)
1649                                 ext4_print_free_blocks(mpd->inode);
1650                 }
1651                 /* invalidate all the pages */
1652                 ext4_da_block_invalidatepages(mpd);
1653
1654                 /* Mark this page range as having been completed */
1655                 mpd->io_done = 1;
1656                 return;
1657         }
1658         BUG_ON(blks == 0);
1659
1660         mapp = &map;
1661         if (map.m_flags & EXT4_MAP_NEW) {
1662                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
1663                 int i;
1664
1665                 for (i = 0; i < map.m_len; i++)
1666                         unmap_underlying_metadata(bdev, map.m_pblk + i);
1667
1668                 if (ext4_should_order_data(mpd->inode)) {
1669                         err = ext4_jbd2_file_inode(handle, mpd->inode);
1670                         if (err) {
1671                                 /* Only if the journal is aborted */
1672                                 mpd->retval = err;
1673                                 goto submit_io;
1674                         }
1675                 }
1676         }
1677
1678         /*
1679          * Update on-disk size along with block allocation.
1680          */
1681         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
1682         if (disksize > i_size_read(mpd->inode))
1683                 disksize = i_size_read(mpd->inode);
1684         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
1685                 ext4_update_i_disksize(mpd->inode, disksize);
1686                 err = ext4_mark_inode_dirty(handle, mpd->inode);
1687                 if (err)
1688                         ext4_error(mpd->inode->i_sb,
1689                                    "Failed to mark inode %lu dirty",
1690                                    mpd->inode->i_ino);
1691         }
1692
1693 submit_io:
1694         mpage_da_submit_io(mpd, mapp);
1695         mpd->io_done = 1;
1696 }
1697
1698 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1699                 (1 << BH_Delay) | (1 << BH_Unwritten))
1700
1701 /*
1702  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1703  *
1704  * @mpd->lbh - extent of blocks
1705  * @logical - logical number of the block in the file
1706  * @bh - bh of the block (used to access block's state)
1707  *
1708  * the function is used to collect contig. blocks in same state
1709  */
1710 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1711                                    sector_t logical, size_t b_size,
1712                                    unsigned long b_state)
1713 {
1714         sector_t next;
1715         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
1716
1717         /*
1718          * XXX Don't go larger than mballoc is willing to allocate
1719          * This is a stopgap solution.  We eventually need to fold
1720          * mpage_da_submit_io() into this function and then call
1721          * ext4_map_blocks() multiple times in a loop
1722          */
1723         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
1724                 goto flush_it;
1725
1726         /* check if thereserved journal credits might overflow */
1727         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
1728                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1729                         /*
1730                          * With non-extent format we are limited by the journal
1731                          * credit available.  Total credit needed to insert
1732                          * nrblocks contiguous blocks is dependent on the
1733                          * nrblocks.  So limit nrblocks.
1734                          */
1735                         goto flush_it;
1736                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1737                                 EXT4_MAX_TRANS_DATA) {
1738                         /*
1739                          * Adding the new buffer_head would make it cross the
1740                          * allowed limit for which we have journal credit
1741                          * reserved. So limit the new bh->b_size
1742                          */
1743                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1744                                                 mpd->inode->i_blkbits;
1745                         /* we will do mpage_da_submit_io in the next loop */
1746                 }
1747         }
1748         /*
1749          * First block in the extent
1750          */
1751         if (mpd->b_size == 0) {
1752                 mpd->b_blocknr = logical;
1753                 mpd->b_size = b_size;
1754                 mpd->b_state = b_state & BH_FLAGS;
1755                 return;
1756         }
1757
1758         next = mpd->b_blocknr + nrblocks;
1759         /*
1760          * Can we merge the block to our big extent?
1761          */
1762         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
1763                 mpd->b_size += b_size;
1764                 return;
1765         }
1766
1767 flush_it:
1768         /*
1769          * We couldn't merge the block to our extent, so we
1770          * need to flush current  extent and start new one
1771          */
1772         mpage_da_map_and_submit(mpd);
1773         return;
1774 }
1775
1776 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1777 {
1778         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1779 }
1780
1781 /*
1782  * This function is grabs code from the very beginning of
1783  * ext4_map_blocks, but assumes that the caller is from delayed write
1784  * time. This function looks up the requested blocks and sets the
1785  * buffer delay bit under the protection of i_data_sem.
1786  */
1787 static int ext4_da_map_blocks(struct inode *inode, sector_t iblock,
1788                               struct ext4_map_blocks *map,
1789                               struct buffer_head *bh)
1790 {
1791         int retval;
1792         sector_t invalid_block = ~((sector_t) 0xffff);
1793
1794         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1795                 invalid_block = ~0;
1796
1797         map->m_flags = 0;
1798         ext_debug("ext4_da_map_blocks(): inode %lu, max_blocks %u,"
1799                   "logical block %lu\n", inode->i_ino, map->m_len,
1800                   (unsigned long) map->m_lblk);
1801         /*
1802          * Try to see if we can get the block without requesting a new
1803          * file system block.
1804          */
1805         down_read((&EXT4_I(inode)->i_data_sem));
1806         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
1807                 retval = ext4_ext_map_blocks(NULL, inode, map, 0);
1808         else
1809                 retval = ext4_ind_map_blocks(NULL, inode, map, 0);
1810
1811         if (retval == 0) {
1812                 /*
1813                  * XXX: __block_prepare_write() unmaps passed block,
1814                  * is it OK?
1815                  */
1816                 /* If the block was allocated from previously allocated cluster,
1817                  * then we dont need to reserve it again. */
1818                 if (!(map->m_flags & EXT4_MAP_FROM_CLUSTER)) {
1819                         retval = ext4_da_reserve_space(inode, iblock);
1820                         if (retval)
1821                                 /* not enough space to reserve */
1822                                 goto out_unlock;
1823                 }
1824
1825                 /* Clear EXT4_MAP_FROM_CLUSTER flag since its purpose is served
1826                  * and it should not appear on the bh->b_state.
1827                  */
1828                 map->m_flags &= ~EXT4_MAP_FROM_CLUSTER;
1829
1830                 map_bh(bh, inode->i_sb, invalid_block);
1831                 set_buffer_new(bh);
1832                 set_buffer_delay(bh);
1833         }
1834
1835 out_unlock:
1836         up_read((&EXT4_I(inode)->i_data_sem));
1837
1838         return retval;
1839 }
1840
1841 /*
1842  * This is a special get_blocks_t callback which is used by
1843  * ext4_da_write_begin().  It will either return mapped block or
1844  * reserve space for a single block.
1845  *
1846  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1847  * We also have b_blocknr = -1 and b_bdev initialized properly
1848  *
1849  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1850  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1851  * initialized properly.
1852  */
1853 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1854                                   struct buffer_head *bh, int create)
1855 {
1856         struct ext4_map_blocks map;
1857         int ret = 0;
1858
1859         BUG_ON(create == 0);
1860         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1861
1862         map.m_lblk = iblock;
1863         map.m_len = 1;
1864
1865         /*
1866          * first, we need to know whether the block is allocated already
1867          * preallocated blocks are unmapped but should treated
1868          * the same as allocated blocks.
1869          */
1870         ret = ext4_da_map_blocks(inode, iblock, &map, bh);
1871         if (ret <= 0)
1872                 return ret;
1873
1874         map_bh(bh, inode->i_sb, map.m_pblk);
1875         ext4_update_bh_state(bh, map.m_flags);
1876
1877         if (buffer_unwritten(bh)) {
1878                 /* A delayed write to unwritten bh should be marked
1879                  * new and mapped.  Mapped ensures that we don't do
1880                  * get_block multiple times when we write to the same
1881                  * offset and new ensures that we do proper zero out
1882                  * for partial write.
1883                  */
1884                 set_buffer_new(bh);
1885                 set_buffer_mapped(bh);
1886         }
1887         return 0;
1888 }
1889
1890 /*
1891  * This function is used as a standard get_block_t calback function
1892  * when there is no desire to allocate any blocks.  It is used as a
1893  * callback function for block_write_begin() and block_write_full_page().
1894  * These functions should only try to map a single block at a time.
1895  *
1896  * Since this function doesn't do block allocations even if the caller
1897  * requests it by passing in create=1, it is critically important that
1898  * any caller checks to make sure that any buffer heads are returned
1899  * by this function are either all already mapped or marked for
1900  * delayed allocation before calling  block_write_full_page().  Otherwise,
1901  * b_blocknr could be left unitialized, and the page write functions will
1902  * be taken by surprise.
1903  */
1904 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
1905                                    struct buffer_head *bh_result, int create)
1906 {
1907         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
1908         return _ext4_get_block(inode, iblock, bh_result, 0);
1909 }
1910
1911 static int bget_one(handle_t *handle, struct buffer_head *bh)
1912 {
1913         get_bh(bh);
1914         return 0;
1915 }
1916
1917 static int bput_one(handle_t *handle, struct buffer_head *bh)
1918 {
1919         put_bh(bh);
1920         return 0;
1921 }
1922
1923 static int __ext4_journalled_writepage(struct page *page,
1924                                        unsigned int len)
1925 {
1926         struct address_space *mapping = page->mapping;
1927         struct inode *inode = mapping->host;
1928         struct buffer_head *page_bufs;
1929         handle_t *handle = NULL;
1930         int ret = 0;
1931         int err;
1932
1933         ClearPageChecked(page);
1934         page_bufs = page_buffers(page);
1935         BUG_ON(!page_bufs);
1936         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
1937         /*
1938          * We need to release the page lock before we start the
1939          * journal, so grab a reference so the page won't disappear
1940          * out from under us.
1941          */
1942         get_page(page);
1943         unlock_page(page);
1944
1945         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1946         if (IS_ERR(handle)) {
1947                 ret = PTR_ERR(handle);
1948                 put_page(page);
1949                 goto out_no_pagelock;
1950         }
1951         BUG_ON(!ext4_handle_valid(handle));
1952
1953         lock_page(page);
1954         put_page(page);
1955         if (page->mapping != mapping) {
1956                 /* The page got truncated from under us */
1957                 ext4_journal_stop(handle);
1958                 ret = 0;
1959                 goto out;
1960         }
1961
1962         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1963                                 do_journal_get_write_access);
1964
1965         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1966                                 write_end_fn);
1967         if (ret == 0)
1968                 ret = err;
1969         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1970         err = ext4_journal_stop(handle);
1971         if (!ret)
1972                 ret = err;
1973
1974         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
1975         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1976 out:
1977         unlock_page(page);
1978 out_no_pagelock:
1979         return ret;
1980 }
1981
1982 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
1983 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
1984
1985 /*
1986  * Note that we don't need to start a transaction unless we're journaling data
1987  * because we should have holes filled from ext4_page_mkwrite(). We even don't
1988  * need to file the inode to the transaction's list in ordered mode because if
1989  * we are writing back data added by write(), the inode is already there and if
1990  * we are writing back data modified via mmap(), no one guarantees in which
1991  * transaction the data will hit the disk. In case we are journaling data, we
1992  * cannot start transaction directly because transaction start ranks above page
1993  * lock so we have to do some magic.
1994  *
1995  * This function can get called via...
1996  *   - ext4_da_writepages after taking page lock (have journal handle)
1997  *   - journal_submit_inode_data_buffers (no journal handle)
1998  *   - shrink_page_list via pdflush (no journal handle)
1999  *   - grab_page_cache when doing write_begin (have journal handle)
2000  *
2001  * We don't do any block allocation in this function. If we have page with
2002  * multiple blocks we need to write those buffer_heads that are mapped. This
2003  * is important for mmaped based write. So if we do with blocksize 1K
2004  * truncate(f, 1024);
2005  * a = mmap(f, 0, 4096);
2006  * a[0] = 'a';
2007  * truncate(f, 4096);
2008  * we have in the page first buffer_head mapped via page_mkwrite call back
2009  * but other bufer_heads would be unmapped but dirty(dirty done via the
2010  * do_wp_page). So writepage should write the first block. If we modify
2011  * the mmap area beyond 1024 we will again get a page_fault and the
2012  * page_mkwrite callback will do the block allocation and mark the
2013  * buffer_heads mapped.
2014  *
2015  * We redirty the page if we have any buffer_heads that is either delay or
2016  * unwritten in the page.
2017  *
2018  * We can get recursively called as show below.
2019  *
2020  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
2021  *              ext4_writepage()
2022  *
2023  * But since we don't do any block allocation we should not deadlock.
2024  * Page also have the dirty flag cleared so we don't get recurive page_lock.
2025  */
2026 static int ext4_writepage(struct page *page,
2027                           struct writeback_control *wbc)
2028 {
2029         int ret = 0, commit_write = 0;
2030         loff_t size;
2031         unsigned int len;
2032         struct buffer_head *page_bufs = NULL;
2033         struct inode *inode = page->mapping->host;
2034
2035         trace_ext4_writepage(page);
2036         size = i_size_read(inode);
2037         if (page->index == size >> PAGE_CACHE_SHIFT)
2038                 len = size & ~PAGE_CACHE_MASK;
2039         else
2040                 len = PAGE_CACHE_SIZE;
2041
2042         /*
2043          * If the page does not have buffers (for whatever reason),
2044          * try to create them using __block_write_begin.  If this
2045          * fails, redirty the page and move on.
2046          */
2047         if (!page_has_buffers(page)) {
2048                 if (__block_write_begin(page, 0, len,
2049                                         noalloc_get_block_write)) {
2050                 redirty_page:
2051                         redirty_page_for_writepage(wbc, page);
2052                         unlock_page(page);
2053                         return 0;
2054                 }
2055                 commit_write = 1;
2056         }
2057         page_bufs = page_buffers(page);
2058         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
2059                               ext4_bh_delay_or_unwritten)) {
2060                 /*
2061                  * We don't want to do block allocation, so redirty
2062                  * the page and return.  We may reach here when we do
2063                  * a journal commit via journal_submit_inode_data_buffers.
2064                  * We can also reach here via shrink_page_list but it
2065                  * should never be for direct reclaim so warn if that
2066                  * happens
2067                  */
2068                 WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
2069                                                                 PF_MEMALLOC);
2070                 goto redirty_page;
2071         }
2072         if (commit_write)
2073                 /* now mark the buffer_heads as dirty and uptodate */
2074                 block_commit_write(page, 0, len);
2075
2076         if (PageChecked(page) && ext4_should_journal_data(inode))
2077                 /*
2078                  * It's mmapped pagecache.  Add buffers and journal it.  There
2079                  * doesn't seem much point in redirtying the page here.
2080                  */
2081                 return __ext4_journalled_writepage(page, len);
2082
2083         if (buffer_uninit(page_bufs)) {
2084                 ext4_set_bh_endio(page_bufs, inode);
2085                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
2086                                             wbc, ext4_end_io_buffer_write);
2087         } else
2088                 ret = block_write_full_page(page, noalloc_get_block_write,
2089                                             wbc);
2090
2091         return ret;
2092 }
2093
2094 /*
2095  * This is called via ext4_da_writepages() to
2096  * calculate the total number of credits to reserve to fit
2097  * a single extent allocation into a single transaction,
2098  * ext4_da_writpeages() will loop calling this before
2099  * the block allocation.
2100  */
2101
2102 static int ext4_da_writepages_trans_blocks(struct inode *inode)
2103 {
2104         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
2105
2106         /*
2107          * With non-extent format the journal credit needed to
2108          * insert nrblocks contiguous block is dependent on
2109          * number of contiguous block. So we will limit
2110          * number of contiguous block to a sane value
2111          */
2112         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
2113             (max_blocks > EXT4_MAX_TRANS_DATA))
2114                 max_blocks = EXT4_MAX_TRANS_DATA;
2115
2116         return ext4_chunk_trans_blocks(inode, max_blocks);
2117 }
2118
2119 /*
2120  * write_cache_pages_da - walk the list of dirty pages of the given
2121  * address space and accumulate pages that need writing, and call
2122  * mpage_da_map_and_submit to map a single contiguous memory region
2123  * and then write them.
2124  */
2125 static int write_cache_pages_da(struct address_space *mapping,
2126                                 struct writeback_control *wbc,
2127                                 struct mpage_da_data *mpd,
2128                                 pgoff_t *done_index)
2129 {
2130         struct buffer_head      *bh, *head;
2131         struct inode            *inode = mapping->host;
2132         struct pagevec          pvec;
2133         unsigned int            nr_pages;
2134         sector_t                logical;
2135         pgoff_t                 index, end;
2136         long                    nr_to_write = wbc->nr_to_write;
2137         int                     i, tag, ret = 0;
2138
2139         memset(mpd, 0, sizeof(struct mpage_da_data));
2140         mpd->wbc = wbc;
2141         mpd->inode = inode;
2142         pagevec_init(&pvec, 0);
2143         index = wbc->range_start >> PAGE_CACHE_SHIFT;
2144         end = wbc->range_end >> PAGE_CACHE_SHIFT;
2145
2146         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2147                 tag = PAGECACHE_TAG_TOWRITE;
2148         else
2149                 tag = PAGECACHE_TAG_DIRTY;
2150
2151         *done_index = index;
2152         while (index <= end) {
2153                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
2154                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
2155                 if (nr_pages == 0)
2156                         return 0;
2157
2158                 for (i = 0; i < nr_pages; i++) {
2159                         struct page *page = pvec.pages[i];
2160
2161                         /*
2162                          * At this point, the page may be truncated or
2163                          * invalidated (changing page->mapping to NULL), or
2164                          * even swizzled back from swapper_space to tmpfs file
2165                          * mapping. However, page->index will not change
2166                          * because we have a reference on the page.
2167                          */
2168                         if (page->index > end)
2169                                 goto out;
2170
2171                         *done_index = page->index + 1;
2172
2173                         /*
2174                          * If we can't merge this page, and we have
2175                          * accumulated an contiguous region, write it
2176                          */
2177                         if ((mpd->next_page != page->index) &&
2178                             (mpd->next_page != mpd->first_page)) {
2179                                 mpage_da_map_and_submit(mpd);
2180                                 goto ret_extent_tail;
2181                         }
2182
2183                         lock_page(page);
2184
2185                         /*
2186                          * If the page is no longer dirty, or its
2187                          * mapping no longer corresponds to inode we
2188                          * are writing (which means it has been
2189                          * truncated or invalidated), or the page is
2190                          * already under writeback and we are not
2191                          * doing a data integrity writeback, skip the page
2192                          */
2193                         if (!PageDirty(page) ||
2194                             (PageWriteback(page) &&
2195                              (wbc->sync_mode == WB_SYNC_NONE)) ||
2196                             unlikely(page->mapping != mapping)) {
2197                                 unlock_page(page);
2198                                 continue;
2199                         }
2200
2201                         wait_on_page_writeback(page);
2202                         BUG_ON(PageWriteback(page));
2203
2204                         if (mpd->next_page != page->index)
2205                                 mpd->first_page = page->index;
2206                         mpd->next_page = page->index + 1;
2207                         logical = (sector_t) page->index <<
2208                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
2209
2210                         if (!page_has_buffers(page)) {
2211                                 mpage_add_bh_to_extent(mpd, logical,
2212                                                        PAGE_CACHE_SIZE,
2213                                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
2214                                 if (mpd->io_done)
2215                                         goto ret_extent_tail;
2216                         } else {
2217                                 /*
2218                                  * Page with regular buffer heads,
2219                                  * just add all dirty ones
2220                                  */
2221                                 head = page_buffers(page);
2222                                 bh = head;
2223                                 do {
2224                                         BUG_ON(buffer_locked(bh));
2225                                         /*
2226                                          * We need to try to allocate
2227                                          * unmapped blocks in the same page.
2228                                          * Otherwise we won't make progress
2229                                          * with the page in ext4_writepage
2230                                          */
2231                                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
2232                                                 mpage_add_bh_to_extent(mpd, logical,
2233                                                                        bh->b_size,
2234                                                                        bh->b_state);
2235                                                 if (mpd->io_done)
2236                                                         goto ret_extent_tail;
2237                                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
2238                                                 /*
2239                                                  * mapped dirty buffer. We need
2240                                                  * to update the b_state
2241                                                  * because we look at b_state
2242                                                  * in mpage_da_map_blocks.  We
2243                                                  * don't update b_size because
2244                                                  * if we find an unmapped
2245                                                  * buffer_head later we need to
2246                                                  * use the b_state flag of that
2247                                                  * buffer_head.
2248                                                  */
2249                                                 if (mpd->b_size == 0)
2250                                                         mpd->b_state = bh->b_state & BH_FLAGS;
2251                                         }
2252                                         logical++;
2253                                 } while ((bh = bh->b_this_page) != head);
2254                         }
2255
2256                         if (nr_to_write > 0) {
2257                                 nr_to_write--;
2258                                 if (nr_to_write == 0 &&
2259                                     wbc->sync_mode == WB_SYNC_NONE)
2260                                         /*
2261                                          * We stop writing back only if we are
2262                                          * not doing integrity sync. In case of
2263                                          * integrity sync we have to keep going
2264                                          * because someone may be concurrently
2265                                          * dirtying pages, and we might have
2266                                          * synced a lot of newly appeared dirty
2267                                          * pages, but have not synced all of the
2268                                          * old dirty pages.
2269                                          */
2270                                         goto out;
2271                         }
2272                 }
2273                 pagevec_release(&pvec);
2274                 cond_resched();
2275         }
2276         return 0;
2277 ret_extent_tail:
2278         ret = MPAGE_DA_EXTENT_TAIL;
2279 out:
2280         pagevec_release(&pvec);
2281         cond_resched();
2282         return ret;
2283 }
2284
2285
2286 static int ext4_da_writepages(struct address_space *mapping,
2287                               struct writeback_control *wbc)
2288 {
2289         pgoff_t index;
2290         int range_whole = 0;
2291         handle_t *handle = NULL;
2292         struct mpage_da_data mpd;
2293         struct inode *inode = mapping->host;
2294         int pages_written = 0;
2295         unsigned int max_pages;
2296         int range_cyclic, cycled = 1, io_done = 0;
2297         int needed_blocks, ret = 0;
2298         long desired_nr_to_write, nr_to_writebump = 0;
2299         loff_t range_start = wbc->range_start;
2300         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2301         pgoff_t done_index = 0;
2302         pgoff_t end;
2303         struct blk_plug plug;
2304
2305         trace_ext4_da_writepages(inode, wbc);
2306
2307         /*
2308          * No pages to write? This is mainly a kludge to avoid starting
2309          * a transaction for special inodes like journal inode on last iput()
2310          * because that could violate lock ordering on umount
2311          */
2312         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2313                 return 0;
2314
2315         /*
2316          * If the filesystem has aborted, it is read-only, so return
2317          * right away instead of dumping stack traces later on that
2318          * will obscure the real source of the problem.  We test
2319          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2320          * the latter could be true if the filesystem is mounted
2321          * read-only, and in that case, ext4_da_writepages should
2322          * *never* be called, so if that ever happens, we would want
2323          * the stack trace.
2324          */
2325         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2326                 return -EROFS;
2327
2328         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2329                 range_whole = 1;
2330
2331         range_cyclic = wbc->range_cyclic;
2332         if (wbc->range_cyclic) {
2333                 index = mapping->writeback_index;
2334                 if (index)
2335                         cycled = 0;
2336                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2337                 wbc->range_end  = LLONG_MAX;
2338                 wbc->range_cyclic = 0;
2339                 end = -1;
2340         } else {
2341                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2342                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2343         }
2344
2345         /*
2346          * This works around two forms of stupidity.  The first is in
2347          * the writeback code, which caps the maximum number of pages
2348          * written to be 1024 pages.  This is wrong on multiple
2349          * levels; different architectues have a different page size,
2350          * which changes the maximum amount of data which gets
2351          * written.  Secondly, 4 megabytes is way too small.  XFS
2352          * forces this value to be 16 megabytes by multiplying
2353          * nr_to_write parameter by four, and then relies on its
2354          * allocator to allocate larger extents to make them
2355          * contiguous.  Unfortunately this brings us to the second
2356          * stupidity, which is that ext4's mballoc code only allocates
2357          * at most 2048 blocks.  So we force contiguous writes up to
2358          * the number of dirty blocks in the inode, or
2359          * sbi->max_writeback_mb_bump whichever is smaller.
2360          */
2361         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2362         if (!range_cyclic && range_whole) {
2363                 if (wbc->nr_to_write == LONG_MAX)
2364                         desired_nr_to_write = wbc->nr_to_write;
2365                 else
2366                         desired_nr_to_write = wbc->nr_to_write * 8;
2367         } else
2368                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2369                                                            max_pages);
2370         if (desired_nr_to_write > max_pages)
2371                 desired_nr_to_write = max_pages;
2372
2373         if (wbc->nr_to_write < desired_nr_to_write) {
2374                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2375                 wbc->nr_to_write = desired_nr_to_write;
2376         }
2377
2378 retry:
2379         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2380                 tag_pages_for_writeback(mapping, index, end);
2381
2382         blk_start_plug(&plug);
2383         while (!ret && wbc->nr_to_write > 0) {
2384
2385                 /*
2386                  * we  insert one extent at a time. So we need
2387                  * credit needed for single extent allocation.
2388                  * journalled mode is currently not supported
2389                  * by delalloc
2390                  */
2391                 BUG_ON(ext4_should_journal_data(inode));
2392                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2393
2394                 /* start a new transaction*/
2395                 handle = ext4_journal_start(inode, needed_blocks);
2396                 if (IS_ERR(handle)) {
2397                         ret = PTR_ERR(handle);
2398                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2399                                "%ld pages, ino %lu; err %d", __func__,
2400                                 wbc->nr_to_write, inode->i_ino, ret);
2401                         blk_finish_plug(&plug);
2402                         goto out_writepages;
2403                 }
2404
2405                 /*
2406                  * Now call write_cache_pages_da() to find the next
2407                  * contiguous region of logical blocks that need
2408                  * blocks to be allocated by ext4 and submit them.
2409                  */
2410                 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
2411                 /*
2412                  * If we have a contiguous extent of pages and we
2413                  * haven't done the I/O yet, map the blocks and submit
2414                  * them for I/O.
2415                  */
2416                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2417                         mpage_da_map_and_submit(&mpd);
2418                         ret = MPAGE_DA_EXTENT_TAIL;
2419                 }
2420                 trace_ext4_da_write_pages(inode, &mpd);
2421                 wbc->nr_to_write -= mpd.pages_written;
2422
2423                 ext4_journal_stop(handle);
2424
2425                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2426                         /* commit the transaction which would
2427                          * free blocks released in the transaction
2428                          * and try again
2429                          */
2430                         jbd2_journal_force_commit_nested(sbi->s_journal);
2431                         ret = 0;
2432                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2433                         /*
2434                          * Got one extent now try with rest of the pages.
2435                          * If mpd.retval is set -EIO, journal is aborted.
2436                          * So we don't need to write any more.
2437                          */
2438                         pages_written += mpd.pages_written;
2439                         ret = mpd.retval;
2440                         io_done = 1;
2441                 } else if (wbc->nr_to_write)
2442                         /*
2443                          * There is no more writeout needed
2444                          * or we requested for a noblocking writeout
2445                          * and we found the device congested
2446                          */
2447                         break;
2448         }
2449         blk_finish_plug(&plug);
2450         if (!io_done && !cycled) {
2451                 cycled = 1;
2452                 index = 0;
2453                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2454                 wbc->range_end  = mapping->writeback_index - 1;
2455                 goto retry;
2456         }
2457
2458         /* Update index */
2459         wbc->range_cyclic = range_cyclic;
2460         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2461                 /*
2462                  * set the writeback_index so that range_cyclic
2463                  * mode will write it back later
2464                  */
2465                 mapping->writeback_index = done_index;
2466
2467 out_writepages:
2468         wbc->nr_to_write -= nr_to_writebump;
2469         wbc->range_start = range_start;
2470         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2471         return ret;
2472 }
2473
2474 #define FALL_BACK_TO_NONDELALLOC 1
2475 static int ext4_nonda_switch(struct super_block *sb)
2476 {
2477         s64 free_blocks, dirty_blocks;
2478         struct ext4_sb_info *sbi = EXT4_SB(sb);
2479
2480         /*
2481          * switch to non delalloc mode if we are running low
2482          * on free block. The free block accounting via percpu
2483          * counters can get slightly wrong with percpu_counter_batch getting
2484          * accumulated on each CPU without updating global counters
2485          * Delalloc need an accurate free block accounting. So switch
2486          * to non delalloc when we are near to error range.
2487          */
2488         free_blocks  = EXT4_C2B(sbi,
2489                 percpu_counter_read_positive(&sbi->s_freeclusters_counter));
2490         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyclusters_counter);
2491         /*
2492          * Start pushing delalloc when 1/2 of free blocks are dirty.
2493          */
2494         if (dirty_blocks && (free_blocks < 2 * dirty_blocks) &&
2495             !writeback_in_progress(sb->s_bdi) &&
2496             down_read_trylock(&sb->s_umount)) {
2497                 writeback_inodes_sb(sb, WB_REASON_FS_FREE_SPACE);
2498                 up_read(&sb->s_umount);
2499         }
2500
2501         if (2 * free_blocks < 3 * dirty_blocks ||
2502                 free_blocks < (dirty_blocks + EXT4_FREECLUSTERS_WATERMARK)) {
2503                 /*
2504                  * free block count is less than 150% of dirty blocks
2505                  * or free blocks is less than watermark
2506                  */
2507                 return 1;
2508         }
2509         return 0;
2510 }
2511
2512 /* We always reserve for an inode update; the superblock could be there too */
2513 static int ext4_da_write_credits(struct inode *inode, loff_t pos, unsigned len)
2514 {
2515         if (likely(EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
2516                                 EXT4_FEATURE_RO_COMPAT_LARGE_FILE)))
2517                 return 1;
2518
2519         if (pos + len <= 0x7fffffffULL)
2520                 return 1;
2521
2522         /* We might need to update the superblock to set LARGE_FILE */
2523         return 2;
2524 }
2525
2526 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2527                                loff_t pos, unsigned len, unsigned flags,
2528                                struct page **pagep, void **fsdata)
2529 {
2530         int ret, retries = 0;
2531         struct page *page;
2532         pgoff_t index;
2533         struct inode *inode = mapping->host;
2534         handle_t *handle;
2535
2536         index = pos >> PAGE_CACHE_SHIFT;
2537
2538         if (ext4_nonda_switch(inode->i_sb)) {
2539                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2540                 return ext4_write_begin(file, mapping, pos,
2541                                         len, flags, pagep, fsdata);
2542         }
2543         *fsdata = (void *)0;
2544         trace_ext4_da_write_begin(inode, pos, len, flags);
2545 retry:
2546         /*
2547          * With delayed allocation, we don't log the i_disksize update
2548          * if there is delayed block allocation. But we still need
2549          * to journalling the i_disksize update if writes to the end
2550          * of file which has an already mapped buffer.
2551          */
2552         handle = ext4_journal_start(inode,
2553                                 ext4_da_write_credits(inode, pos, len));
2554         if (IS_ERR(handle)) {
2555                 ret = PTR_ERR(handle);
2556                 goto out;
2557         }
2558         /* We cannot recurse into the filesystem as the transaction is already
2559          * started */
2560         flags |= AOP_FLAG_NOFS;
2561
2562         page = grab_cache_page_write_begin(mapping, index, flags);
2563         if (!page) {
2564                 ext4_journal_stop(handle);
2565                 ret = -ENOMEM;
2566                 goto out;
2567         }
2568         *pagep = page;
2569
2570         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2571         if (ret < 0) {
2572                 unlock_page(page);
2573                 ext4_journal_stop(handle);
2574                 page_cache_release(page);
2575                 /*
2576                  * block_write_begin may have instantiated a few blocks
2577                  * outside i_size.  Trim these off again. Don't need
2578                  * i_size_read because we hold i_mutex.
2579                  */
2580                 if (pos + len > inode->i_size)
2581                         ext4_truncate_failed_write(inode);
2582         }
2583
2584         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2585                 goto retry;
2586 out:
2587         return ret;
2588 }
2589
2590 /*
2591  * Check if we should update i_disksize
2592  * when write to the end of file but not require block allocation
2593  */
2594 static int ext4_da_should_update_i_disksize(struct page *page,
2595                                             unsigned long offset)
2596 {
2597         struct buffer_head *bh;
2598         struct inode *inode = page->mapping->host;
2599         unsigned int idx;
2600         int i;
2601
2602         bh = page_buffers(page);
2603         idx = offset >> inode->i_blkbits;
2604
2605         for (i = 0; i < idx; i++)
2606                 bh = bh->b_this_page;
2607
2608         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2609                 return 0;
2610         return 1;
2611 }
2612
2613 static int ext4_da_write_end(struct file *file,
2614                              struct address_space *mapping,
2615                              loff_t pos, unsigned len, unsigned copied,
2616                              struct page *page, void *fsdata)
2617 {
2618         struct inode *inode = mapping->host;
2619         int ret = 0, ret2;
2620         handle_t *handle = ext4_journal_current_handle();
2621         loff_t new_i_size;
2622         unsigned long start, end;
2623         int write_mode = (int)(unsigned long)fsdata;
2624
2625         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2626                 switch (ext4_inode_journal_mode(inode)) {
2627                 case EXT4_INODE_ORDERED_DATA_MODE:
2628                         return ext4_ordered_write_end(file, mapping, pos,
2629                                         len, copied, page, fsdata);
2630                 case EXT4_INODE_WRITEBACK_DATA_MODE:
2631                         return ext4_writeback_write_end(file, mapping, pos,
2632                                         len, copied, page, fsdata);
2633                 default:
2634                         BUG();
2635                 }
2636         }
2637
2638         trace_ext4_da_write_end(inode, pos, len, copied);
2639         start = pos & (PAGE_CACHE_SIZE - 1);
2640         end = start + copied - 1;
2641
2642         /*
2643          * generic_write_end() will run mark_inode_dirty() if i_size
2644          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2645          * into that.
2646          */
2647
2648         new_i_size = pos + copied;
2649         if (copied && new_i_size > EXT4_I(inode)->i_disksize) {
2650                 if (ext4_da_should_update_i_disksize(page, end)) {
2651                         down_write(&EXT4_I(inode)->i_data_sem);
2652                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2653                                 /*
2654                                  * Updating i_disksize when extending file
2655                                  * without needing block allocation
2656                                  */
2657                                 if (ext4_should_order_data(inode))
2658                                         ret = ext4_jbd2_file_inode(handle,
2659                                                                    inode);
2660
2661                                 EXT4_I(inode)->i_disksize = new_i_size;
2662                         }
2663                         up_write(&EXT4_I(inode)->i_data_sem);
2664                         /* We need to mark inode dirty even if
2665                          * new_i_size is less that inode->i_size
2666                          * bu greater than i_disksize.(hint delalloc)
2667                          */
2668                         ext4_mark_inode_dirty(handle, inode);
2669                 }
2670         }
2671         ret2 = generic_write_end(file, mapping, pos, len, copied,
2672                                                         page, fsdata);
2673         copied = ret2;
2674         if (ret2 < 0)
2675                 ret = ret2;
2676         ret2 = ext4_journal_stop(handle);
2677         if (!ret)
2678                 ret = ret2;
2679
2680         return ret ? ret : copied;
2681 }
2682
2683 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2684 {
2685         /*
2686          * Drop reserved blocks
2687          */
2688         BUG_ON(!PageLocked(page));
2689         if (!page_has_buffers(page))
2690                 goto out;
2691
2692         ext4_da_page_release_reservation(page, offset);
2693
2694 out:
2695         ext4_invalidatepage(page, offset);
2696
2697         return;
2698 }
2699
2700 /*
2701  * Force all delayed allocation blocks to be allocated for a given inode.
2702  */
2703 int ext4_alloc_da_blocks(struct inode *inode)
2704 {
2705         trace_ext4_alloc_da_blocks(inode);
2706
2707         if (!EXT4_I(inode)->i_reserved_data_blocks &&
2708             !EXT4_I(inode)->i_reserved_meta_blocks)
2709                 return 0;
2710
2711         /*
2712          * We do something simple for now.  The filemap_flush() will
2713          * also start triggering a write of the data blocks, which is
2714          * not strictly speaking necessary (and for users of
2715          * laptop_mode, not even desirable).  However, to do otherwise
2716          * would require replicating code paths in:
2717          *
2718          * ext4_da_writepages() ->
2719          *    write_cache_pages() ---> (via passed in callback function)
2720          *        __mpage_da_writepage() -->
2721          *           mpage_add_bh_to_extent()
2722          *           mpage_da_map_blocks()
2723          *
2724          * The problem is that write_cache_pages(), located in
2725          * mm/page-writeback.c, marks pages clean in preparation for
2726          * doing I/O, which is not desirable if we're not planning on
2727          * doing I/O at all.
2728          *
2729          * We could call write_cache_pages(), and then redirty all of
2730          * the pages by calling redirty_page_for_writepage() but that
2731          * would be ugly in the extreme.  So instead we would need to
2732          * replicate parts of the code in the above functions,
2733          * simplifying them because we wouldn't actually intend to
2734          * write out the pages, but rather only collect contiguous
2735          * logical block extents, call the multi-block allocator, and
2736          * then update the buffer heads with the block allocations.
2737          *
2738          * For now, though, we'll cheat by calling filemap_flush(),
2739          * which will map the blocks, and start the I/O, but not
2740          * actually wait for the I/O to complete.
2741          */
2742         return filemap_flush(inode->i_mapping);
2743 }
2744
2745 /*
2746  * bmap() is special.  It gets used by applications such as lilo and by
2747  * the swapper to find the on-disk block of a specific piece of data.
2748  *
2749  * Naturally, this is dangerous if the block concerned is still in the
2750  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2751  * filesystem and enables swap, then they may get a nasty shock when the
2752  * data getting swapped to that swapfile suddenly gets overwritten by
2753  * the original zero's written out previously to the journal and
2754  * awaiting writeback in the kernel's buffer cache.
2755  *
2756  * So, if we see any bmap calls here on a modified, data-journaled file,
2757  * take extra steps to flush any blocks which might be in the cache.
2758  */
2759 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2760 {
2761         struct inode *inode = mapping->host;
2762         journal_t *journal;
2763         int err;
2764
2765         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2766                         test_opt(inode->i_sb, DELALLOC)) {
2767                 /*
2768                  * With delalloc we want to sync the file
2769                  * so that we can make sure we allocate
2770                  * blocks for file
2771                  */
2772                 filemap_write_and_wait(mapping);
2773         }
2774
2775         if (EXT4_JOURNAL(inode) &&
2776             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2777                 /*
2778                  * This is a REALLY heavyweight approach, but the use of
2779                  * bmap on dirty files is expected to be extremely rare:
2780                  * only if we run lilo or swapon on a freshly made file
2781                  * do we expect this to happen.
2782                  *
2783                  * (bmap requires CAP_SYS_RAWIO so this does not
2784                  * represent an unprivileged user DOS attack --- we'd be
2785                  * in trouble if mortal users could trigger this path at
2786                  * will.)
2787                  *
2788                  * NB. EXT4_STATE_JDATA is not set on files other than
2789                  * regular files.  If somebody wants to bmap a directory
2790                  * or symlink and gets confused because the buffer
2791                  * hasn't yet been flushed to disk, they deserve
2792                  * everything they get.
2793                  */
2794
2795                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2796                 journal = EXT4_JOURNAL(inode);
2797                 jbd2_journal_lock_updates(journal);
2798                 err = jbd2_journal_flush(journal);
2799                 jbd2_journal_unlock_updates(journal);
2800
2801                 if (err)
2802                         return 0;
2803         }
2804
2805         return generic_block_bmap(mapping, block, ext4_get_block);
2806 }
2807
2808 static int ext4_readpage(struct file *file, struct page *page)
2809 {
2810         trace_ext4_readpage(page);
2811         return mpage_readpage(page, ext4_get_block);
2812 }
2813
2814 static int
2815 ext4_readpages(struct file *file, struct address_space *mapping,
2816                 struct list_head *pages, unsigned nr_pages)
2817 {
2818         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2819 }
2820
2821 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
2822 {
2823         struct buffer_head *head, *bh;
2824         unsigned int curr_off = 0;
2825
2826         if (!page_has_buffers(page))
2827                 return;
2828         head = bh = page_buffers(page);
2829         do {
2830                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
2831                                         && bh->b_private) {
2832                         ext4_free_io_end(bh->b_private);
2833                         bh->b_private = NULL;
2834                         bh->b_end_io = NULL;
2835                 }
2836                 curr_off = curr_off + bh->b_size;
2837                 bh = bh->b_this_page;
2838         } while (bh != head);
2839 }
2840
2841 static void ext4_invalidatepage(struct page *page, unsigned long offset)
2842 {
2843         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2844
2845         trace_ext4_invalidatepage(page, offset);
2846
2847         /*
2848          * free any io_end structure allocated for buffers to be discarded
2849          */
2850         if (ext4_should_dioread_nolock(page->mapping->host))
2851                 ext4_invalidatepage_free_endio(page, offset);
2852         /*
2853          * If it's a full truncate we just forget about the pending dirtying
2854          */
2855         if (offset == 0)
2856                 ClearPageChecked(page);
2857
2858         if (journal)
2859                 jbd2_journal_invalidatepage(journal, page, offset);
2860         else
2861                 block_invalidatepage(page, offset);
2862 }
2863
2864 static int ext4_releasepage(struct page *page, gfp_t wait)
2865 {
2866         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2867
2868         trace_ext4_releasepage(page);
2869
2870         WARN_ON(PageChecked(page));
2871         if (!page_has_buffers(page))
2872                 return 0;
2873         if (journal)
2874                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
2875         else
2876                 return try_to_free_buffers(page);
2877 }
2878
2879 /*
2880  * ext4_get_block used when preparing for a DIO write or buffer write.
2881  * We allocate an uinitialized extent if blocks haven't been allocated.
2882  * The extent will be converted to initialized after the IO is complete.
2883  */
2884 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
2885                    struct buffer_head *bh_result, int create)
2886 {
2887         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
2888                    inode->i_ino, create);
2889         return _ext4_get_block(inode, iblock, bh_result,
2890                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
2891 }
2892
2893 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
2894                             ssize_t size, void *private, int ret,
2895                             bool is_async)
2896 {
2897         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
2898         ext4_io_end_t *io_end = iocb->private;
2899         struct workqueue_struct *wq;
2900         unsigned long flags;
2901         struct ext4_inode_info *ei;
2902
2903         /* if not async direct IO or dio with 0 bytes write, just return */
2904         if (!io_end || !size)
2905                 goto out;
2906
2907         ext_debug("ext4_end_io_dio(): io_end 0x%p"
2908                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
2909                   iocb->private, io_end->inode->i_ino, iocb, offset,
2910                   size);
2911
2912         iocb->private = NULL;
2913
2914         /* if not aio dio with unwritten extents, just free io and return */
2915         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2916                 ext4_free_io_end(io_end);
2917 out:
2918                 inode_dio_done(inode);
2919                 if (is_async)
2920                         aio_complete(iocb, ret, 0);
2921                 return;
2922         }
2923
2924         io_end->offset = offset;
2925         io_end->size = size;
2926         if (is_async) {
2927                 io_end->iocb = iocb;
2928                 io_end->result = ret;
2929         }
2930         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
2931
2932         /* Add the io_end to per-inode completed aio dio list*/
2933         ei = EXT4_I(io_end->inode);
2934         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
2935         list_add_tail(&io_end->list, &ei->i_completed_io_list);
2936         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
2937
2938         /* queue the work to convert unwritten extents to written */
2939         queue_work(wq, &io_end->work);
2940 }
2941
2942 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
2943 {
2944         ext4_io_end_t *io_end = bh->b_private;
2945         struct workqueue_struct *wq;
2946         struct inode *inode;
2947         unsigned long flags;
2948
2949         if (!test_clear_buffer_uninit(bh) || !io_end)
2950                 goto out;
2951
2952         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
2953                 printk("sb umounted, discard end_io request for inode %lu\n",
2954                         io_end->inode->i_ino);
2955                 ext4_free_io_end(io_end);
2956                 goto out;
2957         }
2958
2959         /*
2960          * It may be over-defensive here to check EXT4_IO_END_UNWRITTEN now,
2961          * but being more careful is always safe for the future change.
2962          */
2963         inode = io_end->inode;
2964         ext4_set_io_unwritten_flag(inode, io_end);
2965
2966         /* Add the io_end to per-inode completed io list*/
2967         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
2968         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
2969         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
2970
2971         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
2972         /* queue the work to convert unwritten extents to written */
2973         queue_work(wq, &io_end->work);
2974 out:
2975         bh->b_private = NULL;
2976         bh->b_end_io = NULL;
2977         clear_buffer_uninit(bh);
2978         end_buffer_async_write(bh, uptodate);
2979 }
2980
2981 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
2982 {
2983         ext4_io_end_t *io_end;
2984         struct page *page = bh->b_page;
2985         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
2986         size_t size = bh->b_size;
2987
2988 retry:
2989         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
2990         if (!io_end) {
2991                 pr_warn_ratelimited("%s: allocation fail\n", __func__);
2992                 schedule();
2993                 goto retry;
2994         }
2995         io_end->offset = offset;
2996         io_end->size = size;
2997         /*
2998          * We need to hold a reference to the page to make sure it
2999          * doesn't get evicted before ext4_end_io_work() has a chance
3000          * to convert the extent from written to unwritten.
3001          */
3002         io_end->page = page;
3003         get_page(io_end->page);
3004
3005         bh->b_private = io_end;
3006         bh->b_end_io = ext4_end_io_buffer_write;
3007         return 0;
3008 }
3009
3010 /*
3011  * For ext4 extent files, ext4 will do direct-io write to holes,
3012  * preallocated extents, and those write extend the file, no need to
3013  * fall back to buffered IO.
3014  *
3015  * For holes, we fallocate those blocks, mark them as uninitialized
3016  * If those blocks were preallocated, we mark sure they are splited, but
3017  * still keep the range to write as uninitialized.
3018  *
3019  * The unwrritten extents will be converted to written when DIO is completed.
3020  * For async direct IO, since the IO may still pending when return, we
3021  * set up an end_io call back function, which will do the conversion
3022  * when async direct IO completed.
3023  *
3024  * If the O_DIRECT write will extend the file then add this inode to the
3025  * orphan list.  So recovery will truncate it back to the original size
3026  * if the machine crashes during the write.
3027  *
3028  */
3029 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
3030                               const struct iovec *iov, loff_t offset,
3031                               unsigned long nr_segs)
3032 {
3033         struct file *file = iocb->ki_filp;
3034         struct inode *inode = file->f_mapping->host;
3035         ssize_t ret;
3036         size_t count = iov_length(iov, nr_segs);
3037
3038         loff_t final_size = offset + count;
3039         if (rw == WRITE && final_size <= inode->i_size) {
3040                 /*
3041                  * We could direct write to holes and fallocate.
3042                  *
3043                  * Allocated blocks to fill the hole are marked as uninitialized
3044                  * to prevent parallel buffered read to expose the stale data
3045                  * before DIO complete the data IO.
3046                  *
3047                  * As to previously fallocated extents, ext4 get_block
3048                  * will just simply mark the buffer mapped but still
3049                  * keep the extents uninitialized.
3050                  *
3051                  * for non AIO case, we will convert those unwritten extents
3052                  * to written after return back from blockdev_direct_IO.
3053                  *
3054                  * for async DIO, the conversion needs to be defered when
3055                  * the IO is completed. The ext4 end_io callback function
3056                  * will be called to take care of the conversion work.
3057                  * Here for async case, we allocate an io_end structure to