Merge branch 'for-linus' of git://git.kernel.dk/linux-block
[pandora-kernel.git] / fs / ext4 / inode.c
1 /*
2  *  linux/fs/ext4/inode.c
3  *
4  * Copyright (C) 1992, 1993, 1994, 1995
5  * Remy Card (card@masi.ibp.fr)
6  * Laboratoire MASI - Institut Blaise Pascal
7  * Universite Pierre et Marie Curie (Paris VI)
8  *
9  *  from
10  *
11  *  linux/fs/minix/inode.c
12  *
13  *  Copyright (C) 1991, 1992  Linus Torvalds
14  *
15  *  64-bit file support on 64-bit platforms by Jakub Jelinek
16  *      (jj@sunsite.ms.mff.cuni.cz)
17  *
18  *  Assorted race fixes, rewrite of ext4_get_block() by Al Viro, 2000
19  */
20
21 #include <linux/module.h>
22 #include <linux/fs.h>
23 #include <linux/time.h>
24 #include <linux/jbd2.h>
25 #include <linux/highuid.h>
26 #include <linux/pagemap.h>
27 #include <linux/quotaops.h>
28 #include <linux/string.h>
29 #include <linux/buffer_head.h>
30 #include <linux/writeback.h>
31 #include <linux/pagevec.h>
32 #include <linux/mpage.h>
33 #include <linux/namei.h>
34 #include <linux/uio.h>
35 #include <linux/bio.h>
36 #include <linux/workqueue.h>
37 #include <linux/kernel.h>
38 #include <linux/printk.h>
39 #include <linux/slab.h>
40 #include <linux/ratelimit.h>
41
42 #include "ext4_jbd2.h"
43 #include "xattr.h"
44 #include "acl.h"
45 #include "ext4_extents.h"
46 #include "truncate.h"
47
48 #include <trace/events/ext4.h>
49
50 #define MPAGE_DA_EXTENT_TAIL 0x01
51
52 static inline int ext4_begin_ordered_truncate(struct inode *inode,
53                                               loff_t new_size)
54 {
55         trace_ext4_begin_ordered_truncate(inode, new_size);
56         /*
57          * If jinode is zero, then we never opened the file for
58          * writing, so there's no need to call
59          * jbd2_journal_begin_ordered_truncate() since there's no
60          * outstanding writes we need to flush.
61          */
62         if (!EXT4_I(inode)->jinode)
63                 return 0;
64         return jbd2_journal_begin_ordered_truncate(EXT4_JOURNAL(inode),
65                                                    EXT4_I(inode)->jinode,
66                                                    new_size);
67 }
68
69 static void ext4_invalidatepage(struct page *page, unsigned long offset);
70 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
71                                    struct buffer_head *bh_result, int create);
72 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
73 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
74 static int __ext4_journalled_writepage(struct page *page, unsigned int len);
75 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
76
77 /*
78  * Test whether an inode is a fast symlink.
79  */
80 static int ext4_inode_is_fast_symlink(struct inode *inode)
81 {
82         int ea_blocks = EXT4_I(inode)->i_file_acl ?
83                 (inode->i_sb->s_blocksize >> 9) : 0;
84
85         return (S_ISLNK(inode->i_mode) && inode->i_blocks - ea_blocks == 0);
86 }
87
88 /*
89  * Restart the transaction associated with *handle.  This does a commit,
90  * so before we call here everything must be consistently dirtied against
91  * this transaction.
92  */
93 int ext4_truncate_restart_trans(handle_t *handle, struct inode *inode,
94                                  int nblocks)
95 {
96         int ret;
97
98         /*
99          * Drop i_data_sem to avoid deadlock with ext4_map_blocks.  At this
100          * moment, get_block can be called only for blocks inside i_size since
101          * page cache has been already dropped and writes are blocked by
102          * i_mutex. So we can safely drop the i_data_sem here.
103          */
104         BUG_ON(EXT4_JOURNAL(inode) == NULL);
105         jbd_debug(2, "restarting handle %p\n", handle);
106         up_write(&EXT4_I(inode)->i_data_sem);
107         ret = ext4_journal_restart(handle, nblocks);
108         down_write(&EXT4_I(inode)->i_data_sem);
109         ext4_discard_preallocations(inode);
110
111         return ret;
112 }
113
114 /*
115  * Called at the last iput() if i_nlink is zero.
116  */
117 void ext4_evict_inode(struct inode *inode)
118 {
119         handle_t *handle;
120         int err;
121
122         trace_ext4_evict_inode(inode);
123
124         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 (ext4_should_journal_data(inode) &&
146                     (S_ISLNK(inode->i_mode) || S_ISREG(inode->i_mode))) {
147                         journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
148                         tid_t commit_tid = EXT4_I(inode)->i_datasync_tid;
149
150                         jbd2_log_start_commit(journal, commit_tid);
151                         jbd2_log_wait_commit(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                 dquot_initialize(inode);
160
161         if (ext4_should_order_data(inode))
162                 ext4_begin_ordered_truncate(inode, 0);
163         truncate_inode_pages(&inode->i_data, 0);
164
165         if (is_bad_inode(inode))
166                 goto no_delete;
167
168         handle = ext4_journal_start(inode, ext4_blocks_for_truncate(inode)+3);
169         if (IS_ERR(handle)) {
170                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
171                 /*
172                  * If we're going to skip the normal cleanup, we still need to
173                  * make sure that the in-core orphan linked list is properly
174                  * cleaned up.
175                  */
176                 ext4_orphan_del(NULL, inode);
177                 goto no_delete;
178         }
179
180         if (IS_SYNC(inode))
181                 ext4_handle_sync(handle);
182         inode->i_size = 0;
183         err = ext4_mark_inode_dirty(handle, inode);
184         if (err) {
185                 ext4_warning(inode->i_sb,
186                              "couldn't mark inode dirty (err %d)", err);
187                 goto stop_handle;
188         }
189         if (inode->i_blocks)
190                 ext4_truncate(inode);
191
192         /*
193          * ext4_ext_truncate() doesn't reserve any slop when it
194          * restarts journal transactions; therefore there may not be
195          * enough credits left in the handle to remove the inode from
196          * the orphan list and set the dtime field.
197          */
198         if (!ext4_handle_has_enough_credits(handle, 3)) {
199                 err = ext4_journal_extend(handle, 3);
200                 if (err > 0)
201                         err = ext4_journal_restart(handle, 3);
202                 if (err != 0) {
203                         ext4_warning(inode->i_sb,
204                                      "couldn't extend journal (err %d)", err);
205                 stop_handle:
206                         ext4_journal_stop(handle);
207                         ext4_orphan_del(NULL, inode);
208                         goto no_delete;
209                 }
210         }
211
212         /*
213          * Kill off the orphan record which ext4_truncate created.
214          * AKPM: I think this can be inside the above `if'.
215          * Note that ext4_orphan_del() has to be able to cope with the
216          * deletion of a non-existent orphan - this is because we don't
217          * know if ext4_truncate() actually created an orphan record.
218          * (Well, we could do this if we need to, but heck - it works)
219          */
220         ext4_orphan_del(handle, inode);
221         EXT4_I(inode)->i_dtime  = get_seconds();
222
223         /*
224          * One subtle ordering requirement: if anything has gone wrong
225          * (transaction abort, IO errors, whatever), then we can still
226          * do these next steps (the fs will already have been marked as
227          * having errors), but we can't free the inode if the mark_dirty
228          * fails.
229          */
230         if (ext4_mark_inode_dirty(handle, inode))
231                 /* If that failed, just do the required in-core inode clear. */
232                 ext4_clear_inode(inode);
233         else
234                 ext4_free_inode(handle, inode);
235         ext4_journal_stop(handle);
236         return;
237 no_delete:
238         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
239 }
240
241 #ifdef CONFIG_QUOTA
242 qsize_t *ext4_get_reserved_space(struct inode *inode)
243 {
244         return &EXT4_I(inode)->i_reserved_quota;
245 }
246 #endif
247
248 /*
249  * Calculate the number of metadata blocks need to reserve
250  * to allocate a block located at @lblock
251  */
252 static int ext4_calc_metadata_amount(struct inode *inode, ext4_lblk_t lblock)
253 {
254         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
255                 return ext4_ext_calc_metadata_amount(inode, lblock);
256
257         return ext4_ind_calc_metadata_amount(inode, lblock);
258 }
259
260 /*
261  * Called with i_data_sem down, which is important since we can call
262  * ext4_discard_preallocations() from here.
263  */
264 void ext4_da_update_reserve_space(struct inode *inode,
265                                         int used, int quota_claim)
266 {
267         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
268         struct ext4_inode_info *ei = EXT4_I(inode);
269
270         spin_lock(&ei->i_block_reservation_lock);
271         trace_ext4_da_update_reserve_space(inode, used);
272         if (unlikely(used > ei->i_reserved_data_blocks)) {
273                 ext4_msg(inode->i_sb, KERN_NOTICE, "%s: ino %lu, used %d "
274                          "with only %d reserved data blocks\n",
275                          __func__, inode->i_ino, used,
276                          ei->i_reserved_data_blocks);
277                 WARN_ON(1);
278                 used = ei->i_reserved_data_blocks;
279         }
280
281         /* Update per-inode reservations */
282         ei->i_reserved_data_blocks -= used;
283         ei->i_reserved_meta_blocks -= ei->i_allocated_meta_blocks;
284         percpu_counter_sub(&sbi->s_dirtyblocks_counter,
285                            used + ei->i_allocated_meta_blocks);
286         ei->i_allocated_meta_blocks = 0;
287
288         if (ei->i_reserved_data_blocks == 0) {
289                 /*
290                  * We can release all of the reserved metadata blocks
291                  * only when we have written all of the delayed
292                  * allocation blocks.
293                  */
294                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
295                                    ei->i_reserved_meta_blocks);
296                 ei->i_reserved_meta_blocks = 0;
297                 ei->i_da_metadata_calc_len = 0;
298         }
299         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
300
301         /* Update quota subsystem for data blocks */
302         if (quota_claim)
303                 dquot_claim_block(inode, used);
304         else {
305                 /*
306                  * We did fallocate with an offset that is already delayed
307                  * allocated. So on delayed allocated writeback we should
308                  * not re-claim the quota for fallocated blocks.
309                  */
310                 dquot_release_reservation_block(inode, used);
311         }
312
313         /*
314          * If we have done all the pending block allocations and if
315          * there aren't any writers on the inode, we can discard the
316          * inode's preallocations.
317          */
318         if ((ei->i_reserved_data_blocks == 0) &&
319             (atomic_read(&inode->i_writecount) == 0))
320                 ext4_discard_preallocations(inode);
321 }
322
323 static int __check_block_validity(struct inode *inode, const char *func,
324                                 unsigned int line,
325                                 struct ext4_map_blocks *map)
326 {
327         if (!ext4_data_block_valid(EXT4_SB(inode->i_sb), map->m_pblk,
328                                    map->m_len)) {
329                 ext4_error_inode(inode, func, line, map->m_pblk,
330                                  "lblock %lu mapped to illegal pblock "
331                                  "(length %d)", (unsigned long) map->m_lblk,
332                                  map->m_len);
333                 return -EIO;
334         }
335         return 0;
336 }
337
338 #define check_block_validity(inode, map)        \
339         __check_block_validity((inode), __func__, __LINE__, (map))
340
341 /*
342  * Return the number of contiguous dirty pages in a given inode
343  * starting at page frame idx.
344  */
345 static pgoff_t ext4_num_dirty_pages(struct inode *inode, pgoff_t idx,
346                                     unsigned int max_pages)
347 {
348         struct address_space *mapping = inode->i_mapping;
349         pgoff_t index;
350         struct pagevec pvec;
351         pgoff_t num = 0;
352         int i, nr_pages, done = 0;
353
354         if (max_pages == 0)
355                 return 0;
356         pagevec_init(&pvec, 0);
357         while (!done) {
358                 index = idx;
359                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
360                                               PAGECACHE_TAG_DIRTY,
361                                               (pgoff_t)PAGEVEC_SIZE);
362                 if (nr_pages == 0)
363                         break;
364                 for (i = 0; i < nr_pages; i++) {
365                         struct page *page = pvec.pages[i];
366                         struct buffer_head *bh, *head;
367
368                         lock_page(page);
369                         if (unlikely(page->mapping != mapping) ||
370                             !PageDirty(page) ||
371                             PageWriteback(page) ||
372                             page->index != idx) {
373                                 done = 1;
374                                 unlock_page(page);
375                                 break;
376                         }
377                         if (page_has_buffers(page)) {
378                                 bh = head = page_buffers(page);
379                                 do {
380                                         if (!buffer_delay(bh) &&
381                                             !buffer_unwritten(bh))
382                                                 done = 1;
383                                         bh = bh->b_this_page;
384                                 } while (!done && (bh != head));
385                         }
386                         unlock_page(page);
387                         if (done)
388                                 break;
389                         idx++;
390                         num++;
391                         if (num >= max_pages) {
392                                 done = 1;
393                                 break;
394                         }
395                 }
396                 pagevec_release(&pvec);
397         }
398         return num;
399 }
400
401 /*
402  * The ext4_map_blocks() function tries to look up the requested blocks,
403  * and returns if the blocks are already mapped.
404  *
405  * Otherwise it takes the write lock of the i_data_sem and allocate blocks
406  * and store the allocated blocks in the result buffer head and mark it
407  * mapped.
408  *
409  * If file type is extents based, it will call ext4_ext_map_blocks(),
410  * Otherwise, call with ext4_ind_map_blocks() to handle indirect mapping
411  * based files
412  *
413  * On success, it returns the number of blocks being mapped or allocate.
414  * if create==0 and the blocks are pre-allocated and uninitialized block,
415  * the result buffer head is unmapped. If the create ==1, it will make sure
416  * the buffer head is mapped.
417  *
418  * It returns 0 if plain look up failed (blocks have not been allocated), in
419  * that casem, buffer head is unmapped
420  *
421  * It returns the error in case of allocation failure.
422  */
423 int ext4_map_blocks(handle_t *handle, struct inode *inode,
424                     struct ext4_map_blocks *map, int flags)
425 {
426         int retval;
427
428         map->m_flags = 0;
429         ext_debug("ext4_map_blocks(): inode %lu, flag %d, max_blocks %u,"
430                   "logical block %lu\n", inode->i_ino, flags, map->m_len,
431                   (unsigned long) map->m_lblk);
432         /*
433          * Try to see if we can get the block without requesting a new
434          * file system block.
435          */
436         down_read((&EXT4_I(inode)->i_data_sem));
437         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
438                 retval = ext4_ext_map_blocks(handle, inode, map, 0);
439         } else {
440                 retval = ext4_ind_map_blocks(handle, inode, map, 0);
441         }
442         up_read((&EXT4_I(inode)->i_data_sem));
443
444         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
445                 int ret = check_block_validity(inode, map);
446                 if (ret != 0)
447                         return ret;
448         }
449
450         /* If it is only a block(s) look up */
451         if ((flags & EXT4_GET_BLOCKS_CREATE) == 0)
452                 return retval;
453
454         /*
455          * Returns if the blocks have already allocated
456          *
457          * Note that if blocks have been preallocated
458          * ext4_ext_get_block() returns th create = 0
459          * with buffer head unmapped.
460          */
461         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED)
462                 return retval;
463
464         /*
465          * When we call get_blocks without the create flag, the
466          * BH_Unwritten flag could have gotten set if the blocks
467          * requested were part of a uninitialized extent.  We need to
468          * clear this flag now that we are committed to convert all or
469          * part of the uninitialized extent to be an initialized
470          * extent.  This is because we need to avoid the combination
471          * of BH_Unwritten and BH_Mapped flags being simultaneously
472          * set on the buffer_head.
473          */
474         map->m_flags &= ~EXT4_MAP_UNWRITTEN;
475
476         /*
477          * New blocks allocate and/or writing to uninitialized extent
478          * will possibly result in updating i_data, so we take
479          * the write lock of i_data_sem, and call get_blocks()
480          * with create == 1 flag.
481          */
482         down_write((&EXT4_I(inode)->i_data_sem));
483
484         /*
485          * if the caller is from delayed allocation writeout path
486          * we have already reserved fs blocks for allocation
487          * let the underlying get_block() function know to
488          * avoid double accounting
489          */
490         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
491                 ext4_set_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
492         /*
493          * We need to check for EXT4 here because migrate
494          * could have changed the inode type in between
495          */
496         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) {
497                 retval = ext4_ext_map_blocks(handle, inode, map, flags);
498         } else {
499                 retval = ext4_ind_map_blocks(handle, inode, map, flags);
500
501                 if (retval > 0 && map->m_flags & EXT4_MAP_NEW) {
502                         /*
503                          * We allocated new blocks which will result in
504                          * i_data's format changing.  Force the migrate
505                          * to fail by clearing migrate flags
506                          */
507                         ext4_clear_inode_state(inode, EXT4_STATE_EXT_MIGRATE);
508                 }
509
510                 /*
511                  * Update reserved blocks/metadata blocks after successful
512                  * block allocation which had been deferred till now. We don't
513                  * support fallocate for non extent files. So we can update
514                  * reserve space here.
515                  */
516                 if ((retval > 0) &&
517                         (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE))
518                         ext4_da_update_reserve_space(inode, retval, 1);
519         }
520         if (flags & EXT4_GET_BLOCKS_DELALLOC_RESERVE)
521                 ext4_clear_inode_state(inode, EXT4_STATE_DELALLOC_RESERVED);
522
523         up_write((&EXT4_I(inode)->i_data_sem));
524         if (retval > 0 && map->m_flags & EXT4_MAP_MAPPED) {
525                 int ret = check_block_validity(inode, map);
526                 if (ret != 0)
527                         return ret;
528         }
529         return retval;
530 }
531
532 /* Maximum number of blocks we map for direct IO at once. */
533 #define DIO_MAX_BLOCKS 4096
534
535 static int _ext4_get_block(struct inode *inode, sector_t iblock,
536                            struct buffer_head *bh, int flags)
537 {
538         handle_t *handle = ext4_journal_current_handle();
539         struct ext4_map_blocks map;
540         int ret = 0, started = 0;
541         int dio_credits;
542
543         map.m_lblk = iblock;
544         map.m_len = bh->b_size >> inode->i_blkbits;
545
546         if (flags && !handle) {
547                 /* Direct IO write... */
548                 if (map.m_len > DIO_MAX_BLOCKS)
549                         map.m_len = DIO_MAX_BLOCKS;
550                 dio_credits = ext4_chunk_trans_blocks(inode, map.m_len);
551                 handle = ext4_journal_start(inode, dio_credits);
552                 if (IS_ERR(handle)) {
553                         ret = PTR_ERR(handle);
554                         return ret;
555                 }
556                 started = 1;
557         }
558
559         ret = ext4_map_blocks(handle, inode, &map, flags);
560         if (ret > 0) {
561                 map_bh(bh, inode->i_sb, map.m_pblk);
562                 bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
563                 bh->b_size = inode->i_sb->s_blocksize * map.m_len;
564                 ret = 0;
565         }
566         if (started)
567                 ext4_journal_stop(handle);
568         return ret;
569 }
570
571 int ext4_get_block(struct inode *inode, sector_t iblock,
572                    struct buffer_head *bh, int create)
573 {
574         return _ext4_get_block(inode, iblock, bh,
575                                create ? EXT4_GET_BLOCKS_CREATE : 0);
576 }
577
578 /*
579  * `handle' can be NULL if create is zero
580  */
581 struct buffer_head *ext4_getblk(handle_t *handle, struct inode *inode,
582                                 ext4_lblk_t block, int create, int *errp)
583 {
584         struct ext4_map_blocks map;
585         struct buffer_head *bh;
586         int fatal = 0, err;
587
588         J_ASSERT(handle != NULL || create == 0);
589
590         map.m_lblk = block;
591         map.m_len = 1;
592         err = ext4_map_blocks(handle, inode, &map,
593                               create ? EXT4_GET_BLOCKS_CREATE : 0);
594
595         if (err < 0)
596                 *errp = err;
597         if (err <= 0)
598                 return NULL;
599         *errp = 0;
600
601         bh = sb_getblk(inode->i_sb, map.m_pblk);
602         if (!bh) {
603                 *errp = -EIO;
604                 return NULL;
605         }
606         if (map.m_flags & EXT4_MAP_NEW) {
607                 J_ASSERT(create != 0);
608                 J_ASSERT(handle != NULL);
609
610                 /*
611                  * Now that we do not always journal data, we should
612                  * keep in mind whether this should always journal the
613                  * new buffer as metadata.  For now, regular file
614                  * writes use ext4_get_block instead, so it's not a
615                  * problem.
616                  */
617                 lock_buffer(bh);
618                 BUFFER_TRACE(bh, "call get_create_access");
619                 fatal = ext4_journal_get_create_access(handle, bh);
620                 if (!fatal && !buffer_uptodate(bh)) {
621                         memset(bh->b_data, 0, inode->i_sb->s_blocksize);
622                         set_buffer_uptodate(bh);
623                 }
624                 unlock_buffer(bh);
625                 BUFFER_TRACE(bh, "call ext4_handle_dirty_metadata");
626                 err = ext4_handle_dirty_metadata(handle, inode, bh);
627                 if (!fatal)
628                         fatal = err;
629         } else {
630                 BUFFER_TRACE(bh, "not a new buffer");
631         }
632         if (fatal) {
633                 *errp = fatal;
634                 brelse(bh);
635                 bh = NULL;
636         }
637         return bh;
638 }
639
640 struct buffer_head *ext4_bread(handle_t *handle, struct inode *inode,
641                                ext4_lblk_t block, int create, int *err)
642 {
643         struct buffer_head *bh;
644
645         bh = ext4_getblk(handle, inode, block, create, err);
646         if (!bh)
647                 return bh;
648         if (buffer_uptodate(bh))
649                 return bh;
650         ll_rw_block(READ | REQ_META | REQ_PRIO, 1, &bh);
651         wait_on_buffer(bh);
652         if (buffer_uptodate(bh))
653                 return bh;
654         put_bh(bh);
655         *err = -EIO;
656         return NULL;
657 }
658
659 static int walk_page_buffers(handle_t *handle,
660                              struct buffer_head *head,
661                              unsigned from,
662                              unsigned to,
663                              int *partial,
664                              int (*fn)(handle_t *handle,
665                                        struct buffer_head *bh))
666 {
667         struct buffer_head *bh;
668         unsigned block_start, block_end;
669         unsigned blocksize = head->b_size;
670         int err, ret = 0;
671         struct buffer_head *next;
672
673         for (bh = head, block_start = 0;
674              ret == 0 && (bh != head || !block_start);
675              block_start = block_end, bh = next) {
676                 next = bh->b_this_page;
677                 block_end = block_start + blocksize;
678                 if (block_end <= from || block_start >= to) {
679                         if (partial && !buffer_uptodate(bh))
680                                 *partial = 1;
681                         continue;
682                 }
683                 err = (*fn)(handle, bh);
684                 if (!ret)
685                         ret = err;
686         }
687         return ret;
688 }
689
690 /*
691  * To preserve ordering, it is essential that the hole instantiation and
692  * the data write be encapsulated in a single transaction.  We cannot
693  * close off a transaction and start a new one between the ext4_get_block()
694  * and the commit_write().  So doing the jbd2_journal_start at the start of
695  * prepare_write() is the right place.
696  *
697  * Also, this function can nest inside ext4_writepage() ->
698  * block_write_full_page(). In that case, we *know* that ext4_writepage()
699  * has generated enough buffer credits to do the whole page.  So we won't
700  * block on the journal in that case, which is good, because the caller may
701  * be PF_MEMALLOC.
702  *
703  * By accident, ext4 can be reentered when a transaction is open via
704  * quota file writes.  If we were to commit the transaction while thus
705  * reentered, there can be a deadlock - we would be holding a quota
706  * lock, and the commit would never complete if another thread had a
707  * transaction open and was blocking on the quota lock - a ranking
708  * violation.
709  *
710  * So what we do is to rely on the fact that jbd2_journal_stop/journal_start
711  * will _not_ run commit under these circumstances because handle->h_ref
712  * is elevated.  We'll still have enough credits for the tiny quotafile
713  * write.
714  */
715 static int do_journal_get_write_access(handle_t *handle,
716                                        struct buffer_head *bh)
717 {
718         int dirty = buffer_dirty(bh);
719         int ret;
720
721         if (!buffer_mapped(bh) || buffer_freed(bh))
722                 return 0;
723         /*
724          * __block_write_begin() could have dirtied some buffers. Clean
725          * the dirty bit as jbd2_journal_get_write_access() could complain
726          * otherwise about fs integrity issues. Setting of the dirty bit
727          * by __block_write_begin() isn't a real problem here as we clear
728          * the bit before releasing a page lock and thus writeback cannot
729          * ever write the buffer.
730          */
731         if (dirty)
732                 clear_buffer_dirty(bh);
733         ret = ext4_journal_get_write_access(handle, bh);
734         if (!ret && dirty)
735                 ret = ext4_handle_dirty_metadata(handle, NULL, bh);
736         return ret;
737 }
738
739 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
740                    struct buffer_head *bh_result, int create);
741 static int ext4_write_begin(struct file *file, struct address_space *mapping,
742                             loff_t pos, unsigned len, unsigned flags,
743                             struct page **pagep, void **fsdata)
744 {
745         struct inode *inode = mapping->host;
746         int ret, needed_blocks;
747         handle_t *handle;
748         int retries = 0;
749         struct page *page;
750         pgoff_t index;
751         unsigned from, to;
752
753         trace_ext4_write_begin(inode, pos, len, flags);
754         /*
755          * Reserve one block more for addition to orphan list in case
756          * we allocate blocks but write fails for some reason
757          */
758         needed_blocks = ext4_writepage_trans_blocks(inode) + 1;
759         index = pos >> PAGE_CACHE_SHIFT;
760         from = pos & (PAGE_CACHE_SIZE - 1);
761         to = from + len;
762
763 retry:
764         handle = ext4_journal_start(inode, needed_blocks);
765         if (IS_ERR(handle)) {
766                 ret = PTR_ERR(handle);
767                 goto out;
768         }
769
770         /* We cannot recurse into the filesystem as the transaction is already
771          * started */
772         flags |= AOP_FLAG_NOFS;
773
774         page = grab_cache_page_write_begin(mapping, index, flags);
775         if (!page) {
776                 ext4_journal_stop(handle);
777                 ret = -ENOMEM;
778                 goto out;
779         }
780         *pagep = page;
781
782         if (ext4_should_dioread_nolock(inode))
783                 ret = __block_write_begin(page, pos, len, ext4_get_block_write);
784         else
785                 ret = __block_write_begin(page, pos, len, ext4_get_block);
786
787         if (!ret && ext4_should_journal_data(inode)) {
788                 ret = walk_page_buffers(handle, page_buffers(page),
789                                 from, to, NULL, do_journal_get_write_access);
790         }
791
792         if (ret) {
793                 unlock_page(page);
794                 page_cache_release(page);
795                 /*
796                  * __block_write_begin may have instantiated a few blocks
797                  * outside i_size.  Trim these off again. Don't need
798                  * i_size_read because we hold i_mutex.
799                  *
800                  * Add inode to orphan list in case we crash before
801                  * truncate finishes
802                  */
803                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
804                         ext4_orphan_add(handle, inode);
805
806                 ext4_journal_stop(handle);
807                 if (pos + len > inode->i_size) {
808                         ext4_truncate_failed_write(inode);
809                         /*
810                          * If truncate failed early the inode might
811                          * still be on the orphan list; we need to
812                          * make sure the inode is removed from the
813                          * orphan list in that case.
814                          */
815                         if (inode->i_nlink)
816                                 ext4_orphan_del(NULL, inode);
817                 }
818         }
819
820         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
821                 goto retry;
822 out:
823         return ret;
824 }
825
826 /* For write_end() in data=journal mode */
827 static int write_end_fn(handle_t *handle, struct buffer_head *bh)
828 {
829         if (!buffer_mapped(bh) || buffer_freed(bh))
830                 return 0;
831         set_buffer_uptodate(bh);
832         return ext4_handle_dirty_metadata(handle, NULL, bh);
833 }
834
835 static int ext4_generic_write_end(struct file *file,
836                                   struct address_space *mapping,
837                                   loff_t pos, unsigned len, unsigned copied,
838                                   struct page *page, void *fsdata)
839 {
840         int i_size_changed = 0;
841         struct inode *inode = mapping->host;
842         handle_t *handle = ext4_journal_current_handle();
843
844         copied = block_write_end(file, mapping, pos, len, copied, page, fsdata);
845
846         /*
847          * No need to use i_size_read() here, the i_size
848          * cannot change under us because we hold i_mutex.
849          *
850          * But it's important to update i_size while still holding page lock:
851          * page writeout could otherwise come in and zero beyond i_size.
852          */
853         if (pos + copied > inode->i_size) {
854                 i_size_write(inode, pos + copied);
855                 i_size_changed = 1;
856         }
857
858         if (pos + copied >  EXT4_I(inode)->i_disksize) {
859                 /* We need to mark inode dirty even if
860                  * new_i_size is less that inode->i_size
861                  * bu greater than i_disksize.(hint delalloc)
862                  */
863                 ext4_update_i_disksize(inode, (pos + copied));
864                 i_size_changed = 1;
865         }
866         unlock_page(page);
867         page_cache_release(page);
868
869         /*
870          * Don't mark the inode dirty under page lock. First, it unnecessarily
871          * makes the holding time of page lock longer. Second, it forces lock
872          * ordering of page lock and transaction start for journaling
873          * filesystems.
874          */
875         if (i_size_changed)
876                 ext4_mark_inode_dirty(handle, inode);
877
878         return copied;
879 }
880
881 /*
882  * We need to pick up the new inode size which generic_commit_write gave us
883  * `file' can be NULL - eg, when called from page_symlink().
884  *
885  * ext4 never places buffers on inode->i_mapping->private_list.  metadata
886  * buffers are managed internally.
887  */
888 static int ext4_ordered_write_end(struct file *file,
889                                   struct address_space *mapping,
890                                   loff_t pos, unsigned len, unsigned copied,
891                                   struct page *page, void *fsdata)
892 {
893         handle_t *handle = ext4_journal_current_handle();
894         struct inode *inode = mapping->host;
895         int ret = 0, ret2;
896
897         trace_ext4_ordered_write_end(inode, pos, len, copied);
898         ret = ext4_jbd2_file_inode(handle, inode);
899
900         if (ret == 0) {
901                 ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
902                                                         page, fsdata);
903                 copied = ret2;
904                 if (pos + len > inode->i_size && ext4_can_truncate(inode))
905                         /* if we have allocated more blocks and copied
906                          * less. We will have blocks allocated outside
907                          * inode->i_size. So truncate them
908                          */
909                         ext4_orphan_add(handle, inode);
910                 if (ret2 < 0)
911                         ret = ret2;
912         }
913         ret2 = ext4_journal_stop(handle);
914         if (!ret)
915                 ret = ret2;
916
917         if (pos + len > inode->i_size) {
918                 ext4_truncate_failed_write(inode);
919                 /*
920                  * If truncate failed early the inode might still be
921                  * on the orphan list; we need to make sure the inode
922                  * is removed from the orphan list in that case.
923                  */
924                 if (inode->i_nlink)
925                         ext4_orphan_del(NULL, inode);
926         }
927
928
929         return ret ? ret : copied;
930 }
931
932 static int ext4_writeback_write_end(struct file *file,
933                                     struct address_space *mapping,
934                                     loff_t pos, unsigned len, unsigned copied,
935                                     struct page *page, void *fsdata)
936 {
937         handle_t *handle = ext4_journal_current_handle();
938         struct inode *inode = mapping->host;
939         int ret = 0, ret2;
940
941         trace_ext4_writeback_write_end(inode, pos, len, copied);
942         ret2 = ext4_generic_write_end(file, mapping, pos, len, copied,
943                                                         page, fsdata);
944         copied = ret2;
945         if (pos + len > inode->i_size && ext4_can_truncate(inode))
946                 /* if we have allocated more blocks and copied
947                  * less. We will have blocks allocated outside
948                  * inode->i_size. So truncate them
949                  */
950                 ext4_orphan_add(handle, inode);
951
952         if (ret2 < 0)
953                 ret = ret2;
954
955         ret2 = ext4_journal_stop(handle);
956         if (!ret)
957                 ret = ret2;
958
959         if (pos + len > inode->i_size) {
960                 ext4_truncate_failed_write(inode);
961                 /*
962                  * If truncate failed early the inode might still be
963                  * on the orphan list; we need to make sure the inode
964                  * is removed from the orphan list in that case.
965                  */
966                 if (inode->i_nlink)
967                         ext4_orphan_del(NULL, inode);
968         }
969
970         return ret ? ret : copied;
971 }
972
973 static int ext4_journalled_write_end(struct file *file,
974                                      struct address_space *mapping,
975                                      loff_t pos, unsigned len, unsigned copied,
976                                      struct page *page, void *fsdata)
977 {
978         handle_t *handle = ext4_journal_current_handle();
979         struct inode *inode = mapping->host;
980         int ret = 0, ret2;
981         int partial = 0;
982         unsigned from, to;
983         loff_t new_i_size;
984
985         trace_ext4_journalled_write_end(inode, pos, len, copied);
986         from = pos & (PAGE_CACHE_SIZE - 1);
987         to = from + len;
988
989         BUG_ON(!ext4_handle_valid(handle));
990
991         if (copied < len) {
992                 if (!PageUptodate(page))
993                         copied = 0;
994                 page_zero_new_buffers(page, from+copied, to);
995         }
996
997         ret = walk_page_buffers(handle, page_buffers(page), from,
998                                 to, &partial, write_end_fn);
999         if (!partial)
1000                 SetPageUptodate(page);
1001         new_i_size = pos + copied;
1002         if (new_i_size > inode->i_size)
1003                 i_size_write(inode, pos+copied);
1004         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1005         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1006         if (new_i_size > EXT4_I(inode)->i_disksize) {
1007                 ext4_update_i_disksize(inode, new_i_size);
1008                 ret2 = ext4_mark_inode_dirty(handle, inode);
1009                 if (!ret)
1010                         ret = ret2;
1011         }
1012
1013         unlock_page(page);
1014         page_cache_release(page);
1015         if (pos + len > inode->i_size && ext4_can_truncate(inode))
1016                 /* if we have allocated more blocks and copied
1017                  * less. We will have blocks allocated outside
1018                  * inode->i_size. So truncate them
1019                  */
1020                 ext4_orphan_add(handle, inode);
1021
1022         ret2 = ext4_journal_stop(handle);
1023         if (!ret)
1024                 ret = ret2;
1025         if (pos + len > inode->i_size) {
1026                 ext4_truncate_failed_write(inode);
1027                 /*
1028                  * If truncate failed early the inode might still be
1029                  * on the orphan list; we need to make sure the inode
1030                  * is removed from the orphan list in that case.
1031                  */
1032                 if (inode->i_nlink)
1033                         ext4_orphan_del(NULL, inode);
1034         }
1035
1036         return ret ? ret : copied;
1037 }
1038
1039 /*
1040  * Reserve a single block located at lblock
1041  */
1042 static int ext4_da_reserve_space(struct inode *inode, ext4_lblk_t lblock)
1043 {
1044         int retries = 0;
1045         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1046         struct ext4_inode_info *ei = EXT4_I(inode);
1047         unsigned long md_needed;
1048         int ret;
1049
1050         /*
1051          * recalculate the amount of metadata blocks to reserve
1052          * in order to allocate nrblocks
1053          * worse case is one extent per block
1054          */
1055 repeat:
1056         spin_lock(&ei->i_block_reservation_lock);
1057         md_needed = ext4_calc_metadata_amount(inode, lblock);
1058         trace_ext4_da_reserve_space(inode, md_needed);
1059         spin_unlock(&ei->i_block_reservation_lock);
1060
1061         /*
1062          * We will charge metadata quota at writeout time; this saves
1063          * us from metadata over-estimation, though we may go over by
1064          * a small amount in the end.  Here we just reserve for data.
1065          */
1066         ret = dquot_reserve_block(inode, 1);
1067         if (ret)
1068                 return ret;
1069         /*
1070          * We do still charge estimated metadata to the sb though;
1071          * we cannot afford to run out of free blocks.
1072          */
1073         if (ext4_claim_free_blocks(sbi, md_needed + 1, 0)) {
1074                 dquot_release_reservation_block(inode, 1);
1075                 if (ext4_should_retry_alloc(inode->i_sb, &retries)) {
1076                         yield();
1077                         goto repeat;
1078                 }
1079                 return -ENOSPC;
1080         }
1081         spin_lock(&ei->i_block_reservation_lock);
1082         ei->i_reserved_data_blocks++;
1083         ei->i_reserved_meta_blocks += md_needed;
1084         spin_unlock(&ei->i_block_reservation_lock);
1085
1086         return 0;       /* success */
1087 }
1088
1089 static void ext4_da_release_space(struct inode *inode, int to_free)
1090 {
1091         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1092         struct ext4_inode_info *ei = EXT4_I(inode);
1093
1094         if (!to_free)
1095                 return;         /* Nothing to release, exit */
1096
1097         spin_lock(&EXT4_I(inode)->i_block_reservation_lock);
1098
1099         trace_ext4_da_release_space(inode, to_free);
1100         if (unlikely(to_free > ei->i_reserved_data_blocks)) {
1101                 /*
1102                  * if there aren't enough reserved blocks, then the
1103                  * counter is messed up somewhere.  Since this
1104                  * function is called from invalidate page, it's
1105                  * harmless to return without any action.
1106                  */
1107                 ext4_msg(inode->i_sb, KERN_NOTICE, "ext4_da_release_space: "
1108                          "ino %lu, to_free %d with only %d reserved "
1109                          "data blocks\n", inode->i_ino, to_free,
1110                          ei->i_reserved_data_blocks);
1111                 WARN_ON(1);
1112                 to_free = ei->i_reserved_data_blocks;
1113         }
1114         ei->i_reserved_data_blocks -= to_free;
1115
1116         if (ei->i_reserved_data_blocks == 0) {
1117                 /*
1118                  * We can release all of the reserved metadata blocks
1119                  * only when we have written all of the delayed
1120                  * allocation blocks.
1121                  */
1122                 percpu_counter_sub(&sbi->s_dirtyblocks_counter,
1123                                    ei->i_reserved_meta_blocks);
1124                 ei->i_reserved_meta_blocks = 0;
1125                 ei->i_da_metadata_calc_len = 0;
1126         }
1127
1128         /* update fs dirty data blocks counter */
1129         percpu_counter_sub(&sbi->s_dirtyblocks_counter, to_free);
1130
1131         spin_unlock(&EXT4_I(inode)->i_block_reservation_lock);
1132
1133         dquot_release_reservation_block(inode, to_free);
1134 }
1135
1136 static void ext4_da_page_release_reservation(struct page *page,
1137                                              unsigned long offset)
1138 {
1139         int to_release = 0;
1140         struct buffer_head *head, *bh;
1141         unsigned int curr_off = 0;
1142
1143         head = page_buffers(page);
1144         bh = head;
1145         do {
1146                 unsigned int next_off = curr_off + bh->b_size;
1147
1148                 if ((offset <= curr_off) && (buffer_delay(bh))) {
1149                         to_release++;
1150                         clear_buffer_delay(bh);
1151                 }
1152                 curr_off = next_off;
1153         } while ((bh = bh->b_this_page) != head);
1154         ext4_da_release_space(page->mapping->host, to_release);
1155 }
1156
1157 /*
1158  * Delayed allocation stuff
1159  */
1160
1161 /*
1162  * mpage_da_submit_io - walks through extent of pages and try to write
1163  * them with writepage() call back
1164  *
1165  * @mpd->inode: inode
1166  * @mpd->first_page: first page of the extent
1167  * @mpd->next_page: page after the last page of the extent
1168  *
1169  * By the time mpage_da_submit_io() is called we expect all blocks
1170  * to be allocated. this may be wrong if allocation failed.
1171  *
1172  * As pages are already locked by write_cache_pages(), we can't use it
1173  */
1174 static int mpage_da_submit_io(struct mpage_da_data *mpd,
1175                               struct ext4_map_blocks *map)
1176 {
1177         struct pagevec pvec;
1178         unsigned long index, end;
1179         int ret = 0, err, nr_pages, i;
1180         struct inode *inode = mpd->inode;
1181         struct address_space *mapping = inode->i_mapping;
1182         loff_t size = i_size_read(inode);
1183         unsigned int len, block_start;
1184         struct buffer_head *bh, *page_bufs = NULL;
1185         int journal_data = ext4_should_journal_data(inode);
1186         sector_t pblock = 0, cur_logical = 0;
1187         struct ext4_io_submit io_submit;
1188
1189         BUG_ON(mpd->next_page <= mpd->first_page);
1190         memset(&io_submit, 0, sizeof(io_submit));
1191         /*
1192          * We need to start from the first_page to the next_page - 1
1193          * to make sure we also write the mapped dirty buffer_heads.
1194          * If we look at mpd->b_blocknr we would only be looking
1195          * at the currently mapped buffer_heads.
1196          */
1197         index = mpd->first_page;
1198         end = mpd->next_page - 1;
1199
1200         pagevec_init(&pvec, 0);
1201         while (index <= end) {
1202                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1203                 if (nr_pages == 0)
1204                         break;
1205                 for (i = 0; i < nr_pages; i++) {
1206                         int commit_write = 0, skip_page = 0;
1207                         struct page *page = pvec.pages[i];
1208
1209                         index = page->index;
1210                         if (index > end)
1211                                 break;
1212
1213                         if (index == size >> PAGE_CACHE_SHIFT)
1214                                 len = size & ~PAGE_CACHE_MASK;
1215                         else
1216                                 len = PAGE_CACHE_SIZE;
1217                         if (map) {
1218                                 cur_logical = index << (PAGE_CACHE_SHIFT -
1219                                                         inode->i_blkbits);
1220                                 pblock = map->m_pblk + (cur_logical -
1221                                                         map->m_lblk);
1222                         }
1223                         index++;
1224
1225                         BUG_ON(!PageLocked(page));
1226                         BUG_ON(PageWriteback(page));
1227
1228                         /*
1229                          * If the page does not have buffers (for
1230                          * whatever reason), try to create them using
1231                          * __block_write_begin.  If this fails,
1232                          * skip the page and move on.
1233                          */
1234                         if (!page_has_buffers(page)) {
1235                                 if (__block_write_begin(page, 0, len,
1236                                                 noalloc_get_block_write)) {
1237                                 skip_page:
1238                                         unlock_page(page);
1239                                         continue;
1240                                 }
1241                                 commit_write = 1;
1242                         }
1243
1244                         bh = page_bufs = page_buffers(page);
1245                         block_start = 0;
1246                         do {
1247                                 if (!bh)
1248                                         goto skip_page;
1249                                 if (map && (cur_logical >= map->m_lblk) &&
1250                                     (cur_logical <= (map->m_lblk +
1251                                                      (map->m_len - 1)))) {
1252                                         if (buffer_delay(bh)) {
1253                                                 clear_buffer_delay(bh);
1254                                                 bh->b_blocknr = pblock;
1255                                         }
1256                                         if (buffer_unwritten(bh) ||
1257                                             buffer_mapped(bh))
1258                                                 BUG_ON(bh->b_blocknr != pblock);
1259                                         if (map->m_flags & EXT4_MAP_UNINIT)
1260                                                 set_buffer_uninit(bh);
1261                                         clear_buffer_unwritten(bh);
1262                                 }
1263
1264                                 /* skip page if block allocation undone */
1265                                 if (buffer_delay(bh) || buffer_unwritten(bh))
1266                                         skip_page = 1;
1267                                 bh = bh->b_this_page;
1268                                 block_start += bh->b_size;
1269                                 cur_logical++;
1270                                 pblock++;
1271                         } while (bh != page_bufs);
1272
1273                         if (skip_page)
1274                                 goto skip_page;
1275
1276                         if (commit_write)
1277                                 /* mark the buffer_heads as dirty & uptodate */
1278                                 block_commit_write(page, 0, len);
1279
1280                         clear_page_dirty_for_io(page);
1281                         /*
1282                          * Delalloc doesn't support data journalling,
1283                          * but eventually maybe we'll lift this
1284                          * restriction.
1285                          */
1286                         if (unlikely(journal_data && PageChecked(page)))
1287                                 err = __ext4_journalled_writepage(page, len);
1288                         else if (test_opt(inode->i_sb, MBLK_IO_SUBMIT))
1289                                 err = ext4_bio_write_page(&io_submit, page,
1290                                                           len, mpd->wbc);
1291                         else if (buffer_uninit(page_bufs)) {
1292                                 ext4_set_bh_endio(page_bufs, inode);
1293                                 err = block_write_full_page_endio(page,
1294                                         noalloc_get_block_write,
1295                                         mpd->wbc, ext4_end_io_buffer_write);
1296                         } else
1297                                 err = block_write_full_page(page,
1298                                         noalloc_get_block_write, mpd->wbc);
1299
1300                         if (!err)
1301                                 mpd->pages_written++;
1302                         /*
1303                          * In error case, we have to continue because
1304                          * remaining pages are still locked
1305                          */
1306                         if (ret == 0)
1307                                 ret = err;
1308                 }
1309                 pagevec_release(&pvec);
1310         }
1311         ext4_io_submit(&io_submit);
1312         return ret;
1313 }
1314
1315 static void ext4_da_block_invalidatepages(struct mpage_da_data *mpd)
1316 {
1317         int nr_pages, i;
1318         pgoff_t index, end;
1319         struct pagevec pvec;
1320         struct inode *inode = mpd->inode;
1321         struct address_space *mapping = inode->i_mapping;
1322
1323         index = mpd->first_page;
1324         end   = mpd->next_page - 1;
1325         while (index <= end) {
1326                 nr_pages = pagevec_lookup(&pvec, mapping, index, PAGEVEC_SIZE);
1327                 if (nr_pages == 0)
1328                         break;
1329                 for (i = 0; i < nr_pages; i++) {
1330                         struct page *page = pvec.pages[i];
1331                         if (page->index > end)
1332                                 break;
1333                         BUG_ON(!PageLocked(page));
1334                         BUG_ON(PageWriteback(page));
1335                         block_invalidatepage(page, 0);
1336                         ClearPageUptodate(page);
1337                         unlock_page(page);
1338                 }
1339                 index = pvec.pages[nr_pages - 1]->index + 1;
1340                 pagevec_release(&pvec);
1341         }
1342         return;
1343 }
1344
1345 static void ext4_print_free_blocks(struct inode *inode)
1346 {
1347         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1348         printk(KERN_CRIT "Total free blocks count %lld\n",
1349                ext4_count_free_blocks(inode->i_sb));
1350         printk(KERN_CRIT "Free/Dirty block details\n");
1351         printk(KERN_CRIT "free_blocks=%lld\n",
1352                (long long) percpu_counter_sum(&sbi->s_freeblocks_counter));
1353         printk(KERN_CRIT "dirty_blocks=%lld\n",
1354                (long long) percpu_counter_sum(&sbi->s_dirtyblocks_counter));
1355         printk(KERN_CRIT "Block reservation details\n");
1356         printk(KERN_CRIT "i_reserved_data_blocks=%u\n",
1357                EXT4_I(inode)->i_reserved_data_blocks);
1358         printk(KERN_CRIT "i_reserved_meta_blocks=%u\n",
1359                EXT4_I(inode)->i_reserved_meta_blocks);
1360         return;
1361 }
1362
1363 /*
1364  * mpage_da_map_and_submit - go through given space, map them
1365  *       if necessary, and then submit them for I/O
1366  *
1367  * @mpd - bh describing space
1368  *
1369  * The function skips space we know is already mapped to disk blocks.
1370  *
1371  */
1372 static void mpage_da_map_and_submit(struct mpage_da_data *mpd)
1373 {
1374         int err, blks, get_blocks_flags;
1375         struct ext4_map_blocks map, *mapp = NULL;
1376         sector_t next = mpd->b_blocknr;
1377         unsigned max_blocks = mpd->b_size >> mpd->inode->i_blkbits;
1378         loff_t disksize = EXT4_I(mpd->inode)->i_disksize;
1379         handle_t *handle = NULL;
1380
1381         /*
1382          * If the blocks are mapped already, or we couldn't accumulate
1383          * any blocks, then proceed immediately to the submission stage.
1384          */
1385         if ((mpd->b_size == 0) ||
1386             ((mpd->b_state  & (1 << BH_Mapped)) &&
1387              !(mpd->b_state & (1 << BH_Delay)) &&
1388              !(mpd->b_state & (1 << BH_Unwritten))))
1389                 goto submit_io;
1390
1391         handle = ext4_journal_current_handle();
1392         BUG_ON(!handle);
1393
1394         /*
1395          * Call ext4_map_blocks() to allocate any delayed allocation
1396          * blocks, or to convert an uninitialized extent to be
1397          * initialized (in the case where we have written into
1398          * one or more preallocated blocks).
1399          *
1400          * We pass in the magic EXT4_GET_BLOCKS_DELALLOC_RESERVE to
1401          * indicate that we are on the delayed allocation path.  This
1402          * affects functions in many different parts of the allocation
1403          * call path.  This flag exists primarily because we don't
1404          * want to change *many* call functions, so ext4_map_blocks()
1405          * will set the EXT4_STATE_DELALLOC_RESERVED flag once the
1406          * inode's allocation semaphore is taken.
1407          *
1408          * If the blocks in questions were delalloc blocks, set
1409          * EXT4_GET_BLOCKS_DELALLOC_RESERVE so the delalloc accounting
1410          * variables are updated after the blocks have been allocated.
1411          */
1412         map.m_lblk = next;
1413         map.m_len = max_blocks;
1414         get_blocks_flags = EXT4_GET_BLOCKS_CREATE;
1415         if (ext4_should_dioread_nolock(mpd->inode))
1416                 get_blocks_flags |= EXT4_GET_BLOCKS_IO_CREATE_EXT;
1417         if (mpd->b_state & (1 << BH_Delay))
1418                 get_blocks_flags |= EXT4_GET_BLOCKS_DELALLOC_RESERVE;
1419
1420         blks = ext4_map_blocks(handle, mpd->inode, &map, get_blocks_flags);
1421         if (blks < 0) {
1422                 struct super_block *sb = mpd->inode->i_sb;
1423
1424                 err = blks;
1425                 /*
1426                  * If get block returns EAGAIN or ENOSPC and there
1427                  * appears to be free blocks we will just let
1428                  * mpage_da_submit_io() unlock all of the pages.
1429                  */
1430                 if (err == -EAGAIN)
1431                         goto submit_io;
1432
1433                 if (err == -ENOSPC &&
1434                     ext4_count_free_blocks(sb)) {
1435                         mpd->retval = err;
1436                         goto submit_io;
1437                 }
1438
1439                 /*
1440                  * get block failure will cause us to loop in
1441                  * writepages, because a_ops->writepage won't be able
1442                  * to make progress. The page will be redirtied by
1443                  * writepage and writepages will again try to write
1444                  * the same.
1445                  */
1446                 if (!(EXT4_SB(sb)->s_mount_flags & EXT4_MF_FS_ABORTED)) {
1447                         ext4_msg(sb, KERN_CRIT,
1448                                  "delayed block allocation failed for inode %lu "
1449                                  "at logical offset %llu with max blocks %zd "
1450                                  "with error %d", mpd->inode->i_ino,
1451                                  (unsigned long long) next,
1452                                  mpd->b_size >> mpd->inode->i_blkbits, err);
1453                         ext4_msg(sb, KERN_CRIT,
1454                                 "This should not happen!! Data will be lost\n");
1455                         if (err == -ENOSPC)
1456                                 ext4_print_free_blocks(mpd->inode);
1457                 }
1458                 /* invalidate all the pages */
1459                 ext4_da_block_invalidatepages(mpd);
1460
1461                 /* Mark this page range as having been completed */
1462                 mpd->io_done = 1;
1463                 return;
1464         }
1465         BUG_ON(blks == 0);
1466
1467         mapp = &map;
1468         if (map.m_flags & EXT4_MAP_NEW) {
1469                 struct block_device *bdev = mpd->inode->i_sb->s_bdev;
1470                 int i;
1471
1472                 for (i = 0; i < map.m_len; i++)
1473                         unmap_underlying_metadata(bdev, map.m_pblk + i);
1474         }
1475
1476         if (ext4_should_order_data(mpd->inode)) {
1477                 err = ext4_jbd2_file_inode(handle, mpd->inode);
1478                 if (err)
1479                         /* This only happens if the journal is aborted */
1480                         return;
1481         }
1482
1483         /*
1484          * Update on-disk size along with block allocation.
1485          */
1486         disksize = ((loff_t) next + blks) << mpd->inode->i_blkbits;
1487         if (disksize > i_size_read(mpd->inode))
1488                 disksize = i_size_read(mpd->inode);
1489         if (disksize > EXT4_I(mpd->inode)->i_disksize) {
1490                 ext4_update_i_disksize(mpd->inode, disksize);
1491                 err = ext4_mark_inode_dirty(handle, mpd->inode);
1492                 if (err)
1493                         ext4_error(mpd->inode->i_sb,
1494                                    "Failed to mark inode %lu dirty",
1495                                    mpd->inode->i_ino);
1496         }
1497
1498 submit_io:
1499         mpage_da_submit_io(mpd, mapp);
1500         mpd->io_done = 1;
1501 }
1502
1503 #define BH_FLAGS ((1 << BH_Uptodate) | (1 << BH_Mapped) | \
1504                 (1 << BH_Delay) | (1 << BH_Unwritten))
1505
1506 /*
1507  * mpage_add_bh_to_extent - try to add one more block to extent of blocks
1508  *
1509  * @mpd->lbh - extent of blocks
1510  * @logical - logical number of the block in the file
1511  * @bh - bh of the block (used to access block's state)
1512  *
1513  * the function is used to collect contig. blocks in same state
1514  */
1515 static void mpage_add_bh_to_extent(struct mpage_da_data *mpd,
1516                                    sector_t logical, size_t b_size,
1517                                    unsigned long b_state)
1518 {
1519         sector_t next;
1520         int nrblocks = mpd->b_size >> mpd->inode->i_blkbits;
1521
1522         /*
1523          * XXX Don't go larger than mballoc is willing to allocate
1524          * This is a stopgap solution.  We eventually need to fold
1525          * mpage_da_submit_io() into this function and then call
1526          * ext4_map_blocks() multiple times in a loop
1527          */
1528         if (nrblocks >= 8*1024*1024/mpd->inode->i_sb->s_blocksize)
1529                 goto flush_it;
1530
1531         /* check if thereserved journal credits might overflow */
1532         if (!(ext4_test_inode_flag(mpd->inode, EXT4_INODE_EXTENTS))) {
1533                 if (nrblocks >= EXT4_MAX_TRANS_DATA) {
1534                         /*
1535                          * With non-extent format we are limited by the journal
1536                          * credit available.  Total credit needed to insert
1537                          * nrblocks contiguous blocks is dependent on the
1538                          * nrblocks.  So limit nrblocks.
1539                          */
1540                         goto flush_it;
1541                 } else if ((nrblocks + (b_size >> mpd->inode->i_blkbits)) >
1542                                 EXT4_MAX_TRANS_DATA) {
1543                         /*
1544                          * Adding the new buffer_head would make it cross the
1545                          * allowed limit for which we have journal credit
1546                          * reserved. So limit the new bh->b_size
1547                          */
1548                         b_size = (EXT4_MAX_TRANS_DATA - nrblocks) <<
1549                                                 mpd->inode->i_blkbits;
1550                         /* we will do mpage_da_submit_io in the next loop */
1551                 }
1552         }
1553         /*
1554          * First block in the extent
1555          */
1556         if (mpd->b_size == 0) {
1557                 mpd->b_blocknr = logical;
1558                 mpd->b_size = b_size;
1559                 mpd->b_state = b_state & BH_FLAGS;
1560                 return;
1561         }
1562
1563         next = mpd->b_blocknr + nrblocks;
1564         /*
1565          * Can we merge the block to our big extent?
1566          */
1567         if (logical == next && (b_state & BH_FLAGS) == mpd->b_state) {
1568                 mpd->b_size += b_size;
1569                 return;
1570         }
1571
1572 flush_it:
1573         /*
1574          * We couldn't merge the block to our extent, so we
1575          * need to flush current  extent and start new one
1576          */
1577         mpage_da_map_and_submit(mpd);
1578         return;
1579 }
1580
1581 static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh)
1582 {
1583         return (buffer_delay(bh) || buffer_unwritten(bh)) && buffer_dirty(bh);
1584 }
1585
1586 /*
1587  * This is a special get_blocks_t callback which is used by
1588  * ext4_da_write_begin().  It will either return mapped block or
1589  * reserve space for a single block.
1590  *
1591  * For delayed buffer_head we have BH_Mapped, BH_New, BH_Delay set.
1592  * We also have b_blocknr = -1 and b_bdev initialized properly
1593  *
1594  * For unwritten buffer_head we have BH_Mapped, BH_New, BH_Unwritten set.
1595  * We also have b_blocknr = physicalblock mapping unwritten extent and b_bdev
1596  * initialized properly.
1597  */
1598 static int ext4_da_get_block_prep(struct inode *inode, sector_t iblock,
1599                                   struct buffer_head *bh, int create)
1600 {
1601         struct ext4_map_blocks map;
1602         int ret = 0;
1603         sector_t invalid_block = ~((sector_t) 0xffff);
1604
1605         if (invalid_block < ext4_blocks_count(EXT4_SB(inode->i_sb)->s_es))
1606                 invalid_block = ~0;
1607
1608         BUG_ON(create == 0);
1609         BUG_ON(bh->b_size != inode->i_sb->s_blocksize);
1610
1611         map.m_lblk = iblock;
1612         map.m_len = 1;
1613
1614         /*
1615          * first, we need to know whether the block is allocated already
1616          * preallocated blocks are unmapped but should treated
1617          * the same as allocated blocks.
1618          */
1619         ret = ext4_map_blocks(NULL, inode, &map, 0);
1620         if (ret < 0)
1621                 return ret;
1622         if (ret == 0) {
1623                 if (buffer_delay(bh))
1624                         return 0; /* Not sure this could or should happen */
1625                 /*
1626                  * XXX: __block_write_begin() unmaps passed block, is it OK?
1627                  */
1628                 ret = ext4_da_reserve_space(inode, iblock);
1629                 if (ret)
1630                         /* not enough space to reserve */
1631                         return ret;
1632
1633                 map_bh(bh, inode->i_sb, invalid_block);
1634                 set_buffer_new(bh);
1635                 set_buffer_delay(bh);
1636                 return 0;
1637         }
1638
1639         map_bh(bh, inode->i_sb, map.m_pblk);
1640         bh->b_state = (bh->b_state & ~EXT4_MAP_FLAGS) | map.m_flags;
1641
1642         if (buffer_unwritten(bh)) {
1643                 /* A delayed write to unwritten bh should be marked
1644                  * new and mapped.  Mapped ensures that we don't do
1645                  * get_block multiple times when we write to the same
1646                  * offset and new ensures that we do proper zero out
1647                  * for partial write.
1648                  */
1649                 set_buffer_new(bh);
1650                 set_buffer_mapped(bh);
1651         }
1652         return 0;
1653 }
1654
1655 /*
1656  * This function is used as a standard get_block_t calback function
1657  * when there is no desire to allocate any blocks.  It is used as a
1658  * callback function for block_write_begin() and block_write_full_page().
1659  * These functions should only try to map a single block at a time.
1660  *
1661  * Since this function doesn't do block allocations even if the caller
1662  * requests it by passing in create=1, it is critically important that
1663  * any caller checks to make sure that any buffer heads are returned
1664  * by this function are either all already mapped or marked for
1665  * delayed allocation before calling  block_write_full_page().  Otherwise,
1666  * b_blocknr could be left unitialized, and the page write functions will
1667  * be taken by surprise.
1668  */
1669 static int noalloc_get_block_write(struct inode *inode, sector_t iblock,
1670                                    struct buffer_head *bh_result, int create)
1671 {
1672         BUG_ON(bh_result->b_size != inode->i_sb->s_blocksize);
1673         return _ext4_get_block(inode, iblock, bh_result, 0);
1674 }
1675
1676 static int bget_one(handle_t *handle, struct buffer_head *bh)
1677 {
1678         get_bh(bh);
1679         return 0;
1680 }
1681
1682 static int bput_one(handle_t *handle, struct buffer_head *bh)
1683 {
1684         put_bh(bh);
1685         return 0;
1686 }
1687
1688 static int __ext4_journalled_writepage(struct page *page,
1689                                        unsigned int len)
1690 {
1691         struct address_space *mapping = page->mapping;
1692         struct inode *inode = mapping->host;
1693         struct buffer_head *page_bufs;
1694         handle_t *handle = NULL;
1695         int ret = 0;
1696         int err;
1697
1698         ClearPageChecked(page);
1699         page_bufs = page_buffers(page);
1700         BUG_ON(!page_bufs);
1701         walk_page_buffers(handle, page_bufs, 0, len, NULL, bget_one);
1702         /* As soon as we unlock the page, it can go away, but we have
1703          * references to buffers so we are safe */
1704         unlock_page(page);
1705
1706         handle = ext4_journal_start(inode, ext4_writepage_trans_blocks(inode));
1707         if (IS_ERR(handle)) {
1708                 ret = PTR_ERR(handle);
1709                 goto out;
1710         }
1711
1712         BUG_ON(!ext4_handle_valid(handle));
1713
1714         ret = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1715                                 do_journal_get_write_access);
1716
1717         err = walk_page_buffers(handle, page_bufs, 0, len, NULL,
1718                                 write_end_fn);
1719         if (ret == 0)
1720                 ret = err;
1721         EXT4_I(inode)->i_datasync_tid = handle->h_transaction->t_tid;
1722         err = ext4_journal_stop(handle);
1723         if (!ret)
1724                 ret = err;
1725
1726         walk_page_buffers(handle, page_bufs, 0, len, NULL, bput_one);
1727         ext4_set_inode_state(inode, EXT4_STATE_JDATA);
1728 out:
1729         return ret;
1730 }
1731
1732 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode);
1733 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate);
1734
1735 /*
1736  * Note that we don't need to start a transaction unless we're journaling data
1737  * because we should have holes filled from ext4_page_mkwrite(). We even don't
1738  * need to file the inode to the transaction's list in ordered mode because if
1739  * we are writing back data added by write(), the inode is already there and if
1740  * we are writing back data modified via mmap(), no one guarantees in which
1741  * transaction the data will hit the disk. In case we are journaling data, we
1742  * cannot start transaction directly because transaction start ranks above page
1743  * lock so we have to do some magic.
1744  *
1745  * This function can get called via...
1746  *   - ext4_da_writepages after taking page lock (have journal handle)
1747  *   - journal_submit_inode_data_buffers (no journal handle)
1748  *   - shrink_page_list via pdflush (no journal handle)
1749  *   - grab_page_cache when doing write_begin (have journal handle)
1750  *
1751  * We don't do any block allocation in this function. If we have page with
1752  * multiple blocks we need to write those buffer_heads that are mapped. This
1753  * is important for mmaped based write. So if we do with blocksize 1K
1754  * truncate(f, 1024);
1755  * a = mmap(f, 0, 4096);
1756  * a[0] = 'a';
1757  * truncate(f, 4096);
1758  * we have in the page first buffer_head mapped via page_mkwrite call back
1759  * but other bufer_heads would be unmapped but dirty(dirty done via the
1760  * do_wp_page). So writepage should write the first block. If we modify
1761  * the mmap area beyond 1024 we will again get a page_fault and the
1762  * page_mkwrite callback will do the block allocation and mark the
1763  * buffer_heads mapped.
1764  *
1765  * We redirty the page if we have any buffer_heads that is either delay or
1766  * unwritten in the page.
1767  *
1768  * We can get recursively called as show below.
1769  *
1770  *      ext4_writepage() -> kmalloc() -> __alloc_pages() -> page_launder() ->
1771  *              ext4_writepage()
1772  *
1773  * But since we don't do any block allocation we should not deadlock.
1774  * Page also have the dirty flag cleared so we don't get recurive page_lock.
1775  */
1776 static int ext4_writepage(struct page *page,
1777                           struct writeback_control *wbc)
1778 {
1779         int ret = 0, commit_write = 0;
1780         loff_t size;
1781         unsigned int len;
1782         struct buffer_head *page_bufs = NULL;
1783         struct inode *inode = page->mapping->host;
1784
1785         trace_ext4_writepage(page);
1786         size = i_size_read(inode);
1787         if (page->index == size >> PAGE_CACHE_SHIFT)
1788                 len = size & ~PAGE_CACHE_MASK;
1789         else
1790                 len = PAGE_CACHE_SIZE;
1791
1792         /*
1793          * If the page does not have buffers (for whatever reason),
1794          * try to create them using __block_write_begin.  If this
1795          * fails, redirty the page and move on.
1796          */
1797         if (!page_has_buffers(page)) {
1798                 if (__block_write_begin(page, 0, len,
1799                                         noalloc_get_block_write)) {
1800                 redirty_page:
1801                         redirty_page_for_writepage(wbc, page);
1802                         unlock_page(page);
1803                         return 0;
1804                 }
1805                 commit_write = 1;
1806         }
1807         page_bufs = page_buffers(page);
1808         if (walk_page_buffers(NULL, page_bufs, 0, len, NULL,
1809                               ext4_bh_delay_or_unwritten)) {
1810                 /*
1811                  * We don't want to do block allocation, so redirty
1812                  * the page and return.  We may reach here when we do
1813                  * a journal commit via journal_submit_inode_data_buffers.
1814                  * We can also reach here via shrink_page_list
1815                  */
1816                 goto redirty_page;
1817         }
1818         if (commit_write)
1819                 /* now mark the buffer_heads as dirty and uptodate */
1820                 block_commit_write(page, 0, len);
1821
1822         if (PageChecked(page) && ext4_should_journal_data(inode))
1823                 /*
1824                  * It's mmapped pagecache.  Add buffers and journal it.  There
1825                  * doesn't seem much point in redirtying the page here.
1826                  */
1827                 return __ext4_journalled_writepage(page, len);
1828
1829         if (buffer_uninit(page_bufs)) {
1830                 ext4_set_bh_endio(page_bufs, inode);
1831                 ret = block_write_full_page_endio(page, noalloc_get_block_write,
1832                                             wbc, ext4_end_io_buffer_write);
1833         } else
1834                 ret = block_write_full_page(page, noalloc_get_block_write,
1835                                             wbc);
1836
1837         return ret;
1838 }
1839
1840 /*
1841  * This is called via ext4_da_writepages() to
1842  * calculate the total number of credits to reserve to fit
1843  * a single extent allocation into a single transaction,
1844  * ext4_da_writpeages() will loop calling this before
1845  * the block allocation.
1846  */
1847
1848 static int ext4_da_writepages_trans_blocks(struct inode *inode)
1849 {
1850         int max_blocks = EXT4_I(inode)->i_reserved_data_blocks;
1851
1852         /*
1853          * With non-extent format the journal credit needed to
1854          * insert nrblocks contiguous block is dependent on
1855          * number of contiguous block. So we will limit
1856          * number of contiguous block to a sane value
1857          */
1858         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)) &&
1859             (max_blocks > EXT4_MAX_TRANS_DATA))
1860                 max_blocks = EXT4_MAX_TRANS_DATA;
1861
1862         return ext4_chunk_trans_blocks(inode, max_blocks);
1863 }
1864
1865 /*
1866  * write_cache_pages_da - walk the list of dirty pages of the given
1867  * address space and accumulate pages that need writing, and call
1868  * mpage_da_map_and_submit to map a single contiguous memory region
1869  * and then write them.
1870  */
1871 static int write_cache_pages_da(struct address_space *mapping,
1872                                 struct writeback_control *wbc,
1873                                 struct mpage_da_data *mpd,
1874                                 pgoff_t *done_index)
1875 {
1876         struct buffer_head      *bh, *head;
1877         struct inode            *inode = mapping->host;
1878         struct pagevec          pvec;
1879         unsigned int            nr_pages;
1880         sector_t                logical;
1881         pgoff_t                 index, end;
1882         long                    nr_to_write = wbc->nr_to_write;
1883         int                     i, tag, ret = 0;
1884
1885         memset(mpd, 0, sizeof(struct mpage_da_data));
1886         mpd->wbc = wbc;
1887         mpd->inode = inode;
1888         pagevec_init(&pvec, 0);
1889         index = wbc->range_start >> PAGE_CACHE_SHIFT;
1890         end = wbc->range_end >> PAGE_CACHE_SHIFT;
1891
1892         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
1893                 tag = PAGECACHE_TAG_TOWRITE;
1894         else
1895                 tag = PAGECACHE_TAG_DIRTY;
1896
1897         *done_index = index;
1898         while (index <= end) {
1899                 nr_pages = pagevec_lookup_tag(&pvec, mapping, &index, tag,
1900                               min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
1901                 if (nr_pages == 0)
1902                         return 0;
1903
1904                 for (i = 0; i < nr_pages; i++) {
1905                         struct page *page = pvec.pages[i];
1906
1907                         /*
1908                          * At this point, the page may be truncated or
1909                          * invalidated (changing page->mapping to NULL), or
1910                          * even swizzled back from swapper_space to tmpfs file
1911                          * mapping. However, page->index will not change
1912                          * because we have a reference on the page.
1913                          */
1914                         if (page->index > end)
1915                                 goto out;
1916
1917                         *done_index = page->index + 1;
1918
1919                         /*
1920                          * If we can't merge this page, and we have
1921                          * accumulated an contiguous region, write it
1922                          */
1923                         if ((mpd->next_page != page->index) &&
1924                             (mpd->next_page != mpd->first_page)) {
1925                                 mpage_da_map_and_submit(mpd);
1926                                 goto ret_extent_tail;
1927                         }
1928
1929                         lock_page(page);
1930
1931                         /*
1932                          * If the page is no longer dirty, or its
1933                          * mapping no longer corresponds to inode we
1934                          * are writing (which means it has been
1935                          * truncated or invalidated), or the page is
1936                          * already under writeback and we are not
1937                          * doing a data integrity writeback, skip the page
1938                          */
1939                         if (!PageDirty(page) ||
1940                             (PageWriteback(page) &&
1941                              (wbc->sync_mode == WB_SYNC_NONE)) ||
1942                             unlikely(page->mapping != mapping)) {
1943                                 unlock_page(page);
1944                                 continue;
1945                         }
1946
1947                         wait_on_page_writeback(page);
1948                         BUG_ON(PageWriteback(page));
1949
1950                         if (mpd->next_page != page->index)
1951                                 mpd->first_page = page->index;
1952                         mpd->next_page = page->index + 1;
1953                         logical = (sector_t) page->index <<
1954                                 (PAGE_CACHE_SHIFT - inode->i_blkbits);
1955
1956                         if (!page_has_buffers(page)) {
1957                                 mpage_add_bh_to_extent(mpd, logical,
1958                                                        PAGE_CACHE_SIZE,
1959                                                        (1 << BH_Dirty) | (1 << BH_Uptodate));
1960                                 if (mpd->io_done)
1961                                         goto ret_extent_tail;
1962                         } else {
1963                                 /*
1964                                  * Page with regular buffer heads,
1965                                  * just add all dirty ones
1966                                  */
1967                                 head = page_buffers(page);
1968                                 bh = head;
1969                                 do {
1970                                         BUG_ON(buffer_locked(bh));
1971                                         /*
1972                                          * We need to try to allocate
1973                                          * unmapped blocks in the same page.
1974                                          * Otherwise we won't make progress
1975                                          * with the page in ext4_writepage
1976                                          */
1977                                         if (ext4_bh_delay_or_unwritten(NULL, bh)) {
1978                                                 mpage_add_bh_to_extent(mpd, logical,
1979                                                                        bh->b_size,
1980                                                                        bh->b_state);
1981                                                 if (mpd->io_done)
1982                                                         goto ret_extent_tail;
1983                                         } else if (buffer_dirty(bh) && (buffer_mapped(bh))) {
1984                                                 /*
1985                                                  * mapped dirty buffer. We need
1986                                                  * to update the b_state
1987                                                  * because we look at b_state
1988                                                  * in mpage_da_map_blocks.  We
1989                                                  * don't update b_size because
1990                                                  * if we find an unmapped
1991                                                  * buffer_head later we need to
1992                                                  * use the b_state flag of that
1993                                                  * buffer_head.
1994                                                  */
1995                                                 if (mpd->b_size == 0)
1996                                                         mpd->b_state = bh->b_state & BH_FLAGS;
1997                                         }
1998                                         logical++;
1999                                 } while ((bh = bh->b_this_page) != head);
2000                         }
2001
2002                         if (nr_to_write > 0) {
2003                                 nr_to_write--;
2004                                 if (nr_to_write == 0 &&
2005                                     wbc->sync_mode == WB_SYNC_NONE)
2006                                         /*
2007                                          * We stop writing back only if we are
2008                                          * not doing integrity sync. In case of
2009                                          * integrity sync we have to keep going
2010                                          * because someone may be concurrently
2011                                          * dirtying pages, and we might have
2012                                          * synced a lot of newly appeared dirty
2013                                          * pages, but have not synced all of the
2014                                          * old dirty pages.
2015                                          */
2016                                         goto out;
2017                         }
2018                 }
2019                 pagevec_release(&pvec);
2020                 cond_resched();
2021         }
2022         return 0;
2023 ret_extent_tail:
2024         ret = MPAGE_DA_EXTENT_TAIL;
2025 out:
2026         pagevec_release(&pvec);
2027         cond_resched();
2028         return ret;
2029 }
2030
2031
2032 static int ext4_da_writepages(struct address_space *mapping,
2033                               struct writeback_control *wbc)
2034 {
2035         pgoff_t index;
2036         int range_whole = 0;
2037         handle_t *handle = NULL;
2038         struct mpage_da_data mpd;
2039         struct inode *inode = mapping->host;
2040         int pages_written = 0;
2041         unsigned int max_pages;
2042         int range_cyclic, cycled = 1, io_done = 0;
2043         int needed_blocks, ret = 0;
2044         long desired_nr_to_write, nr_to_writebump = 0;
2045         loff_t range_start = wbc->range_start;
2046         struct ext4_sb_info *sbi = EXT4_SB(mapping->host->i_sb);
2047         pgoff_t done_index = 0;
2048         pgoff_t end;
2049
2050         trace_ext4_da_writepages(inode, wbc);
2051
2052         /*
2053          * No pages to write? This is mainly a kludge to avoid starting
2054          * a transaction for special inodes like journal inode on last iput()
2055          * because that could violate lock ordering on umount
2056          */
2057         if (!mapping->nrpages || !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
2058                 return 0;
2059
2060         /*
2061          * If the filesystem has aborted, it is read-only, so return
2062          * right away instead of dumping stack traces later on that
2063          * will obscure the real source of the problem.  We test
2064          * EXT4_MF_FS_ABORTED instead of sb->s_flag's MS_RDONLY because
2065          * the latter could be true if the filesystem is mounted
2066          * read-only, and in that case, ext4_da_writepages should
2067          * *never* be called, so if that ever happens, we would want
2068          * the stack trace.
2069          */
2070         if (unlikely(sbi->s_mount_flags & EXT4_MF_FS_ABORTED))
2071                 return -EROFS;
2072
2073         if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2074                 range_whole = 1;
2075
2076         range_cyclic = wbc->range_cyclic;
2077         if (wbc->range_cyclic) {
2078                 index = mapping->writeback_index;
2079                 if (index)
2080                         cycled = 0;
2081                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2082                 wbc->range_end  = LLONG_MAX;
2083                 wbc->range_cyclic = 0;
2084                 end = -1;
2085         } else {
2086                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
2087                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
2088         }
2089
2090         /*
2091          * This works around two forms of stupidity.  The first is in
2092          * the writeback code, which caps the maximum number of pages
2093          * written to be 1024 pages.  This is wrong on multiple
2094          * levels; different architectues have a different page size,
2095          * which changes the maximum amount of data which gets
2096          * written.  Secondly, 4 megabytes is way too small.  XFS
2097          * forces this value to be 16 megabytes by multiplying
2098          * nr_to_write parameter by four, and then relies on its
2099          * allocator to allocate larger extents to make them
2100          * contiguous.  Unfortunately this brings us to the second
2101          * stupidity, which is that ext4's mballoc code only allocates
2102          * at most 2048 blocks.  So we force contiguous writes up to
2103          * the number of dirty blocks in the inode, or
2104          * sbi->max_writeback_mb_bump whichever is smaller.
2105          */
2106         max_pages = sbi->s_max_writeback_mb_bump << (20 - PAGE_CACHE_SHIFT);
2107         if (!range_cyclic && range_whole) {
2108                 if (wbc->nr_to_write == LONG_MAX)
2109                         desired_nr_to_write = wbc->nr_to_write;
2110                 else
2111                         desired_nr_to_write = wbc->nr_to_write * 8;
2112         } else
2113                 desired_nr_to_write = ext4_num_dirty_pages(inode, index,
2114                                                            max_pages);
2115         if (desired_nr_to_write > max_pages)
2116                 desired_nr_to_write = max_pages;
2117
2118         if (wbc->nr_to_write < desired_nr_to_write) {
2119                 nr_to_writebump = desired_nr_to_write - wbc->nr_to_write;
2120                 wbc->nr_to_write = desired_nr_to_write;
2121         }
2122
2123 retry:
2124         if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
2125                 tag_pages_for_writeback(mapping, index, end);
2126
2127         while (!ret && wbc->nr_to_write > 0) {
2128
2129                 /*
2130                  * we  insert one extent at a time. So we need
2131                  * credit needed for single extent allocation.
2132                  * journalled mode is currently not supported
2133                  * by delalloc
2134                  */
2135                 BUG_ON(ext4_should_journal_data(inode));
2136                 needed_blocks = ext4_da_writepages_trans_blocks(inode);
2137
2138                 /* start a new transaction*/
2139                 handle = ext4_journal_start(inode, needed_blocks);
2140                 if (IS_ERR(handle)) {
2141                         ret = PTR_ERR(handle);
2142                         ext4_msg(inode->i_sb, KERN_CRIT, "%s: jbd2_start: "
2143                                "%ld pages, ino %lu; err %d", __func__,
2144                                 wbc->nr_to_write, inode->i_ino, ret);
2145                         goto out_writepages;
2146                 }
2147
2148                 /*
2149                  * Now call write_cache_pages_da() to find the next
2150                  * contiguous region of logical blocks that need
2151                  * blocks to be allocated by ext4 and submit them.
2152                  */
2153                 ret = write_cache_pages_da(mapping, wbc, &mpd, &done_index);
2154                 /*
2155                  * If we have a contiguous extent of pages and we
2156                  * haven't done the I/O yet, map the blocks and submit
2157                  * them for I/O.
2158                  */
2159                 if (!mpd.io_done && mpd.next_page != mpd.first_page) {
2160                         mpage_da_map_and_submit(&mpd);
2161                         ret = MPAGE_DA_EXTENT_TAIL;
2162                 }
2163                 trace_ext4_da_write_pages(inode, &mpd);
2164                 wbc->nr_to_write -= mpd.pages_written;
2165
2166                 ext4_journal_stop(handle);
2167
2168                 if ((mpd.retval == -ENOSPC) && sbi->s_journal) {
2169                         /* commit the transaction which would
2170                          * free blocks released in the transaction
2171                          * and try again
2172                          */
2173                         jbd2_journal_force_commit_nested(sbi->s_journal);
2174                         ret = 0;
2175                 } else if (ret == MPAGE_DA_EXTENT_TAIL) {
2176                         /*
2177                          * got one extent now try with
2178                          * rest of the pages
2179                          */
2180                         pages_written += mpd.pages_written;
2181                         ret = 0;
2182                         io_done = 1;
2183                 } else if (wbc->nr_to_write)
2184                         /*
2185                          * There is no more writeout needed
2186                          * or we requested for a noblocking writeout
2187                          * and we found the device congested
2188                          */
2189                         break;
2190         }
2191         if (!io_done && !cycled) {
2192                 cycled = 1;
2193                 index = 0;
2194                 wbc->range_start = index << PAGE_CACHE_SHIFT;
2195                 wbc->range_end  = mapping->writeback_index - 1;
2196                 goto retry;
2197         }
2198
2199         /* Update index */
2200         wbc->range_cyclic = range_cyclic;
2201         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2202                 /*
2203                  * set the writeback_index so that range_cyclic
2204                  * mode will write it back later
2205                  */
2206                 mapping->writeback_index = done_index;
2207
2208 out_writepages:
2209         wbc->nr_to_write -= nr_to_writebump;
2210         wbc->range_start = range_start;
2211         trace_ext4_da_writepages_result(inode, wbc, ret, pages_written);
2212         return ret;
2213 }
2214
2215 #define FALL_BACK_TO_NONDELALLOC 1
2216 static int ext4_nonda_switch(struct super_block *sb)
2217 {
2218         s64 free_blocks, dirty_blocks;
2219         struct ext4_sb_info *sbi = EXT4_SB(sb);
2220
2221         /*
2222          * switch to non delalloc mode if we are running low
2223          * on free block. The free block accounting via percpu
2224          * counters can get slightly wrong with percpu_counter_batch getting
2225          * accumulated on each CPU without updating global counters
2226          * Delalloc need an accurate free block accounting. So switch
2227          * to non delalloc when we are near to error range.
2228          */
2229         free_blocks  = percpu_counter_read_positive(&sbi->s_freeblocks_counter);
2230         dirty_blocks = percpu_counter_read_positive(&sbi->s_dirtyblocks_counter);
2231         if (2 * free_blocks < 3 * dirty_blocks ||
2232                 free_blocks < (dirty_blocks + EXT4_FREEBLOCKS_WATERMARK)) {
2233                 /*
2234                  * free block count is less than 150% of dirty blocks
2235                  * or free blocks is less than watermark
2236                  */
2237                 return 1;
2238         }
2239         /*
2240          * Even if we don't switch but are nearing capacity,
2241          * start pushing delalloc when 1/2 of free blocks are dirty.
2242          */
2243         if (free_blocks < 2 * dirty_blocks)
2244                 writeback_inodes_sb_if_idle(sb);
2245
2246         return 0;
2247 }
2248
2249 static int ext4_da_write_begin(struct file *file, struct address_space *mapping,
2250                                loff_t pos, unsigned len, unsigned flags,
2251                                struct page **pagep, void **fsdata)
2252 {
2253         int ret, retries = 0;
2254         struct page *page;
2255         pgoff_t index;
2256         struct inode *inode = mapping->host;
2257         handle_t *handle;
2258
2259         index = pos >> PAGE_CACHE_SHIFT;
2260
2261         if (ext4_nonda_switch(inode->i_sb)) {
2262                 *fsdata = (void *)FALL_BACK_TO_NONDELALLOC;
2263                 return ext4_write_begin(file, mapping, pos,
2264                                         len, flags, pagep, fsdata);
2265         }
2266         *fsdata = (void *)0;
2267         trace_ext4_da_write_begin(inode, pos, len, flags);
2268 retry:
2269         /*
2270          * With delayed allocation, we don't log the i_disksize update
2271          * if there is delayed block allocation. But we still need
2272          * to journalling the i_disksize update if writes to the end
2273          * of file which has an already mapped buffer.
2274          */
2275         handle = ext4_journal_start(inode, 1);
2276         if (IS_ERR(handle)) {
2277                 ret = PTR_ERR(handle);
2278                 goto out;
2279         }
2280         /* We cannot recurse into the filesystem as the transaction is already
2281          * started */
2282         flags |= AOP_FLAG_NOFS;
2283
2284         page = grab_cache_page_write_begin(mapping, index, flags);
2285         if (!page) {
2286                 ext4_journal_stop(handle);
2287                 ret = -ENOMEM;
2288                 goto out;
2289         }
2290         *pagep = page;
2291
2292         ret = __block_write_begin(page, pos, len, ext4_da_get_block_prep);
2293         if (ret < 0) {
2294                 unlock_page(page);
2295                 ext4_journal_stop(handle);
2296                 page_cache_release(page);
2297                 /*
2298                  * block_write_begin may have instantiated a few blocks
2299                  * outside i_size.  Trim these off again. Don't need
2300                  * i_size_read because we hold i_mutex.
2301                  */
2302                 if (pos + len > inode->i_size)
2303                         ext4_truncate_failed_write(inode);
2304         }
2305
2306         if (ret == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
2307                 goto retry;
2308 out:
2309         return ret;
2310 }
2311
2312 /*
2313  * Check if we should update i_disksize
2314  * when write to the end of file but not require block allocation
2315  */
2316 static int ext4_da_should_update_i_disksize(struct page *page,
2317                                             unsigned long offset)
2318 {
2319         struct buffer_head *bh;
2320         struct inode *inode = page->mapping->host;
2321         unsigned int idx;
2322         int i;
2323
2324         bh = page_buffers(page);
2325         idx = offset >> inode->i_blkbits;
2326
2327         for (i = 0; i < idx; i++)
2328                 bh = bh->b_this_page;
2329
2330         if (!buffer_mapped(bh) || (buffer_delay(bh)) || buffer_unwritten(bh))
2331                 return 0;
2332         return 1;
2333 }
2334
2335 static int ext4_da_write_end(struct file *file,
2336                              struct address_space *mapping,
2337                              loff_t pos, unsigned len, unsigned copied,
2338                              struct page *page, void *fsdata)
2339 {
2340         struct inode *inode = mapping->host;
2341         int ret = 0, ret2;
2342         handle_t *handle = ext4_journal_current_handle();
2343         loff_t new_i_size;
2344         unsigned long start, end;
2345         int write_mode = (int)(unsigned long)fsdata;
2346
2347         if (write_mode == FALL_BACK_TO_NONDELALLOC) {
2348                 if (ext4_should_order_data(inode)) {
2349                         return ext4_ordered_write_end(file, mapping, pos,
2350                                         len, copied, page, fsdata);
2351                 } else if (ext4_should_writeback_data(inode)) {
2352                         return ext4_writeback_write_end(file, mapping, pos,
2353                                         len, copied, page, fsdata);
2354                 } else {
2355                         BUG();
2356                 }
2357         }
2358
2359         trace_ext4_da_write_end(inode, pos, len, copied);
2360         start = pos & (PAGE_CACHE_SIZE - 1);
2361         end = start + copied - 1;
2362
2363         /*
2364          * generic_write_end() will run mark_inode_dirty() if i_size
2365          * changes.  So let's piggyback the i_disksize mark_inode_dirty
2366          * into that.
2367          */
2368
2369         new_i_size = pos + copied;
2370         if (new_i_size > EXT4_I(inode)->i_disksize) {
2371                 if (ext4_da_should_update_i_disksize(page, end)) {
2372                         down_write(&EXT4_I(inode)->i_data_sem);
2373                         if (new_i_size > EXT4_I(inode)->i_disksize) {
2374                                 /*
2375                                  * Updating i_disksize when extending file
2376                                  * without needing block allocation
2377                                  */
2378                                 if (ext4_should_order_data(inode))
2379                                         ret = ext4_jbd2_file_inode(handle,
2380                                                                    inode);
2381
2382                                 EXT4_I(inode)->i_disksize = new_i_size;
2383                         }
2384                         up_write(&EXT4_I(inode)->i_data_sem);
2385                         /* We need to mark inode dirty even if
2386                          * new_i_size is less that inode->i_size
2387                          * bu greater than i_disksize.(hint delalloc)
2388                          */
2389                         ext4_mark_inode_dirty(handle, inode);
2390                 }
2391         }
2392         ret2 = generic_write_end(file, mapping, pos, len, copied,
2393                                                         page, fsdata);
2394         copied = ret2;
2395         if (ret2 < 0)
2396                 ret = ret2;
2397         ret2 = ext4_journal_stop(handle);
2398         if (!ret)
2399                 ret = ret2;
2400
2401         return ret ? ret : copied;
2402 }
2403
2404 static void ext4_da_invalidatepage(struct page *page, unsigned long offset)
2405 {
2406         /*
2407          * Drop reserved blocks
2408          */
2409         BUG_ON(!PageLocked(page));
2410         if (!page_has_buffers(page))
2411                 goto out;
2412
2413         ext4_da_page_release_reservation(page, offset);
2414
2415 out:
2416         ext4_invalidatepage(page, offset);
2417
2418         return;
2419 }
2420
2421 /*
2422  * Force all delayed allocation blocks to be allocated for a given inode.
2423  */
2424 int ext4_alloc_da_blocks(struct inode *inode)
2425 {
2426         trace_ext4_alloc_da_blocks(inode);
2427
2428         if (!EXT4_I(inode)->i_reserved_data_blocks &&
2429             !EXT4_I(inode)->i_reserved_meta_blocks)
2430                 return 0;
2431
2432         /*
2433          * We do something simple for now.  The filemap_flush() will
2434          * also start triggering a write of the data blocks, which is
2435          * not strictly speaking necessary (and for users of
2436          * laptop_mode, not even desirable).  However, to do otherwise
2437          * would require replicating code paths in:
2438          *
2439          * ext4_da_writepages() ->
2440          *    write_cache_pages() ---> (via passed in callback function)
2441          *        __mpage_da_writepage() -->
2442          *           mpage_add_bh_to_extent()
2443          *           mpage_da_map_blocks()
2444          *
2445          * The problem is that write_cache_pages(), located in
2446          * mm/page-writeback.c, marks pages clean in preparation for
2447          * doing I/O, which is not desirable if we're not planning on
2448          * doing I/O at all.
2449          *
2450          * We could call write_cache_pages(), and then redirty all of
2451          * the pages by calling redirty_page_for_writepage() but that
2452          * would be ugly in the extreme.  So instead we would need to
2453          * replicate parts of the code in the above functions,
2454          * simplifying them because we wouldn't actually intend to
2455          * write out the pages, but rather only collect contiguous
2456          * logical block extents, call the multi-block allocator, and
2457          * then update the buffer heads with the block allocations.
2458          *
2459          * For now, though, we'll cheat by calling filemap_flush(),
2460          * which will map the blocks, and start the I/O, but not
2461          * actually wait for the I/O to complete.
2462          */
2463         return filemap_flush(inode->i_mapping);
2464 }
2465
2466 /*
2467  * bmap() is special.  It gets used by applications such as lilo and by
2468  * the swapper to find the on-disk block of a specific piece of data.
2469  *
2470  * Naturally, this is dangerous if the block concerned is still in the
2471  * journal.  If somebody makes a swapfile on an ext4 data-journaling
2472  * filesystem and enables swap, then they may get a nasty shock when the
2473  * data getting swapped to that swapfile suddenly gets overwritten by
2474  * the original zero's written out previously to the journal and
2475  * awaiting writeback in the kernel's buffer cache.
2476  *
2477  * So, if we see any bmap calls here on a modified, data-journaled file,
2478  * take extra steps to flush any blocks which might be in the cache.
2479  */
2480 static sector_t ext4_bmap(struct address_space *mapping, sector_t block)
2481 {
2482         struct inode *inode = mapping->host;
2483         journal_t *journal;
2484         int err;
2485
2486         if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY) &&
2487                         test_opt(inode->i_sb, DELALLOC)) {
2488                 /*
2489                  * With delalloc we want to sync the file
2490                  * so that we can make sure we allocate
2491                  * blocks for file
2492                  */
2493                 filemap_write_and_wait(mapping);
2494         }
2495
2496         if (EXT4_JOURNAL(inode) &&
2497             ext4_test_inode_state(inode, EXT4_STATE_JDATA)) {
2498                 /*
2499                  * This is a REALLY heavyweight approach, but the use of
2500                  * bmap on dirty files is expected to be extremely rare:
2501                  * only if we run lilo or swapon on a freshly made file
2502                  * do we expect this to happen.
2503                  *
2504                  * (bmap requires CAP_SYS_RAWIO so this does not
2505                  * represent an unprivileged user DOS attack --- we'd be
2506                  * in trouble if mortal users could trigger this path at
2507                  * will.)
2508                  *
2509                  * NB. EXT4_STATE_JDATA is not set on files other than
2510                  * regular files.  If somebody wants to bmap a directory
2511                  * or symlink and gets confused because the buffer
2512                  * hasn't yet been flushed to disk, they deserve
2513                  * everything they get.
2514                  */
2515
2516                 ext4_clear_inode_state(inode, EXT4_STATE_JDATA);
2517                 journal = EXT4_JOURNAL(inode);
2518                 jbd2_journal_lock_updates(journal);
2519                 err = jbd2_journal_flush(journal);
2520                 jbd2_journal_unlock_updates(journal);
2521
2522                 if (err)
2523                         return 0;
2524         }
2525
2526         return generic_block_bmap(mapping, block, ext4_get_block);
2527 }
2528
2529 static int ext4_readpage(struct file *file, struct page *page)
2530 {
2531         trace_ext4_readpage(page);
2532         return mpage_readpage(page, ext4_get_block);
2533 }
2534
2535 static int
2536 ext4_readpages(struct file *file, struct address_space *mapping,
2537                 struct list_head *pages, unsigned nr_pages)
2538 {
2539         return mpage_readpages(mapping, pages, nr_pages, ext4_get_block);
2540 }
2541
2542 static void ext4_invalidatepage_free_endio(struct page *page, unsigned long offset)
2543 {
2544         struct buffer_head *head, *bh;
2545         unsigned int curr_off = 0;
2546
2547         if (!page_has_buffers(page))
2548                 return;
2549         head = bh = page_buffers(page);
2550         do {
2551                 if (offset <= curr_off && test_clear_buffer_uninit(bh)
2552                                         && bh->b_private) {
2553                         ext4_free_io_end(bh->b_private);
2554                         bh->b_private = NULL;
2555                         bh->b_end_io = NULL;
2556                 }
2557                 curr_off = curr_off + bh->b_size;
2558                 bh = bh->b_this_page;
2559         } while (bh != head);
2560 }
2561
2562 static void ext4_invalidatepage(struct page *page, unsigned long offset)
2563 {
2564         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2565
2566         trace_ext4_invalidatepage(page, offset);
2567
2568         /*
2569          * free any io_end structure allocated for buffers to be discarded
2570          */
2571         if (ext4_should_dioread_nolock(page->mapping->host))
2572                 ext4_invalidatepage_free_endio(page, offset);
2573         /*
2574          * If it's a full truncate we just forget about the pending dirtying
2575          */
2576         if (offset == 0)
2577                 ClearPageChecked(page);
2578
2579         if (journal)
2580                 jbd2_journal_invalidatepage(journal, page, offset);
2581         else
2582                 block_invalidatepage(page, offset);
2583 }
2584
2585 static int ext4_releasepage(struct page *page, gfp_t wait)
2586 {
2587         journal_t *journal = EXT4_JOURNAL(page->mapping->host);
2588
2589         trace_ext4_releasepage(page);
2590
2591         WARN_ON(PageChecked(page));
2592         if (!page_has_buffers(page))
2593                 return 0;
2594         if (journal)
2595                 return jbd2_journal_try_to_free_buffers(journal, page, wait);
2596         else
2597                 return try_to_free_buffers(page);
2598 }
2599
2600 /*
2601  * ext4_get_block used when preparing for a DIO write or buffer write.
2602  * We allocate an uinitialized extent if blocks haven't been allocated.
2603  * The extent will be converted to initialized after the IO is complete.
2604  */
2605 static int ext4_get_block_write(struct inode *inode, sector_t iblock,
2606                    struct buffer_head *bh_result, int create)
2607 {
2608         ext4_debug("ext4_get_block_write: inode %lu, create flag %d\n",
2609                    inode->i_ino, create);
2610         return _ext4_get_block(inode, iblock, bh_result,
2611                                EXT4_GET_BLOCKS_IO_CREATE_EXT);
2612 }
2613
2614 static void ext4_end_io_dio(struct kiocb *iocb, loff_t offset,
2615                             ssize_t size, void *private, int ret,
2616                             bool is_async)
2617 {
2618         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
2619         ext4_io_end_t *io_end = iocb->private;
2620         struct workqueue_struct *wq;
2621         unsigned long flags;
2622         struct ext4_inode_info *ei;
2623
2624         /* if not async direct IO or dio with 0 bytes write, just return */
2625         if (!io_end || !size)
2626                 goto out;
2627
2628         ext_debug("ext4_end_io_dio(): io_end 0x%p"
2629                   "for inode %lu, iocb 0x%p, offset %llu, size %llu\n",
2630                   iocb->private, io_end->inode->i_ino, iocb, offset,
2631                   size);
2632
2633         /* if not aio dio with unwritten extents, just free io and return */
2634         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2635                 ext4_free_io_end(io_end);
2636                 iocb->private = NULL;
2637 out:
2638                 if (is_async)
2639                         aio_complete(iocb, ret, 0);
2640                 inode_dio_done(inode);
2641                 return;
2642         }
2643
2644         io_end->offset = offset;
2645         io_end->size = size;
2646         if (is_async) {
2647                 io_end->iocb = iocb;
2648                 io_end->result = ret;
2649         }
2650         wq = EXT4_SB(io_end->inode->i_sb)->dio_unwritten_wq;
2651
2652         /* Add the io_end to per-inode completed aio dio list*/
2653         ei = EXT4_I(io_end->inode);
2654         spin_lock_irqsave(&ei->i_completed_io_lock, flags);
2655         list_add_tail(&io_end->list, &ei->i_completed_io_list);
2656         spin_unlock_irqrestore(&ei->i_completed_io_lock, flags);
2657
2658         /* queue the work to convert unwritten extents to written */
2659         queue_work(wq, &io_end->work);
2660         iocb->private = NULL;
2661
2662         /* XXX: probably should move into the real I/O completion handler */
2663         inode_dio_done(inode);
2664 }
2665
2666 static void ext4_end_io_buffer_write(struct buffer_head *bh, int uptodate)
2667 {
2668         ext4_io_end_t *io_end = bh->b_private;
2669         struct workqueue_struct *wq;
2670         struct inode *inode;
2671         unsigned long flags;
2672
2673         if (!test_clear_buffer_uninit(bh) || !io_end)
2674                 goto out;
2675
2676         if (!(io_end->inode->i_sb->s_flags & MS_ACTIVE)) {
2677                 printk("sb umounted, discard end_io request for inode %lu\n",
2678                         io_end->inode->i_ino);
2679                 ext4_free_io_end(io_end);
2680                 goto out;
2681         }
2682
2683         /*
2684          * It may be over-defensive here to check EXT4_IO_END_UNWRITTEN now,
2685          * but being more careful is always safe for the future change.
2686          */
2687         inode = io_end->inode;
2688         if (!(io_end->flag & EXT4_IO_END_UNWRITTEN)) {
2689                 io_end->flag |= EXT4_IO_END_UNWRITTEN;
2690                 atomic_inc(&EXT4_I(inode)->i_aiodio_unwritten);
2691         }
2692
2693         /* Add the io_end to per-inode completed io list*/
2694         spin_lock_irqsave(&EXT4_I(inode)->i_completed_io_lock, flags);
2695         list_add_tail(&io_end->list, &EXT4_I(inode)->i_completed_io_list);
2696         spin_unlock_irqrestore(&EXT4_I(inode)->i_completed_io_lock, flags);
2697
2698         wq = EXT4_SB(inode->i_sb)->dio_unwritten_wq;
2699         /* queue the work to convert unwritten extents to written */
2700         queue_work(wq, &io_end->work);
2701 out:
2702         bh->b_private = NULL;
2703         bh->b_end_io = NULL;
2704         clear_buffer_uninit(bh);
2705         end_buffer_async_write(bh, uptodate);
2706 }
2707
2708 static int ext4_set_bh_endio(struct buffer_head *bh, struct inode *inode)
2709 {
2710         ext4_io_end_t *io_end;
2711         struct page *page = bh->b_page;
2712         loff_t offset = (sector_t)page->index << PAGE_CACHE_SHIFT;
2713         size_t size = bh->b_size;
2714
2715 retry:
2716         io_end = ext4_init_io_end(inode, GFP_ATOMIC);
2717         if (!io_end) {
2718                 pr_warn_ratelimited("%s: allocation fail\n", __func__);
2719                 schedule();
2720                 goto retry;
2721         }
2722         io_end->offset = offset;
2723         io_end->size = size;
2724         /*
2725          * We need to hold a reference to the page to make sure it
2726          * doesn't get evicted before ext4_end_io_work() has a chance
2727          * to convert the extent from written to unwritten.
2728          */
2729         io_end->page = page;
2730         get_page(io_end->page);
2731
2732         bh->b_private = io_end;
2733         bh->b_end_io = ext4_end_io_buffer_write;
2734         return 0;
2735 }
2736
2737 /*
2738  * For ext4 extent files, ext4 will do direct-io write to holes,
2739  * preallocated extents, and those write extend the file, no need to
2740  * fall back to buffered IO.
2741  *
2742  * For holes, we fallocate those blocks, mark them as uninitialized
2743  * If those blocks were preallocated, we mark sure they are splited, but
2744  * still keep the range to write as uninitialized.
2745  *
2746  * The unwrritten extents will be converted to written when DIO is completed.
2747  * For async direct IO, since the IO may still pending when return, we
2748  * set up an end_io call back function, which will do the conversion
2749  * when async direct IO completed.
2750  *
2751  * If the O_DIRECT write will extend the file then add this inode to the
2752  * orphan list.  So recovery will truncate it back to the original size
2753  * if the machine crashes during the write.
2754  *
2755  */
2756 static ssize_t ext4_ext_direct_IO(int rw, struct kiocb *iocb,
2757                               const struct iovec *iov, loff_t offset,
2758                               unsigned long nr_segs)
2759 {
2760         struct file *file = iocb->ki_filp;
2761         struct inode *inode = file->f_mapping->host;
2762         ssize_t ret;
2763         size_t count = iov_length(iov, nr_segs);
2764
2765         loff_t final_size = offset + count;
2766         if (rw == WRITE && final_size <= inode->i_size) {
2767                 /*
2768                  * We could direct write to holes and fallocate.
2769                  *
2770                  * Allocated blocks to fill the hole are marked as uninitialized
2771                  * to prevent parallel buffered read to expose the stale data
2772                  * before DIO complete the data IO.
2773                  *
2774                  * As to previously fallocated extents, ext4 get_block
2775                  * will just simply mark the buffer mapped but still
2776                  * keep the extents uninitialized.
2777                  *
2778                  * for non AIO case, we will convert those unwritten extents
2779                  * to written after return back from blockdev_direct_IO.
2780                  *
2781                  * for async DIO, the conversion needs to be defered when
2782                  * the IO is completed. The ext4 end_io callback function
2783                  * will be called to take care of the conversion work.
2784                  * Here for async case, we allocate an io_end structure to
2785                  * hook to the iocb.
2786                  */
2787                 iocb->private = NULL;
2788                 EXT4_I(inode)->cur_aio_dio = NULL;
2789                 if (!is_sync_kiocb(iocb)) {
2790                         iocb->private = ext4_init_io_end(inode, GFP_NOFS);
2791                         if (!iocb->private)
2792                                 return -ENOMEM;
2793                         /*
2794                          * we save the io structure for current async
2795                          * direct IO, so that later ext4_map_blocks()
2796                          * could flag the io structure whether there
2797                          * is a unwritten extents needs to be converted
2798                          * when IO is completed.
2799                          */
2800                         EXT4_I(inode)->cur_aio_dio = iocb->private;
2801                 }
2802
2803                 ret = __blockdev_direct_IO(rw, iocb, inode,
2804                                          inode->i_sb->s_bdev, iov,
2805                                          offset, nr_segs,
2806                                          ext4_get_block_write,
2807                                          ext4_end_io_dio,
2808                                          NULL,
2809                                          DIO_LOCKING | DIO_SKIP_HOLES);
2810                 if (iocb->private)
2811                         EXT4_I(inode)->cur_aio_dio = NULL;
2812                 /*
2813                  * The io_end structure takes a reference to the inode,
2814                  * that structure needs to be destroyed and the
2815                  * reference to the inode need to be dropped, when IO is
2816                  * complete, even with 0 byte write, or failed.
2817                  *
2818                  * In the successful AIO DIO case, the io_end structure will be
2819                  * desctroyed and the reference to the inode will be dropped
2820                  * after the end_io call back function is called.
2821                  *
2822                  * In the case there is 0 byte write, or error case, since
2823                  * VFS direct IO won't invoke the end_io call back function,
2824                  * we need to free the end_io structure here.
2825                  */
2826                 if (ret != -EIOCBQUEUED && ret <= 0 && iocb->private) {
2827                         ext4_free_io_end(iocb->private);
2828                         iocb->private = NULL;
2829                 } else if (ret > 0 && ext4_test_inode_state(inode,
2830                                                 EXT4_STATE_DIO_UNWRITTEN)) {
2831                         int err;
2832                         /*
2833                          * for non AIO case, since the IO is already
2834                          * completed, we could do the conversion right here
2835                          */
2836                         err = ext4_convert_unwritten_extents(inode,
2837                                                              offset, ret);
2838                         if (err < 0)
2839                                 ret = err;
2840                         ext4_clear_inode_state(inode, EXT4_STATE_DIO_UNWRITTEN);
2841                 }
2842                 return ret;
2843         }
2844
2845         /* for write the the end of file case, we fall back to old way */
2846         return ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2847 }
2848
2849 static ssize_t ext4_direct_IO(int rw, struct kiocb *iocb,
2850                               const struct iovec *iov, loff_t offset,
2851                               unsigned long nr_segs)
2852 {
2853         struct file *file = iocb->ki_filp;
2854         struct inode *inode = file->f_mapping->host;
2855         ssize_t ret;
2856
2857         trace_ext4_direct_IO_enter(inode, offset, iov_length(iov, nr_segs), rw);
2858         if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
2859                 ret = ext4_ext_direct_IO(rw, iocb, iov, offset, nr_segs);
2860         else
2861                 ret = ext4_ind_direct_IO(rw, iocb, iov, offset, nr_segs);
2862         trace_ext4_direct_IO_exit(inode, offset,
2863                                 iov_length(iov, nr_segs), rw, ret);
2864         return ret;
2865 }
2866
2867 /*
2868  * Pages can be marked dirty completely asynchronously from ext4's journalling
2869  * activity.  By filemap_sync_pte(), try_to_unmap_one(), etc.  We cannot do
2870  * much here because ->set_page_dirty is called under VFS locks.  The page is
2871  * not necessarily locked.
2872  *
2873  * We cannot just dirty the page and leave attached buffers clean, because the
2874  * buffers' dirty state is "definitive".  We cannot just set the buffers dirty
2875  * or jbddirty because all the journalling code will explode.
2876  *
2877  * So what we do is to mark the page "pending dirty" and next time writepage
2878  * is called, propagate that into the buffers appropriately.
2879  */
2880 static int ext4_journalled_set_page_dirty(struct page *page)
2881 {
2882         SetPageChecked(page);
2883         return __set_page_dirty_nobuffers(page);
2884 }
2885
2886 static const struct address_space_operations ext4_ordered_aops = {
2887         .readpage               = ext4_readpage,
2888         .readpages              = ext4_readpages,
2889         .writepage              = ext4_writepage,
2890         .write_begin            = ext4_write_begin,
2891         .write_end              = ext4_ordered_write_end,
2892         .bmap                   = ext4_bmap,
2893         .invalidatepage         = ext4_invalidatepage,
2894         .releasepage            = ext4_releasepage,
2895         .direct_IO              = ext4_direct_IO,
2896         .migratepage            = buffer_migrate_page,
2897         .is_partially_uptodate  = block_is_partially_uptodate,
2898         .error_remove_page      = generic_error_remove_page,
2899 };
2900
2901 static const struct address_space_operations ext4_writeback_aops = {
2902         .readpage               = ext4_readpage,
2903         .readpages              = ext4_readpages,
2904         .writepage              = ext4_writepage,
2905         .write_begin            = ext4_write_begin,
2906         .write_end              = ext4_writeback_write_end,
2907         .bmap                   = ext4_bmap,
2908         .invalidatepage         = ext4_invalidatepage,
2909         .releasepage            = ext4_releasepage,
2910         .direct_IO              = ext4_direct_IO,
2911         .migratepage            = buffer_migrate_page,
2912         .is_partially_uptodate  = block_is_partially_uptodate,
2913         .error_remove_page      = generic_error_remove_page,
2914 };
2915
2916 static const struct address_space_operations ext4_journalled_aops = {
2917         .readpage               = ext4_readpage,
2918         .readpages              = ext4_readpages,
2919         .writepage              = ext4_writepage,
2920         .write_begin            = ext4_write_begin,
2921         .write_end              = ext4_journalled_write_end,
2922         .set_page_dirty         = ext4_journalled_set_page_dirty,
2923         .bmap                   = ext4_bmap,
2924         .invalidatepage         = ext4_invalidatepage,
2925         .releasepage            = ext4_releasepage,
2926         .is_partially_uptodate  = block_is_partially_uptodate,
2927         .error_remove_page      = generic_error_remove_page,
2928 };
2929
2930 static const struct address_space_operations ext4_da_aops = {
2931         .readpage               = ext4_readpage,
2932         .readpages              = ext4_readpages,
2933         .writepage              = ext4_writepage,
2934         .writepages             = ext4_da_writepages,
2935         .write_begin            = ext4_da_write_begin,
2936         .write_end              = ext4_da_write_end,
2937         .bmap                   = ext4_bmap,
2938         .invalidatepage         = ext4_da_invalidatepage,
2939         .releasepage            = ext4_releasepage,
2940         .direct_IO              = ext4_direct_IO,
2941         .migratepage            = buffer_migrate_page,
2942         .is_partially_uptodate  = block_is_partially_uptodate,
2943         .error_remove_page      = generic_error_remove_page,
2944 };
2945
2946 void ext4_set_aops(struct inode *inode)
2947 {
2948         if (ext4_should_order_data(inode) &&
2949                 test_opt(inode->i_sb, DELALLOC))
2950                 inode->i_mapping->a_ops = &ext4_da_aops;
2951         else if (ext4_should_order_data(inode))
2952                 inode->i_mapping->a_ops = &ext4_ordered_aops;
2953         else if (ext4_should_writeback_data(inode) &&
2954                  test_opt(inode->i_sb, DELALLOC))
2955                 inode->i_mapping->a_ops = &ext4_da_aops;
2956         else if (ext4_should_writeback_data(inode))
2957                 inode->i_mapping->a_ops = &ext4_writeback_aops;
2958         else
2959                 inode->i_mapping->a_ops = &ext4_journalled_aops;
2960 }
2961
2962 /*
2963  * ext4_block_truncate_page() zeroes out a mapping from file offset `from'
2964  * up to the end of the block which corresponds to `from'.
2965  * This required during truncate. We need to physically zero the tail end
2966  * of that block so it doesn't yield old data if the file is later grown.
2967  */
2968 int ext4_block_truncate_page(handle_t *handle,
2969                 struct address_space *mapping, loff_t from)
2970 {
2971         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2972         unsigned length;
2973         unsigned blocksize;
2974         struct inode *inode = mapping->host;
2975
2976         blocksize = inode->i_sb->s_blocksize;
2977         length = blocksize - (offset & (blocksize - 1));
2978
2979         return ext4_block_zero_page_range(handle, mapping, from, length);
2980 }
2981
2982 /*
2983  * ext4_block_zero_page_range() zeros out a mapping of length 'length'
2984  * starting from file offset 'from'.  The range to be zero'd must
2985  * be contained with in one block.  If the specified range exceeds
2986  * the end of the block it will be shortened to end of the block
2987  * that cooresponds to 'from'
2988  */
2989 int ext4_block_zero_page_range(handle_t *handle,
2990                 struct address_space *mapping, loff_t from, loff_t length)
2991 {
2992         ext4_fsblk_t index = from >> PAGE_CACHE_SHIFT;
2993         unsigned offset = from & (PAGE_CACHE_SIZE-1);
2994         unsigned blocksize, max, pos;
2995         ext4_lblk_t iblock;
2996         struct inode *inode = mapping->host;
2997         struct buffer_head *bh;
2998         struct page *page;
2999         int err = 0;
3000
3001         page = find_or_create_page(mapping, from >> PAGE_CACHE_SHIFT,
3002                                    mapping_gfp_mask(mapping) & ~__GFP_FS);
3003         if (!page)
3004                 return -EINVAL;
3005
3006         blocksize = inode->i_sb->s_blocksize;
3007         max = blocksize - (offset & (blocksize - 1));
3008
3009         /*
3010          * correct length if it does not fall between
3011          * 'from' and the end of the block
3012          */
3013         if (length > max || length < 0)
3014                 length = max;
3015
3016         iblock = index << (PAGE_CACHE_SHIFT - inode->i_sb->s_blocksize_bits);
3017
3018         if (!page_has_buffers(page))
3019                 create_empty_buffers(page, blocksize, 0);
3020
3021         /* Find the buffer that contains "offset" */
3022         bh = page_buffers(page);
3023         pos = blocksize;
3024         while (offset >= pos) {
3025                 bh = bh->b_this_page;
3026                 iblock++;
3027                 pos += blocksize;
3028         }
3029
3030         err = 0;
3031         if (buffer_freed(bh)) {
3032                 BUFFER_TRACE(bh, "freed: skip");
3033                 goto unlock;
3034         }
3035
3036         if (!buffer_mapped(bh)) {
3037                 BUFFER_TRACE(bh, "unmapped");
3038                 ext4_get_block(inode, iblock, bh, 0);
3039                 /* unmapped? It's a hole - nothing to do */
3040                 if (!buffer_mapped(bh)) {
3041                         BUFFER_TRACE(bh, "still unmapped");
3042                         goto unlock;
3043                 }
3044         }
3045
3046         /* Ok, it's mapped. Make sure it's up-to-date */
3047         if (PageUptodate(page))
3048                 set_buffer_uptodate(bh);
3049
3050         if (!buffer_uptodate(bh)) {
3051                 err = -EIO;
3052                 ll_rw_block(READ, 1, &bh);
3053                 wait_on_buffer(bh);
3054                 /* Uhhuh. Read error. Complain and punt. */
3055                 if (!buffer_uptodate(bh))
3056                         goto unlock;
3057         }
3058
3059         if (ext4_should_journal_data(inode)) {
3060                 BUFFER_TRACE(bh, "get write access");
3061                 err = ext4_journal_get_write_access(handle, bh);
3062                 if (err)
3063                         goto unlock;
3064         }
3065
3066         zero_user(page, offset, length);
3067