btrfs: use btrfs_ino to access inode number
[linux-2.6.git] / fs / btrfs / file.c
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18
19 #include <linux/fs.h>
20 #include <linux/pagemap.h>
21 #include <linux/highmem.h>
22 #include <linux/time.h>
23 #include <linux/init.h>
24 #include <linux/string.h>
25 #include <linux/backing-dev.h>
26 #include <linux/mpage.h>
27 #include <linux/falloc.h>
28 #include <linux/swap.h>
29 #include <linux/writeback.h>
30 #include <linux/statfs.h>
31 #include <linux/compat.h>
32 #include <linux/slab.h>
33 #include "ctree.h"
34 #include "disk-io.h"
35 #include "transaction.h"
36 #include "btrfs_inode.h"
37 #include "ioctl.h"
38 #include "print-tree.h"
39 #include "tree-log.h"
40 #include "locking.h"
41 #include "compat.h"
42
43 /*
44  * when auto defrag is enabled we
45  * queue up these defrag structs to remember which
46  * inodes need defragging passes
47  */
48 struct inode_defrag {
49         struct rb_node rb_node;
50         /* objectid */
51         u64 ino;
52         /*
53          * transid where the defrag was added, we search for
54          * extents newer than this
55          */
56         u64 transid;
57
58         /* root objectid */
59         u64 root;
60
61         /* last offset we were able to defrag */
62         u64 last_offset;
63
64         /* if we've wrapped around back to zero once already */
65         int cycled;
66 };
67
68 /* pop a record for an inode into the defrag tree.  The lock
69  * must be held already
70  *
71  * If you're inserting a record for an older transid than an
72  * existing record, the transid already in the tree is lowered
73  *
74  * If an existing record is found the defrag item you
75  * pass in is freed
76  */
77 static int __btrfs_add_inode_defrag(struct inode *inode,
78                                     struct inode_defrag *defrag)
79 {
80         struct btrfs_root *root = BTRFS_I(inode)->root;
81         struct inode_defrag *entry;
82         struct rb_node **p;
83         struct rb_node *parent = NULL;
84
85         p = &root->fs_info->defrag_inodes.rb_node;
86         while (*p) {
87                 parent = *p;
88                 entry = rb_entry(parent, struct inode_defrag, rb_node);
89
90                 if (defrag->ino < entry->ino)
91                         p = &parent->rb_left;
92                 else if (defrag->ino > entry->ino)
93                         p = &parent->rb_right;
94                 else {
95                         /* if we're reinserting an entry for
96                          * an old defrag run, make sure to
97                          * lower the transid of our existing record
98                          */
99                         if (defrag->transid < entry->transid)
100                                 entry->transid = defrag->transid;
101                         if (defrag->last_offset > entry->last_offset)
102                                 entry->last_offset = defrag->last_offset;
103                         goto exists;
104                 }
105         }
106         BTRFS_I(inode)->in_defrag = 1;
107         rb_link_node(&defrag->rb_node, parent, p);
108         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
109         return 0;
110
111 exists:
112         kfree(defrag);
113         return 0;
114
115 }
116
117 /*
118  * insert a defrag record for this inode if auto defrag is
119  * enabled
120  */
121 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
122                            struct inode *inode)
123 {
124         struct btrfs_root *root = BTRFS_I(inode)->root;
125         struct inode_defrag *defrag;
126         int ret = 0;
127         u64 transid;
128
129         if (!btrfs_test_opt(root, AUTO_DEFRAG))
130                 return 0;
131
132         if (root->fs_info->closing)
133                 return 0;
134
135         if (BTRFS_I(inode)->in_defrag)
136                 return 0;
137
138         if (trans)
139                 transid = trans->transid;
140         else
141                 transid = BTRFS_I(inode)->root->last_trans;
142
143         defrag = kzalloc(sizeof(*defrag), GFP_NOFS);
144         if (!defrag)
145                 return -ENOMEM;
146
147         defrag->ino = btrfs_ino(inode);
148         defrag->transid = transid;
149         defrag->root = root->root_key.objectid;
150
151         spin_lock(&root->fs_info->defrag_inodes_lock);
152         if (!BTRFS_I(inode)->in_defrag)
153                 ret = __btrfs_add_inode_defrag(inode, defrag);
154         spin_unlock(&root->fs_info->defrag_inodes_lock);
155         return ret;
156 }
157
158 /*
159  * must be called with the defrag_inodes lock held
160  */
161 struct inode_defrag *btrfs_find_defrag_inode(struct btrfs_fs_info *info, u64 ino,
162                                              struct rb_node **next)
163 {
164         struct inode_defrag *entry = NULL;
165         struct rb_node *p;
166         struct rb_node *parent = NULL;
167
168         p = info->defrag_inodes.rb_node;
169         while (p) {
170                 parent = p;
171                 entry = rb_entry(parent, struct inode_defrag, rb_node);
172
173                 if (ino < entry->ino)
174                         p = parent->rb_left;
175                 else if (ino > entry->ino)
176                         p = parent->rb_right;
177                 else
178                         return entry;
179         }
180
181         if (next) {
182                 while (parent && ino > entry->ino) {
183                         parent = rb_next(parent);
184                         entry = rb_entry(parent, struct inode_defrag, rb_node);
185                 }
186                 *next = parent;
187         }
188         return NULL;
189 }
190
191 /*
192  * run through the list of inodes in the FS that need
193  * defragging
194  */
195 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
196 {
197         struct inode_defrag *defrag;
198         struct btrfs_root *inode_root;
199         struct inode *inode;
200         struct rb_node *n;
201         struct btrfs_key key;
202         struct btrfs_ioctl_defrag_range_args range;
203         u64 first_ino = 0;
204         int num_defrag;
205         int defrag_batch = 1024;
206
207         memset(&range, 0, sizeof(range));
208         range.len = (u64)-1;
209
210         atomic_inc(&fs_info->defrag_running);
211         spin_lock(&fs_info->defrag_inodes_lock);
212         while(1) {
213                 n = NULL;
214
215                 /* find an inode to defrag */
216                 defrag = btrfs_find_defrag_inode(fs_info, first_ino, &n);
217                 if (!defrag) {
218                         if (n)
219                                 defrag = rb_entry(n, struct inode_defrag, rb_node);
220                         else if (first_ino) {
221                                 first_ino = 0;
222                                 continue;
223                         } else {
224                                 break;
225                         }
226                 }
227
228                 /* remove it from the rbtree */
229                 first_ino = defrag->ino + 1;
230                 rb_erase(&defrag->rb_node, &fs_info->defrag_inodes);
231
232                 if (fs_info->closing)
233                         goto next_free;
234
235                 spin_unlock(&fs_info->defrag_inodes_lock);
236
237                 /* get the inode */
238                 key.objectid = defrag->root;
239                 btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
240                 key.offset = (u64)-1;
241                 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
242                 if (IS_ERR(inode_root))
243                         goto next;
244
245                 key.objectid = defrag->ino;
246                 btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
247                 key.offset = 0;
248
249                 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
250                 if (IS_ERR(inode))
251                         goto next;
252
253                 /* do a chunk of defrag */
254                 BTRFS_I(inode)->in_defrag = 0;
255                 range.start = defrag->last_offset;
256                 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
257                                                defrag_batch);
258                 /*
259                  * if we filled the whole defrag batch, there
260                  * must be more work to do.  Queue this defrag
261                  * again
262                  */
263                 if (num_defrag == defrag_batch) {
264                         defrag->last_offset = range.start;
265                         __btrfs_add_inode_defrag(inode, defrag);
266                         /*
267                          * we don't want to kfree defrag, we added it back to
268                          * the rbtree
269                          */
270                         defrag = NULL;
271                 } else if (defrag->last_offset && !defrag->cycled) {
272                         /*
273                          * we didn't fill our defrag batch, but
274                          * we didn't start at zero.  Make sure we loop
275                          * around to the start of the file.
276                          */
277                         defrag->last_offset = 0;
278                         defrag->cycled = 1;
279                         __btrfs_add_inode_defrag(inode, defrag);
280                         defrag = NULL;
281                 }
282
283                 iput(inode);
284 next:
285                 spin_lock(&fs_info->defrag_inodes_lock);
286 next_free:
287                 kfree(defrag);
288         }
289         spin_unlock(&fs_info->defrag_inodes_lock);
290
291         atomic_dec(&fs_info->defrag_running);
292
293         /*
294          * during unmount, we use the transaction_wait queue to
295          * wait for the defragger to stop
296          */
297         wake_up(&fs_info->transaction_wait);
298         return 0;
299 }
300
301 /* simple helper to fault in pages and copy.  This should go away
302  * and be replaced with calls into generic code.
303  */
304 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
305                                          size_t write_bytes,
306                                          struct page **prepared_pages,
307                                          struct iov_iter *i)
308 {
309         size_t copied = 0;
310         size_t total_copied = 0;
311         int pg = 0;
312         int offset = pos & (PAGE_CACHE_SIZE - 1);
313
314         while (write_bytes > 0) {
315                 size_t count = min_t(size_t,
316                                      PAGE_CACHE_SIZE - offset, write_bytes);
317                 struct page *page = prepared_pages[pg];
318                 /*
319                  * Copy data from userspace to the current page
320                  *
321                  * Disable pagefault to avoid recursive lock since
322                  * the pages are already locked
323                  */
324                 pagefault_disable();
325                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
326                 pagefault_enable();
327
328                 /* Flush processor's dcache for this page */
329                 flush_dcache_page(page);
330
331                 /*
332                  * if we get a partial write, we can end up with
333                  * partially up to date pages.  These add
334                  * a lot of complexity, so make sure they don't
335                  * happen by forcing this copy to be retried.
336                  *
337                  * The rest of the btrfs_file_write code will fall
338                  * back to page at a time copies after we return 0.
339                  */
340                 if (!PageUptodate(page) && copied < count)
341                         copied = 0;
342
343                 iov_iter_advance(i, copied);
344                 write_bytes -= copied;
345                 total_copied += copied;
346
347                 /* Return to btrfs_file_aio_write to fault page */
348                 if (unlikely(copied == 0))
349                         break;
350
351                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
352                         offset += copied;
353                 } else {
354                         pg++;
355                         offset = 0;
356                 }
357         }
358         return total_copied;
359 }
360
361 /*
362  * unlocks pages after btrfs_file_write is done with them
363  */
364 void btrfs_drop_pages(struct page **pages, size_t num_pages)
365 {
366         size_t i;
367         for (i = 0; i < num_pages; i++) {
368                 /* page checked is some magic around finding pages that
369                  * have been modified without going through btrfs_set_page_dirty
370                  * clear it here
371                  */
372                 ClearPageChecked(pages[i]);
373                 unlock_page(pages[i]);
374                 mark_page_accessed(pages[i]);
375                 page_cache_release(pages[i]);
376         }
377 }
378
379 /*
380  * after copy_from_user, pages need to be dirtied and we need to make
381  * sure holes are created between the current EOF and the start of
382  * any next extents (if required).
383  *
384  * this also makes the decision about creating an inline extent vs
385  * doing real data extents, marking pages dirty and delalloc as required.
386  */
387 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
388                       struct page **pages, size_t num_pages,
389                       loff_t pos, size_t write_bytes,
390                       struct extent_state **cached)
391 {
392         int err = 0;
393         int i;
394         u64 num_bytes;
395         u64 start_pos;
396         u64 end_of_last_block;
397         u64 end_pos = pos + write_bytes;
398         loff_t isize = i_size_read(inode);
399
400         start_pos = pos & ~((u64)root->sectorsize - 1);
401         num_bytes = (write_bytes + pos - start_pos +
402                     root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
403
404         end_of_last_block = start_pos + num_bytes - 1;
405         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
406                                         cached);
407         if (err)
408                 return err;
409
410         for (i = 0; i < num_pages; i++) {
411                 struct page *p = pages[i];
412                 SetPageUptodate(p);
413                 ClearPageChecked(p);
414                 set_page_dirty(p);
415         }
416
417         /*
418          * we've only changed i_size in ram, and we haven't updated
419          * the disk i_size.  There is no need to log the inode
420          * at this time.
421          */
422         if (end_pos > isize)
423                 i_size_write(inode, end_pos);
424         return 0;
425 }
426
427 /*
428  * this drops all the extents in the cache that intersect the range
429  * [start, end].  Existing extents are split as required.
430  */
431 int btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
432                             int skip_pinned)
433 {
434         struct extent_map *em;
435         struct extent_map *split = NULL;
436         struct extent_map *split2 = NULL;
437         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
438         u64 len = end - start + 1;
439         int ret;
440         int testend = 1;
441         unsigned long flags;
442         int compressed = 0;
443
444         WARN_ON(end < start);
445         if (end == (u64)-1) {
446                 len = (u64)-1;
447                 testend = 0;
448         }
449         while (1) {
450                 if (!split)
451                         split = alloc_extent_map();
452                 if (!split2)
453                         split2 = alloc_extent_map();
454                 BUG_ON(!split || !split2);
455
456                 write_lock(&em_tree->lock);
457                 em = lookup_extent_mapping(em_tree, start, len);
458                 if (!em) {
459                         write_unlock(&em_tree->lock);
460                         break;
461                 }
462                 flags = em->flags;
463                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
464                         if (testend && em->start + em->len >= start + len) {
465                                 free_extent_map(em);
466                                 write_unlock(&em_tree->lock);
467                                 break;
468                         }
469                         start = em->start + em->len;
470                         if (testend)
471                                 len = start + len - (em->start + em->len);
472                         free_extent_map(em);
473                         write_unlock(&em_tree->lock);
474                         continue;
475                 }
476                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
477                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
478                 remove_extent_mapping(em_tree, em);
479
480                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
481                     em->start < start) {
482                         split->start = em->start;
483                         split->len = start - em->start;
484                         split->orig_start = em->orig_start;
485                         split->block_start = em->block_start;
486
487                         if (compressed)
488                                 split->block_len = em->block_len;
489                         else
490                                 split->block_len = split->len;
491
492                         split->bdev = em->bdev;
493                         split->flags = flags;
494                         split->compress_type = em->compress_type;
495                         ret = add_extent_mapping(em_tree, split);
496                         BUG_ON(ret);
497                         free_extent_map(split);
498                         split = split2;
499                         split2 = NULL;
500                 }
501                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
502                     testend && em->start + em->len > start + len) {
503                         u64 diff = start + len - em->start;
504
505                         split->start = start + len;
506                         split->len = em->start + em->len - (start + len);
507                         split->bdev = em->bdev;
508                         split->flags = flags;
509                         split->compress_type = em->compress_type;
510
511                         if (compressed) {
512                                 split->block_len = em->block_len;
513                                 split->block_start = em->block_start;
514                                 split->orig_start = em->orig_start;
515                         } else {
516                                 split->block_len = split->len;
517                                 split->block_start = em->block_start + diff;
518                                 split->orig_start = split->start;
519                         }
520
521                         ret = add_extent_mapping(em_tree, split);
522                         BUG_ON(ret);
523                         free_extent_map(split);
524                         split = NULL;
525                 }
526                 write_unlock(&em_tree->lock);
527
528                 /* once for us */
529                 free_extent_map(em);
530                 /* once for the tree*/
531                 free_extent_map(em);
532         }
533         if (split)
534                 free_extent_map(split);
535         if (split2)
536                 free_extent_map(split2);
537         return 0;
538 }
539
540 /*
541  * this is very complex, but the basic idea is to drop all extents
542  * in the range start - end.  hint_block is filled in with a block number
543  * that would be a good hint to the block allocator for this file.
544  *
545  * If an extent intersects the range but is not entirely inside the range
546  * it is either truncated or split.  Anything entirely inside the range
547  * is deleted from the tree.
548  */
549 int btrfs_drop_extents(struct btrfs_trans_handle *trans, struct inode *inode,
550                        u64 start, u64 end, u64 *hint_byte, int drop_cache)
551 {
552         struct btrfs_root *root = BTRFS_I(inode)->root;
553         struct extent_buffer *leaf;
554         struct btrfs_file_extent_item *fi;
555         struct btrfs_path *path;
556         struct btrfs_key key;
557         struct btrfs_key new_key;
558         u64 ino = btrfs_ino(inode);
559         u64 search_start = start;
560         u64 disk_bytenr = 0;
561         u64 num_bytes = 0;
562         u64 extent_offset = 0;
563         u64 extent_end = 0;
564         int del_nr = 0;
565         int del_slot = 0;
566         int extent_type;
567         int recow;
568         int ret;
569
570         if (drop_cache)
571                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
572
573         path = btrfs_alloc_path();
574         if (!path)
575                 return -ENOMEM;
576
577         while (1) {
578                 recow = 0;
579                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
580                                                search_start, -1);
581                 if (ret < 0)
582                         break;
583                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
584                         leaf = path->nodes[0];
585                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
586                         if (key.objectid == ino &&
587                             key.type == BTRFS_EXTENT_DATA_KEY)
588                                 path->slots[0]--;
589                 }
590                 ret = 0;
591 next_slot:
592                 leaf = path->nodes[0];
593                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
594                         BUG_ON(del_nr > 0);
595                         ret = btrfs_next_leaf(root, path);
596                         if (ret < 0)
597                                 break;
598                         if (ret > 0) {
599                                 ret = 0;
600                                 break;
601                         }
602                         leaf = path->nodes[0];
603                         recow = 1;
604                 }
605
606                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
607                 if (key.objectid > ino ||
608                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
609                         break;
610
611                 fi = btrfs_item_ptr(leaf, path->slots[0],
612                                     struct btrfs_file_extent_item);
613                 extent_type = btrfs_file_extent_type(leaf, fi);
614
615                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
616                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
617                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
618                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
619                         extent_offset = btrfs_file_extent_offset(leaf, fi);
620                         extent_end = key.offset +
621                                 btrfs_file_extent_num_bytes(leaf, fi);
622                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
623                         extent_end = key.offset +
624                                 btrfs_file_extent_inline_len(leaf, fi);
625                 } else {
626                         WARN_ON(1);
627                         extent_end = search_start;
628                 }
629
630                 if (extent_end <= search_start) {
631                         path->slots[0]++;
632                         goto next_slot;
633                 }
634
635                 search_start = max(key.offset, start);
636                 if (recow) {
637                         btrfs_release_path(path);
638                         continue;
639                 }
640
641                 /*
642                  *     | - range to drop - |
643                  *  | -------- extent -------- |
644                  */
645                 if (start > key.offset && end < extent_end) {
646                         BUG_ON(del_nr > 0);
647                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
648
649                         memcpy(&new_key, &key, sizeof(new_key));
650                         new_key.offset = start;
651                         ret = btrfs_duplicate_item(trans, root, path,
652                                                    &new_key);
653                         if (ret == -EAGAIN) {
654                                 btrfs_release_path(path);
655                                 continue;
656                         }
657                         if (ret < 0)
658                                 break;
659
660                         leaf = path->nodes[0];
661                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
662                                             struct btrfs_file_extent_item);
663                         btrfs_set_file_extent_num_bytes(leaf, fi,
664                                                         start - key.offset);
665
666                         fi = btrfs_item_ptr(leaf, path->slots[0],
667                                             struct btrfs_file_extent_item);
668
669                         extent_offset += start - key.offset;
670                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
671                         btrfs_set_file_extent_num_bytes(leaf, fi,
672                                                         extent_end - start);
673                         btrfs_mark_buffer_dirty(leaf);
674
675                         if (disk_bytenr > 0) {
676                                 ret = btrfs_inc_extent_ref(trans, root,
677                                                 disk_bytenr, num_bytes, 0,
678                                                 root->root_key.objectid,
679                                                 new_key.objectid,
680                                                 start - extent_offset);
681                                 BUG_ON(ret);
682                                 *hint_byte = disk_bytenr;
683                         }
684                         key.offset = start;
685                 }
686                 /*
687                  *  | ---- range to drop ----- |
688                  *      | -------- extent -------- |
689                  */
690                 if (start <= key.offset && end < extent_end) {
691                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
692
693                         memcpy(&new_key, &key, sizeof(new_key));
694                         new_key.offset = end;
695                         btrfs_set_item_key_safe(trans, root, path, &new_key);
696
697                         extent_offset += end - key.offset;
698                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
699                         btrfs_set_file_extent_num_bytes(leaf, fi,
700                                                         extent_end - end);
701                         btrfs_mark_buffer_dirty(leaf);
702                         if (disk_bytenr > 0) {
703                                 inode_sub_bytes(inode, end - key.offset);
704                                 *hint_byte = disk_bytenr;
705                         }
706                         break;
707                 }
708
709                 search_start = extent_end;
710                 /*
711                  *       | ---- range to drop ----- |
712                  *  | -------- extent -------- |
713                  */
714                 if (start > key.offset && end >= extent_end) {
715                         BUG_ON(del_nr > 0);
716                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
717
718                         btrfs_set_file_extent_num_bytes(leaf, fi,
719                                                         start - key.offset);
720                         btrfs_mark_buffer_dirty(leaf);
721                         if (disk_bytenr > 0) {
722                                 inode_sub_bytes(inode, extent_end - start);
723                                 *hint_byte = disk_bytenr;
724                         }
725                         if (end == extent_end)
726                                 break;
727
728                         path->slots[0]++;
729                         goto next_slot;
730                 }
731
732                 /*
733                  *  | ---- range to drop ----- |
734                  *    | ------ extent ------ |
735                  */
736                 if (start <= key.offset && end >= extent_end) {
737                         if (del_nr == 0) {
738                                 del_slot = path->slots[0];
739                                 del_nr = 1;
740                         } else {
741                                 BUG_ON(del_slot + del_nr != path->slots[0]);
742                                 del_nr++;
743                         }
744
745                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
746                                 inode_sub_bytes(inode,
747                                                 extent_end - key.offset);
748                                 extent_end = ALIGN(extent_end,
749                                                    root->sectorsize);
750                         } else if (disk_bytenr > 0) {
751                                 ret = btrfs_free_extent(trans, root,
752                                                 disk_bytenr, num_bytes, 0,
753                                                 root->root_key.objectid,
754                                                 key.objectid, key.offset -
755                                                 extent_offset);
756                                 BUG_ON(ret);
757                                 inode_sub_bytes(inode,
758                                                 extent_end - key.offset);
759                                 *hint_byte = disk_bytenr;
760                         }
761
762                         if (end == extent_end)
763                                 break;
764
765                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
766                                 path->slots[0]++;
767                                 goto next_slot;
768                         }
769
770                         ret = btrfs_del_items(trans, root, path, del_slot,
771                                               del_nr);
772                         BUG_ON(ret);
773
774                         del_nr = 0;
775                         del_slot = 0;
776
777                         btrfs_release_path(path);
778                         continue;
779                 }
780
781                 BUG_ON(1);
782         }
783
784         if (del_nr > 0) {
785                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
786                 BUG_ON(ret);
787         }
788
789         btrfs_free_path(path);
790         return ret;
791 }
792
793 static int extent_mergeable(struct extent_buffer *leaf, int slot,
794                             u64 objectid, u64 bytenr, u64 orig_offset,
795                             u64 *start, u64 *end)
796 {
797         struct btrfs_file_extent_item *fi;
798         struct btrfs_key key;
799         u64 extent_end;
800
801         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
802                 return 0;
803
804         btrfs_item_key_to_cpu(leaf, &key, slot);
805         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
806                 return 0;
807
808         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
809         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
810             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
811             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
812             btrfs_file_extent_compression(leaf, fi) ||
813             btrfs_file_extent_encryption(leaf, fi) ||
814             btrfs_file_extent_other_encoding(leaf, fi))
815                 return 0;
816
817         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
818         if ((*start && *start != key.offset) || (*end && *end != extent_end))
819                 return 0;
820
821         *start = key.offset;
822         *end = extent_end;
823         return 1;
824 }
825
826 /*
827  * Mark extent in the range start - end as written.
828  *
829  * This changes extent type from 'pre-allocated' to 'regular'. If only
830  * part of extent is marked as written, the extent will be split into
831  * two or three.
832  */
833 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
834                               struct inode *inode, u64 start, u64 end)
835 {
836         struct btrfs_root *root = BTRFS_I(inode)->root;
837         struct extent_buffer *leaf;
838         struct btrfs_path *path;
839         struct btrfs_file_extent_item *fi;
840         struct btrfs_key key;
841         struct btrfs_key new_key;
842         u64 bytenr;
843         u64 num_bytes;
844         u64 extent_end;
845         u64 orig_offset;
846         u64 other_start;
847         u64 other_end;
848         u64 split;
849         int del_nr = 0;
850         int del_slot = 0;
851         int recow;
852         int ret;
853         u64 ino = btrfs_ino(inode);
854
855         btrfs_drop_extent_cache(inode, start, end - 1, 0);
856
857         path = btrfs_alloc_path();
858         BUG_ON(!path);
859 again:
860         recow = 0;
861         split = start;
862         key.objectid = ino;
863         key.type = BTRFS_EXTENT_DATA_KEY;
864         key.offset = split;
865
866         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
867         if (ret < 0)
868                 goto out;
869         if (ret > 0 && path->slots[0] > 0)
870                 path->slots[0]--;
871
872         leaf = path->nodes[0];
873         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
874         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
875         fi = btrfs_item_ptr(leaf, path->slots[0],
876                             struct btrfs_file_extent_item);
877         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
878                BTRFS_FILE_EXTENT_PREALLOC);
879         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
880         BUG_ON(key.offset > start || extent_end < end);
881
882         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
883         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
884         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
885         memcpy(&new_key, &key, sizeof(new_key));
886
887         if (start == key.offset && end < extent_end) {
888                 other_start = 0;
889                 other_end = start;
890                 if (extent_mergeable(leaf, path->slots[0] - 1,
891                                      ino, bytenr, orig_offset,
892                                      &other_start, &other_end)) {
893                         new_key.offset = end;
894                         btrfs_set_item_key_safe(trans, root, path, &new_key);
895                         fi = btrfs_item_ptr(leaf, path->slots[0],
896                                             struct btrfs_file_extent_item);
897                         btrfs_set_file_extent_num_bytes(leaf, fi,
898                                                         extent_end - end);
899                         btrfs_set_file_extent_offset(leaf, fi,
900                                                      end - orig_offset);
901                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
902                                             struct btrfs_file_extent_item);
903                         btrfs_set_file_extent_num_bytes(leaf, fi,
904                                                         end - other_start);
905                         btrfs_mark_buffer_dirty(leaf);
906                         goto out;
907                 }
908         }
909
910         if (start > key.offset && end == extent_end) {
911                 other_start = end;
912                 other_end = 0;
913                 if (extent_mergeable(leaf, path->slots[0] + 1,
914                                      ino, bytenr, orig_offset,
915                                      &other_start, &other_end)) {
916                         fi = btrfs_item_ptr(leaf, path->slots[0],
917                                             struct btrfs_file_extent_item);
918                         btrfs_set_file_extent_num_bytes(leaf, fi,
919                                                         start - key.offset);
920                         path->slots[0]++;
921                         new_key.offset = start;
922                         btrfs_set_item_key_safe(trans, root, path, &new_key);
923
924                         fi = btrfs_item_ptr(leaf, path->slots[0],
925                                             struct btrfs_file_extent_item);
926                         btrfs_set_file_extent_num_bytes(leaf, fi,
927                                                         other_end - start);
928                         btrfs_set_file_extent_offset(leaf, fi,
929                                                      start - orig_offset);
930                         btrfs_mark_buffer_dirty(leaf);
931                         goto out;
932                 }
933         }
934
935         while (start > key.offset || end < extent_end) {
936                 if (key.offset == start)
937                         split = end;
938
939                 new_key.offset = split;
940                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
941                 if (ret == -EAGAIN) {
942                         btrfs_release_path(path);
943                         goto again;
944                 }
945                 BUG_ON(ret < 0);
946
947                 leaf = path->nodes[0];
948                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
949                                     struct btrfs_file_extent_item);
950                 btrfs_set_file_extent_num_bytes(leaf, fi,
951                                                 split - key.offset);
952
953                 fi = btrfs_item_ptr(leaf, path->slots[0],
954                                     struct btrfs_file_extent_item);
955
956                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
957                 btrfs_set_file_extent_num_bytes(leaf, fi,
958                                                 extent_end - split);
959                 btrfs_mark_buffer_dirty(leaf);
960
961                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
962                                            root->root_key.objectid,
963                                            ino, orig_offset);
964                 BUG_ON(ret);
965
966                 if (split == start) {
967                         key.offset = start;
968                 } else {
969                         BUG_ON(start != key.offset);
970                         path->slots[0]--;
971                         extent_end = end;
972                 }
973                 recow = 1;
974         }
975
976         other_start = end;
977         other_end = 0;
978         if (extent_mergeable(leaf, path->slots[0] + 1,
979                              ino, bytenr, orig_offset,
980                              &other_start, &other_end)) {
981                 if (recow) {
982                         btrfs_release_path(path);
983                         goto again;
984                 }
985                 extent_end = other_end;
986                 del_slot = path->slots[0] + 1;
987                 del_nr++;
988                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
989                                         0, root->root_key.objectid,
990                                         ino, orig_offset);
991                 BUG_ON(ret);
992         }
993         other_start = 0;
994         other_end = start;
995         if (extent_mergeable(leaf, path->slots[0] - 1,
996                              ino, bytenr, orig_offset,
997                              &other_start, &other_end)) {
998                 if (recow) {
999                         btrfs_release_path(path);
1000                         goto again;
1001                 }
1002                 key.offset = other_start;
1003                 del_slot = path->slots[0];
1004                 del_nr++;
1005                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1006                                         0, root->root_key.objectid,
1007                                         ino, orig_offset);
1008                 BUG_ON(ret);
1009         }
1010         if (del_nr == 0) {
1011                 fi = btrfs_item_ptr(leaf, path->slots[0],
1012                            struct btrfs_file_extent_item);
1013                 btrfs_set_file_extent_type(leaf, fi,
1014                                            BTRFS_FILE_EXTENT_REG);
1015                 btrfs_mark_buffer_dirty(leaf);
1016         } else {
1017                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1018                            struct btrfs_file_extent_item);
1019                 btrfs_set_file_extent_type(leaf, fi,
1020                                            BTRFS_FILE_EXTENT_REG);
1021                 btrfs_set_file_extent_num_bytes(leaf, fi,
1022                                                 extent_end - key.offset);
1023                 btrfs_mark_buffer_dirty(leaf);
1024
1025                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1026                 BUG_ON(ret);
1027         }
1028 out:
1029         btrfs_free_path(path);
1030         return 0;
1031 }
1032
1033 /*
1034  * on error we return an unlocked page and the error value
1035  * on success we return a locked page and 0
1036  */
1037 static int prepare_uptodate_page(struct page *page, u64 pos)
1038 {
1039         int ret = 0;
1040
1041         if ((pos & (PAGE_CACHE_SIZE - 1)) && !PageUptodate(page)) {
1042                 ret = btrfs_readpage(NULL, page);
1043                 if (ret)
1044                         return ret;
1045                 lock_page(page);
1046                 if (!PageUptodate(page)) {
1047                         unlock_page(page);
1048                         return -EIO;
1049                 }
1050         }
1051         return 0;
1052 }
1053
1054 /*
1055  * this gets pages into the page cache and locks them down, it also properly
1056  * waits for data=ordered extents to finish before allowing the pages to be
1057  * modified.
1058  */
1059 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1060                          struct page **pages, size_t num_pages,
1061                          loff_t pos, unsigned long first_index,
1062                          unsigned long last_index, size_t write_bytes)
1063 {
1064         struct extent_state *cached_state = NULL;
1065         int i;
1066         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1067         struct inode *inode = fdentry(file)->d_inode;
1068         int err = 0;
1069         int faili = 0;
1070         u64 start_pos;
1071         u64 last_pos;
1072
1073         start_pos = pos & ~((u64)root->sectorsize - 1);
1074         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1075
1076         if (start_pos > inode->i_size) {
1077                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1078                 if (err)
1079                         return err;
1080         }
1081
1082 again:
1083         for (i = 0; i < num_pages; i++) {
1084                 pages[i] = grab_cache_page(inode->i_mapping, index + i);
1085                 if (!pages[i]) {
1086                         faili = i - 1;
1087                         err = -ENOMEM;
1088                         goto fail;
1089                 }
1090
1091                 if (i == 0)
1092                         err = prepare_uptodate_page(pages[i], pos);
1093                 if (i == num_pages - 1)
1094                         err = prepare_uptodate_page(pages[i],
1095                                                     pos + write_bytes);
1096                 if (err) {
1097                         page_cache_release(pages[i]);
1098                         faili = i - 1;
1099                         goto fail;
1100                 }
1101                 wait_on_page_writeback(pages[i]);
1102         }
1103         err = 0;
1104         if (start_pos < inode->i_size) {
1105                 struct btrfs_ordered_extent *ordered;
1106                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1107                                  start_pos, last_pos - 1, 0, &cached_state,
1108                                  GFP_NOFS);
1109                 ordered = btrfs_lookup_first_ordered_extent(inode,
1110                                                             last_pos - 1);
1111                 if (ordered &&
1112                     ordered->file_offset + ordered->len > start_pos &&
1113                     ordered->file_offset < last_pos) {
1114                         btrfs_put_ordered_extent(ordered);
1115                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1116                                              start_pos, last_pos - 1,
1117                                              &cached_state, GFP_NOFS);
1118                         for (i = 0; i < num_pages; i++) {
1119                                 unlock_page(pages[i]);
1120                                 page_cache_release(pages[i]);
1121                         }
1122                         btrfs_wait_ordered_range(inode, start_pos,
1123                                                  last_pos - start_pos);
1124                         goto again;
1125                 }
1126                 if (ordered)
1127                         btrfs_put_ordered_extent(ordered);
1128
1129                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1130                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1131                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
1132                                   GFP_NOFS);
1133                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1134                                      start_pos, last_pos - 1, &cached_state,
1135                                      GFP_NOFS);
1136         }
1137         for (i = 0; i < num_pages; i++) {
1138                 clear_page_dirty_for_io(pages[i]);
1139                 set_page_extent_mapped(pages[i]);
1140                 WARN_ON(!PageLocked(pages[i]));
1141         }
1142         return 0;
1143 fail:
1144         while (faili >= 0) {
1145                 unlock_page(pages[faili]);
1146                 page_cache_release(pages[faili]);
1147                 faili--;
1148         }
1149         return err;
1150
1151 }
1152
1153 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1154                                                struct iov_iter *i,
1155                                                loff_t pos)
1156 {
1157         struct inode *inode = fdentry(file)->d_inode;
1158         struct btrfs_root *root = BTRFS_I(inode)->root;
1159         struct page **pages = NULL;
1160         unsigned long first_index;
1161         unsigned long last_index;
1162         size_t num_written = 0;
1163         int nrptrs;
1164         int ret = 0;
1165
1166         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1167                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1168                      (sizeof(struct page *)));
1169         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1170         if (!pages)
1171                 return -ENOMEM;
1172
1173         first_index = pos >> PAGE_CACHE_SHIFT;
1174         last_index = (pos + iov_iter_count(i)) >> PAGE_CACHE_SHIFT;
1175
1176         while (iov_iter_count(i) > 0) {
1177                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1178                 size_t write_bytes = min(iov_iter_count(i),
1179                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1180                                          offset);
1181                 size_t num_pages = (write_bytes + offset +
1182                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1183                 size_t dirty_pages;
1184                 size_t copied;
1185
1186                 WARN_ON(num_pages > nrptrs);
1187
1188                 /*
1189                  * Fault pages before locking them in prepare_pages
1190                  * to avoid recursive lock
1191                  */
1192                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1193                         ret = -EFAULT;
1194                         break;
1195                 }
1196
1197                 ret = btrfs_delalloc_reserve_space(inode,
1198                                         num_pages << PAGE_CACHE_SHIFT);
1199                 if (ret)
1200                         break;
1201
1202                 /*
1203                  * This is going to setup the pages array with the number of
1204                  * pages we want, so we don't really need to worry about the
1205                  * contents of pages from loop to loop
1206                  */
1207                 ret = prepare_pages(root, file, pages, num_pages,
1208                                     pos, first_index, last_index,
1209                                     write_bytes);
1210                 if (ret) {
1211                         btrfs_delalloc_release_space(inode,
1212                                         num_pages << PAGE_CACHE_SHIFT);
1213                         break;
1214                 }
1215
1216                 copied = btrfs_copy_from_user(pos, num_pages,
1217                                            write_bytes, pages, i);
1218
1219                 /*
1220                  * if we have trouble faulting in the pages, fall
1221                  * back to one page at a time
1222                  */
1223                 if (copied < write_bytes)
1224                         nrptrs = 1;
1225
1226                 if (copied == 0)
1227                         dirty_pages = 0;
1228                 else
1229                         dirty_pages = (copied + offset +
1230                                        PAGE_CACHE_SIZE - 1) >>
1231                                        PAGE_CACHE_SHIFT;
1232
1233                 /*
1234                  * If we had a short copy we need to release the excess delaloc
1235                  * bytes we reserved.  We need to increment outstanding_extents
1236                  * because btrfs_delalloc_release_space will decrement it, but
1237                  * we still have an outstanding extent for the chunk we actually
1238                  * managed to copy.
1239                  */
1240                 if (num_pages > dirty_pages) {
1241                         if (copied > 0)
1242                                 atomic_inc(
1243                                         &BTRFS_I(inode)->outstanding_extents);
1244                         btrfs_delalloc_release_space(inode,
1245                                         (num_pages - dirty_pages) <<
1246                                         PAGE_CACHE_SHIFT);
1247                 }
1248
1249                 if (copied > 0) {
1250                         ret = btrfs_dirty_pages(root, inode, pages,
1251                                                 dirty_pages, pos, copied,
1252                                                 NULL);
1253                         if (ret) {
1254                                 btrfs_delalloc_release_space(inode,
1255                                         dirty_pages << PAGE_CACHE_SHIFT);
1256                                 btrfs_drop_pages(pages, num_pages);
1257                                 break;
1258                         }
1259                 }
1260
1261                 btrfs_drop_pages(pages, num_pages);
1262
1263                 cond_resched();
1264
1265                 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
1266                                                    dirty_pages);
1267                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1268                         btrfs_btree_balance_dirty(root, 1);
1269                 btrfs_throttle(root);
1270
1271                 pos += copied;
1272                 num_written += copied;
1273         }
1274
1275         kfree(pages);
1276
1277         return num_written ? num_written : ret;
1278 }
1279
1280 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1281                                     const struct iovec *iov,
1282                                     unsigned long nr_segs, loff_t pos,
1283                                     loff_t *ppos, size_t count, size_t ocount)
1284 {
1285         struct file *file = iocb->ki_filp;
1286         struct inode *inode = fdentry(file)->d_inode;
1287         struct iov_iter i;
1288         ssize_t written;
1289         ssize_t written_buffered;
1290         loff_t endbyte;
1291         int err;
1292
1293         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1294                                             count, ocount);
1295
1296         /*
1297          * the generic O_DIRECT will update in-memory i_size after the
1298          * DIOs are done.  But our endio handlers that update the on
1299          * disk i_size never update past the in memory i_size.  So we
1300          * need one more update here to catch any additions to the
1301          * file
1302          */
1303         if (inode->i_size != BTRFS_I(inode)->disk_i_size) {
1304                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
1305                 mark_inode_dirty(inode);
1306         }
1307
1308         if (written < 0 || written == count)
1309                 return written;
1310
1311         pos += written;
1312         count -= written;
1313         iov_iter_init(&i, iov, nr_segs, count, written);
1314         written_buffered = __btrfs_buffered_write(file, &i, pos);
1315         if (written_buffered < 0) {
1316                 err = written_buffered;
1317                 goto out;
1318         }
1319         endbyte = pos + written_buffered - 1;
1320         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1321         if (err)
1322                 goto out;
1323         written += written_buffered;
1324         *ppos = pos + written_buffered;
1325         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1326                                  endbyte >> PAGE_CACHE_SHIFT);
1327 out:
1328         return written ? written : err;
1329 }
1330
1331 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1332                                     const struct iovec *iov,
1333                                     unsigned long nr_segs, loff_t pos)
1334 {
1335         struct file *file = iocb->ki_filp;
1336         struct inode *inode = fdentry(file)->d_inode;
1337         struct btrfs_root *root = BTRFS_I(inode)->root;
1338         loff_t *ppos = &iocb->ki_pos;
1339         ssize_t num_written = 0;
1340         ssize_t err = 0;
1341         size_t count, ocount;
1342
1343         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1344
1345         mutex_lock(&inode->i_mutex);
1346
1347         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1348         if (err) {
1349                 mutex_unlock(&inode->i_mutex);
1350                 goto out;
1351         }
1352         count = ocount;
1353
1354         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1355         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1356         if (err) {
1357                 mutex_unlock(&inode->i_mutex);
1358                 goto out;
1359         }
1360
1361         if (count == 0) {
1362                 mutex_unlock(&inode->i_mutex);
1363                 goto out;
1364         }
1365
1366         err = file_remove_suid(file);
1367         if (err) {
1368                 mutex_unlock(&inode->i_mutex);
1369                 goto out;
1370         }
1371
1372         /*
1373          * If BTRFS flips readonly due to some impossible error
1374          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1375          * although we have opened a file as writable, we have
1376          * to stop this write operation to ensure FS consistency.
1377          */
1378         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
1379                 mutex_unlock(&inode->i_mutex);
1380                 err = -EROFS;
1381                 goto out;
1382         }
1383
1384         file_update_time(file);
1385         BTRFS_I(inode)->sequence++;
1386
1387         if (unlikely(file->f_flags & O_DIRECT)) {
1388                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1389                                                    pos, ppos, count, ocount);
1390         } else {
1391                 struct iov_iter i;
1392
1393                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1394
1395                 num_written = __btrfs_buffered_write(file, &i, pos);
1396                 if (num_written > 0)
1397                         *ppos = pos + num_written;
1398         }
1399
1400         mutex_unlock(&inode->i_mutex);
1401
1402         /*
1403          * we want to make sure fsync finds this change
1404          * but we haven't joined a transaction running right now.
1405          *
1406          * Later on, someone is sure to update the inode and get the
1407          * real transid recorded.
1408          *
1409          * We set last_trans now to the fs_info generation + 1,
1410          * this will either be one more than the running transaction
1411          * or the generation used for the next transaction if there isn't
1412          * one running right now.
1413          */
1414         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1415         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1416                 err = generic_write_sync(file, pos, num_written);
1417                 if (err < 0 && num_written > 0)
1418                         num_written = err;
1419         }
1420 out:
1421         current->backing_dev_info = NULL;
1422         return num_written ? num_written : err;
1423 }
1424
1425 int btrfs_release_file(struct inode *inode, struct file *filp)
1426 {
1427         /*
1428          * ordered_data_close is set by settattr when we are about to truncate
1429          * a file from a non-zero size to a zero size.  This tries to
1430          * flush down new bytes that may have been written if the
1431          * application were using truncate to replace a file in place.
1432          */
1433         if (BTRFS_I(inode)->ordered_data_close) {
1434                 BTRFS_I(inode)->ordered_data_close = 0;
1435                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1436                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1437                         filemap_flush(inode->i_mapping);
1438         }
1439         if (filp->private_data)
1440                 btrfs_ioctl_trans_end(filp);
1441         return 0;
1442 }
1443
1444 /*
1445  * fsync call for both files and directories.  This logs the inode into
1446  * the tree log instead of forcing full commits whenever possible.
1447  *
1448  * It needs to call filemap_fdatawait so that all ordered extent updates are
1449  * in the metadata btree are up to date for copying to the log.
1450  *
1451  * It drops the inode mutex before doing the tree log commit.  This is an
1452  * important optimization for directories because holding the mutex prevents
1453  * new operations on the dir while we write to disk.
1454  */
1455 int btrfs_sync_file(struct file *file, int datasync)
1456 {
1457         struct dentry *dentry = file->f_path.dentry;
1458         struct inode *inode = dentry->d_inode;
1459         struct btrfs_root *root = BTRFS_I(inode)->root;
1460         int ret = 0;
1461         struct btrfs_trans_handle *trans;
1462
1463         trace_btrfs_sync_file(file, datasync);
1464
1465         /* we wait first, since the writeback may change the inode */
1466         root->log_batch++;
1467         /* the VFS called filemap_fdatawrite for us */
1468         btrfs_wait_ordered_range(inode, 0, (u64)-1);
1469         root->log_batch++;
1470
1471         /*
1472          * check the transaction that last modified this inode
1473          * and see if its already been committed
1474          */
1475         if (!BTRFS_I(inode)->last_trans)
1476                 goto out;
1477
1478         /*
1479          * if the last transaction that changed this file was before
1480          * the current transaction, we can bail out now without any
1481          * syncing
1482          */
1483         smp_mb();
1484         if (BTRFS_I(inode)->last_trans <=
1485             root->fs_info->last_trans_committed) {
1486                 BTRFS_I(inode)->last_trans = 0;
1487                 goto out;
1488         }
1489
1490         /*
1491          * ok we haven't committed the transaction yet, lets do a commit
1492          */
1493         if (file->private_data)
1494                 btrfs_ioctl_trans_end(file);
1495
1496         trans = btrfs_start_transaction(root, 0);
1497         if (IS_ERR(trans)) {
1498                 ret = PTR_ERR(trans);
1499                 goto out;
1500         }
1501
1502         ret = btrfs_log_dentry_safe(trans, root, dentry);
1503         if (ret < 0)
1504                 goto out;
1505
1506         /* we've logged all the items and now have a consistent
1507          * version of the file in the log.  It is possible that
1508          * someone will come in and modify the file, but that's
1509          * fine because the log is consistent on disk, and we
1510          * have references to all of the file's extents
1511          *
1512          * It is possible that someone will come in and log the
1513          * file again, but that will end up using the synchronization
1514          * inside btrfs_sync_log to keep things safe.
1515          */
1516         mutex_unlock(&dentry->d_inode->i_mutex);
1517
1518         if (ret != BTRFS_NO_LOG_SYNC) {
1519                 if (ret > 0) {
1520                         ret = btrfs_commit_transaction(trans, root);
1521                 } else {
1522                         ret = btrfs_sync_log(trans, root);
1523                         if (ret == 0)
1524                                 ret = btrfs_end_transaction(trans, root);
1525                         else
1526                                 ret = btrfs_commit_transaction(trans, root);
1527                 }
1528         } else {
1529                 ret = btrfs_end_transaction(trans, root);
1530         }
1531         mutex_lock(&dentry->d_inode->i_mutex);
1532 out:
1533         return ret > 0 ? -EIO : ret;
1534 }
1535
1536 static const struct vm_operations_struct btrfs_file_vm_ops = {
1537         .fault          = filemap_fault,
1538         .page_mkwrite   = btrfs_page_mkwrite,
1539 };
1540
1541 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1542 {
1543         struct address_space *mapping = filp->f_mapping;
1544
1545         if (!mapping->a_ops->readpage)
1546                 return -ENOEXEC;
1547
1548         file_accessed(filp);
1549         vma->vm_ops = &btrfs_file_vm_ops;
1550         vma->vm_flags |= VM_CAN_NONLINEAR;
1551
1552         return 0;
1553 }
1554
1555 static long btrfs_fallocate(struct file *file, int mode,
1556                             loff_t offset, loff_t len)
1557 {
1558         struct inode *inode = file->f_path.dentry->d_inode;
1559         struct extent_state *cached_state = NULL;
1560         u64 cur_offset;
1561         u64 last_byte;
1562         u64 alloc_start;
1563         u64 alloc_end;
1564         u64 alloc_hint = 0;
1565         u64 locked_end;
1566         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1567         struct extent_map *em;
1568         int ret;
1569
1570         alloc_start = offset & ~mask;
1571         alloc_end =  (offset + len + mask) & ~mask;
1572
1573         /* We only support the FALLOC_FL_KEEP_SIZE mode */
1574         if (mode & ~FALLOC_FL_KEEP_SIZE)
1575                 return -EOPNOTSUPP;
1576
1577         /*
1578          * wait for ordered IO before we have any locks.  We'll loop again
1579          * below with the locks held.
1580          */
1581         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
1582
1583         mutex_lock(&inode->i_mutex);
1584         ret = inode_newsize_ok(inode, alloc_end);
1585         if (ret)
1586                 goto out;
1587
1588         if (alloc_start > inode->i_size) {
1589                 ret = btrfs_cont_expand(inode, i_size_read(inode),
1590                                         alloc_start);
1591                 if (ret)
1592                         goto out;
1593         }
1594
1595         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
1596         if (ret)
1597                 goto out;
1598
1599         locked_end = alloc_end - 1;
1600         while (1) {
1601                 struct btrfs_ordered_extent *ordered;
1602
1603                 /* the extent lock is ordered inside the running
1604                  * transaction
1605                  */
1606                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
1607                                  locked_end, 0, &cached_state, GFP_NOFS);
1608                 ordered = btrfs_lookup_first_ordered_extent(inode,
1609                                                             alloc_end - 1);
1610                 if (ordered &&
1611                     ordered->file_offset + ordered->len > alloc_start &&
1612                     ordered->file_offset < alloc_end) {
1613                         btrfs_put_ordered_extent(ordered);
1614                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1615                                              alloc_start, locked_end,
1616                                              &cached_state, GFP_NOFS);
1617                         /*
1618                          * we can't wait on the range with the transaction
1619                          * running or with the extent lock held
1620                          */
1621                         btrfs_wait_ordered_range(inode, alloc_start,
1622                                                  alloc_end - alloc_start);
1623                 } else {
1624                         if (ordered)
1625                                 btrfs_put_ordered_extent(ordered);
1626                         break;
1627                 }
1628         }
1629
1630         cur_offset = alloc_start;
1631         while (1) {
1632                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
1633                                       alloc_end - cur_offset, 0);
1634                 BUG_ON(IS_ERR_OR_NULL(em));
1635                 last_byte = min(extent_map_end(em), alloc_end);
1636                 last_byte = (last_byte + mask) & ~mask;
1637                 if (em->block_start == EXTENT_MAP_HOLE ||
1638                     (cur_offset >= inode->i_size &&
1639                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
1640                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
1641                                                         last_byte - cur_offset,
1642                                                         1 << inode->i_blkbits,
1643                                                         offset + len,
1644                                                         &alloc_hint);
1645                         if (ret < 0) {
1646                                 free_extent_map(em);
1647                                 break;
1648                         }
1649                 }
1650                 free_extent_map(em);
1651
1652                 cur_offset = last_byte;
1653                 if (cur_offset >= alloc_end) {
1654                         ret = 0;
1655                         break;
1656                 }
1657         }
1658         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
1659                              &cached_state, GFP_NOFS);
1660
1661         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
1662 out:
1663         mutex_unlock(&inode->i_mutex);
1664         return ret;
1665 }
1666
1667 const struct file_operations btrfs_file_operations = {
1668         .llseek         = generic_file_llseek,
1669         .read           = do_sync_read,
1670         .write          = do_sync_write,
1671         .aio_read       = generic_file_aio_read,
1672         .splice_read    = generic_file_splice_read,
1673         .aio_write      = btrfs_file_aio_write,
1674         .mmap           = btrfs_file_mmap,
1675         .open           = generic_file_open,
1676         .release        = btrfs_release_file,
1677         .fsync          = btrfs_sync_file,
1678         .fallocate      = btrfs_fallocate,
1679         .unlocked_ioctl = btrfs_ioctl,
1680 #ifdef CONFIG_COMPAT
1681         .compat_ioctl   = btrfs_ioctl,
1682 #endif
1683 };