Merge branch 'alloc_path' of git://git.kernel.org/pub/scm/linux/kernel/git/mfasheh...
[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 (btrfs_fs_closing(root->fs_info))
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 (btrfs_fs_closing(fs_info))
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         if (!path)
859                 return -ENOMEM;
860 again:
861         recow = 0;
862         split = start;
863         key.objectid = ino;
864         key.type = BTRFS_EXTENT_DATA_KEY;
865         key.offset = split;
866
867         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
868         if (ret < 0)
869                 goto out;
870         if (ret > 0 && path->slots[0] > 0)
871                 path->slots[0]--;
872
873         leaf = path->nodes[0];
874         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
875         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
876         fi = btrfs_item_ptr(leaf, path->slots[0],
877                             struct btrfs_file_extent_item);
878         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
879                BTRFS_FILE_EXTENT_PREALLOC);
880         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
881         BUG_ON(key.offset > start || extent_end < end);
882
883         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
884         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
885         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
886         memcpy(&new_key, &key, sizeof(new_key));
887
888         if (start == key.offset && end < extent_end) {
889                 other_start = 0;
890                 other_end = start;
891                 if (extent_mergeable(leaf, path->slots[0] - 1,
892                                      ino, bytenr, orig_offset,
893                                      &other_start, &other_end)) {
894                         new_key.offset = end;
895                         btrfs_set_item_key_safe(trans, root, path, &new_key);
896                         fi = btrfs_item_ptr(leaf, path->slots[0],
897                                             struct btrfs_file_extent_item);
898                         btrfs_set_file_extent_num_bytes(leaf, fi,
899                                                         extent_end - end);
900                         btrfs_set_file_extent_offset(leaf, fi,
901                                                      end - orig_offset);
902                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
903                                             struct btrfs_file_extent_item);
904                         btrfs_set_file_extent_num_bytes(leaf, fi,
905                                                         end - other_start);
906                         btrfs_mark_buffer_dirty(leaf);
907                         goto out;
908                 }
909         }
910
911         if (start > key.offset && end == extent_end) {
912                 other_start = end;
913                 other_end = 0;
914                 if (extent_mergeable(leaf, path->slots[0] + 1,
915                                      ino, bytenr, orig_offset,
916                                      &other_start, &other_end)) {
917                         fi = btrfs_item_ptr(leaf, path->slots[0],
918                                             struct btrfs_file_extent_item);
919                         btrfs_set_file_extent_num_bytes(leaf, fi,
920                                                         start - key.offset);
921                         path->slots[0]++;
922                         new_key.offset = start;
923                         btrfs_set_item_key_safe(trans, root, path, &new_key);
924
925                         fi = btrfs_item_ptr(leaf, path->slots[0],
926                                             struct btrfs_file_extent_item);
927                         btrfs_set_file_extent_num_bytes(leaf, fi,
928                                                         other_end - start);
929                         btrfs_set_file_extent_offset(leaf, fi,
930                                                      start - orig_offset);
931                         btrfs_mark_buffer_dirty(leaf);
932                         goto out;
933                 }
934         }
935
936         while (start > key.offset || end < extent_end) {
937                 if (key.offset == start)
938                         split = end;
939
940                 new_key.offset = split;
941                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
942                 if (ret == -EAGAIN) {
943                         btrfs_release_path(path);
944                         goto again;
945                 }
946                 BUG_ON(ret < 0);
947
948                 leaf = path->nodes[0];
949                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
950                                     struct btrfs_file_extent_item);
951                 btrfs_set_file_extent_num_bytes(leaf, fi,
952                                                 split - key.offset);
953
954                 fi = btrfs_item_ptr(leaf, path->slots[0],
955                                     struct btrfs_file_extent_item);
956
957                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
958                 btrfs_set_file_extent_num_bytes(leaf, fi,
959                                                 extent_end - split);
960                 btrfs_mark_buffer_dirty(leaf);
961
962                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
963                                            root->root_key.objectid,
964                                            ino, orig_offset);
965                 BUG_ON(ret);
966
967                 if (split == start) {
968                         key.offset = start;
969                 } else {
970                         BUG_ON(start != key.offset);
971                         path->slots[0]--;
972                         extent_end = end;
973                 }
974                 recow = 1;
975         }
976
977         other_start = end;
978         other_end = 0;
979         if (extent_mergeable(leaf, path->slots[0] + 1,
980                              ino, bytenr, orig_offset,
981                              &other_start, &other_end)) {
982                 if (recow) {
983                         btrfs_release_path(path);
984                         goto again;
985                 }
986                 extent_end = other_end;
987                 del_slot = path->slots[0] + 1;
988                 del_nr++;
989                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
990                                         0, root->root_key.objectid,
991                                         ino, orig_offset);
992                 BUG_ON(ret);
993         }
994         other_start = 0;
995         other_end = start;
996         if (extent_mergeable(leaf, path->slots[0] - 1,
997                              ino, bytenr, orig_offset,
998                              &other_start, &other_end)) {
999                 if (recow) {
1000                         btrfs_release_path(path);
1001                         goto again;
1002                 }
1003                 key.offset = other_start;
1004                 del_slot = path->slots[0];
1005                 del_nr++;
1006                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1007                                         0, root->root_key.objectid,
1008                                         ino, orig_offset);
1009                 BUG_ON(ret);
1010         }
1011         if (del_nr == 0) {
1012                 fi = btrfs_item_ptr(leaf, path->slots[0],
1013                            struct btrfs_file_extent_item);
1014                 btrfs_set_file_extent_type(leaf, fi,
1015                                            BTRFS_FILE_EXTENT_REG);
1016                 btrfs_mark_buffer_dirty(leaf);
1017         } else {
1018                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1019                            struct btrfs_file_extent_item);
1020                 btrfs_set_file_extent_type(leaf, fi,
1021                                            BTRFS_FILE_EXTENT_REG);
1022                 btrfs_set_file_extent_num_bytes(leaf, fi,
1023                                                 extent_end - key.offset);
1024                 btrfs_mark_buffer_dirty(leaf);
1025
1026                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1027                 BUG_ON(ret);
1028         }
1029 out:
1030         btrfs_free_path(path);
1031         return 0;
1032 }
1033
1034 /*
1035  * on error we return an unlocked page and the error value
1036  * on success we return a locked page and 0
1037  */
1038 static int prepare_uptodate_page(struct page *page, u64 pos)
1039 {
1040         int ret = 0;
1041
1042         if ((pos & (PAGE_CACHE_SIZE - 1)) && !PageUptodate(page)) {
1043                 ret = btrfs_readpage(NULL, page);
1044                 if (ret)
1045                         return ret;
1046                 lock_page(page);
1047                 if (!PageUptodate(page)) {
1048                         unlock_page(page);
1049                         return -EIO;
1050                 }
1051         }
1052         return 0;
1053 }
1054
1055 /*
1056  * this gets pages into the page cache and locks them down, it also properly
1057  * waits for data=ordered extents to finish before allowing the pages to be
1058  * modified.
1059  */
1060 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1061                          struct page **pages, size_t num_pages,
1062                          loff_t pos, unsigned long first_index,
1063                          unsigned long last_index, size_t write_bytes)
1064 {
1065         struct extent_state *cached_state = NULL;
1066         int i;
1067         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1068         struct inode *inode = fdentry(file)->d_inode;
1069         int err = 0;
1070         int faili = 0;
1071         u64 start_pos;
1072         u64 last_pos;
1073
1074         start_pos = pos & ~((u64)root->sectorsize - 1);
1075         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1076
1077         if (start_pos > inode->i_size) {
1078                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1079                 if (err)
1080                         return err;
1081         }
1082
1083 again:
1084         for (i = 0; i < num_pages; i++) {
1085                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1086                                                GFP_NOFS);
1087                 if (!pages[i]) {
1088                         faili = i - 1;
1089                         err = -ENOMEM;
1090                         goto fail;
1091                 }
1092
1093                 if (i == 0)
1094                         err = prepare_uptodate_page(pages[i], pos);
1095                 if (i == num_pages - 1)
1096                         err = prepare_uptodate_page(pages[i],
1097                                                     pos + write_bytes);
1098                 if (err) {
1099                         page_cache_release(pages[i]);
1100                         faili = i - 1;
1101                         goto fail;
1102                 }
1103                 wait_on_page_writeback(pages[i]);
1104         }
1105         err = 0;
1106         if (start_pos < inode->i_size) {
1107                 struct btrfs_ordered_extent *ordered;
1108                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1109                                  start_pos, last_pos - 1, 0, &cached_state,
1110                                  GFP_NOFS);
1111                 ordered = btrfs_lookup_first_ordered_extent(inode,
1112                                                             last_pos - 1);
1113                 if (ordered &&
1114                     ordered->file_offset + ordered->len > start_pos &&
1115                     ordered->file_offset < last_pos) {
1116                         btrfs_put_ordered_extent(ordered);
1117                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1118                                              start_pos, last_pos - 1,
1119                                              &cached_state, GFP_NOFS);
1120                         for (i = 0; i < num_pages; i++) {
1121                                 unlock_page(pages[i]);
1122                                 page_cache_release(pages[i]);
1123                         }
1124                         btrfs_wait_ordered_range(inode, start_pos,
1125                                                  last_pos - start_pos);
1126                         goto again;
1127                 }
1128                 if (ordered)
1129                         btrfs_put_ordered_extent(ordered);
1130
1131                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1132                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1133                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
1134                                   GFP_NOFS);
1135                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1136                                      start_pos, last_pos - 1, &cached_state,
1137                                      GFP_NOFS);
1138         }
1139         for (i = 0; i < num_pages; i++) {
1140                 clear_page_dirty_for_io(pages[i]);
1141                 set_page_extent_mapped(pages[i]);
1142                 WARN_ON(!PageLocked(pages[i]));
1143         }
1144         return 0;
1145 fail:
1146         while (faili >= 0) {
1147                 unlock_page(pages[faili]);
1148                 page_cache_release(pages[faili]);
1149                 faili--;
1150         }
1151         return err;
1152
1153 }
1154
1155 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1156                                                struct iov_iter *i,
1157                                                loff_t pos)
1158 {
1159         struct inode *inode = fdentry(file)->d_inode;
1160         struct btrfs_root *root = BTRFS_I(inode)->root;
1161         struct page **pages = NULL;
1162         unsigned long first_index;
1163         unsigned long last_index;
1164         size_t num_written = 0;
1165         int nrptrs;
1166         int ret = 0;
1167
1168         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1169                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1170                      (sizeof(struct page *)));
1171         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1172         if (!pages)
1173                 return -ENOMEM;
1174
1175         first_index = pos >> PAGE_CACHE_SHIFT;
1176         last_index = (pos + iov_iter_count(i)) >> PAGE_CACHE_SHIFT;
1177
1178         while (iov_iter_count(i) > 0) {
1179                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1180                 size_t write_bytes = min(iov_iter_count(i),
1181                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1182                                          offset);
1183                 size_t num_pages = (write_bytes + offset +
1184                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1185                 size_t dirty_pages;
1186                 size_t copied;
1187
1188                 WARN_ON(num_pages > nrptrs);
1189
1190                 /*
1191                  * Fault pages before locking them in prepare_pages
1192                  * to avoid recursive lock
1193                  */
1194                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1195                         ret = -EFAULT;
1196                         break;
1197                 }
1198
1199                 ret = btrfs_delalloc_reserve_space(inode,
1200                                         num_pages << PAGE_CACHE_SHIFT);
1201                 if (ret)
1202                         break;
1203
1204                 /*
1205                  * This is going to setup the pages array with the number of
1206                  * pages we want, so we don't really need to worry about the
1207                  * contents of pages from loop to loop
1208                  */
1209                 ret = prepare_pages(root, file, pages, num_pages,
1210                                     pos, first_index, last_index,
1211                                     write_bytes);
1212                 if (ret) {
1213                         btrfs_delalloc_release_space(inode,
1214                                         num_pages << PAGE_CACHE_SHIFT);
1215                         break;
1216                 }
1217
1218                 copied = btrfs_copy_from_user(pos, num_pages,
1219                                            write_bytes, pages, i);
1220
1221                 /*
1222                  * if we have trouble faulting in the pages, fall
1223                  * back to one page at a time
1224                  */
1225                 if (copied < write_bytes)
1226                         nrptrs = 1;
1227
1228                 if (copied == 0)
1229                         dirty_pages = 0;
1230                 else
1231                         dirty_pages = (copied + offset +
1232                                        PAGE_CACHE_SIZE - 1) >>
1233                                        PAGE_CACHE_SHIFT;
1234
1235                 /*
1236                  * If we had a short copy we need to release the excess delaloc
1237                  * bytes we reserved.  We need to increment outstanding_extents
1238                  * because btrfs_delalloc_release_space will decrement it, but
1239                  * we still have an outstanding extent for the chunk we actually
1240                  * managed to copy.
1241                  */
1242                 if (num_pages > dirty_pages) {
1243                         if (copied > 0) {
1244                                 spin_lock(&BTRFS_I(inode)->lock);
1245                                 BTRFS_I(inode)->outstanding_extents++;
1246                                 spin_unlock(&BTRFS_I(inode)->lock);
1247                         }
1248                         btrfs_delalloc_release_space(inode,
1249                                         (num_pages - dirty_pages) <<
1250                                         PAGE_CACHE_SHIFT);
1251                 }
1252
1253                 if (copied > 0) {
1254                         ret = btrfs_dirty_pages(root, inode, pages,
1255                                                 dirty_pages, pos, copied,
1256                                                 NULL);
1257                         if (ret) {
1258                                 btrfs_delalloc_release_space(inode,
1259                                         dirty_pages << PAGE_CACHE_SHIFT);
1260                                 btrfs_drop_pages(pages, num_pages);
1261                                 break;
1262                         }
1263                 }
1264
1265                 btrfs_drop_pages(pages, num_pages);
1266
1267                 cond_resched();
1268
1269                 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
1270                                                    dirty_pages);
1271                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1272                         btrfs_btree_balance_dirty(root, 1);
1273                 btrfs_throttle(root);
1274
1275                 pos += copied;
1276                 num_written += copied;
1277         }
1278
1279         kfree(pages);
1280
1281         return num_written ? num_written : ret;
1282 }
1283
1284 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1285                                     const struct iovec *iov,
1286                                     unsigned long nr_segs, loff_t pos,
1287                                     loff_t *ppos, size_t count, size_t ocount)
1288 {
1289         struct file *file = iocb->ki_filp;
1290         struct inode *inode = fdentry(file)->d_inode;
1291         struct iov_iter i;
1292         ssize_t written;
1293         ssize_t written_buffered;
1294         loff_t endbyte;
1295         int err;
1296
1297         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1298                                             count, ocount);
1299
1300         /*
1301          * the generic O_DIRECT will update in-memory i_size after the
1302          * DIOs are done.  But our endio handlers that update the on
1303          * disk i_size never update past the in memory i_size.  So we
1304          * need one more update here to catch any additions to the
1305          * file
1306          */
1307         if (inode->i_size != BTRFS_I(inode)->disk_i_size) {
1308                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
1309                 mark_inode_dirty(inode);
1310         }
1311
1312         if (written < 0 || written == count)
1313                 return written;
1314
1315         pos += written;
1316         count -= written;
1317         iov_iter_init(&i, iov, nr_segs, count, written);
1318         written_buffered = __btrfs_buffered_write(file, &i, pos);
1319         if (written_buffered < 0) {
1320                 err = written_buffered;
1321                 goto out;
1322         }
1323         endbyte = pos + written_buffered - 1;
1324         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1325         if (err)
1326                 goto out;
1327         written += written_buffered;
1328         *ppos = pos + written_buffered;
1329         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1330                                  endbyte >> PAGE_CACHE_SHIFT);
1331 out:
1332         return written ? written : err;
1333 }
1334
1335 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1336                                     const struct iovec *iov,
1337                                     unsigned long nr_segs, loff_t pos)
1338 {
1339         struct file *file = iocb->ki_filp;
1340         struct inode *inode = fdentry(file)->d_inode;
1341         struct btrfs_root *root = BTRFS_I(inode)->root;
1342         loff_t *ppos = &iocb->ki_pos;
1343         ssize_t num_written = 0;
1344         ssize_t err = 0;
1345         size_t count, ocount;
1346
1347         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
1348
1349         mutex_lock(&inode->i_mutex);
1350
1351         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1352         if (err) {
1353                 mutex_unlock(&inode->i_mutex);
1354                 goto out;
1355         }
1356         count = ocount;
1357
1358         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1359         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1360         if (err) {
1361                 mutex_unlock(&inode->i_mutex);
1362                 goto out;
1363         }
1364
1365         if (count == 0) {
1366                 mutex_unlock(&inode->i_mutex);
1367                 goto out;
1368         }
1369
1370         err = file_remove_suid(file);
1371         if (err) {
1372                 mutex_unlock(&inode->i_mutex);
1373                 goto out;
1374         }
1375
1376         /*
1377          * If BTRFS flips readonly due to some impossible error
1378          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1379          * although we have opened a file as writable, we have
1380          * to stop this write operation to ensure FS consistency.
1381          */
1382         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
1383                 mutex_unlock(&inode->i_mutex);
1384                 err = -EROFS;
1385                 goto out;
1386         }
1387
1388         file_update_time(file);
1389         BTRFS_I(inode)->sequence++;
1390
1391         if (unlikely(file->f_flags & O_DIRECT)) {
1392                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1393                                                    pos, ppos, count, ocount);
1394         } else {
1395                 struct iov_iter i;
1396
1397                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1398
1399                 num_written = __btrfs_buffered_write(file, &i, pos);
1400                 if (num_written > 0)
1401                         *ppos = pos + num_written;
1402         }
1403
1404         mutex_unlock(&inode->i_mutex);
1405
1406         /*
1407          * we want to make sure fsync finds this change
1408          * but we haven't joined a transaction running right now.
1409          *
1410          * Later on, someone is sure to update the inode and get the
1411          * real transid recorded.
1412          *
1413          * We set last_trans now to the fs_info generation + 1,
1414          * this will either be one more than the running transaction
1415          * or the generation used for the next transaction if there isn't
1416          * one running right now.
1417          */
1418         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1419         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1420                 err = generic_write_sync(file, pos, num_written);
1421                 if (err < 0 && num_written > 0)
1422                         num_written = err;
1423         }
1424 out:
1425         current->backing_dev_info = NULL;
1426         return num_written ? num_written : err;
1427 }
1428
1429 int btrfs_release_file(struct inode *inode, struct file *filp)
1430 {
1431         /*
1432          * ordered_data_close is set by settattr when we are about to truncate
1433          * a file from a non-zero size to a zero size.  This tries to
1434          * flush down new bytes that may have been written if the
1435          * application were using truncate to replace a file in place.
1436          */
1437         if (BTRFS_I(inode)->ordered_data_close) {
1438                 BTRFS_I(inode)->ordered_data_close = 0;
1439                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1440                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1441                         filemap_flush(inode->i_mapping);
1442         }
1443         if (filp->private_data)
1444                 btrfs_ioctl_trans_end(filp);
1445         return 0;
1446 }
1447
1448 /*
1449  * fsync call for both files and directories.  This logs the inode into
1450  * the tree log instead of forcing full commits whenever possible.
1451  *
1452  * It needs to call filemap_fdatawait so that all ordered extent updates are
1453  * in the metadata btree are up to date for copying to the log.
1454  *
1455  * It drops the inode mutex before doing the tree log commit.  This is an
1456  * important optimization for directories because holding the mutex prevents
1457  * new operations on the dir while we write to disk.
1458  */
1459 int btrfs_sync_file(struct file *file, int datasync)
1460 {
1461         struct dentry *dentry = file->f_path.dentry;
1462         struct inode *inode = dentry->d_inode;
1463         struct btrfs_root *root = BTRFS_I(inode)->root;
1464         int ret = 0;
1465         struct btrfs_trans_handle *trans;
1466
1467         trace_btrfs_sync_file(file, datasync);
1468
1469         /* we wait first, since the writeback may change the inode */
1470         root->log_batch++;
1471         /* the VFS called filemap_fdatawrite for us */
1472         btrfs_wait_ordered_range(inode, 0, (u64)-1);
1473         root->log_batch++;
1474
1475         /*
1476          * check the transaction that last modified this inode
1477          * and see if its already been committed
1478          */
1479         if (!BTRFS_I(inode)->last_trans)
1480                 goto out;
1481
1482         /*
1483          * if the last transaction that changed this file was before
1484          * the current transaction, we can bail out now without any
1485          * syncing
1486          */
1487         smp_mb();
1488         if (BTRFS_I(inode)->last_trans <=
1489             root->fs_info->last_trans_committed) {
1490                 BTRFS_I(inode)->last_trans = 0;
1491                 goto out;
1492         }
1493
1494         /*
1495          * ok we haven't committed the transaction yet, lets do a commit
1496          */
1497         if (file->private_data)
1498                 btrfs_ioctl_trans_end(file);
1499
1500         trans = btrfs_start_transaction(root, 0);
1501         if (IS_ERR(trans)) {
1502                 ret = PTR_ERR(trans);
1503                 goto out;
1504         }
1505
1506         ret = btrfs_log_dentry_safe(trans, root, dentry);
1507         if (ret < 0)
1508                 goto out;
1509
1510         /* we've logged all the items and now have a consistent
1511          * version of the file in the log.  It is possible that
1512          * someone will come in and modify the file, but that's
1513          * fine because the log is consistent on disk, and we
1514          * have references to all of the file's extents
1515          *
1516          * It is possible that someone will come in and log the
1517          * file again, but that will end up using the synchronization
1518          * inside btrfs_sync_log to keep things safe.
1519          */
1520         mutex_unlock(&dentry->d_inode->i_mutex);
1521
1522         if (ret != BTRFS_NO_LOG_SYNC) {
1523                 if (ret > 0) {
1524                         ret = btrfs_commit_transaction(trans, root);
1525                 } else {
1526                         ret = btrfs_sync_log(trans, root);
1527                         if (ret == 0)
1528                                 ret = btrfs_end_transaction(trans, root);
1529                         else
1530                                 ret = btrfs_commit_transaction(trans, root);
1531                 }
1532         } else {
1533                 ret = btrfs_end_transaction(trans, root);
1534         }
1535         mutex_lock(&dentry->d_inode->i_mutex);
1536 out:
1537         return ret > 0 ? -EIO : ret;
1538 }
1539
1540 static const struct vm_operations_struct btrfs_file_vm_ops = {
1541         .fault          = filemap_fault,
1542         .page_mkwrite   = btrfs_page_mkwrite,
1543 };
1544
1545 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1546 {
1547         struct address_space *mapping = filp->f_mapping;
1548
1549         if (!mapping->a_ops->readpage)
1550                 return -ENOEXEC;
1551
1552         file_accessed(filp);
1553         vma->vm_ops = &btrfs_file_vm_ops;
1554         vma->vm_flags |= VM_CAN_NONLINEAR;
1555
1556         return 0;
1557 }
1558
1559 static long btrfs_fallocate(struct file *file, int mode,
1560                             loff_t offset, loff_t len)
1561 {
1562         struct inode *inode = file->f_path.dentry->d_inode;
1563         struct extent_state *cached_state = NULL;
1564         u64 cur_offset;
1565         u64 last_byte;
1566         u64 alloc_start;
1567         u64 alloc_end;
1568         u64 alloc_hint = 0;
1569         u64 locked_end;
1570         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1571         struct extent_map *em;
1572         int ret;
1573
1574         alloc_start = offset & ~mask;
1575         alloc_end =  (offset + len + mask) & ~mask;
1576
1577         /* We only support the FALLOC_FL_KEEP_SIZE mode */
1578         if (mode & ~FALLOC_FL_KEEP_SIZE)
1579                 return -EOPNOTSUPP;
1580
1581         /*
1582          * wait for ordered IO before we have any locks.  We'll loop again
1583          * below with the locks held.
1584          */
1585         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
1586
1587         mutex_lock(&inode->i_mutex);
1588         ret = inode_newsize_ok(inode, alloc_end);
1589         if (ret)
1590                 goto out;
1591
1592         if (alloc_start > inode->i_size) {
1593                 ret = btrfs_cont_expand(inode, i_size_read(inode),
1594                                         alloc_start);
1595                 if (ret)
1596                         goto out;
1597         }
1598
1599         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
1600         if (ret)
1601                 goto out;
1602
1603         locked_end = alloc_end - 1;
1604         while (1) {
1605                 struct btrfs_ordered_extent *ordered;
1606
1607                 /* the extent lock is ordered inside the running
1608                  * transaction
1609                  */
1610                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
1611                                  locked_end, 0, &cached_state, GFP_NOFS);
1612                 ordered = btrfs_lookup_first_ordered_extent(inode,
1613                                                             alloc_end - 1);
1614                 if (ordered &&
1615                     ordered->file_offset + ordered->len > alloc_start &&
1616                     ordered->file_offset < alloc_end) {
1617                         btrfs_put_ordered_extent(ordered);
1618                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1619                                              alloc_start, locked_end,
1620                                              &cached_state, GFP_NOFS);
1621                         /*
1622                          * we can't wait on the range with the transaction
1623                          * running or with the extent lock held
1624                          */
1625                         btrfs_wait_ordered_range(inode, alloc_start,
1626                                                  alloc_end - alloc_start);
1627                 } else {
1628                         if (ordered)
1629                                 btrfs_put_ordered_extent(ordered);
1630                         break;
1631                 }
1632         }
1633
1634         cur_offset = alloc_start;
1635         while (1) {
1636                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
1637                                       alloc_end - cur_offset, 0);
1638                 BUG_ON(IS_ERR_OR_NULL(em));
1639                 last_byte = min(extent_map_end(em), alloc_end);
1640                 last_byte = (last_byte + mask) & ~mask;
1641                 if (em->block_start == EXTENT_MAP_HOLE ||
1642                     (cur_offset >= inode->i_size &&
1643                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
1644                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
1645                                                         last_byte - cur_offset,
1646                                                         1 << inode->i_blkbits,
1647                                                         offset + len,
1648                                                         &alloc_hint);
1649                         if (ret < 0) {
1650                                 free_extent_map(em);
1651                                 break;
1652                         }
1653                 }
1654                 free_extent_map(em);
1655
1656                 cur_offset = last_byte;
1657                 if (cur_offset >= alloc_end) {
1658                         ret = 0;
1659                         break;
1660                 }
1661         }
1662         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
1663                              &cached_state, GFP_NOFS);
1664
1665         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
1666 out:
1667         mutex_unlock(&inode->i_mutex);
1668         return ret;
1669 }
1670
1671 const struct file_operations btrfs_file_operations = {
1672         .llseek         = generic_file_llseek,
1673         .read           = do_sync_read,
1674         .write          = do_sync_write,
1675         .aio_read       = generic_file_aio_read,
1676         .splice_read    = generic_file_splice_read,
1677         .aio_write      = btrfs_file_aio_write,
1678         .mmap           = btrfs_file_mmap,
1679         .open           = generic_file_open,
1680         .release        = btrfs_release_file,
1681         .fsync          = btrfs_sync_file,
1682         .fallocate      = btrfs_fallocate,
1683         .unlocked_ioctl = btrfs_ioctl,
1684 #ifdef CONFIG_COMPAT
1685         .compat_ioctl   = btrfs_ioctl,
1686 #endif
1687 };