btrfs: unlock on error in btrfs_file_llseek()
[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 void __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;
110
111 exists:
112         kfree(defrag);
113         return;
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         u64 transid;
127
128         if (!btrfs_test_opt(root, AUTO_DEFRAG))
129                 return 0;
130
131         if (btrfs_fs_closing(root->fs_info))
132                 return 0;
133
134         if (BTRFS_I(inode)->in_defrag)
135                 return 0;
136
137         if (trans)
138                 transid = trans->transid;
139         else
140                 transid = BTRFS_I(inode)->root->last_trans;
141
142         defrag = kzalloc(sizeof(*defrag), GFP_NOFS);
143         if (!defrag)
144                 return -ENOMEM;
145
146         defrag->ino = btrfs_ino(inode);
147         defrag->transid = transid;
148         defrag->root = root->root_key.objectid;
149
150         spin_lock(&root->fs_info->defrag_inodes_lock);
151         if (!BTRFS_I(inode)->in_defrag)
152                 __btrfs_add_inode_defrag(inode, defrag);
153         spin_unlock(&root->fs_info->defrag_inodes_lock);
154         return 0;
155 }
156
157 /*
158  * must be called with the defrag_inodes lock held
159  */
160 struct inode_defrag *btrfs_find_defrag_inode(struct btrfs_fs_info *info, u64 ino,
161                                              struct rb_node **next)
162 {
163         struct inode_defrag *entry = NULL;
164         struct rb_node *p;
165         struct rb_node *parent = NULL;
166
167         p = info->defrag_inodes.rb_node;
168         while (p) {
169                 parent = p;
170                 entry = rb_entry(parent, struct inode_defrag, rb_node);
171
172                 if (ino < entry->ino)
173                         p = parent->rb_left;
174                 else if (ino > entry->ino)
175                         p = parent->rb_right;
176                 else
177                         return entry;
178         }
179
180         if (next) {
181                 while (parent && ino > entry->ino) {
182                         parent = rb_next(parent);
183                         entry = rb_entry(parent, struct inode_defrag, rb_node);
184                 }
185                 *next = parent;
186         }
187         return NULL;
188 }
189
190 /*
191  * run through the list of inodes in the FS that need
192  * defragging
193  */
194 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
195 {
196         struct inode_defrag *defrag;
197         struct btrfs_root *inode_root;
198         struct inode *inode;
199         struct rb_node *n;
200         struct btrfs_key key;
201         struct btrfs_ioctl_defrag_range_args range;
202         u64 first_ino = 0;
203         int num_defrag;
204         int defrag_batch = 1024;
205
206         memset(&range, 0, sizeof(range));
207         range.len = (u64)-1;
208
209         atomic_inc(&fs_info->defrag_running);
210         spin_lock(&fs_info->defrag_inodes_lock);
211         while(1) {
212                 n = NULL;
213
214                 /* find an inode to defrag */
215                 defrag = btrfs_find_defrag_inode(fs_info, first_ino, &n);
216                 if (!defrag) {
217                         if (n)
218                                 defrag = rb_entry(n, struct inode_defrag, rb_node);
219                         else if (first_ino) {
220                                 first_ino = 0;
221                                 continue;
222                         } else {
223                                 break;
224                         }
225                 }
226
227                 /* remove it from the rbtree */
228                 first_ino = defrag->ino + 1;
229                 rb_erase(&defrag->rb_node, &fs_info->defrag_inodes);
230
231                 if (btrfs_fs_closing(fs_info))
232                         goto next_free;
233
234                 spin_unlock(&fs_info->defrag_inodes_lock);
235
236                 /* get the inode */
237                 key.objectid = defrag->root;
238                 btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
239                 key.offset = (u64)-1;
240                 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
241                 if (IS_ERR(inode_root))
242                         goto next;
243
244                 key.objectid = defrag->ino;
245                 btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
246                 key.offset = 0;
247
248                 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
249                 if (IS_ERR(inode))
250                         goto next;
251
252                 /* do a chunk of defrag */
253                 BTRFS_I(inode)->in_defrag = 0;
254                 range.start = defrag->last_offset;
255                 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
256                                                defrag_batch);
257                 /*
258                  * if we filled the whole defrag batch, there
259                  * must be more work to do.  Queue this defrag
260                  * again
261                  */
262                 if (num_defrag == defrag_batch) {
263                         defrag->last_offset = range.start;
264                         __btrfs_add_inode_defrag(inode, defrag);
265                         /*
266                          * we don't want to kfree defrag, we added it back to
267                          * the rbtree
268                          */
269                         defrag = NULL;
270                 } else if (defrag->last_offset && !defrag->cycled) {
271                         /*
272                          * we didn't fill our defrag batch, but
273                          * we didn't start at zero.  Make sure we loop
274                          * around to the start of the file.
275                          */
276                         defrag->last_offset = 0;
277                         defrag->cycled = 1;
278                         __btrfs_add_inode_defrag(inode, defrag);
279                         defrag = NULL;
280                 }
281
282                 iput(inode);
283 next:
284                 spin_lock(&fs_info->defrag_inodes_lock);
285 next_free:
286                 kfree(defrag);
287         }
288         spin_unlock(&fs_info->defrag_inodes_lock);
289
290         atomic_dec(&fs_info->defrag_running);
291
292         /*
293          * during unmount, we use the transaction_wait queue to
294          * wait for the defragger to stop
295          */
296         wake_up(&fs_info->transaction_wait);
297         return 0;
298 }
299
300 /* simple helper to fault in pages and copy.  This should go away
301  * and be replaced with calls into generic code.
302  */
303 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
304                                          size_t write_bytes,
305                                          struct page **prepared_pages,
306                                          struct iov_iter *i)
307 {
308         size_t copied = 0;
309         size_t total_copied = 0;
310         int pg = 0;
311         int offset = pos & (PAGE_CACHE_SIZE - 1);
312
313         while (write_bytes > 0) {
314                 size_t count = min_t(size_t,
315                                      PAGE_CACHE_SIZE - offset, write_bytes);
316                 struct page *page = prepared_pages[pg];
317                 /*
318                  * Copy data from userspace to the current page
319                  *
320                  * Disable pagefault to avoid recursive lock since
321                  * the pages are already locked
322                  */
323                 pagefault_disable();
324                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
325                 pagefault_enable();
326
327                 /* Flush processor's dcache for this page */
328                 flush_dcache_page(page);
329
330                 /*
331                  * if we get a partial write, we can end up with
332                  * partially up to date pages.  These add
333                  * a lot of complexity, so make sure they don't
334                  * happen by forcing this copy to be retried.
335                  *
336                  * The rest of the btrfs_file_write code will fall
337                  * back to page at a time copies after we return 0.
338                  */
339                 if (!PageUptodate(page) && copied < count)
340                         copied = 0;
341
342                 iov_iter_advance(i, copied);
343                 write_bytes -= copied;
344                 total_copied += copied;
345
346                 /* Return to btrfs_file_aio_write to fault page */
347                 if (unlikely(copied == 0))
348                         break;
349
350                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
351                         offset += copied;
352                 } else {
353                         pg++;
354                         offset = 0;
355                 }
356         }
357         return total_copied;
358 }
359
360 /*
361  * unlocks pages after btrfs_file_write is done with them
362  */
363 void btrfs_drop_pages(struct page **pages, size_t num_pages)
364 {
365         size_t i;
366         for (i = 0; i < num_pages; i++) {
367                 /* page checked is some magic around finding pages that
368                  * have been modified without going through btrfs_set_page_dirty
369                  * clear it here
370                  */
371                 ClearPageChecked(pages[i]);
372                 unlock_page(pages[i]);
373                 mark_page_accessed(pages[i]);
374                 page_cache_release(pages[i]);
375         }
376 }
377
378 /*
379  * after copy_from_user, pages need to be dirtied and we need to make
380  * sure holes are created between the current EOF and the start of
381  * any next extents (if required).
382  *
383  * this also makes the decision about creating an inline extent vs
384  * doing real data extents, marking pages dirty and delalloc as required.
385  */
386 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
387                       struct page **pages, size_t num_pages,
388                       loff_t pos, size_t write_bytes,
389                       struct extent_state **cached)
390 {
391         int err = 0;
392         int i;
393         u64 num_bytes;
394         u64 start_pos;
395         u64 end_of_last_block;
396         u64 end_pos = pos + write_bytes;
397         loff_t isize = i_size_read(inode);
398
399         start_pos = pos & ~((u64)root->sectorsize - 1);
400         num_bytes = (write_bytes + pos - start_pos +
401                     root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
402
403         end_of_last_block = start_pos + num_bytes - 1;
404         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
405                                         cached);
406         if (err)
407                 return err;
408
409         for (i = 0; i < num_pages; i++) {
410                 struct page *p = pages[i];
411                 SetPageUptodate(p);
412                 ClearPageChecked(p);
413                 set_page_dirty(p);
414         }
415
416         /*
417          * we've only changed i_size in ram, and we haven't updated
418          * the disk i_size.  There is no need to log the inode
419          * at this time.
420          */
421         if (end_pos > isize)
422                 i_size_write(inode, end_pos);
423         return 0;
424 }
425
426 /*
427  * this drops all the extents in the cache that intersect the range
428  * [start, end].  Existing extents are split as required.
429  */
430 int btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
431                             int skip_pinned)
432 {
433         struct extent_map *em;
434         struct extent_map *split = NULL;
435         struct extent_map *split2 = NULL;
436         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
437         u64 len = end - start + 1;
438         int ret;
439         int testend = 1;
440         unsigned long flags;
441         int compressed = 0;
442
443         WARN_ON(end < start);
444         if (end == (u64)-1) {
445                 len = (u64)-1;
446                 testend = 0;
447         }
448         while (1) {
449                 if (!split)
450                         split = alloc_extent_map();
451                 if (!split2)
452                         split2 = alloc_extent_map();
453                 BUG_ON(!split || !split2);
454
455                 write_lock(&em_tree->lock);
456                 em = lookup_extent_mapping(em_tree, start, len);
457                 if (!em) {
458                         write_unlock(&em_tree->lock);
459                         break;
460                 }
461                 flags = em->flags;
462                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
463                         if (testend && em->start + em->len >= start + len) {
464                                 free_extent_map(em);
465                                 write_unlock(&em_tree->lock);
466                                 break;
467                         }
468                         start = em->start + em->len;
469                         if (testend)
470                                 len = start + len - (em->start + em->len);
471                         free_extent_map(em);
472                         write_unlock(&em_tree->lock);
473                         continue;
474                 }
475                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
476                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
477                 remove_extent_mapping(em_tree, em);
478
479                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
480                     em->start < start) {
481                         split->start = em->start;
482                         split->len = start - em->start;
483                         split->orig_start = em->orig_start;
484                         split->block_start = em->block_start;
485
486                         if (compressed)
487                                 split->block_len = em->block_len;
488                         else
489                                 split->block_len = split->len;
490
491                         split->bdev = em->bdev;
492                         split->flags = flags;
493                         split->compress_type = em->compress_type;
494                         ret = add_extent_mapping(em_tree, split);
495                         BUG_ON(ret);
496                         free_extent_map(split);
497                         split = split2;
498                         split2 = NULL;
499                 }
500                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
501                     testend && em->start + em->len > start + len) {
502                         u64 diff = start + len - em->start;
503
504                         split->start = start + len;
505                         split->len = em->start + em->len - (start + len);
506                         split->bdev = em->bdev;
507                         split->flags = flags;
508                         split->compress_type = em->compress_type;
509
510                         if (compressed) {
511                                 split->block_len = em->block_len;
512                                 split->block_start = em->block_start;
513                                 split->orig_start = em->orig_start;
514                         } else {
515                                 split->block_len = split->len;
516                                 split->block_start = em->block_start + diff;
517                                 split->orig_start = split->start;
518                         }
519
520                         ret = add_extent_mapping(em_tree, split);
521                         BUG_ON(ret);
522                         free_extent_map(split);
523                         split = NULL;
524                 }
525                 write_unlock(&em_tree->lock);
526
527                 /* once for us */
528                 free_extent_map(em);
529                 /* once for the tree*/
530                 free_extent_map(em);
531         }
532         if (split)
533                 free_extent_map(split);
534         if (split2)
535                 free_extent_map(split2);
536         return 0;
537 }
538
539 /*
540  * this is very complex, but the basic idea is to drop all extents
541  * in the range start - end.  hint_block is filled in with a block number
542  * that would be a good hint to the block allocator for this file.
543  *
544  * If an extent intersects the range but is not entirely inside the range
545  * it is either truncated or split.  Anything entirely inside the range
546  * is deleted from the tree.
547  */
548 int btrfs_drop_extents(struct btrfs_trans_handle *trans, struct inode *inode,
549                        u64 start, u64 end, u64 *hint_byte, int drop_cache)
550 {
551         struct btrfs_root *root = BTRFS_I(inode)->root;
552         struct extent_buffer *leaf;
553         struct btrfs_file_extent_item *fi;
554         struct btrfs_path *path;
555         struct btrfs_key key;
556         struct btrfs_key new_key;
557         u64 ino = btrfs_ino(inode);
558         u64 search_start = start;
559         u64 disk_bytenr = 0;
560         u64 num_bytes = 0;
561         u64 extent_offset = 0;
562         u64 extent_end = 0;
563         int del_nr = 0;
564         int del_slot = 0;
565         int extent_type;
566         int recow;
567         int ret;
568
569         if (drop_cache)
570                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
571
572         path = btrfs_alloc_path();
573         if (!path)
574                 return -ENOMEM;
575
576         while (1) {
577                 recow = 0;
578                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
579                                                search_start, -1);
580                 if (ret < 0)
581                         break;
582                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
583                         leaf = path->nodes[0];
584                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
585                         if (key.objectid == ino &&
586                             key.type == BTRFS_EXTENT_DATA_KEY)
587                                 path->slots[0]--;
588                 }
589                 ret = 0;
590 next_slot:
591                 leaf = path->nodes[0];
592                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
593                         BUG_ON(del_nr > 0);
594                         ret = btrfs_next_leaf(root, path);
595                         if (ret < 0)
596                                 break;
597                         if (ret > 0) {
598                                 ret = 0;
599                                 break;
600                         }
601                         leaf = path->nodes[0];
602                         recow = 1;
603                 }
604
605                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
606                 if (key.objectid > ino ||
607                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
608                         break;
609
610                 fi = btrfs_item_ptr(leaf, path->slots[0],
611                                     struct btrfs_file_extent_item);
612                 extent_type = btrfs_file_extent_type(leaf, fi);
613
614                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
615                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
616                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
617                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
618                         extent_offset = btrfs_file_extent_offset(leaf, fi);
619                         extent_end = key.offset +
620                                 btrfs_file_extent_num_bytes(leaf, fi);
621                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
622                         extent_end = key.offset +
623                                 btrfs_file_extent_inline_len(leaf, fi);
624                 } else {
625                         WARN_ON(1);
626                         extent_end = search_start;
627                 }
628
629                 if (extent_end <= search_start) {
630                         path->slots[0]++;
631                         goto next_slot;
632                 }
633
634                 search_start = max(key.offset, start);
635                 if (recow) {
636                         btrfs_release_path(path);
637                         continue;
638                 }
639
640                 /*
641                  *     | - range to drop - |
642                  *  | -------- extent -------- |
643                  */
644                 if (start > key.offset && end < extent_end) {
645                         BUG_ON(del_nr > 0);
646                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
647
648                         memcpy(&new_key, &key, sizeof(new_key));
649                         new_key.offset = start;
650                         ret = btrfs_duplicate_item(trans, root, path,
651                                                    &new_key);
652                         if (ret == -EAGAIN) {
653                                 btrfs_release_path(path);
654                                 continue;
655                         }
656                         if (ret < 0)
657                                 break;
658
659                         leaf = path->nodes[0];
660                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
661                                             struct btrfs_file_extent_item);
662                         btrfs_set_file_extent_num_bytes(leaf, fi,
663                                                         start - key.offset);
664
665                         fi = btrfs_item_ptr(leaf, path->slots[0],
666                                             struct btrfs_file_extent_item);
667
668                         extent_offset += start - key.offset;
669                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
670                         btrfs_set_file_extent_num_bytes(leaf, fi,
671                                                         extent_end - start);
672                         btrfs_mark_buffer_dirty(leaf);
673
674                         if (disk_bytenr > 0) {
675                                 ret = btrfs_inc_extent_ref(trans, root,
676                                                 disk_bytenr, num_bytes, 0,
677                                                 root->root_key.objectid,
678                                                 new_key.objectid,
679                                                 start - extent_offset);
680                                 BUG_ON(ret);
681                                 *hint_byte = disk_bytenr;
682                         }
683                         key.offset = start;
684                 }
685                 /*
686                  *  | ---- range to drop ----- |
687                  *      | -------- extent -------- |
688                  */
689                 if (start <= key.offset && end < extent_end) {
690                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
691
692                         memcpy(&new_key, &key, sizeof(new_key));
693                         new_key.offset = end;
694                         btrfs_set_item_key_safe(trans, root, path, &new_key);
695
696                         extent_offset += end - key.offset;
697                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
698                         btrfs_set_file_extent_num_bytes(leaf, fi,
699                                                         extent_end - end);
700                         btrfs_mark_buffer_dirty(leaf);
701                         if (disk_bytenr > 0) {
702                                 inode_sub_bytes(inode, end - key.offset);
703                                 *hint_byte = disk_bytenr;
704                         }
705                         break;
706                 }
707
708                 search_start = extent_end;
709                 /*
710                  *       | ---- range to drop ----- |
711                  *  | -------- extent -------- |
712                  */
713                 if (start > key.offset && end >= extent_end) {
714                         BUG_ON(del_nr > 0);
715                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
716
717                         btrfs_set_file_extent_num_bytes(leaf, fi,
718                                                         start - key.offset);
719                         btrfs_mark_buffer_dirty(leaf);
720                         if (disk_bytenr > 0) {
721                                 inode_sub_bytes(inode, extent_end - start);
722                                 *hint_byte = disk_bytenr;
723                         }
724                         if (end == extent_end)
725                                 break;
726
727                         path->slots[0]++;
728                         goto next_slot;
729                 }
730
731                 /*
732                  *  | ---- range to drop ----- |
733                  *    | ------ extent ------ |
734                  */
735                 if (start <= key.offset && end >= extent_end) {
736                         if (del_nr == 0) {
737                                 del_slot = path->slots[0];
738                                 del_nr = 1;
739                         } else {
740                                 BUG_ON(del_slot + del_nr != path->slots[0]);
741                                 del_nr++;
742                         }
743
744                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
745                                 inode_sub_bytes(inode,
746                                                 extent_end - key.offset);
747                                 extent_end = ALIGN(extent_end,
748                                                    root->sectorsize);
749                         } else if (disk_bytenr > 0) {
750                                 ret = btrfs_free_extent(trans, root,
751                                                 disk_bytenr, num_bytes, 0,
752                                                 root->root_key.objectid,
753                                                 key.objectid, key.offset -
754                                                 extent_offset);
755                                 BUG_ON(ret);
756                                 inode_sub_bytes(inode,
757                                                 extent_end - key.offset);
758                                 *hint_byte = disk_bytenr;
759                         }
760
761                         if (end == extent_end)
762                                 break;
763
764                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
765                                 path->slots[0]++;
766                                 goto next_slot;
767                         }
768
769                         ret = btrfs_del_items(trans, root, path, del_slot,
770                                               del_nr);
771                         BUG_ON(ret);
772
773                         del_nr = 0;
774                         del_slot = 0;
775
776                         btrfs_release_path(path);
777                         continue;
778                 }
779
780                 BUG_ON(1);
781         }
782
783         if (del_nr > 0) {
784                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
785                 BUG_ON(ret);
786         }
787
788         btrfs_free_path(path);
789         return ret;
790 }
791
792 static int extent_mergeable(struct extent_buffer *leaf, int slot,
793                             u64 objectid, u64 bytenr, u64 orig_offset,
794                             u64 *start, u64 *end)
795 {
796         struct btrfs_file_extent_item *fi;
797         struct btrfs_key key;
798         u64 extent_end;
799
800         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
801                 return 0;
802
803         btrfs_item_key_to_cpu(leaf, &key, slot);
804         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
805                 return 0;
806
807         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
808         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
809             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
810             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
811             btrfs_file_extent_compression(leaf, fi) ||
812             btrfs_file_extent_encryption(leaf, fi) ||
813             btrfs_file_extent_other_encoding(leaf, fi))
814                 return 0;
815
816         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
817         if ((*start && *start != key.offset) || (*end && *end != extent_end))
818                 return 0;
819
820         *start = key.offset;
821         *end = extent_end;
822         return 1;
823 }
824
825 /*
826  * Mark extent in the range start - end as written.
827  *
828  * This changes extent type from 'pre-allocated' to 'regular'. If only
829  * part of extent is marked as written, the extent will be split into
830  * two or three.
831  */
832 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
833                               struct inode *inode, u64 start, u64 end)
834 {
835         struct btrfs_root *root = BTRFS_I(inode)->root;
836         struct extent_buffer *leaf;
837         struct btrfs_path *path;
838         struct btrfs_file_extent_item *fi;
839         struct btrfs_key key;
840         struct btrfs_key new_key;
841         u64 bytenr;
842         u64 num_bytes;
843         u64 extent_end;
844         u64 orig_offset;
845         u64 other_start;
846         u64 other_end;
847         u64 split;
848         int del_nr = 0;
849         int del_slot = 0;
850         int recow;
851         int ret;
852         u64 ino = btrfs_ino(inode);
853
854         btrfs_drop_extent_cache(inode, start, end - 1, 0);
855
856         path = btrfs_alloc_path();
857         if (!path)
858                 return -ENOMEM;
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                          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] = find_or_create_page(inode->i_mapping, index + i,
1085                                                GFP_NOFS);
1086                 if (!pages[i]) {
1087                         faili = i - 1;
1088                         err = -ENOMEM;
1089                         goto fail;
1090                 }
1091
1092                 if (i == 0)
1093                         err = prepare_uptodate_page(pages[i], pos);
1094                 if (i == num_pages - 1)
1095                         err = prepare_uptodate_page(pages[i],
1096                                                     pos + write_bytes);
1097                 if (err) {
1098                         page_cache_release(pages[i]);
1099                         faili = i - 1;
1100                         goto fail;
1101                 }
1102                 wait_on_page_writeback(pages[i]);
1103         }
1104         err = 0;
1105         if (start_pos < inode->i_size) {
1106                 struct btrfs_ordered_extent *ordered;
1107                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1108                                  start_pos, last_pos - 1, 0, &cached_state,
1109                                  GFP_NOFS);
1110                 ordered = btrfs_lookup_first_ordered_extent(inode,
1111                                                             last_pos - 1);
1112                 if (ordered &&
1113                     ordered->file_offset + ordered->len > start_pos &&
1114                     ordered->file_offset < last_pos) {
1115                         btrfs_put_ordered_extent(ordered);
1116                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1117                                              start_pos, last_pos - 1,
1118                                              &cached_state, GFP_NOFS);
1119                         for (i = 0; i < num_pages; i++) {
1120                                 unlock_page(pages[i]);
1121                                 page_cache_release(pages[i]);
1122                         }
1123                         btrfs_wait_ordered_range(inode, start_pos,
1124                                                  last_pos - start_pos);
1125                         goto again;
1126                 }
1127                 if (ordered)
1128                         btrfs_put_ordered_extent(ordered);
1129
1130                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1131                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1132                                   EXTENT_DO_ACCOUNTING, 0, 0, &cached_state,
1133                                   GFP_NOFS);
1134                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1135                                      start_pos, last_pos - 1, &cached_state,
1136                                      GFP_NOFS);
1137         }
1138         for (i = 0; i < num_pages; i++) {
1139                 clear_page_dirty_for_io(pages[i]);
1140                 set_page_extent_mapped(pages[i]);
1141                 WARN_ON(!PageLocked(pages[i]));
1142         }
1143         return 0;
1144 fail:
1145         while (faili >= 0) {
1146                 unlock_page(pages[faili]);
1147                 page_cache_release(pages[faili]);
1148                 faili--;
1149         }
1150         return err;
1151
1152 }
1153
1154 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1155                                                struct iov_iter *i,
1156                                                loff_t pos)
1157 {
1158         struct inode *inode = fdentry(file)->d_inode;
1159         struct btrfs_root *root = BTRFS_I(inode)->root;
1160         struct page **pages = NULL;
1161         unsigned long first_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
1175         while (iov_iter_count(i) > 0) {
1176                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1177                 size_t write_bytes = min(iov_iter_count(i),
1178                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1179                                          offset);
1180                 size_t num_pages = (write_bytes + offset +
1181                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1182                 size_t dirty_pages;
1183                 size_t copied;
1184
1185                 WARN_ON(num_pages > nrptrs);
1186
1187                 /*
1188                  * Fault pages before locking them in prepare_pages
1189                  * to avoid recursive lock
1190                  */
1191                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1192                         ret = -EFAULT;
1193                         break;
1194                 }
1195
1196                 ret = btrfs_delalloc_reserve_space(inode,
1197                                         num_pages << PAGE_CACHE_SHIFT);
1198                 if (ret)
1199                         break;
1200
1201                 /*
1202                  * This is going to setup the pages array with the number of
1203                  * pages we want, so we don't really need to worry about the
1204                  * contents of pages from loop to loop
1205                  */
1206                 ret = prepare_pages(root, file, pages, num_pages,
1207                                     pos, first_index, write_bytes);
1208                 if (ret) {
1209                         btrfs_delalloc_release_space(inode,
1210                                         num_pages << PAGE_CACHE_SHIFT);
1211                         break;
1212                 }
1213
1214                 copied = btrfs_copy_from_user(pos, num_pages,
1215                                            write_bytes, pages, i);
1216
1217                 /*
1218                  * if we have trouble faulting in the pages, fall
1219                  * back to one page at a time
1220                  */
1221                 if (copied < write_bytes)
1222                         nrptrs = 1;
1223
1224                 if (copied == 0)
1225                         dirty_pages = 0;
1226                 else
1227                         dirty_pages = (copied + offset +
1228                                        PAGE_CACHE_SIZE - 1) >>
1229                                        PAGE_CACHE_SHIFT;
1230
1231                 /*
1232                  * If we had a short copy we need to release the excess delaloc
1233                  * bytes we reserved.  We need to increment outstanding_extents
1234                  * because btrfs_delalloc_release_space will decrement it, but
1235                  * we still have an outstanding extent for the chunk we actually
1236                  * managed to copy.
1237                  */
1238                 if (num_pages > dirty_pages) {
1239                         if (copied > 0) {
1240                                 spin_lock(&BTRFS_I(inode)->lock);
1241                                 BTRFS_I(inode)->outstanding_extents++;
1242                                 spin_unlock(&BTRFS_I(inode)->lock);
1243                         }
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, loff_t start, loff_t end, 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         ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1466         if (ret)
1467                 return ret;
1468         mutex_lock(&inode->i_mutex);
1469
1470         /* we wait first, since the writeback may change the inode */
1471         root->log_batch++;
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                 mutex_unlock(&inode->i_mutex);
1481                 goto out;
1482         }
1483
1484         /*
1485          * if the last transaction that changed this file was before
1486          * the current transaction, we can bail out now without any
1487          * syncing
1488          */
1489         smp_mb();
1490         if (BTRFS_I(inode)->last_trans <=
1491             root->fs_info->last_trans_committed) {
1492                 BTRFS_I(inode)->last_trans = 0;
1493                 mutex_unlock(&inode->i_mutex);
1494                 goto out;
1495         }
1496
1497         /*
1498          * ok we haven't committed the transaction yet, lets do a commit
1499          */
1500         if (file->private_data)
1501                 btrfs_ioctl_trans_end(file);
1502
1503         trans = btrfs_start_transaction(root, 0);
1504         if (IS_ERR(trans)) {
1505                 ret = PTR_ERR(trans);
1506                 mutex_unlock(&inode->i_mutex);
1507                 goto out;
1508         }
1509
1510         ret = btrfs_log_dentry_safe(trans, root, dentry);
1511         if (ret < 0) {
1512                 mutex_unlock(&inode->i_mutex);
1513                 goto out;
1514         }
1515
1516         /* we've logged all the items and now have a consistent
1517          * version of the file in the log.  It is possible that
1518          * someone will come in and modify the file, but that's
1519          * fine because the log is consistent on disk, and we
1520          * have references to all of the file's extents
1521          *
1522          * It is possible that someone will come in and log the
1523          * file again, but that will end up using the synchronization
1524          * inside btrfs_sync_log to keep things safe.
1525          */
1526         mutex_unlock(&inode->i_mutex);
1527
1528         if (ret != BTRFS_NO_LOG_SYNC) {
1529                 if (ret > 0) {
1530                         ret = btrfs_commit_transaction(trans, root);
1531                 } else {
1532                         ret = btrfs_sync_log(trans, root);
1533                         if (ret == 0)
1534                                 ret = btrfs_end_transaction(trans, root);
1535                         else
1536                                 ret = btrfs_commit_transaction(trans, root);
1537                 }
1538         } else {
1539                 ret = btrfs_end_transaction(trans, root);
1540         }
1541 out:
1542         return ret > 0 ? -EIO : ret;
1543 }
1544
1545 static const struct vm_operations_struct btrfs_file_vm_ops = {
1546         .fault          = filemap_fault,
1547         .page_mkwrite   = btrfs_page_mkwrite,
1548 };
1549
1550 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1551 {
1552         struct address_space *mapping = filp->f_mapping;
1553
1554         if (!mapping->a_ops->readpage)
1555                 return -ENOEXEC;
1556
1557         file_accessed(filp);
1558         vma->vm_ops = &btrfs_file_vm_ops;
1559         vma->vm_flags |= VM_CAN_NONLINEAR;
1560
1561         return 0;
1562 }
1563
1564 static long btrfs_fallocate(struct file *file, int mode,
1565                             loff_t offset, loff_t len)
1566 {
1567         struct inode *inode = file->f_path.dentry->d_inode;
1568         struct extent_state *cached_state = NULL;
1569         u64 cur_offset;
1570         u64 last_byte;
1571         u64 alloc_start;
1572         u64 alloc_end;
1573         u64 alloc_hint = 0;
1574         u64 locked_end;
1575         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1576         struct extent_map *em;
1577         int ret;
1578
1579         alloc_start = offset & ~mask;
1580         alloc_end =  (offset + len + mask) & ~mask;
1581
1582         /* We only support the FALLOC_FL_KEEP_SIZE mode */
1583         if (mode & ~FALLOC_FL_KEEP_SIZE)
1584                 return -EOPNOTSUPP;
1585
1586         /*
1587          * wait for ordered IO before we have any locks.  We'll loop again
1588          * below with the locks held.
1589          */
1590         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
1591
1592         mutex_lock(&inode->i_mutex);
1593         ret = inode_newsize_ok(inode, alloc_end);
1594         if (ret)
1595                 goto out;
1596
1597         if (alloc_start > inode->i_size) {
1598                 ret = btrfs_cont_expand(inode, i_size_read(inode),
1599                                         alloc_start);
1600                 if (ret)
1601                         goto out;
1602         }
1603
1604         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
1605         if (ret)
1606                 goto out;
1607
1608         locked_end = alloc_end - 1;
1609         while (1) {
1610                 struct btrfs_ordered_extent *ordered;
1611
1612                 /* the extent lock is ordered inside the running
1613                  * transaction
1614                  */
1615                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
1616                                  locked_end, 0, &cached_state, GFP_NOFS);
1617                 ordered = btrfs_lookup_first_ordered_extent(inode,
1618                                                             alloc_end - 1);
1619                 if (ordered &&
1620                     ordered->file_offset + ordered->len > alloc_start &&
1621                     ordered->file_offset < alloc_end) {
1622                         btrfs_put_ordered_extent(ordered);
1623                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1624                                              alloc_start, locked_end,
1625                                              &cached_state, GFP_NOFS);
1626                         /*
1627                          * we can't wait on the range with the transaction
1628                          * running or with the extent lock held
1629                          */
1630                         btrfs_wait_ordered_range(inode, alloc_start,
1631                                                  alloc_end - alloc_start);
1632                 } else {
1633                         if (ordered)
1634                                 btrfs_put_ordered_extent(ordered);
1635                         break;
1636                 }
1637         }
1638
1639         cur_offset = alloc_start;
1640         while (1) {
1641                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
1642                                       alloc_end - cur_offset, 0);
1643                 BUG_ON(IS_ERR_OR_NULL(em));
1644                 last_byte = min(extent_map_end(em), alloc_end);
1645                 last_byte = (last_byte + mask) & ~mask;
1646                 if (em->block_start == EXTENT_MAP_HOLE ||
1647                     (cur_offset >= inode->i_size &&
1648                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
1649                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
1650                                                         last_byte - cur_offset,
1651                                                         1 << inode->i_blkbits,
1652                                                         offset + len,
1653                                                         &alloc_hint);
1654                         if (ret < 0) {
1655                                 free_extent_map(em);
1656                                 break;
1657                         }
1658                 }
1659                 free_extent_map(em);
1660
1661                 cur_offset = last_byte;
1662                 if (cur_offset >= alloc_end) {
1663                         ret = 0;
1664                         break;
1665                 }
1666         }
1667         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
1668                              &cached_state, GFP_NOFS);
1669
1670         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
1671 out:
1672         mutex_unlock(&inode->i_mutex);
1673         return ret;
1674 }
1675
1676 static int find_desired_extent(struct inode *inode, loff_t *offset, int origin)
1677 {
1678         struct btrfs_root *root = BTRFS_I(inode)->root;
1679         struct extent_map *em;
1680         struct extent_state *cached_state = NULL;
1681         u64 lockstart = *offset;
1682         u64 lockend = i_size_read(inode);
1683         u64 start = *offset;
1684         u64 orig_start = *offset;
1685         u64 len = i_size_read(inode);
1686         u64 last_end = 0;
1687         int ret = 0;
1688
1689         lockend = max_t(u64, root->sectorsize, lockend);
1690         if (lockend <= lockstart)
1691                 lockend = lockstart + root->sectorsize;
1692
1693         len = lockend - lockstart + 1;
1694
1695         len = max_t(u64, len, root->sectorsize);
1696         if (inode->i_size == 0)
1697                 return -ENXIO;
1698
1699         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
1700                          &cached_state, GFP_NOFS);
1701
1702         /*
1703          * Delalloc is such a pain.  If we have a hole and we have pending
1704          * delalloc for a portion of the hole we will get back a hole that
1705          * exists for the entire range since it hasn't been actually written
1706          * yet.  So to take care of this case we need to look for an extent just
1707          * before the position we want in case there is outstanding delalloc
1708          * going on here.
1709          */
1710         if (origin == SEEK_HOLE && start != 0) {
1711                 if (start <= root->sectorsize)
1712                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
1713                                                      root->sectorsize, 0);
1714                 else
1715                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
1716                                                      start - root->sectorsize,
1717                                                      root->sectorsize, 0);
1718                 if (IS_ERR(em)) {
1719                         ret = -ENXIO;
1720                         goto out;
1721                 }
1722                 last_end = em->start + em->len;
1723                 if (em->block_start == EXTENT_MAP_DELALLOC)
1724                         last_end = min_t(u64, last_end, inode->i_size);
1725                 free_extent_map(em);
1726         }
1727
1728         while (1) {
1729                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
1730                 if (IS_ERR(em)) {
1731                         ret = -ENXIO;
1732                         break;
1733                 }
1734
1735                 if (em->block_start == EXTENT_MAP_HOLE) {
1736                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1737                                 if (last_end <= orig_start) {
1738                                         free_extent_map(em);
1739                                         ret = -ENXIO;
1740                                         break;
1741                                 }
1742                         }
1743
1744                         if (origin == SEEK_HOLE) {
1745                                 *offset = start;
1746                                 free_extent_map(em);
1747                                 break;
1748                         }
1749                 } else {
1750                         if (origin == SEEK_DATA) {
1751                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
1752                                         if (start >= inode->i_size) {
1753                                                 free_extent_map(em);
1754                                                 ret = -ENXIO;
1755                                                 break;
1756                                         }
1757                                 }
1758
1759                                 *offset = start;
1760                                 free_extent_map(em);
1761                                 break;
1762                         }
1763                 }
1764
1765                 start = em->start + em->len;
1766                 last_end = em->start + em->len;
1767
1768                 if (em->block_start == EXTENT_MAP_DELALLOC)
1769                         last_end = min_t(u64, last_end, inode->i_size);
1770
1771                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
1772                         free_extent_map(em);
1773                         ret = -ENXIO;
1774                         break;
1775                 }
1776                 free_extent_map(em);
1777                 cond_resched();
1778         }
1779         if (!ret)
1780                 *offset = min(*offset, inode->i_size);
1781 out:
1782         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
1783                              &cached_state, GFP_NOFS);
1784         return ret;
1785 }
1786
1787 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int origin)
1788 {
1789         struct inode *inode = file->f_mapping->host;
1790         int ret;
1791
1792         mutex_lock(&inode->i_mutex);
1793         switch (origin) {
1794         case SEEK_END:
1795         case SEEK_CUR:
1796                 offset = generic_file_llseek_unlocked(file, offset, origin);
1797                 goto out;
1798         case SEEK_DATA:
1799         case SEEK_HOLE:
1800                 ret = find_desired_extent(inode, &offset, origin);
1801                 if (ret) {
1802                         mutex_unlock(&inode->i_mutex);
1803                         return ret;
1804                 }
1805         }
1806
1807         if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
1808                 ret = -EINVAL;
1809                 goto out;
1810         }
1811         if (offset > inode->i_sb->s_maxbytes) {
1812                 ret = -EINVAL;
1813                 goto out;
1814         }
1815
1816         /* Special lock needed here? */
1817         if (offset != file->f_pos) {
1818                 file->f_pos = offset;
1819                 file->f_version = 0;
1820         }
1821 out:
1822         mutex_unlock(&inode->i_mutex);
1823         return offset;
1824 }
1825
1826 const struct file_operations btrfs_file_operations = {
1827         .llseek         = btrfs_file_llseek,
1828         .read           = do_sync_read,
1829         .write          = do_sync_write,
1830         .aio_read       = generic_file_aio_read,
1831         .splice_read    = generic_file_splice_read,
1832         .aio_write      = btrfs_file_aio_write,
1833         .mmap           = btrfs_file_mmap,
1834         .open           = generic_file_open,
1835         .release        = btrfs_release_file,
1836         .fsync          = btrfs_sync_file,
1837         .fallocate      = btrfs_fallocate,
1838         .unlocked_ioctl = btrfs_ioctl,
1839 #ifdef CONFIG_COMPAT
1840         .compat_ioctl   = btrfs_ioctl,
1841 #endif
1842 };