a335e2e6c84d9f30dc726a526dfc0a35aafe480d
[sfrench/cifs-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/compat.h>
31 #include <linux/slab.h>
32 #include <linux/btrfs.h>
33 #include <linux/uio.h>
34 #include <linux/iversion.h>
35 #include "ctree.h"
36 #include "disk-io.h"
37 #include "transaction.h"
38 #include "btrfs_inode.h"
39 #include "print-tree.h"
40 #include "tree-log.h"
41 #include "locking.h"
42 #include "volumes.h"
43 #include "qgroup.h"
44 #include "compression.h"
45
46 static struct kmem_cache *btrfs_inode_defrag_cachep;
47 /*
48  * when auto defrag is enabled we
49  * queue up these defrag structs to remember which
50  * inodes need defragging passes
51  */
52 struct inode_defrag {
53         struct rb_node rb_node;
54         /* objectid */
55         u64 ino;
56         /*
57          * transid where the defrag was added, we search for
58          * extents newer than this
59          */
60         u64 transid;
61
62         /* root objectid */
63         u64 root;
64
65         /* last offset we were able to defrag */
66         u64 last_offset;
67
68         /* if we've wrapped around back to zero once already */
69         int cycled;
70 };
71
72 static int __compare_inode_defrag(struct inode_defrag *defrag1,
73                                   struct inode_defrag *defrag2)
74 {
75         if (defrag1->root > defrag2->root)
76                 return 1;
77         else if (defrag1->root < defrag2->root)
78                 return -1;
79         else if (defrag1->ino > defrag2->ino)
80                 return 1;
81         else if (defrag1->ino < defrag2->ino)
82                 return -1;
83         else
84                 return 0;
85 }
86
87 /* pop a record for an inode into the defrag tree.  The lock
88  * must be held already
89  *
90  * If you're inserting a record for an older transid than an
91  * existing record, the transid already in the tree is lowered
92  *
93  * If an existing record is found the defrag item you
94  * pass in is freed
95  */
96 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
97                                     struct inode_defrag *defrag)
98 {
99         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
100         struct inode_defrag *entry;
101         struct rb_node **p;
102         struct rb_node *parent = NULL;
103         int ret;
104
105         p = &fs_info->defrag_inodes.rb_node;
106         while (*p) {
107                 parent = *p;
108                 entry = rb_entry(parent, struct inode_defrag, rb_node);
109
110                 ret = __compare_inode_defrag(defrag, entry);
111                 if (ret < 0)
112                         p = &parent->rb_left;
113                 else if (ret > 0)
114                         p = &parent->rb_right;
115                 else {
116                         /* if we're reinserting an entry for
117                          * an old defrag run, make sure to
118                          * lower the transid of our existing record
119                          */
120                         if (defrag->transid < entry->transid)
121                                 entry->transid = defrag->transid;
122                         if (defrag->last_offset > entry->last_offset)
123                                 entry->last_offset = defrag->last_offset;
124                         return -EEXIST;
125                 }
126         }
127         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
128         rb_link_node(&defrag->rb_node, parent, p);
129         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
130         return 0;
131 }
132
133 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
134 {
135         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
136                 return 0;
137
138         if (btrfs_fs_closing(fs_info))
139                 return 0;
140
141         return 1;
142 }
143
144 /*
145  * insert a defrag record for this inode if auto defrag is
146  * enabled
147  */
148 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
149                            struct btrfs_inode *inode)
150 {
151         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
152         struct btrfs_root *root = inode->root;
153         struct inode_defrag *defrag;
154         u64 transid;
155         int ret;
156
157         if (!__need_auto_defrag(fs_info))
158                 return 0;
159
160         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
161                 return 0;
162
163         if (trans)
164                 transid = trans->transid;
165         else
166                 transid = inode->root->last_trans;
167
168         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
169         if (!defrag)
170                 return -ENOMEM;
171
172         defrag->ino = btrfs_ino(inode);
173         defrag->transid = transid;
174         defrag->root = root->root_key.objectid;
175
176         spin_lock(&fs_info->defrag_inodes_lock);
177         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
178                 /*
179                  * If we set IN_DEFRAG flag and evict the inode from memory,
180                  * and then re-read this inode, this new inode doesn't have
181                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
182                  */
183                 ret = __btrfs_add_inode_defrag(inode, defrag);
184                 if (ret)
185                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
186         } else {
187                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
188         }
189         spin_unlock(&fs_info->defrag_inodes_lock);
190         return 0;
191 }
192
193 /*
194  * Requeue the defrag object. If there is a defrag object that points to
195  * the same inode in the tree, we will merge them together (by
196  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
197  */
198 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
199                                        struct inode_defrag *defrag)
200 {
201         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
202         int ret;
203
204         if (!__need_auto_defrag(fs_info))
205                 goto out;
206
207         /*
208          * Here we don't check the IN_DEFRAG flag, because we need merge
209          * them together.
210          */
211         spin_lock(&fs_info->defrag_inodes_lock);
212         ret = __btrfs_add_inode_defrag(inode, defrag);
213         spin_unlock(&fs_info->defrag_inodes_lock);
214         if (ret)
215                 goto out;
216         return;
217 out:
218         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
219 }
220
221 /*
222  * pick the defragable inode that we want, if it doesn't exist, we will get
223  * the next one.
224  */
225 static struct inode_defrag *
226 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
227 {
228         struct inode_defrag *entry = NULL;
229         struct inode_defrag tmp;
230         struct rb_node *p;
231         struct rb_node *parent = NULL;
232         int ret;
233
234         tmp.ino = ino;
235         tmp.root = root;
236
237         spin_lock(&fs_info->defrag_inodes_lock);
238         p = fs_info->defrag_inodes.rb_node;
239         while (p) {
240                 parent = p;
241                 entry = rb_entry(parent, struct inode_defrag, rb_node);
242
243                 ret = __compare_inode_defrag(&tmp, entry);
244                 if (ret < 0)
245                         p = parent->rb_left;
246                 else if (ret > 0)
247                         p = parent->rb_right;
248                 else
249                         goto out;
250         }
251
252         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
253                 parent = rb_next(parent);
254                 if (parent)
255                         entry = rb_entry(parent, struct inode_defrag, rb_node);
256                 else
257                         entry = NULL;
258         }
259 out:
260         if (entry)
261                 rb_erase(parent, &fs_info->defrag_inodes);
262         spin_unlock(&fs_info->defrag_inodes_lock);
263         return entry;
264 }
265
266 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
267 {
268         struct inode_defrag *defrag;
269         struct rb_node *node;
270
271         spin_lock(&fs_info->defrag_inodes_lock);
272         node = rb_first(&fs_info->defrag_inodes);
273         while (node) {
274                 rb_erase(node, &fs_info->defrag_inodes);
275                 defrag = rb_entry(node, struct inode_defrag, rb_node);
276                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
277
278                 cond_resched_lock(&fs_info->defrag_inodes_lock);
279
280                 node = rb_first(&fs_info->defrag_inodes);
281         }
282         spin_unlock(&fs_info->defrag_inodes_lock);
283 }
284
285 #define BTRFS_DEFRAG_BATCH      1024
286
287 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
288                                     struct inode_defrag *defrag)
289 {
290         struct btrfs_root *inode_root;
291         struct inode *inode;
292         struct btrfs_key key;
293         struct btrfs_ioctl_defrag_range_args range;
294         int num_defrag;
295         int index;
296         int ret;
297
298         /* get the inode */
299         key.objectid = defrag->root;
300         key.type = BTRFS_ROOT_ITEM_KEY;
301         key.offset = (u64)-1;
302
303         index = srcu_read_lock(&fs_info->subvol_srcu);
304
305         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
306         if (IS_ERR(inode_root)) {
307                 ret = PTR_ERR(inode_root);
308                 goto cleanup;
309         }
310
311         key.objectid = defrag->ino;
312         key.type = BTRFS_INODE_ITEM_KEY;
313         key.offset = 0;
314         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
315         if (IS_ERR(inode)) {
316                 ret = PTR_ERR(inode);
317                 goto cleanup;
318         }
319         srcu_read_unlock(&fs_info->subvol_srcu, index);
320
321         /* do a chunk of defrag */
322         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
323         memset(&range, 0, sizeof(range));
324         range.len = (u64)-1;
325         range.start = defrag->last_offset;
326
327         sb_start_write(fs_info->sb);
328         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
329                                        BTRFS_DEFRAG_BATCH);
330         sb_end_write(fs_info->sb);
331         /*
332          * if we filled the whole defrag batch, there
333          * must be more work to do.  Queue this defrag
334          * again
335          */
336         if (num_defrag == BTRFS_DEFRAG_BATCH) {
337                 defrag->last_offset = range.start;
338                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
339         } else if (defrag->last_offset && !defrag->cycled) {
340                 /*
341                  * we didn't fill our defrag batch, but
342                  * we didn't start at zero.  Make sure we loop
343                  * around to the start of the file.
344                  */
345                 defrag->last_offset = 0;
346                 defrag->cycled = 1;
347                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
348         } else {
349                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
350         }
351
352         iput(inode);
353         return 0;
354 cleanup:
355         srcu_read_unlock(&fs_info->subvol_srcu, index);
356         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
357         return ret;
358 }
359
360 /*
361  * run through the list of inodes in the FS that need
362  * defragging
363  */
364 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
365 {
366         struct inode_defrag *defrag;
367         u64 first_ino = 0;
368         u64 root_objectid = 0;
369
370         atomic_inc(&fs_info->defrag_running);
371         while (1) {
372                 /* Pause the auto defragger. */
373                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
374                              &fs_info->fs_state))
375                         break;
376
377                 if (!__need_auto_defrag(fs_info))
378                         break;
379
380                 /* find an inode to defrag */
381                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
382                                                  first_ino);
383                 if (!defrag) {
384                         if (root_objectid || first_ino) {
385                                 root_objectid = 0;
386                                 first_ino = 0;
387                                 continue;
388                         } else {
389                                 break;
390                         }
391                 }
392
393                 first_ino = defrag->ino + 1;
394                 root_objectid = defrag->root;
395
396                 __btrfs_run_defrag_inode(fs_info, defrag);
397         }
398         atomic_dec(&fs_info->defrag_running);
399
400         /*
401          * during unmount, we use the transaction_wait queue to
402          * wait for the defragger to stop
403          */
404         wake_up(&fs_info->transaction_wait);
405         return 0;
406 }
407
408 /* simple helper to fault in pages and copy.  This should go away
409  * and be replaced with calls into generic code.
410  */
411 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
412                                          struct page **prepared_pages,
413                                          struct iov_iter *i)
414 {
415         size_t copied = 0;
416         size_t total_copied = 0;
417         int pg = 0;
418         int offset = pos & (PAGE_SIZE - 1);
419
420         while (write_bytes > 0) {
421                 size_t count = min_t(size_t,
422                                      PAGE_SIZE - offset, write_bytes);
423                 struct page *page = prepared_pages[pg];
424                 /*
425                  * Copy data from userspace to the current page
426                  */
427                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
428
429                 /* Flush processor's dcache for this page */
430                 flush_dcache_page(page);
431
432                 /*
433                  * if we get a partial write, we can end up with
434                  * partially up to date pages.  These add
435                  * a lot of complexity, so make sure they don't
436                  * happen by forcing this copy to be retried.
437                  *
438                  * The rest of the btrfs_file_write code will fall
439                  * back to page at a time copies after we return 0.
440                  */
441                 if (!PageUptodate(page) && copied < count)
442                         copied = 0;
443
444                 iov_iter_advance(i, copied);
445                 write_bytes -= copied;
446                 total_copied += copied;
447
448                 /* Return to btrfs_file_write_iter to fault page */
449                 if (unlikely(copied == 0))
450                         break;
451
452                 if (copied < PAGE_SIZE - offset) {
453                         offset += copied;
454                 } else {
455                         pg++;
456                         offset = 0;
457                 }
458         }
459         return total_copied;
460 }
461
462 /*
463  * unlocks pages after btrfs_file_write is done with them
464  */
465 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
466 {
467         size_t i;
468         for (i = 0; i < num_pages; i++) {
469                 /* page checked is some magic around finding pages that
470                  * have been modified without going through btrfs_set_page_dirty
471                  * clear it here. There should be no need to mark the pages
472                  * accessed as prepare_pages should have marked them accessed
473                  * in prepare_pages via find_or_create_page()
474                  */
475                 ClearPageChecked(pages[i]);
476                 unlock_page(pages[i]);
477                 put_page(pages[i]);
478         }
479 }
480
481 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
482                                          const u64 start,
483                                          const u64 len,
484                                          struct extent_state **cached_state)
485 {
486         u64 search_start = start;
487         const u64 end = start + len - 1;
488
489         while (search_start < end) {
490                 const u64 search_len = end - search_start + 1;
491                 struct extent_map *em;
492                 u64 em_len;
493                 int ret = 0;
494
495                 em = btrfs_get_extent(inode, NULL, 0, search_start,
496                                       search_len, 0);
497                 if (IS_ERR(em))
498                         return PTR_ERR(em);
499
500                 if (em->block_start != EXTENT_MAP_HOLE)
501                         goto next;
502
503                 em_len = em->len;
504                 if (em->start < search_start)
505                         em_len -= search_start - em->start;
506                 if (em_len > search_len)
507                         em_len = search_len;
508
509                 ret = set_extent_bit(&inode->io_tree, search_start,
510                                      search_start + em_len - 1,
511                                      EXTENT_DELALLOC_NEW,
512                                      NULL, cached_state, GFP_NOFS);
513 next:
514                 search_start = extent_map_end(em);
515                 free_extent_map(em);
516                 if (ret)
517                         return ret;
518         }
519         return 0;
520 }
521
522 /*
523  * after copy_from_user, pages need to be dirtied and we need to make
524  * sure holes are created between the current EOF and the start of
525  * any next extents (if required).
526  *
527  * this also makes the decision about creating an inline extent vs
528  * doing real data extents, marking pages dirty and delalloc as required.
529  */
530 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
531                       size_t num_pages, loff_t pos, size_t write_bytes,
532                       struct extent_state **cached)
533 {
534         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
535         int err = 0;
536         int i;
537         u64 num_bytes;
538         u64 start_pos;
539         u64 end_of_last_block;
540         u64 end_pos = pos + write_bytes;
541         loff_t isize = i_size_read(inode);
542         unsigned int extra_bits = 0;
543
544         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
545         num_bytes = round_up(write_bytes + pos - start_pos,
546                              fs_info->sectorsize);
547
548         end_of_last_block = start_pos + num_bytes - 1;
549
550         if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
551                 if (start_pos >= isize &&
552                     !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
553                         /*
554                          * There can't be any extents following eof in this case
555                          * so just set the delalloc new bit for the range
556                          * directly.
557                          */
558                         extra_bits |= EXTENT_DELALLOC_NEW;
559                 } else {
560                         err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
561                                                             start_pos,
562                                                             num_bytes, cached);
563                         if (err)
564                                 return err;
565                 }
566         }
567
568         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
569                                         extra_bits, cached, 0);
570         if (err)
571                 return err;
572
573         for (i = 0; i < num_pages; i++) {
574                 struct page *p = pages[i];
575                 SetPageUptodate(p);
576                 ClearPageChecked(p);
577                 set_page_dirty(p);
578         }
579
580         /*
581          * we've only changed i_size in ram, and we haven't updated
582          * the disk i_size.  There is no need to log the inode
583          * at this time.
584          */
585         if (end_pos > isize)
586                 i_size_write(inode, end_pos);
587         return 0;
588 }
589
590 /*
591  * this drops all the extents in the cache that intersect the range
592  * [start, end].  Existing extents are split as required.
593  */
594 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
595                              int skip_pinned)
596 {
597         struct extent_map *em;
598         struct extent_map *split = NULL;
599         struct extent_map *split2 = NULL;
600         struct extent_map_tree *em_tree = &inode->extent_tree;
601         u64 len = end - start + 1;
602         u64 gen;
603         int ret;
604         int testend = 1;
605         unsigned long flags;
606         int compressed = 0;
607         bool modified;
608
609         WARN_ON(end < start);
610         if (end == (u64)-1) {
611                 len = (u64)-1;
612                 testend = 0;
613         }
614         while (1) {
615                 int no_splits = 0;
616
617                 modified = false;
618                 if (!split)
619                         split = alloc_extent_map();
620                 if (!split2)
621                         split2 = alloc_extent_map();
622                 if (!split || !split2)
623                         no_splits = 1;
624
625                 write_lock(&em_tree->lock);
626                 em = lookup_extent_mapping(em_tree, start, len);
627                 if (!em) {
628                         write_unlock(&em_tree->lock);
629                         break;
630                 }
631                 flags = em->flags;
632                 gen = em->generation;
633                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
634                         if (testend && em->start + em->len >= start + len) {
635                                 free_extent_map(em);
636                                 write_unlock(&em_tree->lock);
637                                 break;
638                         }
639                         start = em->start + em->len;
640                         if (testend)
641                                 len = start + len - (em->start + em->len);
642                         free_extent_map(em);
643                         write_unlock(&em_tree->lock);
644                         continue;
645                 }
646                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
647                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
648                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
649                 modified = !list_empty(&em->list);
650                 if (no_splits)
651                         goto next;
652
653                 if (em->start < start) {
654                         split->start = em->start;
655                         split->len = start - em->start;
656
657                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
658                                 split->orig_start = em->orig_start;
659                                 split->block_start = em->block_start;
660
661                                 if (compressed)
662                                         split->block_len = em->block_len;
663                                 else
664                                         split->block_len = split->len;
665                                 split->orig_block_len = max(split->block_len,
666                                                 em->orig_block_len);
667                                 split->ram_bytes = em->ram_bytes;
668                         } else {
669                                 split->orig_start = split->start;
670                                 split->block_len = 0;
671                                 split->block_start = em->block_start;
672                                 split->orig_block_len = 0;
673                                 split->ram_bytes = split->len;
674                         }
675
676                         split->generation = gen;
677                         split->bdev = em->bdev;
678                         split->flags = flags;
679                         split->compress_type = em->compress_type;
680                         replace_extent_mapping(em_tree, em, split, modified);
681                         free_extent_map(split);
682                         split = split2;
683                         split2 = NULL;
684                 }
685                 if (testend && em->start + em->len > start + len) {
686                         u64 diff = start + len - em->start;
687
688                         split->start = start + len;
689                         split->len = em->start + em->len - (start + len);
690                         split->bdev = em->bdev;
691                         split->flags = flags;
692                         split->compress_type = em->compress_type;
693                         split->generation = gen;
694
695                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
696                                 split->orig_block_len = max(em->block_len,
697                                                     em->orig_block_len);
698
699                                 split->ram_bytes = em->ram_bytes;
700                                 if (compressed) {
701                                         split->block_len = em->block_len;
702                                         split->block_start = em->block_start;
703                                         split->orig_start = em->orig_start;
704                                 } else {
705                                         split->block_len = split->len;
706                                         split->block_start = em->block_start
707                                                 + diff;
708                                         split->orig_start = em->orig_start;
709                                 }
710                         } else {
711                                 split->ram_bytes = split->len;
712                                 split->orig_start = split->start;
713                                 split->block_len = 0;
714                                 split->block_start = em->block_start;
715                                 split->orig_block_len = 0;
716                         }
717
718                         if (extent_map_in_tree(em)) {
719                                 replace_extent_mapping(em_tree, em, split,
720                                                        modified);
721                         } else {
722                                 ret = add_extent_mapping(em_tree, split,
723                                                          modified);
724                                 ASSERT(ret == 0); /* Logic error */
725                         }
726                         free_extent_map(split);
727                         split = NULL;
728                 }
729 next:
730                 if (extent_map_in_tree(em))
731                         remove_extent_mapping(em_tree, em);
732                 write_unlock(&em_tree->lock);
733
734                 /* once for us */
735                 free_extent_map(em);
736                 /* once for the tree*/
737                 free_extent_map(em);
738         }
739         if (split)
740                 free_extent_map(split);
741         if (split2)
742                 free_extent_map(split2);
743 }
744
745 /*
746  * this is very complex, but the basic idea is to drop all extents
747  * in the range start - end.  hint_block is filled in with a block number
748  * that would be a good hint to the block allocator for this file.
749  *
750  * If an extent intersects the range but is not entirely inside the range
751  * it is either truncated or split.  Anything entirely inside the range
752  * is deleted from the tree.
753  */
754 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
755                          struct btrfs_root *root, struct inode *inode,
756                          struct btrfs_path *path, u64 start, u64 end,
757                          u64 *drop_end, int drop_cache,
758                          int replace_extent,
759                          u32 extent_item_size,
760                          int *key_inserted)
761 {
762         struct btrfs_fs_info *fs_info = root->fs_info;
763         struct extent_buffer *leaf;
764         struct btrfs_file_extent_item *fi;
765         struct btrfs_key key;
766         struct btrfs_key new_key;
767         u64 ino = btrfs_ino(BTRFS_I(inode));
768         u64 search_start = start;
769         u64 disk_bytenr = 0;
770         u64 num_bytes = 0;
771         u64 extent_offset = 0;
772         u64 extent_end = 0;
773         u64 last_end = start;
774         int del_nr = 0;
775         int del_slot = 0;
776         int extent_type;
777         int recow;
778         int ret;
779         int modify_tree = -1;
780         int update_refs;
781         int found = 0;
782         int leafs_visited = 0;
783
784         if (drop_cache)
785                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
786
787         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
788                 modify_tree = 0;
789
790         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
791                        root == fs_info->tree_root);
792         while (1) {
793                 recow = 0;
794                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
795                                                search_start, modify_tree);
796                 if (ret < 0)
797                         break;
798                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
799                         leaf = path->nodes[0];
800                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
801                         if (key.objectid == ino &&
802                             key.type == BTRFS_EXTENT_DATA_KEY)
803                                 path->slots[0]--;
804                 }
805                 ret = 0;
806                 leafs_visited++;
807 next_slot:
808                 leaf = path->nodes[0];
809                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
810                         BUG_ON(del_nr > 0);
811                         ret = btrfs_next_leaf(root, path);
812                         if (ret < 0)
813                                 break;
814                         if (ret > 0) {
815                                 ret = 0;
816                                 break;
817                         }
818                         leafs_visited++;
819                         leaf = path->nodes[0];
820                         recow = 1;
821                 }
822
823                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
824
825                 if (key.objectid > ino)
826                         break;
827                 if (WARN_ON_ONCE(key.objectid < ino) ||
828                     key.type < BTRFS_EXTENT_DATA_KEY) {
829                         ASSERT(del_nr == 0);
830                         path->slots[0]++;
831                         goto next_slot;
832                 }
833                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
834                         break;
835
836                 fi = btrfs_item_ptr(leaf, path->slots[0],
837                                     struct btrfs_file_extent_item);
838                 extent_type = btrfs_file_extent_type(leaf, fi);
839
840                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
841                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
842                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
843                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
844                         extent_offset = btrfs_file_extent_offset(leaf, fi);
845                         extent_end = key.offset +
846                                 btrfs_file_extent_num_bytes(leaf, fi);
847                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
848                         extent_end = key.offset +
849                                 btrfs_file_extent_inline_len(leaf,
850                                                      path->slots[0], fi);
851                 } else {
852                         /* can't happen */
853                         BUG();
854                 }
855
856                 /*
857                  * Don't skip extent items representing 0 byte lengths. They
858                  * used to be created (bug) if while punching holes we hit
859                  * -ENOSPC condition. So if we find one here, just ensure we
860                  * delete it, otherwise we would insert a new file extent item
861                  * with the same key (offset) as that 0 bytes length file
862                  * extent item in the call to setup_items_for_insert() later
863                  * in this function.
864                  */
865                 if (extent_end == key.offset && extent_end >= search_start) {
866                         last_end = extent_end;
867                         goto delete_extent_item;
868                 }
869
870                 if (extent_end <= search_start) {
871                         path->slots[0]++;
872                         goto next_slot;
873                 }
874
875                 found = 1;
876                 search_start = max(key.offset, start);
877                 if (recow || !modify_tree) {
878                         modify_tree = -1;
879                         btrfs_release_path(path);
880                         continue;
881                 }
882
883                 /*
884                  *     | - range to drop - |
885                  *  | -------- extent -------- |
886                  */
887                 if (start > key.offset && end < extent_end) {
888                         BUG_ON(del_nr > 0);
889                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
890                                 ret = -EOPNOTSUPP;
891                                 break;
892                         }
893
894                         memcpy(&new_key, &key, sizeof(new_key));
895                         new_key.offset = start;
896                         ret = btrfs_duplicate_item(trans, root, path,
897                                                    &new_key);
898                         if (ret == -EAGAIN) {
899                                 btrfs_release_path(path);
900                                 continue;
901                         }
902                         if (ret < 0)
903                                 break;
904
905                         leaf = path->nodes[0];
906                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
907                                             struct btrfs_file_extent_item);
908                         btrfs_set_file_extent_num_bytes(leaf, fi,
909                                                         start - key.offset);
910
911                         fi = btrfs_item_ptr(leaf, path->slots[0],
912                                             struct btrfs_file_extent_item);
913
914                         extent_offset += start - key.offset;
915                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
916                         btrfs_set_file_extent_num_bytes(leaf, fi,
917                                                         extent_end - start);
918                         btrfs_mark_buffer_dirty(leaf);
919
920                         if (update_refs && disk_bytenr > 0) {
921                                 ret = btrfs_inc_extent_ref(trans, root,
922                                                 disk_bytenr, num_bytes, 0,
923                                                 root->root_key.objectid,
924                                                 new_key.objectid,
925                                                 start - extent_offset);
926                                 BUG_ON(ret); /* -ENOMEM */
927                         }
928                         key.offset = start;
929                 }
930                 /*
931                  * From here on out we will have actually dropped something, so
932                  * last_end can be updated.
933                  */
934                 last_end = extent_end;
935
936                 /*
937                  *  | ---- range to drop ----- |
938                  *      | -------- extent -------- |
939                  */
940                 if (start <= key.offset && end < extent_end) {
941                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
942                                 ret = -EOPNOTSUPP;
943                                 break;
944                         }
945
946                         memcpy(&new_key, &key, sizeof(new_key));
947                         new_key.offset = end;
948                         btrfs_set_item_key_safe(fs_info, path, &new_key);
949
950                         extent_offset += end - key.offset;
951                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
952                         btrfs_set_file_extent_num_bytes(leaf, fi,
953                                                         extent_end - end);
954                         btrfs_mark_buffer_dirty(leaf);
955                         if (update_refs && disk_bytenr > 0)
956                                 inode_sub_bytes(inode, end - key.offset);
957                         break;
958                 }
959
960                 search_start = extent_end;
961                 /*
962                  *       | ---- range to drop ----- |
963                  *  | -------- extent -------- |
964                  */
965                 if (start > key.offset && end >= extent_end) {
966                         BUG_ON(del_nr > 0);
967                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
968                                 ret = -EOPNOTSUPP;
969                                 break;
970                         }
971
972                         btrfs_set_file_extent_num_bytes(leaf, fi,
973                                                         start - key.offset);
974                         btrfs_mark_buffer_dirty(leaf);
975                         if (update_refs && disk_bytenr > 0)
976                                 inode_sub_bytes(inode, extent_end - start);
977                         if (end == extent_end)
978                                 break;
979
980                         path->slots[0]++;
981                         goto next_slot;
982                 }
983
984                 /*
985                  *  | ---- range to drop ----- |
986                  *    | ------ extent ------ |
987                  */
988                 if (start <= key.offset && end >= extent_end) {
989 delete_extent_item:
990                         if (del_nr == 0) {
991                                 del_slot = path->slots[0];
992                                 del_nr = 1;
993                         } else {
994                                 BUG_ON(del_slot + del_nr != path->slots[0]);
995                                 del_nr++;
996                         }
997
998                         if (update_refs &&
999                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
1000                                 inode_sub_bytes(inode,
1001                                                 extent_end - key.offset);
1002                                 extent_end = ALIGN(extent_end,
1003                                                    fs_info->sectorsize);
1004                         } else if (update_refs && disk_bytenr > 0) {
1005                                 ret = btrfs_free_extent(trans, root,
1006                                                 disk_bytenr, num_bytes, 0,
1007                                                 root->root_key.objectid,
1008                                                 key.objectid, key.offset -
1009                                                 extent_offset);
1010                                 BUG_ON(ret); /* -ENOMEM */
1011                                 inode_sub_bytes(inode,
1012                                                 extent_end - key.offset);
1013                         }
1014
1015                         if (end == extent_end)
1016                                 break;
1017
1018                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1019                                 path->slots[0]++;
1020                                 goto next_slot;
1021                         }
1022
1023                         ret = btrfs_del_items(trans, root, path, del_slot,
1024                                               del_nr);
1025                         if (ret) {
1026                                 btrfs_abort_transaction(trans, ret);
1027                                 break;
1028                         }
1029
1030                         del_nr = 0;
1031                         del_slot = 0;
1032
1033                         btrfs_release_path(path);
1034                         continue;
1035                 }
1036
1037                 BUG_ON(1);
1038         }
1039
1040         if (!ret && del_nr > 0) {
1041                 /*
1042                  * Set path->slots[0] to first slot, so that after the delete
1043                  * if items are move off from our leaf to its immediate left or
1044                  * right neighbor leafs, we end up with a correct and adjusted
1045                  * path->slots[0] for our insertion (if replace_extent != 0).
1046                  */
1047                 path->slots[0] = del_slot;
1048                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1049                 if (ret)
1050                         btrfs_abort_transaction(trans, ret);
1051         }
1052
1053         leaf = path->nodes[0];
1054         /*
1055          * If btrfs_del_items() was called, it might have deleted a leaf, in
1056          * which case it unlocked our path, so check path->locks[0] matches a
1057          * write lock.
1058          */
1059         if (!ret && replace_extent && leafs_visited == 1 &&
1060             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1061              path->locks[0] == BTRFS_WRITE_LOCK) &&
1062             btrfs_leaf_free_space(fs_info, leaf) >=
1063             sizeof(struct btrfs_item) + extent_item_size) {
1064
1065                 key.objectid = ino;
1066                 key.type = BTRFS_EXTENT_DATA_KEY;
1067                 key.offset = start;
1068                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1069                         struct btrfs_key slot_key;
1070
1071                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1072                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1073                                 path->slots[0]++;
1074                 }
1075                 setup_items_for_insert(root, path, &key,
1076                                        &extent_item_size,
1077                                        extent_item_size,
1078                                        sizeof(struct btrfs_item) +
1079                                        extent_item_size, 1);
1080                 *key_inserted = 1;
1081         }
1082
1083         if (!replace_extent || !(*key_inserted))
1084                 btrfs_release_path(path);
1085         if (drop_end)
1086                 *drop_end = found ? min(end, last_end) : end;
1087         return ret;
1088 }
1089
1090 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1091                        struct btrfs_root *root, struct inode *inode, u64 start,
1092                        u64 end, int drop_cache)
1093 {
1094         struct btrfs_path *path;
1095         int ret;
1096
1097         path = btrfs_alloc_path();
1098         if (!path)
1099                 return -ENOMEM;
1100         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1101                                    drop_cache, 0, 0, NULL);
1102         btrfs_free_path(path);
1103         return ret;
1104 }
1105
1106 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1107                             u64 objectid, u64 bytenr, u64 orig_offset,
1108                             u64 *start, u64 *end)
1109 {
1110         struct btrfs_file_extent_item *fi;
1111         struct btrfs_key key;
1112         u64 extent_end;
1113
1114         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1115                 return 0;
1116
1117         btrfs_item_key_to_cpu(leaf, &key, slot);
1118         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1119                 return 0;
1120
1121         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1122         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1123             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1124             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1125             btrfs_file_extent_compression(leaf, fi) ||
1126             btrfs_file_extent_encryption(leaf, fi) ||
1127             btrfs_file_extent_other_encoding(leaf, fi))
1128                 return 0;
1129
1130         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1131         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1132                 return 0;
1133
1134         *start = key.offset;
1135         *end = extent_end;
1136         return 1;
1137 }
1138
1139 /*
1140  * Mark extent in the range start - end as written.
1141  *
1142  * This changes extent type from 'pre-allocated' to 'regular'. If only
1143  * part of extent is marked as written, the extent will be split into
1144  * two or three.
1145  */
1146 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1147                               struct btrfs_inode *inode, u64 start, u64 end)
1148 {
1149         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1150         struct btrfs_root *root = inode->root;
1151         struct extent_buffer *leaf;
1152         struct btrfs_path *path;
1153         struct btrfs_file_extent_item *fi;
1154         struct btrfs_key key;
1155         struct btrfs_key new_key;
1156         u64 bytenr;
1157         u64 num_bytes;
1158         u64 extent_end;
1159         u64 orig_offset;
1160         u64 other_start;
1161         u64 other_end;
1162         u64 split;
1163         int del_nr = 0;
1164         int del_slot = 0;
1165         int recow;
1166         int ret;
1167         u64 ino = btrfs_ino(inode);
1168
1169         path = btrfs_alloc_path();
1170         if (!path)
1171                 return -ENOMEM;
1172 again:
1173         recow = 0;
1174         split = start;
1175         key.objectid = ino;
1176         key.type = BTRFS_EXTENT_DATA_KEY;
1177         key.offset = split;
1178
1179         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1180         if (ret < 0)
1181                 goto out;
1182         if (ret > 0 && path->slots[0] > 0)
1183                 path->slots[0]--;
1184
1185         leaf = path->nodes[0];
1186         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1187         if (key.objectid != ino ||
1188             key.type != BTRFS_EXTENT_DATA_KEY) {
1189                 ret = -EINVAL;
1190                 btrfs_abort_transaction(trans, ret);
1191                 goto out;
1192         }
1193         fi = btrfs_item_ptr(leaf, path->slots[0],
1194                             struct btrfs_file_extent_item);
1195         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1196                 ret = -EINVAL;
1197                 btrfs_abort_transaction(trans, ret);
1198                 goto out;
1199         }
1200         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1201         if (key.offset > start || extent_end < end) {
1202                 ret = -EINVAL;
1203                 btrfs_abort_transaction(trans, ret);
1204                 goto out;
1205         }
1206
1207         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1208         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1209         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1210         memcpy(&new_key, &key, sizeof(new_key));
1211
1212         if (start == key.offset && end < extent_end) {
1213                 other_start = 0;
1214                 other_end = start;
1215                 if (extent_mergeable(leaf, path->slots[0] - 1,
1216                                      ino, bytenr, orig_offset,
1217                                      &other_start, &other_end)) {
1218                         new_key.offset = end;
1219                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1220                         fi = btrfs_item_ptr(leaf, path->slots[0],
1221                                             struct btrfs_file_extent_item);
1222                         btrfs_set_file_extent_generation(leaf, fi,
1223                                                          trans->transid);
1224                         btrfs_set_file_extent_num_bytes(leaf, fi,
1225                                                         extent_end - end);
1226                         btrfs_set_file_extent_offset(leaf, fi,
1227                                                      end - orig_offset);
1228                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1229                                             struct btrfs_file_extent_item);
1230                         btrfs_set_file_extent_generation(leaf, fi,
1231                                                          trans->transid);
1232                         btrfs_set_file_extent_num_bytes(leaf, fi,
1233                                                         end - other_start);
1234                         btrfs_mark_buffer_dirty(leaf);
1235                         goto out;
1236                 }
1237         }
1238
1239         if (start > key.offset && end == extent_end) {
1240                 other_start = end;
1241                 other_end = 0;
1242                 if (extent_mergeable(leaf, path->slots[0] + 1,
1243                                      ino, bytenr, orig_offset,
1244                                      &other_start, &other_end)) {
1245                         fi = btrfs_item_ptr(leaf, path->slots[0],
1246                                             struct btrfs_file_extent_item);
1247                         btrfs_set_file_extent_num_bytes(leaf, fi,
1248                                                         start - key.offset);
1249                         btrfs_set_file_extent_generation(leaf, fi,
1250                                                          trans->transid);
1251                         path->slots[0]++;
1252                         new_key.offset = start;
1253                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1254
1255                         fi = btrfs_item_ptr(leaf, path->slots[0],
1256                                             struct btrfs_file_extent_item);
1257                         btrfs_set_file_extent_generation(leaf, fi,
1258                                                          trans->transid);
1259                         btrfs_set_file_extent_num_bytes(leaf, fi,
1260                                                         other_end - start);
1261                         btrfs_set_file_extent_offset(leaf, fi,
1262                                                      start - orig_offset);
1263                         btrfs_mark_buffer_dirty(leaf);
1264                         goto out;
1265                 }
1266         }
1267
1268         while (start > key.offset || end < extent_end) {
1269                 if (key.offset == start)
1270                         split = end;
1271
1272                 new_key.offset = split;
1273                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1274                 if (ret == -EAGAIN) {
1275                         btrfs_release_path(path);
1276                         goto again;
1277                 }
1278                 if (ret < 0) {
1279                         btrfs_abort_transaction(trans, ret);
1280                         goto out;
1281                 }
1282
1283                 leaf = path->nodes[0];
1284                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1285                                     struct btrfs_file_extent_item);
1286                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1287                 btrfs_set_file_extent_num_bytes(leaf, fi,
1288                                                 split - key.offset);
1289
1290                 fi = btrfs_item_ptr(leaf, path->slots[0],
1291                                     struct btrfs_file_extent_item);
1292
1293                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1294                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1295                 btrfs_set_file_extent_num_bytes(leaf, fi,
1296                                                 extent_end - split);
1297                 btrfs_mark_buffer_dirty(leaf);
1298
1299                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes,
1300                                            0, root->root_key.objectid,
1301                                            ino, orig_offset);
1302                 if (ret) {
1303                         btrfs_abort_transaction(trans, ret);
1304                         goto out;
1305                 }
1306
1307                 if (split == start) {
1308                         key.offset = start;
1309                 } else {
1310                         if (start != key.offset) {
1311                                 ret = -EINVAL;
1312                                 btrfs_abort_transaction(trans, ret);
1313                                 goto out;
1314                         }
1315                         path->slots[0]--;
1316                         extent_end = end;
1317                 }
1318                 recow = 1;
1319         }
1320
1321         other_start = end;
1322         other_end = 0;
1323         if (extent_mergeable(leaf, path->slots[0] + 1,
1324                              ino, bytenr, orig_offset,
1325                              &other_start, &other_end)) {
1326                 if (recow) {
1327                         btrfs_release_path(path);
1328                         goto again;
1329                 }
1330                 extent_end = other_end;
1331                 del_slot = path->slots[0] + 1;
1332                 del_nr++;
1333                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1334                                         0, root->root_key.objectid,
1335                                         ino, orig_offset);
1336                 if (ret) {
1337                         btrfs_abort_transaction(trans, ret);
1338                         goto out;
1339                 }
1340         }
1341         other_start = 0;
1342         other_end = start;
1343         if (extent_mergeable(leaf, path->slots[0] - 1,
1344                              ino, bytenr, orig_offset,
1345                              &other_start, &other_end)) {
1346                 if (recow) {
1347                         btrfs_release_path(path);
1348                         goto again;
1349                 }
1350                 key.offset = other_start;
1351                 del_slot = path->slots[0];
1352                 del_nr++;
1353                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1354                                         0, root->root_key.objectid,
1355                                         ino, orig_offset);
1356                 if (ret) {
1357                         btrfs_abort_transaction(trans, ret);
1358                         goto out;
1359                 }
1360         }
1361         if (del_nr == 0) {
1362                 fi = btrfs_item_ptr(leaf, path->slots[0],
1363                            struct btrfs_file_extent_item);
1364                 btrfs_set_file_extent_type(leaf, fi,
1365                                            BTRFS_FILE_EXTENT_REG);
1366                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1367                 btrfs_mark_buffer_dirty(leaf);
1368         } else {
1369                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1370                            struct btrfs_file_extent_item);
1371                 btrfs_set_file_extent_type(leaf, fi,
1372                                            BTRFS_FILE_EXTENT_REG);
1373                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1374                 btrfs_set_file_extent_num_bytes(leaf, fi,
1375                                                 extent_end - key.offset);
1376                 btrfs_mark_buffer_dirty(leaf);
1377
1378                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1379                 if (ret < 0) {
1380                         btrfs_abort_transaction(trans, ret);
1381                         goto out;
1382                 }
1383         }
1384 out:
1385         btrfs_free_path(path);
1386         return 0;
1387 }
1388
1389 /*
1390  * on error we return an unlocked page and the error value
1391  * on success we return a locked page and 0
1392  */
1393 static int prepare_uptodate_page(struct inode *inode,
1394                                  struct page *page, u64 pos,
1395                                  bool force_uptodate)
1396 {
1397         int ret = 0;
1398
1399         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1400             !PageUptodate(page)) {
1401                 ret = btrfs_readpage(NULL, page);
1402                 if (ret)
1403                         return ret;
1404                 lock_page(page);
1405                 if (!PageUptodate(page)) {
1406                         unlock_page(page);
1407                         return -EIO;
1408                 }
1409                 if (page->mapping != inode->i_mapping) {
1410                         unlock_page(page);
1411                         return -EAGAIN;
1412                 }
1413         }
1414         return 0;
1415 }
1416
1417 /*
1418  * this just gets pages into the page cache and locks them down.
1419  */
1420 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1421                                   size_t num_pages, loff_t pos,
1422                                   size_t write_bytes, bool force_uptodate)
1423 {
1424         int i;
1425         unsigned long index = pos >> PAGE_SHIFT;
1426         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1427         int err = 0;
1428         int faili;
1429
1430         for (i = 0; i < num_pages; i++) {
1431 again:
1432                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1433                                                mask | __GFP_WRITE);
1434                 if (!pages[i]) {
1435                         faili = i - 1;
1436                         err = -ENOMEM;
1437                         goto fail;
1438                 }
1439
1440                 if (i == 0)
1441                         err = prepare_uptodate_page(inode, pages[i], pos,
1442                                                     force_uptodate);
1443                 if (!err && i == num_pages - 1)
1444                         err = prepare_uptodate_page(inode, pages[i],
1445                                                     pos + write_bytes, false);
1446                 if (err) {
1447                         put_page(pages[i]);
1448                         if (err == -EAGAIN) {
1449                                 err = 0;
1450                                 goto again;
1451                         }
1452                         faili = i - 1;
1453                         goto fail;
1454                 }
1455                 wait_on_page_writeback(pages[i]);
1456         }
1457
1458         return 0;
1459 fail:
1460         while (faili >= 0) {
1461                 unlock_page(pages[faili]);
1462                 put_page(pages[faili]);
1463                 faili--;
1464         }
1465         return err;
1466
1467 }
1468
1469 /*
1470  * This function locks the extent and properly waits for data=ordered extents
1471  * to finish before allowing the pages to be modified if need.
1472  *
1473  * The return value:
1474  * 1 - the extent is locked
1475  * 0 - the extent is not locked, and everything is OK
1476  * -EAGAIN - need re-prepare the pages
1477  * the other < 0 number - Something wrong happens
1478  */
1479 static noinline int
1480 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1481                                 size_t num_pages, loff_t pos,
1482                                 size_t write_bytes,
1483                                 u64 *lockstart, u64 *lockend,
1484                                 struct extent_state **cached_state)
1485 {
1486         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1487         u64 start_pos;
1488         u64 last_pos;
1489         int i;
1490         int ret = 0;
1491
1492         start_pos = round_down(pos, fs_info->sectorsize);
1493         last_pos = start_pos
1494                 + round_up(pos + write_bytes - start_pos,
1495                            fs_info->sectorsize) - 1;
1496
1497         if (start_pos < inode->vfs_inode.i_size) {
1498                 struct btrfs_ordered_extent *ordered;
1499
1500                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1501                                 cached_state);
1502                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1503                                                      last_pos - start_pos + 1);
1504                 if (ordered &&
1505                     ordered->file_offset + ordered->len > start_pos &&
1506                     ordered->file_offset <= last_pos) {
1507                         unlock_extent_cached(&inode->io_tree, start_pos,
1508                                         last_pos, cached_state);
1509                         for (i = 0; i < num_pages; i++) {
1510                                 unlock_page(pages[i]);
1511                                 put_page(pages[i]);
1512                         }
1513                         btrfs_start_ordered_extent(&inode->vfs_inode,
1514                                         ordered, 1);
1515                         btrfs_put_ordered_extent(ordered);
1516                         return -EAGAIN;
1517                 }
1518                 if (ordered)
1519                         btrfs_put_ordered_extent(ordered);
1520                 clear_extent_bit(&inode->io_tree, start_pos, last_pos,
1521                                  EXTENT_DIRTY | EXTENT_DELALLOC |
1522                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1523                                  0, 0, cached_state);
1524                 *lockstart = start_pos;
1525                 *lockend = last_pos;
1526                 ret = 1;
1527         }
1528
1529         for (i = 0; i < num_pages; i++) {
1530                 if (clear_page_dirty_for_io(pages[i]))
1531                         account_page_redirty(pages[i]);
1532                 set_page_extent_mapped(pages[i]);
1533                 WARN_ON(!PageLocked(pages[i]));
1534         }
1535
1536         return ret;
1537 }
1538
1539 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1540                                     size_t *write_bytes)
1541 {
1542         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1543         struct btrfs_root *root = inode->root;
1544         struct btrfs_ordered_extent *ordered;
1545         u64 lockstart, lockend;
1546         u64 num_bytes;
1547         int ret;
1548
1549         ret = btrfs_start_write_no_snapshotting(root);
1550         if (!ret)
1551                 return -ENOSPC;
1552
1553         lockstart = round_down(pos, fs_info->sectorsize);
1554         lockend = round_up(pos + *write_bytes,
1555                            fs_info->sectorsize) - 1;
1556
1557         while (1) {
1558                 lock_extent(&inode->io_tree, lockstart, lockend);
1559                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1560                                                      lockend - lockstart + 1);
1561                 if (!ordered) {
1562                         break;
1563                 }
1564                 unlock_extent(&inode->io_tree, lockstart, lockend);
1565                 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1566                 btrfs_put_ordered_extent(ordered);
1567         }
1568
1569         num_bytes = lockend - lockstart + 1;
1570         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1571                         NULL, NULL, NULL);
1572         if (ret <= 0) {
1573                 ret = 0;
1574                 btrfs_end_write_no_snapshotting(root);
1575         } else {
1576                 *write_bytes = min_t(size_t, *write_bytes ,
1577                                      num_bytes - pos + lockstart);
1578         }
1579
1580         unlock_extent(&inode->io_tree, lockstart, lockend);
1581
1582         return ret;
1583 }
1584
1585 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1586                                                struct iov_iter *i,
1587                                                loff_t pos)
1588 {
1589         struct inode *inode = file_inode(file);
1590         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1591         struct btrfs_root *root = BTRFS_I(inode)->root;
1592         struct page **pages = NULL;
1593         struct extent_state *cached_state = NULL;
1594         struct extent_changeset *data_reserved = NULL;
1595         u64 release_bytes = 0;
1596         u64 lockstart;
1597         u64 lockend;
1598         size_t num_written = 0;
1599         int nrptrs;
1600         int ret = 0;
1601         bool only_release_metadata = false;
1602         bool force_page_uptodate = false;
1603
1604         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1605                         PAGE_SIZE / (sizeof(struct page *)));
1606         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1607         nrptrs = max(nrptrs, 8);
1608         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1609         if (!pages)
1610                 return -ENOMEM;
1611
1612         while (iov_iter_count(i) > 0) {
1613                 size_t offset = pos & (PAGE_SIZE - 1);
1614                 size_t sector_offset;
1615                 size_t write_bytes = min(iov_iter_count(i),
1616                                          nrptrs * (size_t)PAGE_SIZE -
1617                                          offset);
1618                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1619                                                 PAGE_SIZE);
1620                 size_t reserve_bytes;
1621                 size_t dirty_pages;
1622                 size_t copied;
1623                 size_t dirty_sectors;
1624                 size_t num_sectors;
1625                 int extents_locked;
1626
1627                 WARN_ON(num_pages > nrptrs);
1628
1629                 /*
1630                  * Fault pages before locking them in prepare_pages
1631                  * to avoid recursive lock
1632                  */
1633                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1634                         ret = -EFAULT;
1635                         break;
1636                 }
1637
1638                 sector_offset = pos & (fs_info->sectorsize - 1);
1639                 reserve_bytes = round_up(write_bytes + sector_offset,
1640                                 fs_info->sectorsize);
1641
1642                 extent_changeset_release(data_reserved);
1643                 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1644                                                   write_bytes);
1645                 if (ret < 0) {
1646                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1647                                                       BTRFS_INODE_PREALLOC)) &&
1648                             check_can_nocow(BTRFS_I(inode), pos,
1649                                         &write_bytes) > 0) {
1650                                 /*
1651                                  * For nodata cow case, no need to reserve
1652                                  * data space.
1653                                  */
1654                                 only_release_metadata = true;
1655                                 /*
1656                                  * our prealloc extent may be smaller than
1657                                  * write_bytes, so scale down.
1658                                  */
1659                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1660                                                          PAGE_SIZE);
1661                                 reserve_bytes = round_up(write_bytes +
1662                                                          sector_offset,
1663                                                          fs_info->sectorsize);
1664                         } else {
1665                                 break;
1666                         }
1667                 }
1668
1669                 WARN_ON(reserve_bytes == 0);
1670                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1671                                 reserve_bytes);
1672                 if (ret) {
1673                         if (!only_release_metadata)
1674                                 btrfs_free_reserved_data_space(inode,
1675                                                 data_reserved, pos,
1676                                                 write_bytes);
1677                         else
1678                                 btrfs_end_write_no_snapshotting(root);
1679                         break;
1680                 }
1681
1682                 release_bytes = reserve_bytes;
1683 again:
1684                 /*
1685                  * This is going to setup the pages array with the number of
1686                  * pages we want, so we don't really need to worry about the
1687                  * contents of pages from loop to loop
1688                  */
1689                 ret = prepare_pages(inode, pages, num_pages,
1690                                     pos, write_bytes,
1691                                     force_page_uptodate);
1692                 if (ret) {
1693                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1694                                                        reserve_bytes);
1695                         break;
1696                 }
1697
1698                 extents_locked = lock_and_cleanup_extent_if_need(
1699                                 BTRFS_I(inode), pages,
1700                                 num_pages, pos, write_bytes, &lockstart,
1701                                 &lockend, &cached_state);
1702                 if (extents_locked < 0) {
1703                         if (extents_locked == -EAGAIN)
1704                                 goto again;
1705                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1706                                                        reserve_bytes);
1707                         ret = extents_locked;
1708                         break;
1709                 }
1710
1711                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1712
1713                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1714                 dirty_sectors = round_up(copied + sector_offset,
1715                                         fs_info->sectorsize);
1716                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1717
1718                 /*
1719                  * if we have trouble faulting in the pages, fall
1720                  * back to one page at a time
1721                  */
1722                 if (copied < write_bytes)
1723                         nrptrs = 1;
1724
1725                 if (copied == 0) {
1726                         force_page_uptodate = true;
1727                         dirty_sectors = 0;
1728                         dirty_pages = 0;
1729                 } else {
1730                         force_page_uptodate = false;
1731                         dirty_pages = DIV_ROUND_UP(copied + offset,
1732                                                    PAGE_SIZE);
1733                 }
1734
1735                 if (num_sectors > dirty_sectors) {
1736                         /* release everything except the sectors we dirtied */
1737                         release_bytes -= dirty_sectors <<
1738                                                 fs_info->sb->s_blocksize_bits;
1739                         if (only_release_metadata) {
1740                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1741                                                                 release_bytes);
1742                         } else {
1743                                 u64 __pos;
1744
1745                                 __pos = round_down(pos,
1746                                                    fs_info->sectorsize) +
1747                                         (dirty_pages << PAGE_SHIFT);
1748                                 btrfs_delalloc_release_space(inode,
1749                                                 data_reserved, __pos,
1750                                                 release_bytes);
1751                         }
1752                 }
1753
1754                 release_bytes = round_up(copied + sector_offset,
1755                                         fs_info->sectorsize);
1756
1757                 if (copied > 0)
1758                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1759                                                 pos, copied, &cached_state);
1760                 if (extents_locked)
1761                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1762                                              lockstart, lockend, &cached_state);
1763                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1764                 if (ret) {
1765                         btrfs_drop_pages(pages, num_pages);
1766                         break;
1767                 }
1768
1769                 release_bytes = 0;
1770                 if (only_release_metadata)
1771                         btrfs_end_write_no_snapshotting(root);
1772
1773                 if (only_release_metadata && copied > 0) {
1774                         lockstart = round_down(pos,
1775                                                fs_info->sectorsize);
1776                         lockend = round_up(pos + copied,
1777                                            fs_info->sectorsize) - 1;
1778
1779                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1780                                        lockend, EXTENT_NORESERVE, NULL,
1781                                        NULL, GFP_NOFS);
1782                         only_release_metadata = false;
1783                 }
1784
1785                 btrfs_drop_pages(pages, num_pages);
1786
1787                 cond_resched();
1788
1789                 balance_dirty_pages_ratelimited(inode->i_mapping);
1790                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1791                         btrfs_btree_balance_dirty(fs_info);
1792
1793                 pos += copied;
1794                 num_written += copied;
1795         }
1796
1797         kfree(pages);
1798
1799         if (release_bytes) {
1800                 if (only_release_metadata) {
1801                         btrfs_end_write_no_snapshotting(root);
1802                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1803                                         release_bytes);
1804                 } else {
1805                         btrfs_delalloc_release_space(inode, data_reserved,
1806                                         round_down(pos, fs_info->sectorsize),
1807                                         release_bytes);
1808                 }
1809         }
1810
1811         extent_changeset_free(data_reserved);
1812         return num_written ? num_written : ret;
1813 }
1814
1815 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1816 {
1817         struct file *file = iocb->ki_filp;
1818         struct inode *inode = file_inode(file);
1819         loff_t pos = iocb->ki_pos;
1820         ssize_t written;
1821         ssize_t written_buffered;
1822         loff_t endbyte;
1823         int err;
1824
1825         written = generic_file_direct_write(iocb, from);
1826
1827         if (written < 0 || !iov_iter_count(from))
1828                 return written;
1829
1830         pos += written;
1831         written_buffered = __btrfs_buffered_write(file, from, pos);
1832         if (written_buffered < 0) {
1833                 err = written_buffered;
1834                 goto out;
1835         }
1836         /*
1837          * Ensure all data is persisted. We want the next direct IO read to be
1838          * able to read what was just written.
1839          */
1840         endbyte = pos + written_buffered - 1;
1841         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1842         if (err)
1843                 goto out;
1844         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1845         if (err)
1846                 goto out;
1847         written += written_buffered;
1848         iocb->ki_pos = pos + written_buffered;
1849         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1850                                  endbyte >> PAGE_SHIFT);
1851 out:
1852         return written ? written : err;
1853 }
1854
1855 static void update_time_for_write(struct inode *inode)
1856 {
1857         struct timespec now;
1858
1859         if (IS_NOCMTIME(inode))
1860                 return;
1861
1862         now = current_time(inode);
1863         if (!timespec_equal(&inode->i_mtime, &now))
1864                 inode->i_mtime = now;
1865
1866         if (!timespec_equal(&inode->i_ctime, &now))
1867                 inode->i_ctime = now;
1868
1869         if (IS_I_VERSION(inode))
1870                 inode_inc_iversion(inode);
1871 }
1872
1873 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1874                                     struct iov_iter *from)
1875 {
1876         struct file *file = iocb->ki_filp;
1877         struct inode *inode = file_inode(file);
1878         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1879         struct btrfs_root *root = BTRFS_I(inode)->root;
1880         u64 start_pos;
1881         u64 end_pos;
1882         ssize_t num_written = 0;
1883         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1884         ssize_t err;
1885         loff_t pos;
1886         size_t count = iov_iter_count(from);
1887         loff_t oldsize;
1888         int clean_page = 0;
1889
1890         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1891             (iocb->ki_flags & IOCB_NOWAIT))
1892                 return -EOPNOTSUPP;
1893
1894         if (!inode_trylock(inode)) {
1895                 if (iocb->ki_flags & IOCB_NOWAIT)
1896                         return -EAGAIN;
1897                 inode_lock(inode);
1898         }
1899
1900         err = generic_write_checks(iocb, from);
1901         if (err <= 0) {
1902                 inode_unlock(inode);
1903                 return err;
1904         }
1905
1906         pos = iocb->ki_pos;
1907         if (iocb->ki_flags & IOCB_NOWAIT) {
1908                 /*
1909                  * We will allocate space in case nodatacow is not set,
1910                  * so bail
1911                  */
1912                 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1913                                               BTRFS_INODE_PREALLOC)) ||
1914                     check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1915                         inode_unlock(inode);
1916                         return -EAGAIN;
1917                 }
1918         }
1919
1920         current->backing_dev_info = inode_to_bdi(inode);
1921         err = file_remove_privs(file);
1922         if (err) {
1923                 inode_unlock(inode);
1924                 goto out;
1925         }
1926
1927         /*
1928          * If BTRFS flips readonly due to some impossible error
1929          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1930          * although we have opened a file as writable, we have
1931          * to stop this write operation to ensure FS consistency.
1932          */
1933         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1934                 inode_unlock(inode);
1935                 err = -EROFS;
1936                 goto out;
1937         }
1938
1939         /*
1940          * We reserve space for updating the inode when we reserve space for the
1941          * extent we are going to write, so we will enospc out there.  We don't
1942          * need to start yet another transaction to update the inode as we will
1943          * update the inode when we finish writing whatever data we write.
1944          */
1945         update_time_for_write(inode);
1946
1947         start_pos = round_down(pos, fs_info->sectorsize);
1948         oldsize = i_size_read(inode);
1949         if (start_pos > oldsize) {
1950                 /* Expand hole size to cover write data, preventing empty gap */
1951                 end_pos = round_up(pos + count,
1952                                    fs_info->sectorsize);
1953                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1954                 if (err) {
1955                         inode_unlock(inode);
1956                         goto out;
1957                 }
1958                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1959                         clean_page = 1;
1960         }
1961
1962         if (sync)
1963                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1964
1965         if (iocb->ki_flags & IOCB_DIRECT) {
1966                 num_written = __btrfs_direct_write(iocb, from);
1967         } else {
1968                 num_written = __btrfs_buffered_write(file, from, pos);
1969                 if (num_written > 0)
1970                         iocb->ki_pos = pos + num_written;
1971                 if (clean_page)
1972                         pagecache_isize_extended(inode, oldsize,
1973                                                 i_size_read(inode));
1974         }
1975
1976         inode_unlock(inode);
1977
1978         /*
1979          * We also have to set last_sub_trans to the current log transid,
1980          * otherwise subsequent syncs to a file that's been synced in this
1981          * transaction will appear to have already occurred.
1982          */
1983         spin_lock(&BTRFS_I(inode)->lock);
1984         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1985         spin_unlock(&BTRFS_I(inode)->lock);
1986         if (num_written > 0)
1987                 num_written = generic_write_sync(iocb, num_written);
1988
1989         if (sync)
1990                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1991 out:
1992         current->backing_dev_info = NULL;
1993         return num_written ? num_written : err;
1994 }
1995
1996 int btrfs_release_file(struct inode *inode, struct file *filp)
1997 {
1998         struct btrfs_file_private *private = filp->private_data;
1999
2000         if (private && private->trans)
2001                 btrfs_ioctl_trans_end(filp);
2002         if (private && private->filldir_buf)
2003                 kfree(private->filldir_buf);
2004         kfree(private);
2005         filp->private_data = NULL;
2006
2007         /*
2008          * ordered_data_close is set by settattr when we are about to truncate
2009          * a file from a non-zero size to a zero size.  This tries to
2010          * flush down new bytes that may have been written if the
2011          * application were using truncate to replace a file in place.
2012          */
2013         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2014                                &BTRFS_I(inode)->runtime_flags))
2015                         filemap_flush(inode->i_mapping);
2016         return 0;
2017 }
2018
2019 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2020 {
2021         int ret;
2022         struct blk_plug plug;
2023
2024         /*
2025          * This is only called in fsync, which would do synchronous writes, so
2026          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2027          * multiple disks using raid profile, a large IO can be split to
2028          * several segments of stripe length (currently 64K).
2029          */
2030         blk_start_plug(&plug);
2031         atomic_inc(&BTRFS_I(inode)->sync_writers);
2032         ret = btrfs_fdatawrite_range(inode, start, end);
2033         atomic_dec(&BTRFS_I(inode)->sync_writers);
2034         blk_finish_plug(&plug);
2035
2036         return ret;
2037 }
2038
2039 /*
2040  * fsync call for both files and directories.  This logs the inode into
2041  * the tree log instead of forcing full commits whenever possible.
2042  *
2043  * It needs to call filemap_fdatawait so that all ordered extent updates are
2044  * in the metadata btree are up to date for copying to the log.
2045  *
2046  * It drops the inode mutex before doing the tree log commit.  This is an
2047  * important optimization for directories because holding the mutex prevents
2048  * new operations on the dir while we write to disk.
2049  */
2050 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2051 {
2052         struct dentry *dentry = file_dentry(file);
2053         struct inode *inode = d_inode(dentry);
2054         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2055         struct btrfs_root *root = BTRFS_I(inode)->root;
2056         struct btrfs_trans_handle *trans;
2057         struct btrfs_log_ctx ctx;
2058         int ret = 0, err;
2059         bool full_sync = false;
2060         u64 len;
2061
2062         /*
2063          * The range length can be represented by u64, we have to do the typecasts
2064          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2065          */
2066         len = (u64)end - (u64)start + 1;
2067         trace_btrfs_sync_file(file, datasync);
2068
2069         btrfs_init_log_ctx(&ctx, inode);
2070
2071         /*
2072          * We write the dirty pages in the range and wait until they complete
2073          * out of the ->i_mutex. If so, we can flush the dirty pages by
2074          * multi-task, and make the performance up.  See
2075          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2076          */
2077         ret = start_ordered_ops(inode, start, end);
2078         if (ret)
2079                 goto out;
2080
2081         inode_lock(inode);
2082         atomic_inc(&root->log_batch);
2083         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2084                              &BTRFS_I(inode)->runtime_flags);
2085         /*
2086          * We might have have had more pages made dirty after calling
2087          * start_ordered_ops and before acquiring the inode's i_mutex.
2088          */
2089         if (full_sync) {
2090                 /*
2091                  * For a full sync, we need to make sure any ordered operations
2092                  * start and finish before we start logging the inode, so that
2093                  * all extents are persisted and the respective file extent
2094                  * items are in the fs/subvol btree.
2095                  */
2096                 ret = btrfs_wait_ordered_range(inode, start, len);
2097         } else {
2098                 /*
2099                  * Start any new ordered operations before starting to log the
2100                  * inode. We will wait for them to finish in btrfs_sync_log().
2101                  *
2102                  * Right before acquiring the inode's mutex, we might have new
2103                  * writes dirtying pages, which won't immediately start the
2104                  * respective ordered operations - that is done through the
2105                  * fill_delalloc callbacks invoked from the writepage and
2106                  * writepages address space operations. So make sure we start
2107                  * all ordered operations before starting to log our inode. Not
2108                  * doing this means that while logging the inode, writeback
2109                  * could start and invoke writepage/writepages, which would call
2110                  * the fill_delalloc callbacks (cow_file_range,
2111                  * submit_compressed_extents). These callbacks add first an
2112                  * extent map to the modified list of extents and then create
2113                  * the respective ordered operation, which means in
2114                  * tree-log.c:btrfs_log_inode() we might capture all existing
2115                  * ordered operations (with btrfs_get_logged_extents()) before
2116                  * the fill_delalloc callback adds its ordered operation, and by
2117                  * the time we visit the modified list of extent maps (with
2118                  * btrfs_log_changed_extents()), we see and process the extent
2119                  * map they created. We then use the extent map to construct a
2120                  * file extent item for logging without waiting for the
2121                  * respective ordered operation to finish - this file extent
2122                  * item points to a disk location that might not have yet been
2123                  * written to, containing random data - so after a crash a log
2124                  * replay will make our inode have file extent items that point
2125                  * to disk locations containing invalid data, as we returned
2126                  * success to userspace without waiting for the respective
2127                  * ordered operation to finish, because it wasn't captured by
2128                  * btrfs_get_logged_extents().
2129                  */
2130                 ret = start_ordered_ops(inode, start, end);
2131         }
2132         if (ret) {
2133                 inode_unlock(inode);
2134                 goto out;
2135         }
2136         atomic_inc(&root->log_batch);
2137
2138         /*
2139          * If the last transaction that changed this file was before the current
2140          * transaction and we have the full sync flag set in our inode, we can
2141          * bail out now without any syncing.
2142          *
2143          * Note that we can't bail out if the full sync flag isn't set. This is
2144          * because when the full sync flag is set we start all ordered extents
2145          * and wait for them to fully complete - when they complete they update
2146          * the inode's last_trans field through:
2147          *
2148          *     btrfs_finish_ordered_io() ->
2149          *         btrfs_update_inode_fallback() ->
2150          *             btrfs_update_inode() ->
2151          *                 btrfs_set_inode_last_trans()
2152          *
2153          * So we are sure that last_trans is up to date and can do this check to
2154          * bail out safely. For the fast path, when the full sync flag is not
2155          * set in our inode, we can not do it because we start only our ordered
2156          * extents and don't wait for them to complete (that is when
2157          * btrfs_finish_ordered_io runs), so here at this point their last_trans
2158          * value might be less than or equals to fs_info->last_trans_committed,
2159          * and setting a speculative last_trans for an inode when a buffered
2160          * write is made (such as fs_info->generation + 1 for example) would not
2161          * be reliable since after setting the value and before fsync is called
2162          * any number of transactions can start and commit (transaction kthread
2163          * commits the current transaction periodically), and a transaction
2164          * commit does not start nor waits for ordered extents to complete.
2165          */
2166         smp_mb();
2167         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2168             (full_sync && BTRFS_I(inode)->last_trans <=
2169              fs_info->last_trans_committed) ||
2170             (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2171              BTRFS_I(inode)->last_trans
2172              <= fs_info->last_trans_committed)) {
2173                 /*
2174                  * We've had everything committed since the last time we were
2175                  * modified so clear this flag in case it was set for whatever
2176                  * reason, it's no longer relevant.
2177                  */
2178                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2179                           &BTRFS_I(inode)->runtime_flags);
2180                 /*
2181                  * An ordered extent might have started before and completed
2182                  * already with io errors, in which case the inode was not
2183                  * updated and we end up here. So check the inode's mapping
2184                  * for any errors that might have happened since we last
2185                  * checked called fsync.
2186                  */
2187                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2188                 inode_unlock(inode);
2189                 goto out;
2190         }
2191
2192         /*
2193          * ok we haven't committed the transaction yet, lets do a commit
2194          */
2195         if (file->private_data)
2196                 btrfs_ioctl_trans_end(file);
2197
2198         /*
2199          * We use start here because we will need to wait on the IO to complete
2200          * in btrfs_sync_log, which could require joining a transaction (for
2201          * example checking cross references in the nocow path).  If we use join
2202          * here we could get into a situation where we're waiting on IO to
2203          * happen that is blocked on a transaction trying to commit.  With start
2204          * we inc the extwriter counter, so we wait for all extwriters to exit
2205          * before we start blocking join'ers.  This comment is to keep somebody
2206          * from thinking they are super smart and changing this to
2207          * btrfs_join_transaction *cough*Josef*cough*.
2208          */
2209         trans = btrfs_start_transaction(root, 0);
2210         if (IS_ERR(trans)) {
2211                 ret = PTR_ERR(trans);
2212                 inode_unlock(inode);
2213                 goto out;
2214         }
2215         trans->sync = true;
2216
2217         ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2218         if (ret < 0) {
2219                 /* Fallthrough and commit/free transaction. */
2220                 ret = 1;
2221         }
2222
2223         /* we've logged all the items and now have a consistent
2224          * version of the file in the log.  It is possible that
2225          * someone will come in and modify the file, but that's
2226          * fine because the log is consistent on disk, and we
2227          * have references to all of the file's extents
2228          *
2229          * It is possible that someone will come in and log the
2230          * file again, but that will end up using the synchronization
2231          * inside btrfs_sync_log to keep things safe.
2232          */
2233         inode_unlock(inode);
2234
2235         /*
2236          * If any of the ordered extents had an error, just return it to user
2237          * space, so that the application knows some writes didn't succeed and
2238          * can take proper action (retry for e.g.). Blindly committing the
2239          * transaction in this case, would fool userspace that everything was
2240          * successful. And we also want to make sure our log doesn't contain
2241          * file extent items pointing to extents that weren't fully written to -
2242          * just like in the non fast fsync path, where we check for the ordered
2243          * operation's error flag before writing to the log tree and return -EIO
2244          * if any of them had this flag set (btrfs_wait_ordered_range) -
2245          * therefore we need to check for errors in the ordered operations,
2246          * which are indicated by ctx.io_err.
2247          */
2248         if (ctx.io_err) {
2249                 btrfs_end_transaction(trans);
2250                 ret = ctx.io_err;
2251                 goto out;
2252         }
2253
2254         if (ret != BTRFS_NO_LOG_SYNC) {
2255                 if (!ret) {
2256                         ret = btrfs_sync_log(trans, root, &ctx);
2257                         if (!ret) {
2258                                 ret = btrfs_end_transaction(trans);
2259                                 goto out;
2260                         }
2261                 }
2262                 if (!full_sync) {
2263                         ret = btrfs_wait_ordered_range(inode, start, len);
2264                         if (ret) {
2265                                 btrfs_end_transaction(trans);
2266                                 goto out;
2267                         }
2268                 }
2269                 ret = btrfs_commit_transaction(trans);
2270         } else {
2271                 ret = btrfs_end_transaction(trans);
2272         }
2273 out:
2274         ASSERT(list_empty(&ctx.list));
2275         err = file_check_and_advance_wb_err(file);
2276         if (!ret)
2277                 ret = err;
2278         return ret > 0 ? -EIO : ret;
2279 }
2280
2281 static const struct vm_operations_struct btrfs_file_vm_ops = {
2282         .fault          = filemap_fault,
2283         .map_pages      = filemap_map_pages,
2284         .page_mkwrite   = btrfs_page_mkwrite,
2285 };
2286
2287 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2288 {
2289         struct address_space *mapping = filp->f_mapping;
2290
2291         if (!mapping->a_ops->readpage)
2292                 return -ENOEXEC;
2293
2294         file_accessed(filp);
2295         vma->vm_ops = &btrfs_file_vm_ops;
2296
2297         return 0;
2298 }
2299
2300 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2301                           int slot, u64 start, u64 end)
2302 {
2303         struct btrfs_file_extent_item *fi;
2304         struct btrfs_key key;
2305
2306         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2307                 return 0;
2308
2309         btrfs_item_key_to_cpu(leaf, &key, slot);
2310         if (key.objectid != btrfs_ino(inode) ||
2311             key.type != BTRFS_EXTENT_DATA_KEY)
2312                 return 0;
2313
2314         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2315
2316         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2317                 return 0;
2318
2319         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2320                 return 0;
2321
2322         if (key.offset == end)
2323                 return 1;
2324         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2325                 return 1;
2326         return 0;
2327 }
2328
2329 static int fill_holes(struct btrfs_trans_handle *trans,
2330                 struct btrfs_inode *inode,
2331                 struct btrfs_path *path, u64 offset, u64 end)
2332 {
2333         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2334         struct btrfs_root *root = inode->root;
2335         struct extent_buffer *leaf;
2336         struct btrfs_file_extent_item *fi;
2337         struct extent_map *hole_em;
2338         struct extent_map_tree *em_tree = &inode->extent_tree;
2339         struct btrfs_key key;
2340         int ret;
2341
2342         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2343                 goto out;
2344
2345         key.objectid = btrfs_ino(inode);
2346         key.type = BTRFS_EXTENT_DATA_KEY;
2347         key.offset = offset;
2348
2349         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2350         if (ret <= 0) {
2351                 /*
2352                  * We should have dropped this offset, so if we find it then
2353                  * something has gone horribly wrong.
2354                  */
2355                 if (ret == 0)
2356                         ret = -EINVAL;
2357                 return ret;
2358         }
2359
2360         leaf = path->nodes[0];
2361         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2362                 u64 num_bytes;
2363
2364                 path->slots[0]--;
2365                 fi = btrfs_item_ptr(leaf, path->slots[0],
2366                                     struct btrfs_file_extent_item);
2367                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2368                         end - offset;
2369                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2370                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2371                 btrfs_set_file_extent_offset(leaf, fi, 0);
2372                 btrfs_mark_buffer_dirty(leaf);
2373                 goto out;
2374         }
2375
2376         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2377                 u64 num_bytes;
2378
2379                 key.offset = offset;
2380                 btrfs_set_item_key_safe(fs_info, path, &key);
2381                 fi = btrfs_item_ptr(leaf, path->slots[0],
2382                                     struct btrfs_file_extent_item);
2383                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2384                         offset;
2385                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2386                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2387                 btrfs_set_file_extent_offset(leaf, fi, 0);
2388                 btrfs_mark_buffer_dirty(leaf);
2389                 goto out;
2390         }
2391         btrfs_release_path(path);
2392
2393         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2394                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2395         if (ret)
2396                 return ret;
2397
2398 out:
2399         btrfs_release_path(path);
2400
2401         hole_em = alloc_extent_map();
2402         if (!hole_em) {
2403                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2404                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2405         } else {
2406                 hole_em->start = offset;
2407                 hole_em->len = end - offset;
2408                 hole_em->ram_bytes = hole_em->len;
2409                 hole_em->orig_start = offset;
2410
2411                 hole_em->block_start = EXTENT_MAP_HOLE;
2412                 hole_em->block_len = 0;
2413                 hole_em->orig_block_len = 0;
2414                 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2415                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2416                 hole_em->generation = trans->transid;
2417
2418                 do {
2419                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2420                         write_lock(&em_tree->lock);
2421                         ret = add_extent_mapping(em_tree, hole_em, 1);
2422                         write_unlock(&em_tree->lock);
2423                 } while (ret == -EEXIST);
2424                 free_extent_map(hole_em);
2425                 if (ret)
2426                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2427                                         &inode->runtime_flags);
2428         }
2429
2430         return 0;
2431 }
2432
2433 /*
2434  * Find a hole extent on given inode and change start/len to the end of hole
2435  * extent.(hole/vacuum extent whose em->start <= start &&
2436  *         em->start + em->len > start)
2437  * When a hole extent is found, return 1 and modify start/len.
2438  */
2439 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2440 {
2441         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2442         struct extent_map *em;
2443         int ret = 0;
2444
2445         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2446                               round_down(*start, fs_info->sectorsize),
2447                               round_up(*len, fs_info->sectorsize), 0);
2448         if (IS_ERR(em))
2449                 return PTR_ERR(em);
2450
2451         /* Hole or vacuum extent(only exists in no-hole mode) */
2452         if (em->block_start == EXTENT_MAP_HOLE) {
2453                 ret = 1;
2454                 *len = em->start + em->len > *start + *len ?
2455                        0 : *start + *len - em->start - em->len;
2456                 *start = em->start + em->len;
2457         }
2458         free_extent_map(em);
2459         return ret;
2460 }
2461
2462 static int btrfs_punch_hole_lock_range(struct inode *inode,
2463                                        const u64 lockstart,
2464                                        const u64 lockend,
2465                                        struct extent_state **cached_state)
2466 {
2467         while (1) {
2468                 struct btrfs_ordered_extent *ordered;
2469                 int ret;
2470
2471                 truncate_pagecache_range(inode, lockstart, lockend);
2472
2473                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2474                                  cached_state);
2475                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2476
2477                 /*
2478                  * We need to make sure we have no ordered extents in this range
2479                  * and nobody raced in and read a page in this range, if we did
2480                  * we need to try again.
2481                  */
2482                 if ((!ordered ||
2483                     (ordered->file_offset + ordered->len <= lockstart ||
2484                      ordered->file_offset > lockend)) &&
2485                      !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2486                         if (ordered)
2487                                 btrfs_put_ordered_extent(ordered);
2488                         break;
2489                 }
2490                 if (ordered)
2491                         btrfs_put_ordered_extent(ordered);
2492                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2493                                      lockend, cached_state);
2494                 ret = btrfs_wait_ordered_range(inode, lockstart,
2495                                                lockend - lockstart + 1);
2496                 if (ret)
2497                         return ret;
2498         }
2499         return 0;
2500 }
2501
2502 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2503 {
2504         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2505         struct btrfs_root *root = BTRFS_I(inode)->root;
2506         struct extent_state *cached_state = NULL;
2507         struct btrfs_path *path;
2508         struct btrfs_block_rsv *rsv;
2509         struct btrfs_trans_handle *trans;
2510         u64 lockstart;
2511         u64 lockend;
2512         u64 tail_start;
2513         u64 tail_len;
2514         u64 orig_start = offset;
2515         u64 cur_offset;
2516         u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2517         u64 drop_end;
2518         int ret = 0;
2519         int err = 0;
2520         unsigned int rsv_count;
2521         bool same_block;
2522         bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2523         u64 ino_size;
2524         bool truncated_block = false;
2525         bool updated_inode = false;
2526
2527         ret = btrfs_wait_ordered_range(inode, offset, len);
2528         if (ret)
2529                 return ret;
2530
2531         inode_lock(inode);
2532         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2533         ret = find_first_non_hole(inode, &offset, &len);
2534         if (ret < 0)
2535                 goto out_only_mutex;
2536         if (ret && !len) {
2537                 /* Already in a large hole */
2538                 ret = 0;
2539                 goto out_only_mutex;
2540         }
2541
2542         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2543         lockend = round_down(offset + len,
2544                              btrfs_inode_sectorsize(inode)) - 1;
2545         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2546                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2547         /*
2548          * We needn't truncate any block which is beyond the end of the file
2549          * because we are sure there is no data there.
2550          */
2551         /*
2552          * Only do this if we are in the same block and we aren't doing the
2553          * entire block.
2554          */
2555         if (same_block && len < fs_info->sectorsize) {
2556                 if (offset < ino_size) {
2557                         truncated_block = true;
2558                         ret = btrfs_truncate_block(inode, offset, len, 0);
2559                 } else {
2560                         ret = 0;
2561                 }
2562                 goto out_only_mutex;
2563         }
2564
2565         /* zero back part of the first block */
2566         if (offset < ino_size) {
2567                 truncated_block = true;
2568                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2569                 if (ret) {
2570                         inode_unlock(inode);
2571                         return ret;
2572                 }
2573         }
2574
2575         /* Check the aligned pages after the first unaligned page,
2576          * if offset != orig_start, which means the first unaligned page
2577          * including several following pages are already in holes,
2578          * the extra check can be skipped */
2579         if (offset == orig_start) {
2580                 /* after truncate page, check hole again */
2581                 len = offset + len - lockstart;
2582                 offset = lockstart;
2583                 ret = find_first_non_hole(inode, &offset, &len);
2584                 if (ret < 0)
2585                         goto out_only_mutex;
2586                 if (ret && !len) {
2587                         ret = 0;
2588                         goto out_only_mutex;
2589                 }
2590                 lockstart = offset;
2591         }
2592
2593         /* Check the tail unaligned part is in a hole */
2594         tail_start = lockend + 1;
2595         tail_len = offset + len - tail_start;
2596         if (tail_len) {
2597                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2598                 if (unlikely(ret < 0))
2599                         goto out_only_mutex;
2600                 if (!ret) {
2601                         /* zero the front end of the last page */
2602                         if (tail_start + tail_len < ino_size) {
2603                                 truncated_block = true;
2604                                 ret = btrfs_truncate_block(inode,
2605                                                         tail_start + tail_len,
2606                                                         0, 1);
2607                                 if (ret)
2608                                         goto out_only_mutex;
2609                         }
2610                 }
2611         }
2612
2613         if (lockend < lockstart) {
2614                 ret = 0;
2615                 goto out_only_mutex;
2616         }
2617
2618         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2619                                           &cached_state);
2620         if (ret) {
2621                 inode_unlock(inode);
2622                 goto out_only_mutex;
2623         }
2624
2625         path = btrfs_alloc_path();
2626         if (!path) {
2627                 ret = -ENOMEM;
2628                 goto out;
2629         }
2630
2631         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2632         if (!rsv) {
2633                 ret = -ENOMEM;
2634                 goto out_free;
2635         }
2636         rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2637         rsv->failfast = 1;
2638
2639         /*
2640          * 1 - update the inode
2641          * 1 - removing the extents in the range
2642          * 1 - adding the hole extent if no_holes isn't set
2643          */
2644         rsv_count = no_holes ? 2 : 3;
2645         trans = btrfs_start_transaction(root, rsv_count);
2646         if (IS_ERR(trans)) {
2647                 err = PTR_ERR(trans);
2648                 goto out_free;
2649         }
2650
2651         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2652                                       min_size, 0);
2653         BUG_ON(ret);
2654         trans->block_rsv = rsv;
2655
2656         cur_offset = lockstart;
2657         len = lockend - cur_offset;
2658         while (cur_offset < lockend) {
2659                 ret = __btrfs_drop_extents(trans, root, inode, path,
2660                                            cur_offset, lockend + 1,
2661                                            &drop_end, 1, 0, 0, NULL);
2662                 if (ret != -ENOSPC)
2663                         break;
2664
2665                 trans->block_rsv = &fs_info->trans_block_rsv;
2666
2667                 if (cur_offset < drop_end && cur_offset < ino_size) {
2668                         ret = fill_holes(trans, BTRFS_I(inode), path,
2669                                         cur_offset, drop_end);
2670                         if (ret) {
2671                                 /*
2672                                  * If we failed then we didn't insert our hole
2673                                  * entries for the area we dropped, so now the
2674                                  * fs is corrupted, so we must abort the
2675                                  * transaction.
2676                                  */
2677                                 btrfs_abort_transaction(trans, ret);
2678                                 err = ret;
2679                                 break;
2680                         }
2681                 }
2682
2683                 cur_offset = drop_end;
2684
2685                 ret = btrfs_update_inode(trans, root, inode);
2686                 if (ret) {
2687                         err = ret;
2688                         break;
2689                 }
2690
2691                 btrfs_end_transaction(trans);
2692                 btrfs_btree_balance_dirty(fs_info);
2693
2694                 trans = btrfs_start_transaction(root, rsv_count);
2695                 if (IS_ERR(trans)) {
2696                         ret = PTR_ERR(trans);
2697                         trans = NULL;
2698                         break;
2699                 }
2700
2701                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2702                                               rsv, min_size, 0);
2703                 BUG_ON(ret);    /* shouldn't happen */
2704                 trans->block_rsv = rsv;
2705
2706                 ret = find_first_non_hole(inode, &cur_offset, &len);
2707                 if (unlikely(ret < 0))
2708                         break;
2709                 if (ret && !len) {
2710                         ret = 0;
2711                         break;
2712                 }
2713         }
2714
2715         if (ret) {
2716                 err = ret;
2717                 goto out_trans;
2718         }
2719
2720         trans->block_rsv = &fs_info->trans_block_rsv;
2721         /*
2722          * If we are using the NO_HOLES feature we might have had already an
2723          * hole that overlaps a part of the region [lockstart, lockend] and
2724          * ends at (or beyond) lockend. Since we have no file extent items to
2725          * represent holes, drop_end can be less than lockend and so we must
2726          * make sure we have an extent map representing the existing hole (the
2727          * call to __btrfs_drop_extents() might have dropped the existing extent
2728          * map representing the existing hole), otherwise the fast fsync path
2729          * will not record the existence of the hole region
2730          * [existing_hole_start, lockend].
2731          */
2732         if (drop_end <= lockend)
2733                 drop_end = lockend + 1;
2734         /*
2735          * Don't insert file hole extent item if it's for a range beyond eof
2736          * (because it's useless) or if it represents a 0 bytes range (when
2737          * cur_offset == drop_end).
2738          */
2739         if (cur_offset < ino_size && cur_offset < drop_end) {
2740                 ret = fill_holes(trans, BTRFS_I(inode), path,
2741                                 cur_offset, drop_end);
2742                 if (ret) {
2743                         /* Same comment as above. */
2744                         btrfs_abort_transaction(trans, ret);
2745                         err = ret;
2746                         goto out_trans;
2747                 }
2748         }
2749
2750 out_trans:
2751         if (!trans)
2752                 goto out_free;
2753
2754         inode_inc_iversion(inode);
2755         inode->i_mtime = inode->i_ctime = current_time(inode);
2756
2757         trans->block_rsv = &fs_info->trans_block_rsv;
2758         ret = btrfs_update_inode(trans, root, inode);
2759         updated_inode = true;
2760         btrfs_end_transaction(trans);
2761         btrfs_btree_balance_dirty(fs_info);
2762 out_free:
2763         btrfs_free_path(path);
2764         btrfs_free_block_rsv(fs_info, rsv);
2765 out:
2766         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2767                              &cached_state);
2768 out_only_mutex:
2769         if (!updated_inode && truncated_block && !ret && !err) {
2770                 /*
2771                  * If we only end up zeroing part of a page, we still need to
2772                  * update the inode item, so that all the time fields are
2773                  * updated as well as the necessary btrfs inode in memory fields
2774                  * for detecting, at fsync time, if the inode isn't yet in the
2775                  * log tree or it's there but not up to date.
2776                  */
2777                 trans = btrfs_start_transaction(root, 1);
2778                 if (IS_ERR(trans)) {
2779                         err = PTR_ERR(trans);
2780                 } else {
2781                         err = btrfs_update_inode(trans, root, inode);
2782                         ret = btrfs_end_transaction(trans);
2783                 }
2784         }
2785         inode_unlock(inode);
2786         if (ret && !err)
2787                 err = ret;
2788         return err;
2789 }
2790
2791 /* Helper structure to record which range is already reserved */
2792 struct falloc_range {
2793         struct list_head list;
2794         u64 start;
2795         u64 len;
2796 };
2797
2798 /*
2799  * Helper function to add falloc range
2800  *
2801  * Caller should have locked the larger range of extent containing
2802  * [start, len)
2803  */
2804 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2805 {
2806         struct falloc_range *prev = NULL;
2807         struct falloc_range *range = NULL;
2808
2809         if (list_empty(head))
2810                 goto insert;
2811
2812         /*
2813          * As fallocate iterate by bytenr order, we only need to check
2814          * the last range.
2815          */
2816         prev = list_entry(head->prev, struct falloc_range, list);
2817         if (prev->start + prev->len == start) {
2818                 prev->len += len;
2819                 return 0;
2820         }
2821 insert:
2822         range = kmalloc(sizeof(*range), GFP_KERNEL);
2823         if (!range)
2824                 return -ENOMEM;
2825         range->start = start;
2826         range->len = len;
2827         list_add_tail(&range->list, head);
2828         return 0;
2829 }
2830
2831 static int btrfs_fallocate_update_isize(struct inode *inode,
2832                                         const u64 end,
2833                                         const int mode)
2834 {
2835         struct btrfs_trans_handle *trans;
2836         struct btrfs_root *root = BTRFS_I(inode)->root;
2837         int ret;
2838         int ret2;
2839
2840         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2841                 return 0;
2842
2843         trans = btrfs_start_transaction(root, 1);
2844         if (IS_ERR(trans))
2845                 return PTR_ERR(trans);
2846
2847         inode->i_ctime = current_time(inode);
2848         i_size_write(inode, end);
2849         btrfs_ordered_update_i_size(inode, end, NULL);
2850         ret = btrfs_update_inode(trans, root, inode);
2851         ret2 = btrfs_end_transaction(trans);
2852
2853         return ret ? ret : ret2;
2854 }
2855
2856 enum {
2857         RANGE_BOUNDARY_WRITTEN_EXTENT = 0,
2858         RANGE_BOUNDARY_PREALLOC_EXTENT = 1,
2859         RANGE_BOUNDARY_HOLE = 2,
2860 };
2861
2862 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2863                                                  u64 offset)
2864 {
2865         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2866         struct extent_map *em;
2867         int ret;
2868
2869         offset = round_down(offset, sectorsize);
2870         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2871         if (IS_ERR(em))
2872                 return PTR_ERR(em);
2873
2874         if (em->block_start == EXTENT_MAP_HOLE)
2875                 ret = RANGE_BOUNDARY_HOLE;
2876         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2877                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2878         else
2879                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2880
2881         free_extent_map(em);
2882         return ret;
2883 }
2884
2885 static int btrfs_zero_range(struct inode *inode,
2886                             loff_t offset,
2887                             loff_t len,
2888                             const int mode)
2889 {
2890         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2891         struct extent_map *em;
2892         struct extent_changeset *data_reserved = NULL;
2893         int ret;
2894         u64 alloc_hint = 0;
2895         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2896         u64 alloc_start = round_down(offset, sectorsize);
2897         u64 alloc_end = round_up(offset + len, sectorsize);
2898         u64 bytes_to_reserve = 0;
2899         bool space_reserved = false;
2900
2901         inode_dio_wait(inode);
2902
2903         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2904                               alloc_start, alloc_end - alloc_start, 0);
2905         if (IS_ERR(em)) {
2906                 ret = PTR_ERR(em);
2907                 goto out;
2908         }
2909
2910         /*
2911          * Avoid hole punching and extent allocation for some cases. More cases
2912          * could be considered, but these are unlikely common and we keep things
2913          * as simple as possible for now. Also, intentionally, if the target
2914          * range contains one or more prealloc extents together with regular
2915          * extents and holes, we drop all the existing extents and allocate a
2916          * new prealloc extent, so that we get a larger contiguous disk extent.
2917          */
2918         if (em->start <= alloc_start &&
2919             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2920                 const u64 em_end = em->start + em->len;
2921
2922                 if (em_end >= offset + len) {
2923                         /*
2924                          * The whole range is already a prealloc extent,
2925                          * do nothing except updating the inode's i_size if
2926                          * needed.
2927                          */
2928                         free_extent_map(em);
2929                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2930                                                            mode);
2931                         goto out;
2932                 }
2933                 /*
2934                  * Part of the range is already a prealloc extent, so operate
2935                  * only on the remaining part of the range.
2936                  */
2937                 alloc_start = em_end;
2938                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2939                 len = offset + len - alloc_start;
2940                 offset = alloc_start;
2941                 alloc_hint = em->block_start + em->len;
2942         }
2943         free_extent_map(em);
2944
2945         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2946             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2947                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2948                                       alloc_start, sectorsize, 0);
2949                 if (IS_ERR(em)) {
2950                         ret = PTR_ERR(em);
2951                         goto out;
2952                 }
2953
2954                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2955                         free_extent_map(em);
2956                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2957                                                            mode);
2958                         goto out;
2959                 }
2960                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2961                         free_extent_map(em);
2962                         ret = btrfs_truncate_block(inode, offset, len, 0);
2963                         if (!ret)
2964                                 ret = btrfs_fallocate_update_isize(inode,
2965                                                                    offset + len,
2966                                                                    mode);
2967                         return ret;
2968                 }
2969                 free_extent_map(em);
2970                 alloc_start = round_down(offset, sectorsize);
2971                 alloc_end = alloc_start + sectorsize;
2972                 goto reserve_space;
2973         }
2974
2975         alloc_start = round_up(offset, sectorsize);
2976         alloc_end = round_down(offset + len, sectorsize);
2977
2978         /*
2979          * For unaligned ranges, check the pages at the boundaries, they might
2980          * map to an extent, in which case we need to partially zero them, or
2981          * they might map to a hole, in which case we need our allocation range
2982          * to cover them.
2983          */
2984         if (!IS_ALIGNED(offset, sectorsize)) {
2985                 ret = btrfs_zero_range_check_range_boundary(inode, offset);
2986                 if (ret < 0)
2987                         goto out;
2988                 if (ret == RANGE_BOUNDARY_HOLE) {
2989                         alloc_start = round_down(offset, sectorsize);
2990                         ret = 0;
2991                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2992                         ret = btrfs_truncate_block(inode, offset, 0, 0);
2993                         if (ret)
2994                                 goto out;
2995                 } else {
2996                         ret = 0;
2997                 }
2998         }
2999
3000         if (!IS_ALIGNED(offset + len, sectorsize)) {
3001                 ret = btrfs_zero_range_check_range_boundary(inode,
3002                                                             offset + len);
3003                 if (ret < 0)
3004                         goto out;
3005                 if (ret == RANGE_BOUNDARY_HOLE) {
3006                         alloc_end = round_up(offset + len, sectorsize);
3007                         ret = 0;
3008                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3009                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3010                         if (ret)
3011                                 goto out;
3012                 } else {
3013                         ret = 0;
3014                 }
3015         }
3016
3017 reserve_space:
3018         if (alloc_start < alloc_end) {
3019                 struct extent_state *cached_state = NULL;
3020                 const u64 lockstart = alloc_start;
3021                 const u64 lockend = alloc_end - 1;
3022
3023                 bytes_to_reserve = alloc_end - alloc_start;
3024                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3025                                                       bytes_to_reserve);
3026                 if (ret < 0)
3027                         goto out;
3028                 space_reserved = true;
3029                 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3030                                                 alloc_start, bytes_to_reserve);
3031                 if (ret)
3032                         goto out;
3033                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3034                                                   &cached_state);
3035                 if (ret)
3036                         goto out;
3037                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3038                                                 alloc_end - alloc_start,
3039                                                 i_blocksize(inode),
3040                                                 offset + len, &alloc_hint);
3041                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3042                                      lockend, &cached_state);
3043                 /* btrfs_prealloc_file_range releases reserved space on error */
3044                 if (ret) {
3045                         space_reserved = false;
3046                         goto out;
3047                 }
3048         }
3049         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3050  out:
3051         if (ret && space_reserved)
3052                 btrfs_free_reserved_data_space(inode, data_reserved,
3053                                                alloc_start, bytes_to_reserve);
3054         extent_changeset_free(data_reserved);
3055
3056         return ret;
3057 }
3058
3059 static long btrfs_fallocate(struct file *file, int mode,
3060                             loff_t offset, loff_t len)
3061 {
3062         struct inode *inode = file_inode(file);
3063         struct extent_state *cached_state = NULL;
3064         struct extent_changeset *data_reserved = NULL;
3065         struct falloc_range *range;
3066         struct falloc_range *tmp;
3067         struct list_head reserve_list;
3068         u64 cur_offset;
3069         u64 last_byte;
3070         u64 alloc_start;
3071         u64 alloc_end;
3072         u64 alloc_hint = 0;
3073         u64 locked_end;
3074         u64 actual_end = 0;
3075         struct extent_map *em;
3076         int blocksize = btrfs_inode_sectorsize(inode);
3077         int ret;
3078
3079         alloc_start = round_down(offset, blocksize);
3080         alloc_end = round_up(offset + len, blocksize);
3081         cur_offset = alloc_start;
3082
3083         /* Make sure we aren't being give some crap mode */
3084         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3085                      FALLOC_FL_ZERO_RANGE))
3086                 return -EOPNOTSUPP;
3087
3088         if (mode & FALLOC_FL_PUNCH_HOLE)
3089                 return btrfs_punch_hole(inode, offset, len);
3090
3091         /*
3092          * Only trigger disk allocation, don't trigger qgroup reserve
3093          *
3094          * For qgroup space, it will be checked later.
3095          */
3096         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3097                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3098                                                       alloc_end - alloc_start);
3099                 if (ret < 0)
3100                         return ret;
3101         }
3102
3103         inode_lock(inode);
3104
3105         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3106                 ret = inode_newsize_ok(inode, offset + len);
3107                 if (ret)
3108                         goto out;
3109         }
3110
3111         /*
3112          * TODO: Move these two operations after we have checked
3113          * accurate reserved space, or fallocate can still fail but
3114          * with page truncated or size expanded.
3115          *
3116          * But that's a minor problem and won't do much harm BTW.
3117          */
3118         if (alloc_start > inode->i_size) {
3119                 ret = btrfs_cont_expand(inode, i_size_read(inode),
3120                                         alloc_start);
3121                 if (ret)
3122                         goto out;
3123         } else if (offset + len > inode->i_size) {
3124                 /*
3125                  * If we are fallocating from the end of the file onward we
3126                  * need to zero out the end of the block if i_size lands in the
3127                  * middle of a block.
3128                  */
3129                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3130                 if (ret)
3131                         goto out;
3132         }
3133
3134         /*
3135          * wait for ordered IO before we have any locks.  We'll loop again
3136          * below with the locks held.
3137          */
3138         ret = btrfs_wait_ordered_range(inode, alloc_start,
3139                                        alloc_end - alloc_start);
3140         if (ret)
3141                 goto out;
3142
3143         if (mode & FALLOC_FL_ZERO_RANGE) {
3144                 ret = btrfs_zero_range(inode, offset, len, mode);
3145                 inode_unlock(inode);
3146                 return ret;
3147         }
3148
3149         locked_end = alloc_end - 1;
3150         while (1) {
3151                 struct btrfs_ordered_extent *ordered;
3152
3153                 /* the extent lock is ordered inside the running
3154                  * transaction
3155                  */
3156                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3157                                  locked_end, &cached_state);
3158                 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3159
3160                 if (ordered &&
3161                     ordered->file_offset + ordered->len > alloc_start &&
3162                     ordered->file_offset < alloc_end) {
3163                         btrfs_put_ordered_extent(ordered);
3164                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3165                                              alloc_start, locked_end,
3166                                              &cached_state);
3167                         /*
3168                          * we can't wait on the range with the transaction
3169                          * running or with the extent lock held
3170                          */
3171                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3172                                                        alloc_end - alloc_start);
3173                         if (ret)
3174                                 goto out;
3175                 } else {
3176                         if (ordered)
3177                                 btrfs_put_ordered_extent(ordered);
3178                         break;
3179                 }
3180         }
3181
3182         /* First, check if we exceed the qgroup limit */
3183         INIT_LIST_HEAD(&reserve_list);
3184         while (cur_offset < alloc_end) {
3185                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3186                                       alloc_end - cur_offset, 0);
3187                 if (IS_ERR(em)) {
3188                         ret = PTR_ERR(em);
3189                         break;
3190                 }
3191                 last_byte = min(extent_map_end(em), alloc_end);
3192                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3193                 last_byte = ALIGN(last_byte, blocksize);
3194                 if (em->block_start == EXTENT_MAP_HOLE ||
3195                     (cur_offset >= inode->i_size &&
3196                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3197                         ret = add_falloc_range(&reserve_list, cur_offset,
3198                                                last_byte - cur_offset);
3199                         if (ret < 0) {
3200                                 free_extent_map(em);
3201                                 break;
3202                         }
3203                         ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3204                                         cur_offset, last_byte - cur_offset);
3205                         if (ret < 0) {
3206                                 free_extent_map(em);
3207                                 break;
3208                         }
3209                 } else {
3210                         /*
3211                          * Do not need to reserve unwritten extent for this
3212                          * range, free reserved data space first, otherwise
3213                          * it'll result in false ENOSPC error.
3214                          */
3215                         btrfs_free_reserved_data_space(inode, data_reserved,
3216                                         cur_offset, last_byte - cur_offset);
3217                 }
3218                 free_extent_map(em);
3219                 cur_offset = last_byte;
3220         }
3221
3222         /*
3223          * If ret is still 0, means we're OK to fallocate.
3224          * Or just cleanup the list and exit.
3225          */
3226         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3227                 if (!ret)
3228                         ret = btrfs_prealloc_file_range(inode, mode,
3229                                         range->start,
3230                                         range->len, i_blocksize(inode),
3231                                         offset + len, &alloc_hint);
3232                 else
3233                         btrfs_free_reserved_data_space(inode,
3234                                         data_reserved, range->start,
3235                                         range->len);
3236                 list_del(&range->list);
3237                 kfree(range);
3238         }
3239         if (ret < 0)
3240                 goto out_unlock;
3241
3242         /*
3243          * We didn't need to allocate any more space, but we still extended the
3244          * size of the file so we need to update i_size and the inode item.
3245          */
3246         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3247 out_unlock:
3248         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3249                              &cached_state);
3250 out:
3251         inode_unlock(inode);
3252         /* Let go of our reservation. */
3253         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3254                 btrfs_free_reserved_data_space(inode, data_reserved,
3255                                 alloc_start, alloc_end - cur_offset);
3256         extent_changeset_free(data_reserved);
3257         return ret;
3258 }
3259
3260 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3261 {
3262         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3263         struct extent_map *em = NULL;
3264         struct extent_state *cached_state = NULL;
3265         u64 lockstart;
3266         u64 lockend;
3267         u64 start;
3268         u64 len;
3269         int ret = 0;
3270
3271         if (inode->i_size == 0)
3272                 return -ENXIO;
3273
3274         /*
3275          * *offset can be negative, in this case we start finding DATA/HOLE from
3276          * the very start of the file.
3277          */
3278         start = max_t(loff_t, 0, *offset);
3279
3280         lockstart = round_down(start, fs_info->sectorsize);
3281         lockend = round_up(i_size_read(inode),
3282                            fs_info->sectorsize);
3283         if (lockend <= lockstart)
3284                 lockend = lockstart + fs_info->sectorsize;
3285         lockend--;
3286         len = lockend - lockstart + 1;
3287
3288         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3289                          &cached_state);
3290
3291         while (start < inode->i_size) {
3292                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3293                                 start, len, 0);
3294                 if (IS_ERR(em)) {
3295                         ret = PTR_ERR(em);
3296                         em = NULL;
3297                         break;
3298                 }
3299
3300                 if (whence == SEEK_HOLE &&
3301                     (em->block_start == EXTENT_MAP_HOLE ||
3302                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3303                         break;
3304                 else if (whence == SEEK_DATA &&
3305                            (em->block_start != EXTENT_MAP_HOLE &&
3306                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3307                         break;
3308
3309                 start = em->start + em->len;
3310                 free_extent_map(em);
3311                 em = NULL;
3312                 cond_resched();
3313         }
3314         free_extent_map(em);
3315         if (!ret) {
3316                 if (whence == SEEK_DATA && start >= inode->i_size)
3317                         ret = -ENXIO;
3318                 else
3319                         *offset = min_t(loff_t, start, inode->i_size);
3320         }
3321         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3322                              &cached_state);
3323         return ret;
3324 }
3325
3326 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3327 {
3328         struct inode *inode = file->f_mapping->host;
3329         int ret;
3330
3331         inode_lock(inode);
3332         switch (whence) {
3333         case SEEK_END:
3334         case SEEK_CUR:
3335                 offset = generic_file_llseek(file, offset, whence);
3336                 goto out;
3337         case SEEK_DATA:
3338         case SEEK_HOLE:
3339                 if (offset >= i_size_read(inode)) {
3340                         inode_unlock(inode);
3341                         return -ENXIO;
3342                 }
3343
3344                 ret = find_desired_extent(inode, &offset, whence);
3345                 if (ret) {
3346                         inode_unlock(inode);
3347                         return ret;
3348                 }
3349         }
3350
3351         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3352 out:
3353         inode_unlock(inode);
3354         return offset;
3355 }
3356
3357 static int btrfs_file_open(struct inode *inode, struct file *filp)
3358 {
3359         filp->f_mode |= FMODE_NOWAIT;
3360         return generic_file_open(inode, filp);
3361 }
3362
3363 const struct file_operations btrfs_file_operations = {
3364         .llseek         = btrfs_file_llseek,
3365         .read_iter      = generic_file_read_iter,
3366         .splice_read    = generic_file_splice_read,
3367         .write_iter     = btrfs_file_write_iter,
3368         .mmap           = btrfs_file_mmap,
3369         .open           = btrfs_file_open,
3370         .release        = btrfs_release_file,
3371         .fsync          = btrfs_sync_file,
3372         .fallocate      = btrfs_fallocate,
3373         .unlocked_ioctl = btrfs_ioctl,
3374 #ifdef CONFIG_COMPAT
3375         .compat_ioctl   = btrfs_compat_ioctl,
3376 #endif
3377         .clone_file_range = btrfs_clone_file_range,
3378         .dedupe_file_range = btrfs_dedupe_file_range,
3379 };
3380
3381 void __cold btrfs_auto_defrag_exit(void)
3382 {
3383         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3384 }
3385
3386 int __init btrfs_auto_defrag_init(void)
3387 {
3388         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3389                                         sizeof(struct inode_defrag), 0,
3390                                         SLAB_MEM_SPREAD,
3391                                         NULL);
3392         if (!btrfs_inode_defrag_cachep)
3393                 return -ENOMEM;
3394
3395         return 0;
3396 }
3397
3398 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3399 {
3400         int ret;
3401
3402         /*
3403          * So with compression we will find and lock a dirty page and clear the
3404          * first one as dirty, setup an async extent, and immediately return
3405          * with the entire range locked but with nobody actually marked with
3406          * writeback.  So we can't just filemap_write_and_wait_range() and
3407          * expect it to work since it will just kick off a thread to do the
3408          * actual work.  So we need to call filemap_fdatawrite_range _again_
3409          * since it will wait on the page lock, which won't be unlocked until
3410          * after the pages have been marked as writeback and so we're good to go
3411          * from there.  We have to do this otherwise we'll miss the ordered
3412          * extents and that results in badness.  Please Josef, do not think you
3413          * know better and pull this out at some point in the future, it is
3414          * right and you are wrong.
3415          */
3416         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3417         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3418                              &BTRFS_I(inode)->runtime_flags))
3419                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3420
3421         return ret;
3422 }