89c9404fee9abc501da5f3099a70ed80da93959e
[sfrench/cifs-2.6.git] / fs / btrfs / file.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5
6 #include <linux/fs.h>
7 #include <linux/pagemap.h>
8 #include <linux/highmem.h>
9 #include <linux/time.h>
10 #include <linux/init.h>
11 #include <linux/string.h>
12 #include <linux/backing-dev.h>
13 #include <linux/mpage.h>
14 #include <linux/falloc.h>
15 #include <linux/swap.h>
16 #include <linux/writeback.h>
17 #include <linux/compat.h>
18 #include <linux/slab.h>
19 #include <linux/btrfs.h>
20 #include <linux/uio.h>
21 #include <linux/iversion.h>
22 #include "ctree.h"
23 #include "disk-io.h"
24 #include "transaction.h"
25 #include "btrfs_inode.h"
26 #include "print-tree.h"
27 #include "tree-log.h"
28 #include "locking.h"
29 #include "volumes.h"
30 #include "qgroup.h"
31 #include "compression.h"
32
33 static struct kmem_cache *btrfs_inode_defrag_cachep;
34 /*
35  * when auto defrag is enabled we
36  * queue up these defrag structs to remember which
37  * inodes need defragging passes
38  */
39 struct inode_defrag {
40         struct rb_node rb_node;
41         /* objectid */
42         u64 ino;
43         /*
44          * transid where the defrag was added, we search for
45          * extents newer than this
46          */
47         u64 transid;
48
49         /* root objectid */
50         u64 root;
51
52         /* last offset we were able to defrag */
53         u64 last_offset;
54
55         /* if we've wrapped around back to zero once already */
56         int cycled;
57 };
58
59 static int __compare_inode_defrag(struct inode_defrag *defrag1,
60                                   struct inode_defrag *defrag2)
61 {
62         if (defrag1->root > defrag2->root)
63                 return 1;
64         else if (defrag1->root < defrag2->root)
65                 return -1;
66         else if (defrag1->ino > defrag2->ino)
67                 return 1;
68         else if (defrag1->ino < defrag2->ino)
69                 return -1;
70         else
71                 return 0;
72 }
73
74 /* pop a record for an inode into the defrag tree.  The lock
75  * must be held already
76  *
77  * If you're inserting a record for an older transid than an
78  * existing record, the transid already in the tree is lowered
79  *
80  * If an existing record is found the defrag item you
81  * pass in is freed
82  */
83 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
84                                     struct inode_defrag *defrag)
85 {
86         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
87         struct inode_defrag *entry;
88         struct rb_node **p;
89         struct rb_node *parent = NULL;
90         int ret;
91
92         p = &fs_info->defrag_inodes.rb_node;
93         while (*p) {
94                 parent = *p;
95                 entry = rb_entry(parent, struct inode_defrag, rb_node);
96
97                 ret = __compare_inode_defrag(defrag, entry);
98                 if (ret < 0)
99                         p = &parent->rb_left;
100                 else if (ret > 0)
101                         p = &parent->rb_right;
102                 else {
103                         /* if we're reinserting an entry for
104                          * an old defrag run, make sure to
105                          * lower the transid of our existing record
106                          */
107                         if (defrag->transid < entry->transid)
108                                 entry->transid = defrag->transid;
109                         if (defrag->last_offset > entry->last_offset)
110                                 entry->last_offset = defrag->last_offset;
111                         return -EEXIST;
112                 }
113         }
114         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
115         rb_link_node(&defrag->rb_node, parent, p);
116         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
117         return 0;
118 }
119
120 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
121 {
122         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
123                 return 0;
124
125         if (btrfs_fs_closing(fs_info))
126                 return 0;
127
128         return 1;
129 }
130
131 /*
132  * insert a defrag record for this inode if auto defrag is
133  * enabled
134  */
135 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
136                            struct btrfs_inode *inode)
137 {
138         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
139         struct btrfs_root *root = inode->root;
140         struct inode_defrag *defrag;
141         u64 transid;
142         int ret;
143
144         if (!__need_auto_defrag(fs_info))
145                 return 0;
146
147         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
148                 return 0;
149
150         if (trans)
151                 transid = trans->transid;
152         else
153                 transid = inode->root->last_trans;
154
155         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
156         if (!defrag)
157                 return -ENOMEM;
158
159         defrag->ino = btrfs_ino(inode);
160         defrag->transid = transid;
161         defrag->root = root->root_key.objectid;
162
163         spin_lock(&fs_info->defrag_inodes_lock);
164         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
165                 /*
166                  * If we set IN_DEFRAG flag and evict the inode from memory,
167                  * and then re-read this inode, this new inode doesn't have
168                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
169                  */
170                 ret = __btrfs_add_inode_defrag(inode, defrag);
171                 if (ret)
172                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
173         } else {
174                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
175         }
176         spin_unlock(&fs_info->defrag_inodes_lock);
177         return 0;
178 }
179
180 /*
181  * Requeue the defrag object. If there is a defrag object that points to
182  * the same inode in the tree, we will merge them together (by
183  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
184  */
185 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
186                                        struct inode_defrag *defrag)
187 {
188         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
189         int ret;
190
191         if (!__need_auto_defrag(fs_info))
192                 goto out;
193
194         /*
195          * Here we don't check the IN_DEFRAG flag, because we need merge
196          * them together.
197          */
198         spin_lock(&fs_info->defrag_inodes_lock);
199         ret = __btrfs_add_inode_defrag(inode, defrag);
200         spin_unlock(&fs_info->defrag_inodes_lock);
201         if (ret)
202                 goto out;
203         return;
204 out:
205         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
206 }
207
208 /*
209  * pick the defragable inode that we want, if it doesn't exist, we will get
210  * the next one.
211  */
212 static struct inode_defrag *
213 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
214 {
215         struct inode_defrag *entry = NULL;
216         struct inode_defrag tmp;
217         struct rb_node *p;
218         struct rb_node *parent = NULL;
219         int ret;
220
221         tmp.ino = ino;
222         tmp.root = root;
223
224         spin_lock(&fs_info->defrag_inodes_lock);
225         p = fs_info->defrag_inodes.rb_node;
226         while (p) {
227                 parent = p;
228                 entry = rb_entry(parent, struct inode_defrag, rb_node);
229
230                 ret = __compare_inode_defrag(&tmp, entry);
231                 if (ret < 0)
232                         p = parent->rb_left;
233                 else if (ret > 0)
234                         p = parent->rb_right;
235                 else
236                         goto out;
237         }
238
239         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
240                 parent = rb_next(parent);
241                 if (parent)
242                         entry = rb_entry(parent, struct inode_defrag, rb_node);
243                 else
244                         entry = NULL;
245         }
246 out:
247         if (entry)
248                 rb_erase(parent, &fs_info->defrag_inodes);
249         spin_unlock(&fs_info->defrag_inodes_lock);
250         return entry;
251 }
252
253 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
254 {
255         struct inode_defrag *defrag;
256         struct rb_node *node;
257
258         spin_lock(&fs_info->defrag_inodes_lock);
259         node = rb_first(&fs_info->defrag_inodes);
260         while (node) {
261                 rb_erase(node, &fs_info->defrag_inodes);
262                 defrag = rb_entry(node, struct inode_defrag, rb_node);
263                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
264
265                 cond_resched_lock(&fs_info->defrag_inodes_lock);
266
267                 node = rb_first(&fs_info->defrag_inodes);
268         }
269         spin_unlock(&fs_info->defrag_inodes_lock);
270 }
271
272 #define BTRFS_DEFRAG_BATCH      1024
273
274 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
275                                     struct inode_defrag *defrag)
276 {
277         struct btrfs_root *inode_root;
278         struct inode *inode;
279         struct btrfs_key key;
280         struct btrfs_ioctl_defrag_range_args range;
281         int num_defrag;
282         int index;
283         int ret;
284
285         /* get the inode */
286         key.objectid = defrag->root;
287         key.type = BTRFS_ROOT_ITEM_KEY;
288         key.offset = (u64)-1;
289
290         index = srcu_read_lock(&fs_info->subvol_srcu);
291
292         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
293         if (IS_ERR(inode_root)) {
294                 ret = PTR_ERR(inode_root);
295                 goto cleanup;
296         }
297
298         key.objectid = defrag->ino;
299         key.type = BTRFS_INODE_ITEM_KEY;
300         key.offset = 0;
301         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
302         if (IS_ERR(inode)) {
303                 ret = PTR_ERR(inode);
304                 goto cleanup;
305         }
306         srcu_read_unlock(&fs_info->subvol_srcu, index);
307
308         /* do a chunk of defrag */
309         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
310         memset(&range, 0, sizeof(range));
311         range.len = (u64)-1;
312         range.start = defrag->last_offset;
313
314         sb_start_write(fs_info->sb);
315         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
316                                        BTRFS_DEFRAG_BATCH);
317         sb_end_write(fs_info->sb);
318         /*
319          * if we filled the whole defrag batch, there
320          * must be more work to do.  Queue this defrag
321          * again
322          */
323         if (num_defrag == BTRFS_DEFRAG_BATCH) {
324                 defrag->last_offset = range.start;
325                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
326         } else if (defrag->last_offset && !defrag->cycled) {
327                 /*
328                  * we didn't fill our defrag batch, but
329                  * we didn't start at zero.  Make sure we loop
330                  * around to the start of the file.
331                  */
332                 defrag->last_offset = 0;
333                 defrag->cycled = 1;
334                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
335         } else {
336                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
337         }
338
339         iput(inode);
340         return 0;
341 cleanup:
342         srcu_read_unlock(&fs_info->subvol_srcu, index);
343         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
344         return ret;
345 }
346
347 /*
348  * run through the list of inodes in the FS that need
349  * defragging
350  */
351 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
352 {
353         struct inode_defrag *defrag;
354         u64 first_ino = 0;
355         u64 root_objectid = 0;
356
357         atomic_inc(&fs_info->defrag_running);
358         while (1) {
359                 /* Pause the auto defragger. */
360                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
361                              &fs_info->fs_state))
362                         break;
363
364                 if (!__need_auto_defrag(fs_info))
365                         break;
366
367                 /* find an inode to defrag */
368                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
369                                                  first_ino);
370                 if (!defrag) {
371                         if (root_objectid || first_ino) {
372                                 root_objectid = 0;
373                                 first_ino = 0;
374                                 continue;
375                         } else {
376                                 break;
377                         }
378                 }
379
380                 first_ino = defrag->ino + 1;
381                 root_objectid = defrag->root;
382
383                 __btrfs_run_defrag_inode(fs_info, defrag);
384         }
385         atomic_dec(&fs_info->defrag_running);
386
387         /*
388          * during unmount, we use the transaction_wait queue to
389          * wait for the defragger to stop
390          */
391         wake_up(&fs_info->transaction_wait);
392         return 0;
393 }
394
395 /* simple helper to fault in pages and copy.  This should go away
396  * and be replaced with calls into generic code.
397  */
398 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
399                                          struct page **prepared_pages,
400                                          struct iov_iter *i)
401 {
402         size_t copied = 0;
403         size_t total_copied = 0;
404         int pg = 0;
405         int offset = pos & (PAGE_SIZE - 1);
406
407         while (write_bytes > 0) {
408                 size_t count = min_t(size_t,
409                                      PAGE_SIZE - offset, write_bytes);
410                 struct page *page = prepared_pages[pg];
411                 /*
412                  * Copy data from userspace to the current page
413                  */
414                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
415
416                 /* Flush processor's dcache for this page */
417                 flush_dcache_page(page);
418
419                 /*
420                  * if we get a partial write, we can end up with
421                  * partially up to date pages.  These add
422                  * a lot of complexity, so make sure they don't
423                  * happen by forcing this copy to be retried.
424                  *
425                  * The rest of the btrfs_file_write code will fall
426                  * back to page at a time copies after we return 0.
427                  */
428                 if (!PageUptodate(page) && copied < count)
429                         copied = 0;
430
431                 iov_iter_advance(i, copied);
432                 write_bytes -= copied;
433                 total_copied += copied;
434
435                 /* Return to btrfs_file_write_iter to fault page */
436                 if (unlikely(copied == 0))
437                         break;
438
439                 if (copied < PAGE_SIZE - offset) {
440                         offset += copied;
441                 } else {
442                         pg++;
443                         offset = 0;
444                 }
445         }
446         return total_copied;
447 }
448
449 /*
450  * unlocks pages after btrfs_file_write is done with them
451  */
452 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
453 {
454         size_t i;
455         for (i = 0; i < num_pages; i++) {
456                 /* page checked is some magic around finding pages that
457                  * have been modified without going through btrfs_set_page_dirty
458                  * clear it here. There should be no need to mark the pages
459                  * accessed as prepare_pages should have marked them accessed
460                  * in prepare_pages via find_or_create_page()
461                  */
462                 ClearPageChecked(pages[i]);
463                 unlock_page(pages[i]);
464                 put_page(pages[i]);
465         }
466 }
467
468 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
469                                          const u64 start,
470                                          const u64 len,
471                                          struct extent_state **cached_state)
472 {
473         u64 search_start = start;
474         const u64 end = start + len - 1;
475
476         while (search_start < end) {
477                 const u64 search_len = end - search_start + 1;
478                 struct extent_map *em;
479                 u64 em_len;
480                 int ret = 0;
481
482                 em = btrfs_get_extent(inode, NULL, 0, search_start,
483                                       search_len, 0);
484                 if (IS_ERR(em))
485                         return PTR_ERR(em);
486
487                 if (em->block_start != EXTENT_MAP_HOLE)
488                         goto next;
489
490                 em_len = em->len;
491                 if (em->start < search_start)
492                         em_len -= search_start - em->start;
493                 if (em_len > search_len)
494                         em_len = search_len;
495
496                 ret = set_extent_bit(&inode->io_tree, search_start,
497                                      search_start + em_len - 1,
498                                      EXTENT_DELALLOC_NEW,
499                                      NULL, cached_state, GFP_NOFS);
500 next:
501                 search_start = extent_map_end(em);
502                 free_extent_map(em);
503                 if (ret)
504                         return ret;
505         }
506         return 0;
507 }
508
509 /*
510  * after copy_from_user, pages need to be dirtied and we need to make
511  * sure holes are created between the current EOF and the start of
512  * any next extents (if required).
513  *
514  * this also makes the decision about creating an inline extent vs
515  * doing real data extents, marking pages dirty and delalloc as required.
516  */
517 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
518                       size_t num_pages, loff_t pos, size_t write_bytes,
519                       struct extent_state **cached)
520 {
521         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
522         int err = 0;
523         int i;
524         u64 num_bytes;
525         u64 start_pos;
526         u64 end_of_last_block;
527         u64 end_pos = pos + write_bytes;
528         loff_t isize = i_size_read(inode);
529         unsigned int extra_bits = 0;
530
531         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
532         num_bytes = round_up(write_bytes + pos - start_pos,
533                              fs_info->sectorsize);
534
535         end_of_last_block = start_pos + num_bytes - 1;
536
537         if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
538                 if (start_pos >= isize &&
539                     !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
540                         /*
541                          * There can't be any extents following eof in this case
542                          * so just set the delalloc new bit for the range
543                          * directly.
544                          */
545                         extra_bits |= EXTENT_DELALLOC_NEW;
546                 } else {
547                         err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
548                                                             start_pos,
549                                                             num_bytes, cached);
550                         if (err)
551                                 return err;
552                 }
553         }
554
555         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
556                                         extra_bits, cached, 0);
557         if (err)
558                 return err;
559
560         for (i = 0; i < num_pages; i++) {
561                 struct page *p = pages[i];
562                 SetPageUptodate(p);
563                 ClearPageChecked(p);
564                 set_page_dirty(p);
565         }
566
567         /*
568          * we've only changed i_size in ram, and we haven't updated
569          * the disk i_size.  There is no need to log the inode
570          * at this time.
571          */
572         if (end_pos > isize)
573                 i_size_write(inode, end_pos);
574         return 0;
575 }
576
577 /*
578  * this drops all the extents in the cache that intersect the range
579  * [start, end].  Existing extents are split as required.
580  */
581 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
582                              int skip_pinned)
583 {
584         struct extent_map *em;
585         struct extent_map *split = NULL;
586         struct extent_map *split2 = NULL;
587         struct extent_map_tree *em_tree = &inode->extent_tree;
588         u64 len = end - start + 1;
589         u64 gen;
590         int ret;
591         int testend = 1;
592         unsigned long flags;
593         int compressed = 0;
594         bool modified;
595
596         WARN_ON(end < start);
597         if (end == (u64)-1) {
598                 len = (u64)-1;
599                 testend = 0;
600         }
601         while (1) {
602                 int no_splits = 0;
603
604                 modified = false;
605                 if (!split)
606                         split = alloc_extent_map();
607                 if (!split2)
608                         split2 = alloc_extent_map();
609                 if (!split || !split2)
610                         no_splits = 1;
611
612                 write_lock(&em_tree->lock);
613                 em = lookup_extent_mapping(em_tree, start, len);
614                 if (!em) {
615                         write_unlock(&em_tree->lock);
616                         break;
617                 }
618                 flags = em->flags;
619                 gen = em->generation;
620                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
621                         if (testend && em->start + em->len >= start + len) {
622                                 free_extent_map(em);
623                                 write_unlock(&em_tree->lock);
624                                 break;
625                         }
626                         start = em->start + em->len;
627                         if (testend)
628                                 len = start + len - (em->start + em->len);
629                         free_extent_map(em);
630                         write_unlock(&em_tree->lock);
631                         continue;
632                 }
633                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
634                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
635                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
636                 modified = !list_empty(&em->list);
637                 if (no_splits)
638                         goto next;
639
640                 if (em->start < start) {
641                         split->start = em->start;
642                         split->len = start - em->start;
643
644                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
645                                 split->orig_start = em->orig_start;
646                                 split->block_start = em->block_start;
647
648                                 if (compressed)
649                                         split->block_len = em->block_len;
650                                 else
651                                         split->block_len = split->len;
652                                 split->orig_block_len = max(split->block_len,
653                                                 em->orig_block_len);
654                                 split->ram_bytes = em->ram_bytes;
655                         } else {
656                                 split->orig_start = split->start;
657                                 split->block_len = 0;
658                                 split->block_start = em->block_start;
659                                 split->orig_block_len = 0;
660                                 split->ram_bytes = split->len;
661                         }
662
663                         split->generation = gen;
664                         split->bdev = em->bdev;
665                         split->flags = flags;
666                         split->compress_type = em->compress_type;
667                         replace_extent_mapping(em_tree, em, split, modified);
668                         free_extent_map(split);
669                         split = split2;
670                         split2 = NULL;
671                 }
672                 if (testend && em->start + em->len > start + len) {
673                         u64 diff = start + len - em->start;
674
675                         split->start = start + len;
676                         split->len = em->start + em->len - (start + len);
677                         split->bdev = em->bdev;
678                         split->flags = flags;
679                         split->compress_type = em->compress_type;
680                         split->generation = gen;
681
682                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
683                                 split->orig_block_len = max(em->block_len,
684                                                     em->orig_block_len);
685
686                                 split->ram_bytes = em->ram_bytes;
687                                 if (compressed) {
688                                         split->block_len = em->block_len;
689                                         split->block_start = em->block_start;
690                                         split->orig_start = em->orig_start;
691                                 } else {
692                                         split->block_len = split->len;
693                                         split->block_start = em->block_start
694                                                 + diff;
695                                         split->orig_start = em->orig_start;
696                                 }
697                         } else {
698                                 split->ram_bytes = split->len;
699                                 split->orig_start = split->start;
700                                 split->block_len = 0;
701                                 split->block_start = em->block_start;
702                                 split->orig_block_len = 0;
703                         }
704
705                         if (extent_map_in_tree(em)) {
706                                 replace_extent_mapping(em_tree, em, split,
707                                                        modified);
708                         } else {
709                                 ret = add_extent_mapping(em_tree, split,
710                                                          modified);
711                                 ASSERT(ret == 0); /* Logic error */
712                         }
713                         free_extent_map(split);
714                         split = NULL;
715                 }
716 next:
717                 if (extent_map_in_tree(em))
718                         remove_extent_mapping(em_tree, em);
719                 write_unlock(&em_tree->lock);
720
721                 /* once for us */
722                 free_extent_map(em);
723                 /* once for the tree*/
724                 free_extent_map(em);
725         }
726         if (split)
727                 free_extent_map(split);
728         if (split2)
729                 free_extent_map(split2);
730 }
731
732 /*
733  * this is very complex, but the basic idea is to drop all extents
734  * in the range start - end.  hint_block is filled in with a block number
735  * that would be a good hint to the block allocator for this file.
736  *
737  * If an extent intersects the range but is not entirely inside the range
738  * it is either truncated or split.  Anything entirely inside the range
739  * is deleted from the tree.
740  */
741 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
742                          struct btrfs_root *root, struct inode *inode,
743                          struct btrfs_path *path, u64 start, u64 end,
744                          u64 *drop_end, int drop_cache,
745                          int replace_extent,
746                          u32 extent_item_size,
747                          int *key_inserted)
748 {
749         struct btrfs_fs_info *fs_info = root->fs_info;
750         struct extent_buffer *leaf;
751         struct btrfs_file_extent_item *fi;
752         struct btrfs_key key;
753         struct btrfs_key new_key;
754         u64 ino = btrfs_ino(BTRFS_I(inode));
755         u64 search_start = start;
756         u64 disk_bytenr = 0;
757         u64 num_bytes = 0;
758         u64 extent_offset = 0;
759         u64 extent_end = 0;
760         u64 last_end = start;
761         int del_nr = 0;
762         int del_slot = 0;
763         int extent_type;
764         int recow;
765         int ret;
766         int modify_tree = -1;
767         int update_refs;
768         int found = 0;
769         int leafs_visited = 0;
770
771         if (drop_cache)
772                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
773
774         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
775                 modify_tree = 0;
776
777         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
778                        root == fs_info->tree_root);
779         while (1) {
780                 recow = 0;
781                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
782                                                search_start, modify_tree);
783                 if (ret < 0)
784                         break;
785                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
786                         leaf = path->nodes[0];
787                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
788                         if (key.objectid == ino &&
789                             key.type == BTRFS_EXTENT_DATA_KEY)
790                                 path->slots[0]--;
791                 }
792                 ret = 0;
793                 leafs_visited++;
794 next_slot:
795                 leaf = path->nodes[0];
796                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
797                         BUG_ON(del_nr > 0);
798                         ret = btrfs_next_leaf(root, path);
799                         if (ret < 0)
800                                 break;
801                         if (ret > 0) {
802                                 ret = 0;
803                                 break;
804                         }
805                         leafs_visited++;
806                         leaf = path->nodes[0];
807                         recow = 1;
808                 }
809
810                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
811
812                 if (key.objectid > ino)
813                         break;
814                 if (WARN_ON_ONCE(key.objectid < ino) ||
815                     key.type < BTRFS_EXTENT_DATA_KEY) {
816                         ASSERT(del_nr == 0);
817                         path->slots[0]++;
818                         goto next_slot;
819                 }
820                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
821                         break;
822
823                 fi = btrfs_item_ptr(leaf, path->slots[0],
824                                     struct btrfs_file_extent_item);
825                 extent_type = btrfs_file_extent_type(leaf, fi);
826
827                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
828                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
829                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
830                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
831                         extent_offset = btrfs_file_extent_offset(leaf, fi);
832                         extent_end = key.offset +
833                                 btrfs_file_extent_num_bytes(leaf, fi);
834                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
835                         extent_end = key.offset +
836                                 btrfs_file_extent_ram_bytes(leaf, fi);
837                 } else {
838                         /* can't happen */
839                         BUG();
840                 }
841
842                 /*
843                  * Don't skip extent items representing 0 byte lengths. They
844                  * used to be created (bug) if while punching holes we hit
845                  * -ENOSPC condition. So if we find one here, just ensure we
846                  * delete it, otherwise we would insert a new file extent item
847                  * with the same key (offset) as that 0 bytes length file
848                  * extent item in the call to setup_items_for_insert() later
849                  * in this function.
850                  */
851                 if (extent_end == key.offset && extent_end >= search_start) {
852                         last_end = extent_end;
853                         goto delete_extent_item;
854                 }
855
856                 if (extent_end <= search_start) {
857                         path->slots[0]++;
858                         goto next_slot;
859                 }
860
861                 found = 1;
862                 search_start = max(key.offset, start);
863                 if (recow || !modify_tree) {
864                         modify_tree = -1;
865                         btrfs_release_path(path);
866                         continue;
867                 }
868
869                 /*
870                  *     | - range to drop - |
871                  *  | -------- extent -------- |
872                  */
873                 if (start > key.offset && end < extent_end) {
874                         BUG_ON(del_nr > 0);
875                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
876                                 ret = -EOPNOTSUPP;
877                                 break;
878                         }
879
880                         memcpy(&new_key, &key, sizeof(new_key));
881                         new_key.offset = start;
882                         ret = btrfs_duplicate_item(trans, root, path,
883                                                    &new_key);
884                         if (ret == -EAGAIN) {
885                                 btrfs_release_path(path);
886                                 continue;
887                         }
888                         if (ret < 0)
889                                 break;
890
891                         leaf = path->nodes[0];
892                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
893                                             struct btrfs_file_extent_item);
894                         btrfs_set_file_extent_num_bytes(leaf, fi,
895                                                         start - key.offset);
896
897                         fi = btrfs_item_ptr(leaf, path->slots[0],
898                                             struct btrfs_file_extent_item);
899
900                         extent_offset += start - key.offset;
901                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
902                         btrfs_set_file_extent_num_bytes(leaf, fi,
903                                                         extent_end - start);
904                         btrfs_mark_buffer_dirty(leaf);
905
906                         if (update_refs && disk_bytenr > 0) {
907                                 ret = btrfs_inc_extent_ref(trans, root,
908                                                 disk_bytenr, num_bytes, 0,
909                                                 root->root_key.objectid,
910                                                 new_key.objectid,
911                                                 start - extent_offset);
912                                 BUG_ON(ret); /* -ENOMEM */
913                         }
914                         key.offset = start;
915                 }
916                 /*
917                  * From here on out we will have actually dropped something, so
918                  * last_end can be updated.
919                  */
920                 last_end = extent_end;
921
922                 /*
923                  *  | ---- range to drop ----- |
924                  *      | -------- extent -------- |
925                  */
926                 if (start <= key.offset && end < extent_end) {
927                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
928                                 ret = -EOPNOTSUPP;
929                                 break;
930                         }
931
932                         memcpy(&new_key, &key, sizeof(new_key));
933                         new_key.offset = end;
934                         btrfs_set_item_key_safe(fs_info, path, &new_key);
935
936                         extent_offset += end - key.offset;
937                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
938                         btrfs_set_file_extent_num_bytes(leaf, fi,
939                                                         extent_end - end);
940                         btrfs_mark_buffer_dirty(leaf);
941                         if (update_refs && disk_bytenr > 0)
942                                 inode_sub_bytes(inode, end - key.offset);
943                         break;
944                 }
945
946                 search_start = extent_end;
947                 /*
948                  *       | ---- range to drop ----- |
949                  *  | -------- extent -------- |
950                  */
951                 if (start > key.offset && end >= extent_end) {
952                         BUG_ON(del_nr > 0);
953                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
954                                 ret = -EOPNOTSUPP;
955                                 break;
956                         }
957
958                         btrfs_set_file_extent_num_bytes(leaf, fi,
959                                                         start - key.offset);
960                         btrfs_mark_buffer_dirty(leaf);
961                         if (update_refs && disk_bytenr > 0)
962                                 inode_sub_bytes(inode, extent_end - start);
963                         if (end == extent_end)
964                                 break;
965
966                         path->slots[0]++;
967                         goto next_slot;
968                 }
969
970                 /*
971                  *  | ---- range to drop ----- |
972                  *    | ------ extent ------ |
973                  */
974                 if (start <= key.offset && end >= extent_end) {
975 delete_extent_item:
976                         if (del_nr == 0) {
977                                 del_slot = path->slots[0];
978                                 del_nr = 1;
979                         } else {
980                                 BUG_ON(del_slot + del_nr != path->slots[0]);
981                                 del_nr++;
982                         }
983
984                         if (update_refs &&
985                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
986                                 inode_sub_bytes(inode,
987                                                 extent_end - key.offset);
988                                 extent_end = ALIGN(extent_end,
989                                                    fs_info->sectorsize);
990                         } else if (update_refs && disk_bytenr > 0) {
991                                 ret = btrfs_free_extent(trans, root,
992                                                 disk_bytenr, num_bytes, 0,
993                                                 root->root_key.objectid,
994                                                 key.objectid, key.offset -
995                                                 extent_offset);
996                                 BUG_ON(ret); /* -ENOMEM */
997                                 inode_sub_bytes(inode,
998                                                 extent_end - key.offset);
999                         }
1000
1001                         if (end == extent_end)
1002                                 break;
1003
1004                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1005                                 path->slots[0]++;
1006                                 goto next_slot;
1007                         }
1008
1009                         ret = btrfs_del_items(trans, root, path, del_slot,
1010                                               del_nr);
1011                         if (ret) {
1012                                 btrfs_abort_transaction(trans, ret);
1013                                 break;
1014                         }
1015
1016                         del_nr = 0;
1017                         del_slot = 0;
1018
1019                         btrfs_release_path(path);
1020                         continue;
1021                 }
1022
1023                 BUG_ON(1);
1024         }
1025
1026         if (!ret && del_nr > 0) {
1027                 /*
1028                  * Set path->slots[0] to first slot, so that after the delete
1029                  * if items are move off from our leaf to its immediate left or
1030                  * right neighbor leafs, we end up with a correct and adjusted
1031                  * path->slots[0] for our insertion (if replace_extent != 0).
1032                  */
1033                 path->slots[0] = del_slot;
1034                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1035                 if (ret)
1036                         btrfs_abort_transaction(trans, ret);
1037         }
1038
1039         leaf = path->nodes[0];
1040         /*
1041          * If btrfs_del_items() was called, it might have deleted a leaf, in
1042          * which case it unlocked our path, so check path->locks[0] matches a
1043          * write lock.
1044          */
1045         if (!ret && replace_extent && leafs_visited == 1 &&
1046             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1047              path->locks[0] == BTRFS_WRITE_LOCK) &&
1048             btrfs_leaf_free_space(fs_info, leaf) >=
1049             sizeof(struct btrfs_item) + extent_item_size) {
1050
1051                 key.objectid = ino;
1052                 key.type = BTRFS_EXTENT_DATA_KEY;
1053                 key.offset = start;
1054                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1055                         struct btrfs_key slot_key;
1056
1057                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1058                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1059                                 path->slots[0]++;
1060                 }
1061                 setup_items_for_insert(root, path, &key,
1062                                        &extent_item_size,
1063                                        extent_item_size,
1064                                        sizeof(struct btrfs_item) +
1065                                        extent_item_size, 1);
1066                 *key_inserted = 1;
1067         }
1068
1069         if (!replace_extent || !(*key_inserted))
1070                 btrfs_release_path(path);
1071         if (drop_end)
1072                 *drop_end = found ? min(end, last_end) : end;
1073         return ret;
1074 }
1075
1076 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1077                        struct btrfs_root *root, struct inode *inode, u64 start,
1078                        u64 end, int drop_cache)
1079 {
1080         struct btrfs_path *path;
1081         int ret;
1082
1083         path = btrfs_alloc_path();
1084         if (!path)
1085                 return -ENOMEM;
1086         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1087                                    drop_cache, 0, 0, NULL);
1088         btrfs_free_path(path);
1089         return ret;
1090 }
1091
1092 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1093                             u64 objectid, u64 bytenr, u64 orig_offset,
1094                             u64 *start, u64 *end)
1095 {
1096         struct btrfs_file_extent_item *fi;
1097         struct btrfs_key key;
1098         u64 extent_end;
1099
1100         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1101                 return 0;
1102
1103         btrfs_item_key_to_cpu(leaf, &key, slot);
1104         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1105                 return 0;
1106
1107         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1108         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1109             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1110             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1111             btrfs_file_extent_compression(leaf, fi) ||
1112             btrfs_file_extent_encryption(leaf, fi) ||
1113             btrfs_file_extent_other_encoding(leaf, fi))
1114                 return 0;
1115
1116         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1117         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1118                 return 0;
1119
1120         *start = key.offset;
1121         *end = extent_end;
1122         return 1;
1123 }
1124
1125 /*
1126  * Mark extent in the range start - end as written.
1127  *
1128  * This changes extent type from 'pre-allocated' to 'regular'. If only
1129  * part of extent is marked as written, the extent will be split into
1130  * two or three.
1131  */
1132 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1133                               struct btrfs_inode *inode, u64 start, u64 end)
1134 {
1135         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1136         struct btrfs_root *root = inode->root;
1137         struct extent_buffer *leaf;
1138         struct btrfs_path *path;
1139         struct btrfs_file_extent_item *fi;
1140         struct btrfs_key key;
1141         struct btrfs_key new_key;
1142         u64 bytenr;
1143         u64 num_bytes;
1144         u64 extent_end;
1145         u64 orig_offset;
1146         u64 other_start;
1147         u64 other_end;
1148         u64 split;
1149         int del_nr = 0;
1150         int del_slot = 0;
1151         int recow;
1152         int ret;
1153         u64 ino = btrfs_ino(inode);
1154
1155         path = btrfs_alloc_path();
1156         if (!path)
1157                 return -ENOMEM;
1158 again:
1159         recow = 0;
1160         split = start;
1161         key.objectid = ino;
1162         key.type = BTRFS_EXTENT_DATA_KEY;
1163         key.offset = split;
1164
1165         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1166         if (ret < 0)
1167                 goto out;
1168         if (ret > 0 && path->slots[0] > 0)
1169                 path->slots[0]--;
1170
1171         leaf = path->nodes[0];
1172         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1173         if (key.objectid != ino ||
1174             key.type != BTRFS_EXTENT_DATA_KEY) {
1175                 ret = -EINVAL;
1176                 btrfs_abort_transaction(trans, ret);
1177                 goto out;
1178         }
1179         fi = btrfs_item_ptr(leaf, path->slots[0],
1180                             struct btrfs_file_extent_item);
1181         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1182                 ret = -EINVAL;
1183                 btrfs_abort_transaction(trans, ret);
1184                 goto out;
1185         }
1186         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1187         if (key.offset > start || extent_end < end) {
1188                 ret = -EINVAL;
1189                 btrfs_abort_transaction(trans, ret);
1190                 goto out;
1191         }
1192
1193         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1194         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1195         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1196         memcpy(&new_key, &key, sizeof(new_key));
1197
1198         if (start == key.offset && end < extent_end) {
1199                 other_start = 0;
1200                 other_end = start;
1201                 if (extent_mergeable(leaf, path->slots[0] - 1,
1202                                      ino, bytenr, orig_offset,
1203                                      &other_start, &other_end)) {
1204                         new_key.offset = end;
1205                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1206                         fi = btrfs_item_ptr(leaf, path->slots[0],
1207                                             struct btrfs_file_extent_item);
1208                         btrfs_set_file_extent_generation(leaf, fi,
1209                                                          trans->transid);
1210                         btrfs_set_file_extent_num_bytes(leaf, fi,
1211                                                         extent_end - end);
1212                         btrfs_set_file_extent_offset(leaf, fi,
1213                                                      end - orig_offset);
1214                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1215                                             struct btrfs_file_extent_item);
1216                         btrfs_set_file_extent_generation(leaf, fi,
1217                                                          trans->transid);
1218                         btrfs_set_file_extent_num_bytes(leaf, fi,
1219                                                         end - other_start);
1220                         btrfs_mark_buffer_dirty(leaf);
1221                         goto out;
1222                 }
1223         }
1224
1225         if (start > key.offset && end == extent_end) {
1226                 other_start = end;
1227                 other_end = 0;
1228                 if (extent_mergeable(leaf, path->slots[0] + 1,
1229                                      ino, bytenr, orig_offset,
1230                                      &other_start, &other_end)) {
1231                         fi = btrfs_item_ptr(leaf, path->slots[0],
1232                                             struct btrfs_file_extent_item);
1233                         btrfs_set_file_extent_num_bytes(leaf, fi,
1234                                                         start - key.offset);
1235                         btrfs_set_file_extent_generation(leaf, fi,
1236                                                          trans->transid);
1237                         path->slots[0]++;
1238                         new_key.offset = start;
1239                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1240
1241                         fi = btrfs_item_ptr(leaf, path->slots[0],
1242                                             struct btrfs_file_extent_item);
1243                         btrfs_set_file_extent_generation(leaf, fi,
1244                                                          trans->transid);
1245                         btrfs_set_file_extent_num_bytes(leaf, fi,
1246                                                         other_end - start);
1247                         btrfs_set_file_extent_offset(leaf, fi,
1248                                                      start - orig_offset);
1249                         btrfs_mark_buffer_dirty(leaf);
1250                         goto out;
1251                 }
1252         }
1253
1254         while (start > key.offset || end < extent_end) {
1255                 if (key.offset == start)
1256                         split = end;
1257
1258                 new_key.offset = split;
1259                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1260                 if (ret == -EAGAIN) {
1261                         btrfs_release_path(path);
1262                         goto again;
1263                 }
1264                 if (ret < 0) {
1265                         btrfs_abort_transaction(trans, ret);
1266                         goto out;
1267                 }
1268
1269                 leaf = path->nodes[0];
1270                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1271                                     struct btrfs_file_extent_item);
1272                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1273                 btrfs_set_file_extent_num_bytes(leaf, fi,
1274                                                 split - key.offset);
1275
1276                 fi = btrfs_item_ptr(leaf, path->slots[0],
1277                                     struct btrfs_file_extent_item);
1278
1279                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1280                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1281                 btrfs_set_file_extent_num_bytes(leaf, fi,
1282                                                 extent_end - split);
1283                 btrfs_mark_buffer_dirty(leaf);
1284
1285                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes,
1286                                            0, root->root_key.objectid,
1287                                            ino, orig_offset);
1288                 if (ret) {
1289                         btrfs_abort_transaction(trans, ret);
1290                         goto out;
1291                 }
1292
1293                 if (split == start) {
1294                         key.offset = start;
1295                 } else {
1296                         if (start != key.offset) {
1297                                 ret = -EINVAL;
1298                                 btrfs_abort_transaction(trans, ret);
1299                                 goto out;
1300                         }
1301                         path->slots[0]--;
1302                         extent_end = end;
1303                 }
1304                 recow = 1;
1305         }
1306
1307         other_start = end;
1308         other_end = 0;
1309         if (extent_mergeable(leaf, path->slots[0] + 1,
1310                              ino, bytenr, orig_offset,
1311                              &other_start, &other_end)) {
1312                 if (recow) {
1313                         btrfs_release_path(path);
1314                         goto again;
1315                 }
1316                 extent_end = other_end;
1317                 del_slot = path->slots[0] + 1;
1318                 del_nr++;
1319                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1320                                         0, root->root_key.objectid,
1321                                         ino, orig_offset);
1322                 if (ret) {
1323                         btrfs_abort_transaction(trans, ret);
1324                         goto out;
1325                 }
1326         }
1327         other_start = 0;
1328         other_end = start;
1329         if (extent_mergeable(leaf, path->slots[0] - 1,
1330                              ino, bytenr, orig_offset,
1331                              &other_start, &other_end)) {
1332                 if (recow) {
1333                         btrfs_release_path(path);
1334                         goto again;
1335                 }
1336                 key.offset = other_start;
1337                 del_slot = path->slots[0];
1338                 del_nr++;
1339                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1340                                         0, root->root_key.objectid,
1341                                         ino, orig_offset);
1342                 if (ret) {
1343                         btrfs_abort_transaction(trans, ret);
1344                         goto out;
1345                 }
1346         }
1347         if (del_nr == 0) {
1348                 fi = btrfs_item_ptr(leaf, path->slots[0],
1349                            struct btrfs_file_extent_item);
1350                 btrfs_set_file_extent_type(leaf, fi,
1351                                            BTRFS_FILE_EXTENT_REG);
1352                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1353                 btrfs_mark_buffer_dirty(leaf);
1354         } else {
1355                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1356                            struct btrfs_file_extent_item);
1357                 btrfs_set_file_extent_type(leaf, fi,
1358                                            BTRFS_FILE_EXTENT_REG);
1359                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1360                 btrfs_set_file_extent_num_bytes(leaf, fi,
1361                                                 extent_end - key.offset);
1362                 btrfs_mark_buffer_dirty(leaf);
1363
1364                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1365                 if (ret < 0) {
1366                         btrfs_abort_transaction(trans, ret);
1367                         goto out;
1368                 }
1369         }
1370 out:
1371         btrfs_free_path(path);
1372         return 0;
1373 }
1374
1375 /*
1376  * on error we return an unlocked page and the error value
1377  * on success we return a locked page and 0
1378  */
1379 static int prepare_uptodate_page(struct inode *inode,
1380                                  struct page *page, u64 pos,
1381                                  bool force_uptodate)
1382 {
1383         int ret = 0;
1384
1385         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1386             !PageUptodate(page)) {
1387                 ret = btrfs_readpage(NULL, page);
1388                 if (ret)
1389                         return ret;
1390                 lock_page(page);
1391                 if (!PageUptodate(page)) {
1392                         unlock_page(page);
1393                         return -EIO;
1394                 }
1395                 if (page->mapping != inode->i_mapping) {
1396                         unlock_page(page);
1397                         return -EAGAIN;
1398                 }
1399         }
1400         return 0;
1401 }
1402
1403 /*
1404  * this just gets pages into the page cache and locks them down.
1405  */
1406 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1407                                   size_t num_pages, loff_t pos,
1408                                   size_t write_bytes, bool force_uptodate)
1409 {
1410         int i;
1411         unsigned long index = pos >> PAGE_SHIFT;
1412         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1413         int err = 0;
1414         int faili;
1415
1416         for (i = 0; i < num_pages; i++) {
1417 again:
1418                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1419                                                mask | __GFP_WRITE);
1420                 if (!pages[i]) {
1421                         faili = i - 1;
1422                         err = -ENOMEM;
1423                         goto fail;
1424                 }
1425
1426                 if (i == 0)
1427                         err = prepare_uptodate_page(inode, pages[i], pos,
1428                                                     force_uptodate);
1429                 if (!err && i == num_pages - 1)
1430                         err = prepare_uptodate_page(inode, pages[i],
1431                                                     pos + write_bytes, false);
1432                 if (err) {
1433                         put_page(pages[i]);
1434                         if (err == -EAGAIN) {
1435                                 err = 0;
1436                                 goto again;
1437                         }
1438                         faili = i - 1;
1439                         goto fail;
1440                 }
1441                 wait_on_page_writeback(pages[i]);
1442         }
1443
1444         return 0;
1445 fail:
1446         while (faili >= 0) {
1447                 unlock_page(pages[faili]);
1448                 put_page(pages[faili]);
1449                 faili--;
1450         }
1451         return err;
1452
1453 }
1454
1455 /*
1456  * This function locks the extent and properly waits for data=ordered extents
1457  * to finish before allowing the pages to be modified if need.
1458  *
1459  * The return value:
1460  * 1 - the extent is locked
1461  * 0 - the extent is not locked, and everything is OK
1462  * -EAGAIN - need re-prepare the pages
1463  * the other < 0 number - Something wrong happens
1464  */
1465 static noinline int
1466 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1467                                 size_t num_pages, loff_t pos,
1468                                 size_t write_bytes,
1469                                 u64 *lockstart, u64 *lockend,
1470                                 struct extent_state **cached_state)
1471 {
1472         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1473         u64 start_pos;
1474         u64 last_pos;
1475         int i;
1476         int ret = 0;
1477
1478         start_pos = round_down(pos, fs_info->sectorsize);
1479         last_pos = start_pos
1480                 + round_up(pos + write_bytes - start_pos,
1481                            fs_info->sectorsize) - 1;
1482
1483         if (start_pos < inode->vfs_inode.i_size) {
1484                 struct btrfs_ordered_extent *ordered;
1485
1486                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1487                                 cached_state);
1488                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1489                                                      last_pos - start_pos + 1);
1490                 if (ordered &&
1491                     ordered->file_offset + ordered->len > start_pos &&
1492                     ordered->file_offset <= last_pos) {
1493                         unlock_extent_cached(&inode->io_tree, start_pos,
1494                                         last_pos, cached_state);
1495                         for (i = 0; i < num_pages; i++) {
1496                                 unlock_page(pages[i]);
1497                                 put_page(pages[i]);
1498                         }
1499                         btrfs_start_ordered_extent(&inode->vfs_inode,
1500                                         ordered, 1);
1501                         btrfs_put_ordered_extent(ordered);
1502                         return -EAGAIN;
1503                 }
1504                 if (ordered)
1505                         btrfs_put_ordered_extent(ordered);
1506                 clear_extent_bit(&inode->io_tree, start_pos, last_pos,
1507                                  EXTENT_DIRTY | EXTENT_DELALLOC |
1508                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1509                                  0, 0, cached_state);
1510                 *lockstart = start_pos;
1511                 *lockend = last_pos;
1512                 ret = 1;
1513         }
1514
1515         for (i = 0; i < num_pages; i++) {
1516                 if (clear_page_dirty_for_io(pages[i]))
1517                         account_page_redirty(pages[i]);
1518                 set_page_extent_mapped(pages[i]);
1519                 WARN_ON(!PageLocked(pages[i]));
1520         }
1521
1522         return ret;
1523 }
1524
1525 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1526                                     size_t *write_bytes)
1527 {
1528         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1529         struct btrfs_root *root = inode->root;
1530         struct btrfs_ordered_extent *ordered;
1531         u64 lockstart, lockend;
1532         u64 num_bytes;
1533         int ret;
1534
1535         ret = btrfs_start_write_no_snapshotting(root);
1536         if (!ret)
1537                 return -ENOSPC;
1538
1539         lockstart = round_down(pos, fs_info->sectorsize);
1540         lockend = round_up(pos + *write_bytes,
1541                            fs_info->sectorsize) - 1;
1542
1543         while (1) {
1544                 lock_extent(&inode->io_tree, lockstart, lockend);
1545                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1546                                                      lockend - lockstart + 1);
1547                 if (!ordered) {
1548                         break;
1549                 }
1550                 unlock_extent(&inode->io_tree, lockstart, lockend);
1551                 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1552                 btrfs_put_ordered_extent(ordered);
1553         }
1554
1555         num_bytes = lockend - lockstart + 1;
1556         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1557                         NULL, NULL, NULL);
1558         if (ret <= 0) {
1559                 ret = 0;
1560                 btrfs_end_write_no_snapshotting(root);
1561         } else {
1562                 *write_bytes = min_t(size_t, *write_bytes ,
1563                                      num_bytes - pos + lockstart);
1564         }
1565
1566         unlock_extent(&inode->io_tree, lockstart, lockend);
1567
1568         return ret;
1569 }
1570
1571 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1572                                                struct iov_iter *i)
1573 {
1574         struct file *file = iocb->ki_filp;
1575         loff_t pos = iocb->ki_pos;
1576         struct inode *inode = file_inode(file);
1577         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1578         struct btrfs_root *root = BTRFS_I(inode)->root;
1579         struct page **pages = NULL;
1580         struct extent_state *cached_state = NULL;
1581         struct extent_changeset *data_reserved = NULL;
1582         u64 release_bytes = 0;
1583         u64 lockstart;
1584         u64 lockend;
1585         size_t num_written = 0;
1586         int nrptrs;
1587         int ret = 0;
1588         bool only_release_metadata = false;
1589         bool force_page_uptodate = false;
1590
1591         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1592                         PAGE_SIZE / (sizeof(struct page *)));
1593         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1594         nrptrs = max(nrptrs, 8);
1595         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1596         if (!pages)
1597                 return -ENOMEM;
1598
1599         while (iov_iter_count(i) > 0) {
1600                 size_t offset = pos & (PAGE_SIZE - 1);
1601                 size_t sector_offset;
1602                 size_t write_bytes = min(iov_iter_count(i),
1603                                          nrptrs * (size_t)PAGE_SIZE -
1604                                          offset);
1605                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1606                                                 PAGE_SIZE);
1607                 size_t reserve_bytes;
1608                 size_t dirty_pages;
1609                 size_t copied;
1610                 size_t dirty_sectors;
1611                 size_t num_sectors;
1612                 int extents_locked;
1613
1614                 WARN_ON(num_pages > nrptrs);
1615
1616                 /*
1617                  * Fault pages before locking them in prepare_pages
1618                  * to avoid recursive lock
1619                  */
1620                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1621                         ret = -EFAULT;
1622                         break;
1623                 }
1624
1625                 sector_offset = pos & (fs_info->sectorsize - 1);
1626                 reserve_bytes = round_up(write_bytes + sector_offset,
1627                                 fs_info->sectorsize);
1628
1629                 extent_changeset_release(data_reserved);
1630                 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1631                                                   write_bytes);
1632                 if (ret < 0) {
1633                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1634                                                       BTRFS_INODE_PREALLOC)) &&
1635                             check_can_nocow(BTRFS_I(inode), pos,
1636                                         &write_bytes) > 0) {
1637                                 /*
1638                                  * For nodata cow case, no need to reserve
1639                                  * data space.
1640                                  */
1641                                 only_release_metadata = true;
1642                                 /*
1643                                  * our prealloc extent may be smaller than
1644                                  * write_bytes, so scale down.
1645                                  */
1646                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1647                                                          PAGE_SIZE);
1648                                 reserve_bytes = round_up(write_bytes +
1649                                                          sector_offset,
1650                                                          fs_info->sectorsize);
1651                         } else {
1652                                 break;
1653                         }
1654                 }
1655
1656                 WARN_ON(reserve_bytes == 0);
1657                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1658                                 reserve_bytes);
1659                 if (ret) {
1660                         if (!only_release_metadata)
1661                                 btrfs_free_reserved_data_space(inode,
1662                                                 data_reserved, pos,
1663                                                 write_bytes);
1664                         else
1665                                 btrfs_end_write_no_snapshotting(root);
1666                         break;
1667                 }
1668
1669                 release_bytes = reserve_bytes;
1670 again:
1671                 /*
1672                  * This is going to setup the pages array with the number of
1673                  * pages we want, so we don't really need to worry about the
1674                  * contents of pages from loop to loop
1675                  */
1676                 ret = prepare_pages(inode, pages, num_pages,
1677                                     pos, write_bytes,
1678                                     force_page_uptodate);
1679                 if (ret) {
1680                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1681                                                        reserve_bytes, true);
1682                         break;
1683                 }
1684
1685                 extents_locked = lock_and_cleanup_extent_if_need(
1686                                 BTRFS_I(inode), pages,
1687                                 num_pages, pos, write_bytes, &lockstart,
1688                                 &lockend, &cached_state);
1689                 if (extents_locked < 0) {
1690                         if (extents_locked == -EAGAIN)
1691                                 goto again;
1692                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1693                                                        reserve_bytes, true);
1694                         ret = extents_locked;
1695                         break;
1696                 }
1697
1698                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1699
1700                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1701                 dirty_sectors = round_up(copied + sector_offset,
1702                                         fs_info->sectorsize);
1703                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1704
1705                 /*
1706                  * if we have trouble faulting in the pages, fall
1707                  * back to one page at a time
1708                  */
1709                 if (copied < write_bytes)
1710                         nrptrs = 1;
1711
1712                 if (copied == 0) {
1713                         force_page_uptodate = true;
1714                         dirty_sectors = 0;
1715                         dirty_pages = 0;
1716                 } else {
1717                         force_page_uptodate = false;
1718                         dirty_pages = DIV_ROUND_UP(copied + offset,
1719                                                    PAGE_SIZE);
1720                 }
1721
1722                 if (num_sectors > dirty_sectors) {
1723                         /* release everything except the sectors we dirtied */
1724                         release_bytes -= dirty_sectors <<
1725                                                 fs_info->sb->s_blocksize_bits;
1726                         if (only_release_metadata) {
1727                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1728                                                         release_bytes, true);
1729                         } else {
1730                                 u64 __pos;
1731
1732                                 __pos = round_down(pos,
1733                                                    fs_info->sectorsize) +
1734                                         (dirty_pages << PAGE_SHIFT);
1735                                 btrfs_delalloc_release_space(inode,
1736                                                 data_reserved, __pos,
1737                                                 release_bytes, true);
1738                         }
1739                 }
1740
1741                 release_bytes = round_up(copied + sector_offset,
1742                                         fs_info->sectorsize);
1743
1744                 if (copied > 0)
1745                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1746                                                 pos, copied, &cached_state);
1747                 if (extents_locked)
1748                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1749                                              lockstart, lockend, &cached_state);
1750                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes,
1751                                                true);
1752                 if (ret) {
1753                         btrfs_drop_pages(pages, num_pages);
1754                         break;
1755                 }
1756
1757                 release_bytes = 0;
1758                 if (only_release_metadata)
1759                         btrfs_end_write_no_snapshotting(root);
1760
1761                 if (only_release_metadata && copied > 0) {
1762                         lockstart = round_down(pos,
1763                                                fs_info->sectorsize);
1764                         lockend = round_up(pos + copied,
1765                                            fs_info->sectorsize) - 1;
1766
1767                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1768                                        lockend, EXTENT_NORESERVE, NULL,
1769                                        NULL, GFP_NOFS);
1770                         only_release_metadata = false;
1771                 }
1772
1773                 btrfs_drop_pages(pages, num_pages);
1774
1775                 cond_resched();
1776
1777                 balance_dirty_pages_ratelimited(inode->i_mapping);
1778                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1779                         btrfs_btree_balance_dirty(fs_info);
1780
1781                 pos += copied;
1782                 num_written += copied;
1783         }
1784
1785         kfree(pages);
1786
1787         if (release_bytes) {
1788                 if (only_release_metadata) {
1789                         btrfs_end_write_no_snapshotting(root);
1790                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1791                                         release_bytes, true);
1792                 } else {
1793                         btrfs_delalloc_release_space(inode, data_reserved,
1794                                         round_down(pos, fs_info->sectorsize),
1795                                         release_bytes, true);
1796                 }
1797         }
1798
1799         extent_changeset_free(data_reserved);
1800         return num_written ? num_written : ret;
1801 }
1802
1803 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1804 {
1805         struct file *file = iocb->ki_filp;
1806         struct inode *inode = file_inode(file);
1807         loff_t pos;
1808         ssize_t written;
1809         ssize_t written_buffered;
1810         loff_t endbyte;
1811         int err;
1812
1813         written = generic_file_direct_write(iocb, from);
1814
1815         if (written < 0 || !iov_iter_count(from))
1816                 return written;
1817
1818         pos = iocb->ki_pos;
1819         written_buffered = btrfs_buffered_write(iocb, from);
1820         if (written_buffered < 0) {
1821                 err = written_buffered;
1822                 goto out;
1823         }
1824         /*
1825          * Ensure all data is persisted. We want the next direct IO read to be
1826          * able to read what was just written.
1827          */
1828         endbyte = pos + written_buffered - 1;
1829         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1830         if (err)
1831                 goto out;
1832         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1833         if (err)
1834                 goto out;
1835         written += written_buffered;
1836         iocb->ki_pos = pos + written_buffered;
1837         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1838                                  endbyte >> PAGE_SHIFT);
1839 out:
1840         return written ? written : err;
1841 }
1842
1843 static void update_time_for_write(struct inode *inode)
1844 {
1845         struct timespec64 now;
1846
1847         if (IS_NOCMTIME(inode))
1848                 return;
1849
1850         now = current_time(inode);
1851         if (!timespec64_equal(&inode->i_mtime, &now))
1852                 inode->i_mtime = now;
1853
1854         if (!timespec64_equal(&inode->i_ctime, &now))
1855                 inode->i_ctime = now;
1856
1857         if (IS_I_VERSION(inode))
1858                 inode_inc_iversion(inode);
1859 }
1860
1861 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1862                                     struct iov_iter *from)
1863 {
1864         struct file *file = iocb->ki_filp;
1865         struct inode *inode = file_inode(file);
1866         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1867         struct btrfs_root *root = BTRFS_I(inode)->root;
1868         u64 start_pos;
1869         u64 end_pos;
1870         ssize_t num_written = 0;
1871         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1872         ssize_t err;
1873         loff_t pos;
1874         size_t count = iov_iter_count(from);
1875         loff_t oldsize;
1876         int clean_page = 0;
1877
1878         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1879             (iocb->ki_flags & IOCB_NOWAIT))
1880                 return -EOPNOTSUPP;
1881
1882         if (!inode_trylock(inode)) {
1883                 if (iocb->ki_flags & IOCB_NOWAIT)
1884                         return -EAGAIN;
1885                 inode_lock(inode);
1886         }
1887
1888         err = generic_write_checks(iocb, from);
1889         if (err <= 0) {
1890                 inode_unlock(inode);
1891                 return err;
1892         }
1893
1894         pos = iocb->ki_pos;
1895         if (iocb->ki_flags & IOCB_NOWAIT) {
1896                 /*
1897                  * We will allocate space in case nodatacow is not set,
1898                  * so bail
1899                  */
1900                 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1901                                               BTRFS_INODE_PREALLOC)) ||
1902                     check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1903                         inode_unlock(inode);
1904                         return -EAGAIN;
1905                 }
1906         }
1907
1908         current->backing_dev_info = inode_to_bdi(inode);
1909         err = file_remove_privs(file);
1910         if (err) {
1911                 inode_unlock(inode);
1912                 goto out;
1913         }
1914
1915         /*
1916          * If BTRFS flips readonly due to some impossible error
1917          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1918          * although we have opened a file as writable, we have
1919          * to stop this write operation to ensure FS consistency.
1920          */
1921         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1922                 inode_unlock(inode);
1923                 err = -EROFS;
1924                 goto out;
1925         }
1926
1927         /*
1928          * We reserve space for updating the inode when we reserve space for the
1929          * extent we are going to write, so we will enospc out there.  We don't
1930          * need to start yet another transaction to update the inode as we will
1931          * update the inode when we finish writing whatever data we write.
1932          */
1933         update_time_for_write(inode);
1934
1935         start_pos = round_down(pos, fs_info->sectorsize);
1936         oldsize = i_size_read(inode);
1937         if (start_pos > oldsize) {
1938                 /* Expand hole size to cover write data, preventing empty gap */
1939                 end_pos = round_up(pos + count,
1940                                    fs_info->sectorsize);
1941                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1942                 if (err) {
1943                         inode_unlock(inode);
1944                         goto out;
1945                 }
1946                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1947                         clean_page = 1;
1948         }
1949
1950         if (sync)
1951                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1952
1953         if (iocb->ki_flags & IOCB_DIRECT) {
1954                 num_written = __btrfs_direct_write(iocb, from);
1955         } else {
1956                 num_written = btrfs_buffered_write(iocb, from);
1957                 if (num_written > 0)
1958                         iocb->ki_pos = pos + num_written;
1959                 if (clean_page)
1960                         pagecache_isize_extended(inode, oldsize,
1961                                                 i_size_read(inode));
1962         }
1963
1964         inode_unlock(inode);
1965
1966         /*
1967          * We also have to set last_sub_trans to the current log transid,
1968          * otherwise subsequent syncs to a file that's been synced in this
1969          * transaction will appear to have already occurred.
1970          */
1971         spin_lock(&BTRFS_I(inode)->lock);
1972         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1973         spin_unlock(&BTRFS_I(inode)->lock);
1974         if (num_written > 0)
1975                 num_written = generic_write_sync(iocb, num_written);
1976
1977         if (sync)
1978                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1979 out:
1980         current->backing_dev_info = NULL;
1981         return num_written ? num_written : err;
1982 }
1983
1984 int btrfs_release_file(struct inode *inode, struct file *filp)
1985 {
1986         struct btrfs_file_private *private = filp->private_data;
1987
1988         if (private && private->filldir_buf)
1989                 kfree(private->filldir_buf);
1990         kfree(private);
1991         filp->private_data = NULL;
1992
1993         /*
1994          * ordered_data_close is set by settattr when we are about to truncate
1995          * a file from a non-zero size to a zero size.  This tries to
1996          * flush down new bytes that may have been written if the
1997          * application were using truncate to replace a file in place.
1998          */
1999         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2000                                &BTRFS_I(inode)->runtime_flags))
2001                         filemap_flush(inode->i_mapping);
2002         return 0;
2003 }
2004
2005 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2006 {
2007         int ret;
2008         struct blk_plug plug;
2009
2010         /*
2011          * This is only called in fsync, which would do synchronous writes, so
2012          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2013          * multiple disks using raid profile, a large IO can be split to
2014          * several segments of stripe length (currently 64K).
2015          */
2016         blk_start_plug(&plug);
2017         atomic_inc(&BTRFS_I(inode)->sync_writers);
2018         ret = btrfs_fdatawrite_range(inode, start, end);
2019         atomic_dec(&BTRFS_I(inode)->sync_writers);
2020         blk_finish_plug(&plug);
2021
2022         return ret;
2023 }
2024
2025 /*
2026  * fsync call for both files and directories.  This logs the inode into
2027  * the tree log instead of forcing full commits whenever possible.
2028  *
2029  * It needs to call filemap_fdatawait so that all ordered extent updates are
2030  * in the metadata btree are up to date for copying to the log.
2031  *
2032  * It drops the inode mutex before doing the tree log commit.  This is an
2033  * important optimization for directories because holding the mutex prevents
2034  * new operations on the dir while we write to disk.
2035  */
2036 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2037 {
2038         struct dentry *dentry = file_dentry(file);
2039         struct inode *inode = d_inode(dentry);
2040         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2041         struct btrfs_root *root = BTRFS_I(inode)->root;
2042         struct btrfs_trans_handle *trans;
2043         struct btrfs_log_ctx ctx;
2044         int ret = 0, err;
2045         u64 len;
2046
2047         /*
2048          * The range length can be represented by u64, we have to do the typecasts
2049          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2050          */
2051         len = (u64)end - (u64)start + 1;
2052         trace_btrfs_sync_file(file, datasync);
2053
2054         btrfs_init_log_ctx(&ctx, inode);
2055
2056         /*
2057          * We write the dirty pages in the range and wait until they complete
2058          * out of the ->i_mutex. If so, we can flush the dirty pages by
2059          * multi-task, and make the performance up.  See
2060          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2061          */
2062         ret = start_ordered_ops(inode, start, end);
2063         if (ret)
2064                 goto out;
2065
2066         inode_lock(inode);
2067         atomic_inc(&root->log_batch);
2068
2069         /*
2070          * We have to do this here to avoid the priority inversion of waiting on
2071          * IO of a lower priority task while holding a transaciton open.
2072          */
2073         ret = btrfs_wait_ordered_range(inode, start, len);
2074         if (ret) {
2075                 inode_unlock(inode);
2076                 goto out;
2077         }
2078         atomic_inc(&root->log_batch);
2079
2080         smp_mb();
2081         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2082             BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2083                 /*
2084                  * We've had everything committed since the last time we were
2085                  * modified so clear this flag in case it was set for whatever
2086                  * reason, it's no longer relevant.
2087                  */
2088                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2089                           &BTRFS_I(inode)->runtime_flags);
2090                 /*
2091                  * An ordered extent might have started before and completed
2092                  * already with io errors, in which case the inode was not
2093                  * updated and we end up here. So check the inode's mapping
2094                  * for any errors that might have happened since we last
2095                  * checked called fsync.
2096                  */
2097                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2098                 inode_unlock(inode);
2099                 goto out;
2100         }
2101
2102         /*
2103          * We use start here because we will need to wait on the IO to complete
2104          * in btrfs_sync_log, which could require joining a transaction (for
2105          * example checking cross references in the nocow path).  If we use join
2106          * here we could get into a situation where we're waiting on IO to
2107          * happen that is blocked on a transaction trying to commit.  With start
2108          * we inc the extwriter counter, so we wait for all extwriters to exit
2109          * before we start blocking join'ers.  This comment is to keep somebody
2110          * from thinking they are super smart and changing this to
2111          * btrfs_join_transaction *cough*Josef*cough*.
2112          */
2113         trans = btrfs_start_transaction(root, 0);
2114         if (IS_ERR(trans)) {
2115                 ret = PTR_ERR(trans);
2116                 inode_unlock(inode);
2117                 goto out;
2118         }
2119         trans->sync = true;
2120
2121         ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2122         if (ret < 0) {
2123                 /* Fallthrough and commit/free transaction. */
2124                 ret = 1;
2125         }
2126
2127         /* we've logged all the items and now have a consistent
2128          * version of the file in the log.  It is possible that
2129          * someone will come in and modify the file, but that's
2130          * fine because the log is consistent on disk, and we
2131          * have references to all of the file's extents
2132          *
2133          * It is possible that someone will come in and log the
2134          * file again, but that will end up using the synchronization
2135          * inside btrfs_sync_log to keep things safe.
2136          */
2137         inode_unlock(inode);
2138
2139         /*
2140          * If any of the ordered extents had an error, just return it to user
2141          * space, so that the application knows some writes didn't succeed and
2142          * can take proper action (retry for e.g.). Blindly committing the
2143          * transaction in this case, would fool userspace that everything was
2144          * successful. And we also want to make sure our log doesn't contain
2145          * file extent items pointing to extents that weren't fully written to -
2146          * just like in the non fast fsync path, where we check for the ordered
2147          * operation's error flag before writing to the log tree and return -EIO
2148          * if any of them had this flag set (btrfs_wait_ordered_range) -
2149          * therefore we need to check for errors in the ordered operations,
2150          * which are indicated by ctx.io_err.
2151          */
2152         if (ctx.io_err) {
2153                 btrfs_end_transaction(trans);
2154                 ret = ctx.io_err;
2155                 goto out;
2156         }
2157
2158         if (ret != BTRFS_NO_LOG_SYNC) {
2159                 if (!ret) {
2160                         ret = btrfs_sync_log(trans, root, &ctx);
2161                         if (!ret) {
2162                                 ret = btrfs_end_transaction(trans);
2163                                 goto out;
2164                         }
2165                 }
2166                 ret = btrfs_commit_transaction(trans);
2167         } else {
2168                 ret = btrfs_end_transaction(trans);
2169         }
2170 out:
2171         ASSERT(list_empty(&ctx.list));
2172         err = file_check_and_advance_wb_err(file);
2173         if (!ret)
2174                 ret = err;
2175         return ret > 0 ? -EIO : ret;
2176 }
2177
2178 static const struct vm_operations_struct btrfs_file_vm_ops = {
2179         .fault          = filemap_fault,
2180         .map_pages      = filemap_map_pages,
2181         .page_mkwrite   = btrfs_page_mkwrite,
2182 };
2183
2184 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2185 {
2186         struct address_space *mapping = filp->f_mapping;
2187
2188         if (!mapping->a_ops->readpage)
2189                 return -ENOEXEC;
2190
2191         file_accessed(filp);
2192         vma->vm_ops = &btrfs_file_vm_ops;
2193
2194         return 0;
2195 }
2196
2197 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2198                           int slot, u64 start, u64 end)
2199 {
2200         struct btrfs_file_extent_item *fi;
2201         struct btrfs_key key;
2202
2203         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2204                 return 0;
2205
2206         btrfs_item_key_to_cpu(leaf, &key, slot);
2207         if (key.objectid != btrfs_ino(inode) ||
2208             key.type != BTRFS_EXTENT_DATA_KEY)
2209                 return 0;
2210
2211         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2212
2213         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2214                 return 0;
2215
2216         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2217                 return 0;
2218
2219         if (key.offset == end)
2220                 return 1;
2221         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2222                 return 1;
2223         return 0;
2224 }
2225
2226 static int fill_holes(struct btrfs_trans_handle *trans,
2227                 struct btrfs_inode *inode,
2228                 struct btrfs_path *path, u64 offset, u64 end)
2229 {
2230         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2231         struct btrfs_root *root = inode->root;
2232         struct extent_buffer *leaf;
2233         struct btrfs_file_extent_item *fi;
2234         struct extent_map *hole_em;
2235         struct extent_map_tree *em_tree = &inode->extent_tree;
2236         struct btrfs_key key;
2237         int ret;
2238
2239         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2240                 goto out;
2241
2242         key.objectid = btrfs_ino(inode);
2243         key.type = BTRFS_EXTENT_DATA_KEY;
2244         key.offset = offset;
2245
2246         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2247         if (ret <= 0) {
2248                 /*
2249                  * We should have dropped this offset, so if we find it then
2250                  * something has gone horribly wrong.
2251                  */
2252                 if (ret == 0)
2253                         ret = -EINVAL;
2254                 return ret;
2255         }
2256
2257         leaf = path->nodes[0];
2258         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2259                 u64 num_bytes;
2260
2261                 path->slots[0]--;
2262                 fi = btrfs_item_ptr(leaf, path->slots[0],
2263                                     struct btrfs_file_extent_item);
2264                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2265                         end - offset;
2266                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2267                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2268                 btrfs_set_file_extent_offset(leaf, fi, 0);
2269                 btrfs_mark_buffer_dirty(leaf);
2270                 goto out;
2271         }
2272
2273         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2274                 u64 num_bytes;
2275
2276                 key.offset = offset;
2277                 btrfs_set_item_key_safe(fs_info, path, &key);
2278                 fi = btrfs_item_ptr(leaf, path->slots[0],
2279                                     struct btrfs_file_extent_item);
2280                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2281                         offset;
2282                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2283                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2284                 btrfs_set_file_extent_offset(leaf, fi, 0);
2285                 btrfs_mark_buffer_dirty(leaf);
2286                 goto out;
2287         }
2288         btrfs_release_path(path);
2289
2290         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2291                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2292         if (ret)
2293                 return ret;
2294
2295 out:
2296         btrfs_release_path(path);
2297
2298         hole_em = alloc_extent_map();
2299         if (!hole_em) {
2300                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2301                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2302         } else {
2303                 hole_em->start = offset;
2304                 hole_em->len = end - offset;
2305                 hole_em->ram_bytes = hole_em->len;
2306                 hole_em->orig_start = offset;
2307
2308                 hole_em->block_start = EXTENT_MAP_HOLE;
2309                 hole_em->block_len = 0;
2310                 hole_em->orig_block_len = 0;
2311                 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2312                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2313                 hole_em->generation = trans->transid;
2314
2315                 do {
2316                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2317                         write_lock(&em_tree->lock);
2318                         ret = add_extent_mapping(em_tree, hole_em, 1);
2319                         write_unlock(&em_tree->lock);
2320                 } while (ret == -EEXIST);
2321                 free_extent_map(hole_em);
2322                 if (ret)
2323                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2324                                         &inode->runtime_flags);
2325         }
2326
2327         return 0;
2328 }
2329
2330 /*
2331  * Find a hole extent on given inode and change start/len to the end of hole
2332  * extent.(hole/vacuum extent whose em->start <= start &&
2333  *         em->start + em->len > start)
2334  * When a hole extent is found, return 1 and modify start/len.
2335  */
2336 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2337 {
2338         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2339         struct extent_map *em;
2340         int ret = 0;
2341
2342         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2343                               round_down(*start, fs_info->sectorsize),
2344                               round_up(*len, fs_info->sectorsize), 0);
2345         if (IS_ERR(em))
2346                 return PTR_ERR(em);
2347
2348         /* Hole or vacuum extent(only exists in no-hole mode) */
2349         if (em->block_start == EXTENT_MAP_HOLE) {
2350                 ret = 1;
2351                 *len = em->start + em->len > *start + *len ?
2352                        0 : *start + *len - em->start - em->len;
2353                 *start = em->start + em->len;
2354         }
2355         free_extent_map(em);
2356         return ret;
2357 }
2358
2359 static int btrfs_punch_hole_lock_range(struct inode *inode,
2360                                        const u64 lockstart,
2361                                        const u64 lockend,
2362                                        struct extent_state **cached_state)
2363 {
2364         while (1) {
2365                 struct btrfs_ordered_extent *ordered;
2366                 int ret;
2367
2368                 truncate_pagecache_range(inode, lockstart, lockend);
2369
2370                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2371                                  cached_state);
2372                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2373
2374                 /*
2375                  * We need to make sure we have no ordered extents in this range
2376                  * and nobody raced in and read a page in this range, if we did
2377                  * we need to try again.
2378                  */
2379                 if ((!ordered ||
2380                     (ordered->file_offset + ordered->len <= lockstart ||
2381                      ordered->file_offset > lockend)) &&
2382                      !filemap_range_has_page(inode->i_mapping,
2383                                              lockstart, lockend)) {
2384                         if (ordered)
2385                                 btrfs_put_ordered_extent(ordered);
2386                         break;
2387                 }
2388                 if (ordered)
2389                         btrfs_put_ordered_extent(ordered);
2390                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2391                                      lockend, cached_state);
2392                 ret = btrfs_wait_ordered_range(inode, lockstart,
2393                                                lockend - lockstart + 1);
2394                 if (ret)
2395                         return ret;
2396         }
2397         return 0;
2398 }
2399
2400 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2401 {
2402         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2403         struct btrfs_root *root = BTRFS_I(inode)->root;
2404         struct extent_state *cached_state = NULL;
2405         struct btrfs_path *path;
2406         struct btrfs_block_rsv *rsv;
2407         struct btrfs_trans_handle *trans;
2408         u64 lockstart;
2409         u64 lockend;
2410         u64 tail_start;
2411         u64 tail_len;
2412         u64 orig_start = offset;
2413         u64 cur_offset;
2414         u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2415         u64 drop_end;
2416         int ret = 0;
2417         int err = 0;
2418         unsigned int rsv_count;
2419         bool same_block;
2420         bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2421         u64 ino_size;
2422         bool truncated_block = false;
2423         bool updated_inode = false;
2424
2425         ret = btrfs_wait_ordered_range(inode, offset, len);
2426         if (ret)
2427                 return ret;
2428
2429         inode_lock(inode);
2430         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2431         ret = find_first_non_hole(inode, &offset, &len);
2432         if (ret < 0)
2433                 goto out_only_mutex;
2434         if (ret && !len) {
2435                 /* Already in a large hole */
2436                 ret = 0;
2437                 goto out_only_mutex;
2438         }
2439
2440         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2441         lockend = round_down(offset + len,
2442                              btrfs_inode_sectorsize(inode)) - 1;
2443         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2444                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2445         /*
2446          * We needn't truncate any block which is beyond the end of the file
2447          * because we are sure there is no data there.
2448          */
2449         /*
2450          * Only do this if we are in the same block and we aren't doing the
2451          * entire block.
2452          */
2453         if (same_block && len < fs_info->sectorsize) {
2454                 if (offset < ino_size) {
2455                         truncated_block = true;
2456                         ret = btrfs_truncate_block(inode, offset, len, 0);
2457                 } else {
2458                         ret = 0;
2459                 }
2460                 goto out_only_mutex;
2461         }
2462
2463         /* zero back part of the first block */
2464         if (offset < ino_size) {
2465                 truncated_block = true;
2466                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2467                 if (ret) {
2468                         inode_unlock(inode);
2469                         return ret;
2470                 }
2471         }
2472
2473         /* Check the aligned pages after the first unaligned page,
2474          * if offset != orig_start, which means the first unaligned page
2475          * including several following pages are already in holes,
2476          * the extra check can be skipped */
2477         if (offset == orig_start) {
2478                 /* after truncate page, check hole again */
2479                 len = offset + len - lockstart;
2480                 offset = lockstart;
2481                 ret = find_first_non_hole(inode, &offset, &len);
2482                 if (ret < 0)
2483                         goto out_only_mutex;
2484                 if (ret && !len) {
2485                         ret = 0;
2486                         goto out_only_mutex;
2487                 }
2488                 lockstart = offset;
2489         }
2490
2491         /* Check the tail unaligned part is in a hole */
2492         tail_start = lockend + 1;
2493         tail_len = offset + len - tail_start;
2494         if (tail_len) {
2495                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2496                 if (unlikely(ret < 0))
2497                         goto out_only_mutex;
2498                 if (!ret) {
2499                         /* zero the front end of the last page */
2500                         if (tail_start + tail_len < ino_size) {
2501                                 truncated_block = true;
2502                                 ret = btrfs_truncate_block(inode,
2503                                                         tail_start + tail_len,
2504                                                         0, 1);
2505                                 if (ret)
2506                                         goto out_only_mutex;
2507                         }
2508                 }
2509         }
2510
2511         if (lockend < lockstart) {
2512                 ret = 0;
2513                 goto out_only_mutex;
2514         }
2515
2516         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2517                                           &cached_state);
2518         if (ret) {
2519                 inode_unlock(inode);
2520                 goto out_only_mutex;
2521         }
2522
2523         path = btrfs_alloc_path();
2524         if (!path) {
2525                 ret = -ENOMEM;
2526                 goto out;
2527         }
2528
2529         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2530         if (!rsv) {
2531                 ret = -ENOMEM;
2532                 goto out_free;
2533         }
2534         rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2535         rsv->failfast = 1;
2536
2537         /*
2538          * 1 - update the inode
2539          * 1 - removing the extents in the range
2540          * 1 - adding the hole extent if no_holes isn't set
2541          */
2542         rsv_count = no_holes ? 2 : 3;
2543         trans = btrfs_start_transaction(root, rsv_count);
2544         if (IS_ERR(trans)) {
2545                 err = PTR_ERR(trans);
2546                 goto out_free;
2547         }
2548
2549         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2550                                       min_size, 0);
2551         BUG_ON(ret);
2552         trans->block_rsv = rsv;
2553
2554         cur_offset = lockstart;
2555         len = lockend - cur_offset;
2556         while (cur_offset < lockend) {
2557                 ret = __btrfs_drop_extents(trans, root, inode, path,
2558                                            cur_offset, lockend + 1,
2559                                            &drop_end, 1, 0, 0, NULL);
2560                 if (ret != -ENOSPC)
2561                         break;
2562
2563                 trans->block_rsv = &fs_info->trans_block_rsv;
2564
2565                 if (cur_offset < drop_end && cur_offset < ino_size) {
2566                         ret = fill_holes(trans, BTRFS_I(inode), path,
2567                                         cur_offset, drop_end);
2568                         if (ret) {
2569                                 /*
2570                                  * If we failed then we didn't insert our hole
2571                                  * entries for the area we dropped, so now the
2572                                  * fs is corrupted, so we must abort the
2573                                  * transaction.
2574                                  */
2575                                 btrfs_abort_transaction(trans, ret);
2576                                 err = ret;
2577                                 break;
2578                         }
2579                 }
2580
2581                 cur_offset = drop_end;
2582
2583                 ret = btrfs_update_inode(trans, root, inode);
2584                 if (ret) {
2585                         err = ret;
2586                         break;
2587                 }
2588
2589                 btrfs_end_transaction(trans);
2590                 btrfs_btree_balance_dirty(fs_info);
2591
2592                 trans = btrfs_start_transaction(root, rsv_count);
2593                 if (IS_ERR(trans)) {
2594                         ret = PTR_ERR(trans);
2595                         trans = NULL;
2596                         break;
2597                 }
2598
2599                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2600                                               rsv, min_size, 0);
2601                 BUG_ON(ret);    /* shouldn't happen */
2602                 trans->block_rsv = rsv;
2603
2604                 ret = find_first_non_hole(inode, &cur_offset, &len);
2605                 if (unlikely(ret < 0))
2606                         break;
2607                 if (ret && !len) {
2608                         ret = 0;
2609                         break;
2610                 }
2611         }
2612
2613         if (ret) {
2614                 err = ret;
2615                 goto out_trans;
2616         }
2617
2618         trans->block_rsv = &fs_info->trans_block_rsv;
2619         /*
2620          * If we are using the NO_HOLES feature we might have had already an
2621          * hole that overlaps a part of the region [lockstart, lockend] and
2622          * ends at (or beyond) lockend. Since we have no file extent items to
2623          * represent holes, drop_end can be less than lockend and so we must
2624          * make sure we have an extent map representing the existing hole (the
2625          * call to __btrfs_drop_extents() might have dropped the existing extent
2626          * map representing the existing hole), otherwise the fast fsync path
2627          * will not record the existence of the hole region
2628          * [existing_hole_start, lockend].
2629          */
2630         if (drop_end <= lockend)
2631                 drop_end = lockend + 1;
2632         /*
2633          * Don't insert file hole extent item if it's for a range beyond eof
2634          * (because it's useless) or if it represents a 0 bytes range (when
2635          * cur_offset == drop_end).
2636          */
2637         if (cur_offset < ino_size && cur_offset < drop_end) {
2638                 ret = fill_holes(trans, BTRFS_I(inode), path,
2639                                 cur_offset, drop_end);
2640                 if (ret) {
2641                         /* Same comment as above. */
2642                         btrfs_abort_transaction(trans, ret);
2643                         err = ret;
2644                         goto out_trans;
2645                 }
2646         }
2647
2648 out_trans:
2649         if (!trans)
2650                 goto out_free;
2651
2652         inode_inc_iversion(inode);
2653         inode->i_mtime = inode->i_ctime = current_time(inode);
2654
2655         trans->block_rsv = &fs_info->trans_block_rsv;
2656         ret = btrfs_update_inode(trans, root, inode);
2657         updated_inode = true;
2658         btrfs_end_transaction(trans);
2659         btrfs_btree_balance_dirty(fs_info);
2660 out_free:
2661         btrfs_free_path(path);
2662         btrfs_free_block_rsv(fs_info, rsv);
2663 out:
2664         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2665                              &cached_state);
2666 out_only_mutex:
2667         if (!updated_inode && truncated_block && !ret && !err) {
2668                 /*
2669                  * If we only end up zeroing part of a page, we still need to
2670                  * update the inode item, so that all the time fields are
2671                  * updated as well as the necessary btrfs inode in memory fields
2672                  * for detecting, at fsync time, if the inode isn't yet in the
2673                  * log tree or it's there but not up to date.
2674                  */
2675                 trans = btrfs_start_transaction(root, 1);
2676                 if (IS_ERR(trans)) {
2677                         err = PTR_ERR(trans);
2678                 } else {
2679                         err = btrfs_update_inode(trans, root, inode);
2680                         ret = btrfs_end_transaction(trans);
2681                 }
2682         }
2683         inode_unlock(inode);
2684         if (ret && !err)
2685                 err = ret;
2686         return err;
2687 }
2688
2689 /* Helper structure to record which range is already reserved */
2690 struct falloc_range {
2691         struct list_head list;
2692         u64 start;
2693         u64 len;
2694 };
2695
2696 /*
2697  * Helper function to add falloc range
2698  *
2699  * Caller should have locked the larger range of extent containing
2700  * [start, len)
2701  */
2702 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2703 {
2704         struct falloc_range *prev = NULL;
2705         struct falloc_range *range = NULL;
2706
2707         if (list_empty(head))
2708                 goto insert;
2709
2710         /*
2711          * As fallocate iterate by bytenr order, we only need to check
2712          * the last range.
2713          */
2714         prev = list_entry(head->prev, struct falloc_range, list);
2715         if (prev->start + prev->len == start) {
2716                 prev->len += len;
2717                 return 0;
2718         }
2719 insert:
2720         range = kmalloc(sizeof(*range), GFP_KERNEL);
2721         if (!range)
2722                 return -ENOMEM;
2723         range->start = start;
2724         range->len = len;
2725         list_add_tail(&range->list, head);
2726         return 0;
2727 }
2728
2729 static int btrfs_fallocate_update_isize(struct inode *inode,
2730                                         const u64 end,
2731                                         const int mode)
2732 {
2733         struct btrfs_trans_handle *trans;
2734         struct btrfs_root *root = BTRFS_I(inode)->root;
2735         int ret;
2736         int ret2;
2737
2738         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2739                 return 0;
2740
2741         trans = btrfs_start_transaction(root, 1);
2742         if (IS_ERR(trans))
2743                 return PTR_ERR(trans);
2744
2745         inode->i_ctime = current_time(inode);
2746         i_size_write(inode, end);
2747         btrfs_ordered_update_i_size(inode, end, NULL);
2748         ret = btrfs_update_inode(trans, root, inode);
2749         ret2 = btrfs_end_transaction(trans);
2750
2751         return ret ? ret : ret2;
2752 }
2753
2754 enum {
2755         RANGE_BOUNDARY_WRITTEN_EXTENT = 0,
2756         RANGE_BOUNDARY_PREALLOC_EXTENT = 1,
2757         RANGE_BOUNDARY_HOLE = 2,
2758 };
2759
2760 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2761                                                  u64 offset)
2762 {
2763         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2764         struct extent_map *em;
2765         int ret;
2766
2767         offset = round_down(offset, sectorsize);
2768         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2769         if (IS_ERR(em))
2770                 return PTR_ERR(em);
2771
2772         if (em->block_start == EXTENT_MAP_HOLE)
2773                 ret = RANGE_BOUNDARY_HOLE;
2774         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2775                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2776         else
2777                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2778
2779         free_extent_map(em);
2780         return ret;
2781 }
2782
2783 static int btrfs_zero_range(struct inode *inode,
2784                             loff_t offset,
2785                             loff_t len,
2786                             const int mode)
2787 {
2788         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2789         struct extent_map *em;
2790         struct extent_changeset *data_reserved = NULL;
2791         int ret;
2792         u64 alloc_hint = 0;
2793         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2794         u64 alloc_start = round_down(offset, sectorsize);
2795         u64 alloc_end = round_up(offset + len, sectorsize);
2796         u64 bytes_to_reserve = 0;
2797         bool space_reserved = false;
2798
2799         inode_dio_wait(inode);
2800
2801         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2802                               alloc_start, alloc_end - alloc_start, 0);
2803         if (IS_ERR(em)) {
2804                 ret = PTR_ERR(em);
2805                 goto out;
2806         }
2807
2808         /*
2809          * Avoid hole punching and extent allocation for some cases. More cases
2810          * could be considered, but these are unlikely common and we keep things
2811          * as simple as possible for now. Also, intentionally, if the target
2812          * range contains one or more prealloc extents together with regular
2813          * extents and holes, we drop all the existing extents and allocate a
2814          * new prealloc extent, so that we get a larger contiguous disk extent.
2815          */
2816         if (em->start <= alloc_start &&
2817             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2818                 const u64 em_end = em->start + em->len;
2819
2820                 if (em_end >= offset + len) {
2821                         /*
2822                          * The whole range is already a prealloc extent,
2823                          * do nothing except updating the inode's i_size if
2824                          * needed.
2825                          */
2826                         free_extent_map(em);
2827                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2828                                                            mode);
2829                         goto out;
2830                 }
2831                 /*
2832                  * Part of the range is already a prealloc extent, so operate
2833                  * only on the remaining part of the range.
2834                  */
2835                 alloc_start = em_end;
2836                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2837                 len = offset + len - alloc_start;
2838                 offset = alloc_start;
2839                 alloc_hint = em->block_start + em->len;
2840         }
2841         free_extent_map(em);
2842
2843         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2844             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2845                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2846                                       alloc_start, sectorsize, 0);
2847                 if (IS_ERR(em)) {
2848                         ret = PTR_ERR(em);
2849                         goto out;
2850                 }
2851
2852                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2853                         free_extent_map(em);
2854                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2855                                                            mode);
2856                         goto out;
2857                 }
2858                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2859                         free_extent_map(em);
2860                         ret = btrfs_truncate_block(inode, offset, len, 0);
2861                         if (!ret)
2862                                 ret = btrfs_fallocate_update_isize(inode,
2863                                                                    offset + len,
2864                                                                    mode);
2865                         return ret;
2866                 }
2867                 free_extent_map(em);
2868                 alloc_start = round_down(offset, sectorsize);
2869                 alloc_end = alloc_start + sectorsize;
2870                 goto reserve_space;
2871         }
2872
2873         alloc_start = round_up(offset, sectorsize);
2874         alloc_end = round_down(offset + len, sectorsize);
2875
2876         /*
2877          * For unaligned ranges, check the pages at the boundaries, they might
2878          * map to an extent, in which case we need to partially zero them, or
2879          * they might map to a hole, in which case we need our allocation range
2880          * to cover them.
2881          */
2882         if (!IS_ALIGNED(offset, sectorsize)) {
2883                 ret = btrfs_zero_range_check_range_boundary(inode, offset);
2884                 if (ret < 0)
2885                         goto out;
2886                 if (ret == RANGE_BOUNDARY_HOLE) {
2887                         alloc_start = round_down(offset, sectorsize);
2888                         ret = 0;
2889                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2890                         ret = btrfs_truncate_block(inode, offset, 0, 0);
2891                         if (ret)
2892                                 goto out;
2893                 } else {
2894                         ret = 0;
2895                 }
2896         }
2897
2898         if (!IS_ALIGNED(offset + len, sectorsize)) {
2899                 ret = btrfs_zero_range_check_range_boundary(inode,
2900                                                             offset + len);
2901                 if (ret < 0)
2902                         goto out;
2903                 if (ret == RANGE_BOUNDARY_HOLE) {
2904                         alloc_end = round_up(offset + len, sectorsize);
2905                         ret = 0;
2906                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2907                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
2908                         if (ret)
2909                                 goto out;
2910                 } else {
2911                         ret = 0;
2912                 }
2913         }
2914
2915 reserve_space:
2916         if (alloc_start < alloc_end) {
2917                 struct extent_state *cached_state = NULL;
2918                 const u64 lockstart = alloc_start;
2919                 const u64 lockend = alloc_end - 1;
2920
2921                 bytes_to_reserve = alloc_end - alloc_start;
2922                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2923                                                       bytes_to_reserve);
2924                 if (ret < 0)
2925                         goto out;
2926                 space_reserved = true;
2927                 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
2928                                                 alloc_start, bytes_to_reserve);
2929                 if (ret)
2930                         goto out;
2931                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2932                                                   &cached_state);
2933                 if (ret)
2934                         goto out;
2935                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
2936                                                 alloc_end - alloc_start,
2937                                                 i_blocksize(inode),
2938                                                 offset + len, &alloc_hint);
2939                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2940                                      lockend, &cached_state);
2941                 /* btrfs_prealloc_file_range releases reserved space on error */
2942                 if (ret) {
2943                         space_reserved = false;
2944                         goto out;
2945                 }
2946         }
2947         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
2948  out:
2949         if (ret && space_reserved)
2950                 btrfs_free_reserved_data_space(inode, data_reserved,
2951                                                alloc_start, bytes_to_reserve);
2952         extent_changeset_free(data_reserved);
2953
2954         return ret;
2955 }
2956
2957 static long btrfs_fallocate(struct file *file, int mode,
2958                             loff_t offset, loff_t len)
2959 {
2960         struct inode *inode = file_inode(file);
2961         struct extent_state *cached_state = NULL;
2962         struct extent_changeset *data_reserved = NULL;
2963         struct falloc_range *range;
2964         struct falloc_range *tmp;
2965         struct list_head reserve_list;
2966         u64 cur_offset;
2967         u64 last_byte;
2968         u64 alloc_start;
2969         u64 alloc_end;
2970         u64 alloc_hint = 0;
2971         u64 locked_end;
2972         u64 actual_end = 0;
2973         struct extent_map *em;
2974         int blocksize = btrfs_inode_sectorsize(inode);
2975         int ret;
2976
2977         alloc_start = round_down(offset, blocksize);
2978         alloc_end = round_up(offset + len, blocksize);
2979         cur_offset = alloc_start;
2980
2981         /* Make sure we aren't being give some crap mode */
2982         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
2983                      FALLOC_FL_ZERO_RANGE))
2984                 return -EOPNOTSUPP;
2985
2986         if (mode & FALLOC_FL_PUNCH_HOLE)
2987                 return btrfs_punch_hole(inode, offset, len);
2988
2989         /*
2990          * Only trigger disk allocation, don't trigger qgroup reserve
2991          *
2992          * For qgroup space, it will be checked later.
2993          */
2994         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
2995                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2996                                                       alloc_end - alloc_start);
2997                 if (ret < 0)
2998                         return ret;
2999         }
3000
3001         inode_lock(inode);
3002
3003         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3004                 ret = inode_newsize_ok(inode, offset + len);
3005                 if (ret)
3006                         goto out;
3007         }
3008
3009         /*
3010          * TODO: Move these two operations after we have checked
3011          * accurate reserved space, or fallocate can still fail but
3012          * with page truncated or size expanded.
3013          *
3014          * But that's a minor problem and won't do much harm BTW.
3015          */
3016         if (alloc_start > inode->i_size) {
3017                 ret = btrfs_cont_expand(inode, i_size_read(inode),
3018                                         alloc_start);
3019                 if (ret)
3020                         goto out;
3021         } else if (offset + len > inode->i_size) {
3022                 /*
3023                  * If we are fallocating from the end of the file onward we
3024                  * need to zero out the end of the block if i_size lands in the
3025                  * middle of a block.
3026                  */
3027                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3028                 if (ret)
3029                         goto out;
3030         }
3031
3032         /*
3033          * wait for ordered IO before we have any locks.  We'll loop again
3034          * below with the locks held.
3035          */
3036         ret = btrfs_wait_ordered_range(inode, alloc_start,
3037                                        alloc_end - alloc_start);
3038         if (ret)
3039                 goto out;
3040
3041         if (mode & FALLOC_FL_ZERO_RANGE) {
3042                 ret = btrfs_zero_range(inode, offset, len, mode);
3043                 inode_unlock(inode);
3044                 return ret;
3045         }
3046
3047         locked_end = alloc_end - 1;
3048         while (1) {
3049                 struct btrfs_ordered_extent *ordered;
3050
3051                 /* the extent lock is ordered inside the running
3052                  * transaction
3053                  */
3054                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3055                                  locked_end, &cached_state);
3056                 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3057
3058                 if (ordered &&
3059                     ordered->file_offset + ordered->len > alloc_start &&
3060                     ordered->file_offset < alloc_end) {
3061                         btrfs_put_ordered_extent(ordered);
3062                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3063                                              alloc_start, locked_end,
3064                                              &cached_state);
3065                         /*
3066                          * we can't wait on the range with the transaction
3067                          * running or with the extent lock held
3068                          */
3069                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3070                                                        alloc_end - alloc_start);
3071                         if (ret)
3072                                 goto out;
3073                 } else {
3074                         if (ordered)
3075                                 btrfs_put_ordered_extent(ordered);
3076                         break;
3077                 }
3078         }
3079
3080         /* First, check if we exceed the qgroup limit */
3081         INIT_LIST_HEAD(&reserve_list);
3082         while (cur_offset < alloc_end) {
3083                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3084                                       alloc_end - cur_offset, 0);
3085                 if (IS_ERR(em)) {
3086                         ret = PTR_ERR(em);
3087                         break;
3088                 }
3089                 last_byte = min(extent_map_end(em), alloc_end);
3090                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3091                 last_byte = ALIGN(last_byte, blocksize);
3092                 if (em->block_start == EXTENT_MAP_HOLE ||
3093                     (cur_offset >= inode->i_size &&
3094                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3095                         ret = add_falloc_range(&reserve_list, cur_offset,
3096                                                last_byte - cur_offset);
3097                         if (ret < 0) {
3098                                 free_extent_map(em);
3099                                 break;
3100                         }
3101                         ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3102                                         cur_offset, last_byte - cur_offset);
3103                         if (ret < 0) {
3104                                 free_extent_map(em);
3105                                 break;
3106                         }
3107                 } else {
3108                         /*
3109                          * Do not need to reserve unwritten extent for this
3110                          * range, free reserved data space first, otherwise
3111                          * it'll result in false ENOSPC error.
3112                          */
3113                         btrfs_free_reserved_data_space(inode, data_reserved,
3114                                         cur_offset, last_byte - cur_offset);
3115                 }
3116                 free_extent_map(em);
3117                 cur_offset = last_byte;
3118         }
3119
3120         /*
3121          * If ret is still 0, means we're OK to fallocate.
3122          * Or just cleanup the list and exit.
3123          */
3124         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3125                 if (!ret)
3126                         ret = btrfs_prealloc_file_range(inode, mode,
3127                                         range->start,
3128                                         range->len, i_blocksize(inode),
3129                                         offset + len, &alloc_hint);
3130                 else
3131                         btrfs_free_reserved_data_space(inode,
3132                                         data_reserved, range->start,
3133                                         range->len);
3134                 list_del(&range->list);
3135                 kfree(range);
3136         }
3137         if (ret < 0)
3138                 goto out_unlock;
3139
3140         /*
3141          * We didn't need to allocate any more space, but we still extended the
3142          * size of the file so we need to update i_size and the inode item.
3143          */
3144         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3145 out_unlock:
3146         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3147                              &cached_state);
3148 out:
3149         inode_unlock(inode);
3150         /* Let go of our reservation. */
3151         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3152                 btrfs_free_reserved_data_space(inode, data_reserved,
3153                                 alloc_start, alloc_end - cur_offset);
3154         extent_changeset_free(data_reserved);
3155         return ret;
3156 }
3157
3158 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3159 {
3160         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3161         struct extent_map *em = NULL;
3162         struct extent_state *cached_state = NULL;
3163         u64 lockstart;
3164         u64 lockend;
3165         u64 start;
3166         u64 len;
3167         int ret = 0;
3168
3169         if (inode->i_size == 0)
3170                 return -ENXIO;
3171
3172         /*
3173          * *offset can be negative, in this case we start finding DATA/HOLE from
3174          * the very start of the file.
3175          */
3176         start = max_t(loff_t, 0, *offset);
3177
3178         lockstart = round_down(start, fs_info->sectorsize);
3179         lockend = round_up(i_size_read(inode),
3180                            fs_info->sectorsize);
3181         if (lockend <= lockstart)
3182                 lockend = lockstart + fs_info->sectorsize;
3183         lockend--;
3184         len = lockend - lockstart + 1;
3185
3186         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3187                          &cached_state);
3188
3189         while (start < inode->i_size) {
3190                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3191                                 start, len, 0);
3192                 if (IS_ERR(em)) {
3193                         ret = PTR_ERR(em);
3194                         em = NULL;
3195                         break;
3196                 }
3197
3198                 if (whence == SEEK_HOLE &&
3199                     (em->block_start == EXTENT_MAP_HOLE ||
3200                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3201                         break;
3202                 else if (whence == SEEK_DATA &&
3203                            (em->block_start != EXTENT_MAP_HOLE &&
3204                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3205                         break;
3206
3207                 start = em->start + em->len;
3208                 free_extent_map(em);
3209                 em = NULL;
3210                 cond_resched();
3211         }
3212         free_extent_map(em);
3213         if (!ret) {
3214                 if (whence == SEEK_DATA && start >= inode->i_size)
3215                         ret = -ENXIO;
3216                 else
3217                         *offset = min_t(loff_t, start, inode->i_size);
3218         }
3219         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3220                              &cached_state);
3221         return ret;
3222 }
3223
3224 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3225 {
3226         struct inode *inode = file->f_mapping->host;
3227         int ret;
3228
3229         inode_lock(inode);
3230         switch (whence) {
3231         case SEEK_END:
3232         case SEEK_CUR:
3233                 offset = generic_file_llseek(file, offset, whence);
3234                 goto out;
3235         case SEEK_DATA:
3236         case SEEK_HOLE:
3237                 if (offset >= i_size_read(inode)) {
3238                         inode_unlock(inode);
3239                         return -ENXIO;
3240                 }
3241
3242                 ret = find_desired_extent(inode, &offset, whence);
3243                 if (ret) {
3244                         inode_unlock(inode);
3245                         return ret;
3246                 }
3247         }
3248
3249         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3250 out:
3251         inode_unlock(inode);
3252         return offset;
3253 }
3254
3255 static int btrfs_file_open(struct inode *inode, struct file *filp)
3256 {
3257         filp->f_mode |= FMODE_NOWAIT;
3258         return generic_file_open(inode, filp);
3259 }
3260
3261 const struct file_operations btrfs_file_operations = {
3262         .llseek         = btrfs_file_llseek,
3263         .read_iter      = generic_file_read_iter,
3264         .splice_read    = generic_file_splice_read,
3265         .write_iter     = btrfs_file_write_iter,
3266         .mmap           = btrfs_file_mmap,
3267         .open           = btrfs_file_open,
3268         .release        = btrfs_release_file,
3269         .fsync          = btrfs_sync_file,
3270         .fallocate      = btrfs_fallocate,
3271         .unlocked_ioctl = btrfs_ioctl,
3272 #ifdef CONFIG_COMPAT
3273         .compat_ioctl   = btrfs_compat_ioctl,
3274 #endif
3275         .clone_file_range = btrfs_clone_file_range,
3276         .dedupe_file_range = btrfs_dedupe_file_range,
3277 };
3278
3279 void __cold btrfs_auto_defrag_exit(void)
3280 {
3281         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3282 }
3283
3284 int __init btrfs_auto_defrag_init(void)
3285 {
3286         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3287                                         sizeof(struct inode_defrag), 0,
3288                                         SLAB_MEM_SPREAD,
3289                                         NULL);
3290         if (!btrfs_inode_defrag_cachep)
3291                 return -ENOMEM;
3292
3293         return 0;
3294 }
3295
3296 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3297 {
3298         int ret;
3299
3300         /*
3301          * So with compression we will find and lock a dirty page and clear the
3302          * first one as dirty, setup an async extent, and immediately return
3303          * with the entire range locked but with nobody actually marked with
3304          * writeback.  So we can't just filemap_write_and_wait_range() and
3305          * expect it to work since it will just kick off a thread to do the
3306          * actual work.  So we need to call filemap_fdatawrite_range _again_
3307          * since it will wait on the page lock, which won't be unlocked until
3308          * after the pages have been marked as writeback and so we're good to go
3309          * from there.  We have to do this otherwise we'll miss the ordered
3310          * extents and that results in badness.  Please Josef, do not think you
3311          * know better and pull this out at some point in the future, it is
3312          * right and you are wrong.
3313          */
3314         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3315         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3316                              &BTRFS_I(inode)->runtime_flags))
3317                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3318
3319         return ret;
3320 }