Merge tag 'for-4.20-rc4-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave...
[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/time.h>
9 #include <linux/init.h>
10 #include <linux/string.h>
11 #include <linux/backing-dev.h>
12 #include <linux/falloc.h>
13 #include <linux/writeback.h>
14 #include <linux/compat.h>
15 #include <linux/slab.h>
16 #include <linux/btrfs.h>
17 #include <linux/uio.h>
18 #include <linux/iversion.h>
19 #include "ctree.h"
20 #include "disk-io.h"
21 #include "transaction.h"
22 #include "btrfs_inode.h"
23 #include "print-tree.h"
24 #include "tree-log.h"
25 #include "locking.h"
26 #include "volumes.h"
27 #include "qgroup.h"
28 #include "compression.h"
29
30 static struct kmem_cache *btrfs_inode_defrag_cachep;
31 /*
32  * when auto defrag is enabled we
33  * queue up these defrag structs to remember which
34  * inodes need defragging passes
35  */
36 struct inode_defrag {
37         struct rb_node rb_node;
38         /* objectid */
39         u64 ino;
40         /*
41          * transid where the defrag was added, we search for
42          * extents newer than this
43          */
44         u64 transid;
45
46         /* root objectid */
47         u64 root;
48
49         /* last offset we were able to defrag */
50         u64 last_offset;
51
52         /* if we've wrapped around back to zero once already */
53         int cycled;
54 };
55
56 static int __compare_inode_defrag(struct inode_defrag *defrag1,
57                                   struct inode_defrag *defrag2)
58 {
59         if (defrag1->root > defrag2->root)
60                 return 1;
61         else if (defrag1->root < defrag2->root)
62                 return -1;
63         else if (defrag1->ino > defrag2->ino)
64                 return 1;
65         else if (defrag1->ino < defrag2->ino)
66                 return -1;
67         else
68                 return 0;
69 }
70
71 /* pop a record for an inode into the defrag tree.  The lock
72  * must be held already
73  *
74  * If you're inserting a record for an older transid than an
75  * existing record, the transid already in the tree is lowered
76  *
77  * If an existing record is found the defrag item you
78  * pass in is freed
79  */
80 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
81                                     struct inode_defrag *defrag)
82 {
83         struct btrfs_fs_info *fs_info = inode->root->fs_info;
84         struct inode_defrag *entry;
85         struct rb_node **p;
86         struct rb_node *parent = NULL;
87         int ret;
88
89         p = &fs_info->defrag_inodes.rb_node;
90         while (*p) {
91                 parent = *p;
92                 entry = rb_entry(parent, struct inode_defrag, rb_node);
93
94                 ret = __compare_inode_defrag(defrag, entry);
95                 if (ret < 0)
96                         p = &parent->rb_left;
97                 else if (ret > 0)
98                         p = &parent->rb_right;
99                 else {
100                         /* if we're reinserting an entry for
101                          * an old defrag run, make sure to
102                          * lower the transid of our existing record
103                          */
104                         if (defrag->transid < entry->transid)
105                                 entry->transid = defrag->transid;
106                         if (defrag->last_offset > entry->last_offset)
107                                 entry->last_offset = defrag->last_offset;
108                         return -EEXIST;
109                 }
110         }
111         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
112         rb_link_node(&defrag->rb_node, parent, p);
113         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
114         return 0;
115 }
116
117 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
118 {
119         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
120                 return 0;
121
122         if (btrfs_fs_closing(fs_info))
123                 return 0;
124
125         return 1;
126 }
127
128 /*
129  * insert a defrag record for this inode if auto defrag is
130  * enabled
131  */
132 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
133                            struct btrfs_inode *inode)
134 {
135         struct btrfs_root *root = inode->root;
136         struct btrfs_fs_info *fs_info = root->fs_info;
137         struct inode_defrag *defrag;
138         u64 transid;
139         int ret;
140
141         if (!__need_auto_defrag(fs_info))
142                 return 0;
143
144         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
145                 return 0;
146
147         if (trans)
148                 transid = trans->transid;
149         else
150                 transid = inode->root->last_trans;
151
152         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
153         if (!defrag)
154                 return -ENOMEM;
155
156         defrag->ino = btrfs_ino(inode);
157         defrag->transid = transid;
158         defrag->root = root->root_key.objectid;
159
160         spin_lock(&fs_info->defrag_inodes_lock);
161         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
162                 /*
163                  * If we set IN_DEFRAG flag and evict the inode from memory,
164                  * and then re-read this inode, this new inode doesn't have
165                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
166                  */
167                 ret = __btrfs_add_inode_defrag(inode, defrag);
168                 if (ret)
169                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
170         } else {
171                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
172         }
173         spin_unlock(&fs_info->defrag_inodes_lock);
174         return 0;
175 }
176
177 /*
178  * Requeue the defrag object. If there is a defrag object that points to
179  * the same inode in the tree, we will merge them together (by
180  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
181  */
182 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
183                                        struct inode_defrag *defrag)
184 {
185         struct btrfs_fs_info *fs_info = inode->root->fs_info;
186         int ret;
187
188         if (!__need_auto_defrag(fs_info))
189                 goto out;
190
191         /*
192          * Here we don't check the IN_DEFRAG flag, because we need merge
193          * them together.
194          */
195         spin_lock(&fs_info->defrag_inodes_lock);
196         ret = __btrfs_add_inode_defrag(inode, defrag);
197         spin_unlock(&fs_info->defrag_inodes_lock);
198         if (ret)
199                 goto out;
200         return;
201 out:
202         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
203 }
204
205 /*
206  * pick the defragable inode that we want, if it doesn't exist, we will get
207  * the next one.
208  */
209 static struct inode_defrag *
210 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
211 {
212         struct inode_defrag *entry = NULL;
213         struct inode_defrag tmp;
214         struct rb_node *p;
215         struct rb_node *parent = NULL;
216         int ret;
217
218         tmp.ino = ino;
219         tmp.root = root;
220
221         spin_lock(&fs_info->defrag_inodes_lock);
222         p = fs_info->defrag_inodes.rb_node;
223         while (p) {
224                 parent = p;
225                 entry = rb_entry(parent, struct inode_defrag, rb_node);
226
227                 ret = __compare_inode_defrag(&tmp, entry);
228                 if (ret < 0)
229                         p = parent->rb_left;
230                 else if (ret > 0)
231                         p = parent->rb_right;
232                 else
233                         goto out;
234         }
235
236         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
237                 parent = rb_next(parent);
238                 if (parent)
239                         entry = rb_entry(parent, struct inode_defrag, rb_node);
240                 else
241                         entry = NULL;
242         }
243 out:
244         if (entry)
245                 rb_erase(parent, &fs_info->defrag_inodes);
246         spin_unlock(&fs_info->defrag_inodes_lock);
247         return entry;
248 }
249
250 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
251 {
252         struct inode_defrag *defrag;
253         struct rb_node *node;
254
255         spin_lock(&fs_info->defrag_inodes_lock);
256         node = rb_first(&fs_info->defrag_inodes);
257         while (node) {
258                 rb_erase(node, &fs_info->defrag_inodes);
259                 defrag = rb_entry(node, struct inode_defrag, rb_node);
260                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
261
262                 cond_resched_lock(&fs_info->defrag_inodes_lock);
263
264                 node = rb_first(&fs_info->defrag_inodes);
265         }
266         spin_unlock(&fs_info->defrag_inodes_lock);
267 }
268
269 #define BTRFS_DEFRAG_BATCH      1024
270
271 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
272                                     struct inode_defrag *defrag)
273 {
274         struct btrfs_root *inode_root;
275         struct inode *inode;
276         struct btrfs_key key;
277         struct btrfs_ioctl_defrag_range_args range;
278         int num_defrag;
279         int index;
280         int ret;
281
282         /* get the inode */
283         key.objectid = defrag->root;
284         key.type = BTRFS_ROOT_ITEM_KEY;
285         key.offset = (u64)-1;
286
287         index = srcu_read_lock(&fs_info->subvol_srcu);
288
289         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
290         if (IS_ERR(inode_root)) {
291                 ret = PTR_ERR(inode_root);
292                 goto cleanup;
293         }
294
295         key.objectid = defrag->ino;
296         key.type = BTRFS_INODE_ITEM_KEY;
297         key.offset = 0;
298         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
299         if (IS_ERR(inode)) {
300                 ret = PTR_ERR(inode);
301                 goto cleanup;
302         }
303         srcu_read_unlock(&fs_info->subvol_srcu, index);
304
305         /* do a chunk of defrag */
306         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
307         memset(&range, 0, sizeof(range));
308         range.len = (u64)-1;
309         range.start = defrag->last_offset;
310
311         sb_start_write(fs_info->sb);
312         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
313                                        BTRFS_DEFRAG_BATCH);
314         sb_end_write(fs_info->sb);
315         /*
316          * if we filled the whole defrag batch, there
317          * must be more work to do.  Queue this defrag
318          * again
319          */
320         if (num_defrag == BTRFS_DEFRAG_BATCH) {
321                 defrag->last_offset = range.start;
322                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
323         } else if (defrag->last_offset && !defrag->cycled) {
324                 /*
325                  * we didn't fill our defrag batch, but
326                  * we didn't start at zero.  Make sure we loop
327                  * around to the start of the file.
328                  */
329                 defrag->last_offset = 0;
330                 defrag->cycled = 1;
331                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
332         } else {
333                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
334         }
335
336         iput(inode);
337         return 0;
338 cleanup:
339         srcu_read_unlock(&fs_info->subvol_srcu, index);
340         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
341         return ret;
342 }
343
344 /*
345  * run through the list of inodes in the FS that need
346  * defragging
347  */
348 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
349 {
350         struct inode_defrag *defrag;
351         u64 first_ino = 0;
352         u64 root_objectid = 0;
353
354         atomic_inc(&fs_info->defrag_running);
355         while (1) {
356                 /* Pause the auto defragger. */
357                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
358                              &fs_info->fs_state))
359                         break;
360
361                 if (!__need_auto_defrag(fs_info))
362                         break;
363
364                 /* find an inode to defrag */
365                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
366                                                  first_ino);
367                 if (!defrag) {
368                         if (root_objectid || first_ino) {
369                                 root_objectid = 0;
370                                 first_ino = 0;
371                                 continue;
372                         } else {
373                                 break;
374                         }
375                 }
376
377                 first_ino = defrag->ino + 1;
378                 root_objectid = defrag->root;
379
380                 __btrfs_run_defrag_inode(fs_info, defrag);
381         }
382         atomic_dec(&fs_info->defrag_running);
383
384         /*
385          * during unmount, we use the transaction_wait queue to
386          * wait for the defragger to stop
387          */
388         wake_up(&fs_info->transaction_wait);
389         return 0;
390 }
391
392 /* simple helper to fault in pages and copy.  This should go away
393  * and be replaced with calls into generic code.
394  */
395 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
396                                          struct page **prepared_pages,
397                                          struct iov_iter *i)
398 {
399         size_t copied = 0;
400         size_t total_copied = 0;
401         int pg = 0;
402         int offset = pos & (PAGE_SIZE - 1);
403
404         while (write_bytes > 0) {
405                 size_t count = min_t(size_t,
406                                      PAGE_SIZE - offset, write_bytes);
407                 struct page *page = prepared_pages[pg];
408                 /*
409                  * Copy data from userspace to the current page
410                  */
411                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
412
413                 /* Flush processor's dcache for this page */
414                 flush_dcache_page(page);
415
416                 /*
417                  * if we get a partial write, we can end up with
418                  * partially up to date pages.  These add
419                  * a lot of complexity, so make sure they don't
420                  * happen by forcing this copy to be retried.
421                  *
422                  * The rest of the btrfs_file_write code will fall
423                  * back to page at a time copies after we return 0.
424                  */
425                 if (!PageUptodate(page) && copied < count)
426                         copied = 0;
427
428                 iov_iter_advance(i, copied);
429                 write_bytes -= copied;
430                 total_copied += copied;
431
432                 /* Return to btrfs_file_write_iter to fault page */
433                 if (unlikely(copied == 0))
434                         break;
435
436                 if (copied < PAGE_SIZE - offset) {
437                         offset += copied;
438                 } else {
439                         pg++;
440                         offset = 0;
441                 }
442         }
443         return total_copied;
444 }
445
446 /*
447  * unlocks pages after btrfs_file_write is done with them
448  */
449 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
450 {
451         size_t i;
452         for (i = 0; i < num_pages; i++) {
453                 /* page checked is some magic around finding pages that
454                  * have been modified without going through btrfs_set_page_dirty
455                  * clear it here. There should be no need to mark the pages
456                  * accessed as prepare_pages should have marked them accessed
457                  * in prepare_pages via find_or_create_page()
458                  */
459                 ClearPageChecked(pages[i]);
460                 unlock_page(pages[i]);
461                 put_page(pages[i]);
462         }
463 }
464
465 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
466                                          const u64 start,
467                                          const u64 len,
468                                          struct extent_state **cached_state)
469 {
470         u64 search_start = start;
471         const u64 end = start + len - 1;
472
473         while (search_start < end) {
474                 const u64 search_len = end - search_start + 1;
475                 struct extent_map *em;
476                 u64 em_len;
477                 int ret = 0;
478
479                 em = btrfs_get_extent(inode, NULL, 0, search_start,
480                                       search_len, 0);
481                 if (IS_ERR(em))
482                         return PTR_ERR(em);
483
484                 if (em->block_start != EXTENT_MAP_HOLE)
485                         goto next;
486
487                 em_len = em->len;
488                 if (em->start < search_start)
489                         em_len -= search_start - em->start;
490                 if (em_len > search_len)
491                         em_len = search_len;
492
493                 ret = set_extent_bit(&inode->io_tree, search_start,
494                                      search_start + em_len - 1,
495                                      EXTENT_DELALLOC_NEW,
496                                      NULL, cached_state, GFP_NOFS);
497 next:
498                 search_start = extent_map_end(em);
499                 free_extent_map(em);
500                 if (ret)
501                         return ret;
502         }
503         return 0;
504 }
505
506 /*
507  * after copy_from_user, pages need to be dirtied and we need to make
508  * sure holes are created between the current EOF and the start of
509  * any next extents (if required).
510  *
511  * this also makes the decision about creating an inline extent vs
512  * doing real data extents, marking pages dirty and delalloc as required.
513  */
514 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
515                       size_t num_pages, loff_t pos, size_t write_bytes,
516                       struct extent_state **cached)
517 {
518         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
519         int err = 0;
520         int i;
521         u64 num_bytes;
522         u64 start_pos;
523         u64 end_of_last_block;
524         u64 end_pos = pos + write_bytes;
525         loff_t isize = i_size_read(inode);
526         unsigned int extra_bits = 0;
527
528         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
529         num_bytes = round_up(write_bytes + pos - start_pos,
530                              fs_info->sectorsize);
531
532         end_of_last_block = start_pos + num_bytes - 1;
533
534         /*
535          * The pages may have already been dirty, clear out old accounting so
536          * we can set things up properly
537          */
538         clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
539                          EXTENT_DIRTY | EXTENT_DELALLOC |
540                          EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 0, 0, cached);
541
542         if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
543                 if (start_pos >= isize &&
544                     !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
545                         /*
546                          * There can't be any extents following eof in this case
547                          * so just set the delalloc new bit for the range
548                          * directly.
549                          */
550                         extra_bits |= EXTENT_DELALLOC_NEW;
551                 } else {
552                         err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
553                                                             start_pos,
554                                                             num_bytes, cached);
555                         if (err)
556                                 return err;
557                 }
558         }
559
560         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
561                                         extra_bits, cached, 0);
562         if (err)
563                 return err;
564
565         for (i = 0; i < num_pages; i++) {
566                 struct page *p = pages[i];
567                 SetPageUptodate(p);
568                 ClearPageChecked(p);
569                 set_page_dirty(p);
570         }
571
572         /*
573          * we've only changed i_size in ram, and we haven't updated
574          * the disk i_size.  There is no need to log the inode
575          * at this time.
576          */
577         if (end_pos > isize)
578                 i_size_write(inode, end_pos);
579         return 0;
580 }
581
582 /*
583  * this drops all the extents in the cache that intersect the range
584  * [start, end].  Existing extents are split as required.
585  */
586 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
587                              int skip_pinned)
588 {
589         struct extent_map *em;
590         struct extent_map *split = NULL;
591         struct extent_map *split2 = NULL;
592         struct extent_map_tree *em_tree = &inode->extent_tree;
593         u64 len = end - start + 1;
594         u64 gen;
595         int ret;
596         int testend = 1;
597         unsigned long flags;
598         int compressed = 0;
599         bool modified;
600
601         WARN_ON(end < start);
602         if (end == (u64)-1) {
603                 len = (u64)-1;
604                 testend = 0;
605         }
606         while (1) {
607                 int no_splits = 0;
608
609                 modified = false;
610                 if (!split)
611                         split = alloc_extent_map();
612                 if (!split2)
613                         split2 = alloc_extent_map();
614                 if (!split || !split2)
615                         no_splits = 1;
616
617                 write_lock(&em_tree->lock);
618                 em = lookup_extent_mapping(em_tree, start, len);
619                 if (!em) {
620                         write_unlock(&em_tree->lock);
621                         break;
622                 }
623                 flags = em->flags;
624                 gen = em->generation;
625                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
626                         if (testend && em->start + em->len >= start + len) {
627                                 free_extent_map(em);
628                                 write_unlock(&em_tree->lock);
629                                 break;
630                         }
631                         start = em->start + em->len;
632                         if (testend)
633                                 len = start + len - (em->start + em->len);
634                         free_extent_map(em);
635                         write_unlock(&em_tree->lock);
636                         continue;
637                 }
638                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
639                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
640                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
641                 modified = !list_empty(&em->list);
642                 if (no_splits)
643                         goto next;
644
645                 if (em->start < start) {
646                         split->start = em->start;
647                         split->len = start - em->start;
648
649                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
650                                 split->orig_start = em->orig_start;
651                                 split->block_start = em->block_start;
652
653                                 if (compressed)
654                                         split->block_len = em->block_len;
655                                 else
656                                         split->block_len = split->len;
657                                 split->orig_block_len = max(split->block_len,
658                                                 em->orig_block_len);
659                                 split->ram_bytes = em->ram_bytes;
660                         } else {
661                                 split->orig_start = split->start;
662                                 split->block_len = 0;
663                                 split->block_start = em->block_start;
664                                 split->orig_block_len = 0;
665                                 split->ram_bytes = split->len;
666                         }
667
668                         split->generation = gen;
669                         split->bdev = em->bdev;
670                         split->flags = flags;
671                         split->compress_type = em->compress_type;
672                         replace_extent_mapping(em_tree, em, split, modified);
673                         free_extent_map(split);
674                         split = split2;
675                         split2 = NULL;
676                 }
677                 if (testend && em->start + em->len > start + len) {
678                         u64 diff = start + len - em->start;
679
680                         split->start = start + len;
681                         split->len = em->start + em->len - (start + len);
682                         split->bdev = em->bdev;
683                         split->flags = flags;
684                         split->compress_type = em->compress_type;
685                         split->generation = gen;
686
687                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
688                                 split->orig_block_len = max(em->block_len,
689                                                     em->orig_block_len);
690
691                                 split->ram_bytes = em->ram_bytes;
692                                 if (compressed) {
693                                         split->block_len = em->block_len;
694                                         split->block_start = em->block_start;
695                                         split->orig_start = em->orig_start;
696                                 } else {
697                                         split->block_len = split->len;
698                                         split->block_start = em->block_start
699                                                 + diff;
700                                         split->orig_start = em->orig_start;
701                                 }
702                         } else {
703                                 split->ram_bytes = split->len;
704                                 split->orig_start = split->start;
705                                 split->block_len = 0;
706                                 split->block_start = em->block_start;
707                                 split->orig_block_len = 0;
708                         }
709
710                         if (extent_map_in_tree(em)) {
711                                 replace_extent_mapping(em_tree, em, split,
712                                                        modified);
713                         } else {
714                                 ret = add_extent_mapping(em_tree, split,
715                                                          modified);
716                                 ASSERT(ret == 0); /* Logic error */
717                         }
718                         free_extent_map(split);
719                         split = NULL;
720                 }
721 next:
722                 if (extent_map_in_tree(em))
723                         remove_extent_mapping(em_tree, em);
724                 write_unlock(&em_tree->lock);
725
726                 /* once for us */
727                 free_extent_map(em);
728                 /* once for the tree*/
729                 free_extent_map(em);
730         }
731         if (split)
732                 free_extent_map(split);
733         if (split2)
734                 free_extent_map(split2);
735 }
736
737 /*
738  * this is very complex, but the basic idea is to drop all extents
739  * in the range start - end.  hint_block is filled in with a block number
740  * that would be a good hint to the block allocator for this file.
741  *
742  * If an extent intersects the range but is not entirely inside the range
743  * it is either truncated or split.  Anything entirely inside the range
744  * is deleted from the tree.
745  */
746 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
747                          struct btrfs_root *root, struct inode *inode,
748                          struct btrfs_path *path, u64 start, u64 end,
749                          u64 *drop_end, int drop_cache,
750                          int replace_extent,
751                          u32 extent_item_size,
752                          int *key_inserted)
753 {
754         struct btrfs_fs_info *fs_info = root->fs_info;
755         struct extent_buffer *leaf;
756         struct btrfs_file_extent_item *fi;
757         struct btrfs_key key;
758         struct btrfs_key new_key;
759         u64 ino = btrfs_ino(BTRFS_I(inode));
760         u64 search_start = start;
761         u64 disk_bytenr = 0;
762         u64 num_bytes = 0;
763         u64 extent_offset = 0;
764         u64 extent_end = 0;
765         u64 last_end = start;
766         int del_nr = 0;
767         int del_slot = 0;
768         int extent_type;
769         int recow;
770         int ret;
771         int modify_tree = -1;
772         int update_refs;
773         int found = 0;
774         int leafs_visited = 0;
775
776         if (drop_cache)
777                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
778
779         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
780                 modify_tree = 0;
781
782         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
783                        root == fs_info->tree_root);
784         while (1) {
785                 recow = 0;
786                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
787                                                search_start, modify_tree);
788                 if (ret < 0)
789                         break;
790                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
791                         leaf = path->nodes[0];
792                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
793                         if (key.objectid == ino &&
794                             key.type == BTRFS_EXTENT_DATA_KEY)
795                                 path->slots[0]--;
796                 }
797                 ret = 0;
798                 leafs_visited++;
799 next_slot:
800                 leaf = path->nodes[0];
801                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
802                         BUG_ON(del_nr > 0);
803                         ret = btrfs_next_leaf(root, path);
804                         if (ret < 0)
805                                 break;
806                         if (ret > 0) {
807                                 ret = 0;
808                                 break;
809                         }
810                         leafs_visited++;
811                         leaf = path->nodes[0];
812                         recow = 1;
813                 }
814
815                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
816
817                 if (key.objectid > ino)
818                         break;
819                 if (WARN_ON_ONCE(key.objectid < ino) ||
820                     key.type < BTRFS_EXTENT_DATA_KEY) {
821                         ASSERT(del_nr == 0);
822                         path->slots[0]++;
823                         goto next_slot;
824                 }
825                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
826                         break;
827
828                 fi = btrfs_item_ptr(leaf, path->slots[0],
829                                     struct btrfs_file_extent_item);
830                 extent_type = btrfs_file_extent_type(leaf, fi);
831
832                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
833                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
834                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
835                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
836                         extent_offset = btrfs_file_extent_offset(leaf, fi);
837                         extent_end = key.offset +
838                                 btrfs_file_extent_num_bytes(leaf, fi);
839                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
840                         extent_end = key.offset +
841                                 btrfs_file_extent_ram_bytes(leaf, fi);
842                 } else {
843                         /* can't happen */
844                         BUG();
845                 }
846
847                 /*
848                  * Don't skip extent items representing 0 byte lengths. They
849                  * used to be created (bug) if while punching holes we hit
850                  * -ENOSPC condition. So if we find one here, just ensure we
851                  * delete it, otherwise we would insert a new file extent item
852                  * with the same key (offset) as that 0 bytes length file
853                  * extent item in the call to setup_items_for_insert() later
854                  * in this function.
855                  */
856                 if (extent_end == key.offset && extent_end >= search_start) {
857                         last_end = extent_end;
858                         goto delete_extent_item;
859                 }
860
861                 if (extent_end <= search_start) {
862                         path->slots[0]++;
863                         goto next_slot;
864                 }
865
866                 found = 1;
867                 search_start = max(key.offset, start);
868                 if (recow || !modify_tree) {
869                         modify_tree = -1;
870                         btrfs_release_path(path);
871                         continue;
872                 }
873
874                 /*
875                  *     | - range to drop - |
876                  *  | -------- extent -------- |
877                  */
878                 if (start > key.offset && end < extent_end) {
879                         BUG_ON(del_nr > 0);
880                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
881                                 ret = -EOPNOTSUPP;
882                                 break;
883                         }
884
885                         memcpy(&new_key, &key, sizeof(new_key));
886                         new_key.offset = start;
887                         ret = btrfs_duplicate_item(trans, root, path,
888                                                    &new_key);
889                         if (ret == -EAGAIN) {
890                                 btrfs_release_path(path);
891                                 continue;
892                         }
893                         if (ret < 0)
894                                 break;
895
896                         leaf = path->nodes[0];
897                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
898                                             struct btrfs_file_extent_item);
899                         btrfs_set_file_extent_num_bytes(leaf, fi,
900                                                         start - key.offset);
901
902                         fi = btrfs_item_ptr(leaf, path->slots[0],
903                                             struct btrfs_file_extent_item);
904
905                         extent_offset += start - key.offset;
906                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
907                         btrfs_set_file_extent_num_bytes(leaf, fi,
908                                                         extent_end - start);
909                         btrfs_mark_buffer_dirty(leaf);
910
911                         if (update_refs && disk_bytenr > 0) {
912                                 ret = btrfs_inc_extent_ref(trans, root,
913                                                 disk_bytenr, num_bytes, 0,
914                                                 root->root_key.objectid,
915                                                 new_key.objectid,
916                                                 start - extent_offset);
917                                 BUG_ON(ret); /* -ENOMEM */
918                         }
919                         key.offset = start;
920                 }
921                 /*
922                  * From here on out we will have actually dropped something, so
923                  * last_end can be updated.
924                  */
925                 last_end = extent_end;
926
927                 /*
928                  *  | ---- range to drop ----- |
929                  *      | -------- extent -------- |
930                  */
931                 if (start <= key.offset && end < extent_end) {
932                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
933                                 ret = -EOPNOTSUPP;
934                                 break;
935                         }
936
937                         memcpy(&new_key, &key, sizeof(new_key));
938                         new_key.offset = end;
939                         btrfs_set_item_key_safe(fs_info, path, &new_key);
940
941                         extent_offset += end - key.offset;
942                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
943                         btrfs_set_file_extent_num_bytes(leaf, fi,
944                                                         extent_end - end);
945                         btrfs_mark_buffer_dirty(leaf);
946                         if (update_refs && disk_bytenr > 0)
947                                 inode_sub_bytes(inode, end - key.offset);
948                         break;
949                 }
950
951                 search_start = extent_end;
952                 /*
953                  *       | ---- range to drop ----- |
954                  *  | -------- extent -------- |
955                  */
956                 if (start > key.offset && end >= extent_end) {
957                         BUG_ON(del_nr > 0);
958                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
959                                 ret = -EOPNOTSUPP;
960                                 break;
961                         }
962
963                         btrfs_set_file_extent_num_bytes(leaf, fi,
964                                                         start - key.offset);
965                         btrfs_mark_buffer_dirty(leaf);
966                         if (update_refs && disk_bytenr > 0)
967                                 inode_sub_bytes(inode, extent_end - start);
968                         if (end == extent_end)
969                                 break;
970
971                         path->slots[0]++;
972                         goto next_slot;
973                 }
974
975                 /*
976                  *  | ---- range to drop ----- |
977                  *    | ------ extent ------ |
978                  */
979                 if (start <= key.offset && end >= extent_end) {
980 delete_extent_item:
981                         if (del_nr == 0) {
982                                 del_slot = path->slots[0];
983                                 del_nr = 1;
984                         } else {
985                                 BUG_ON(del_slot + del_nr != path->slots[0]);
986                                 del_nr++;
987                         }
988
989                         if (update_refs &&
990                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
991                                 inode_sub_bytes(inode,
992                                                 extent_end - key.offset);
993                                 extent_end = ALIGN(extent_end,
994                                                    fs_info->sectorsize);
995                         } else if (update_refs && disk_bytenr > 0) {
996                                 ret = btrfs_free_extent(trans, root,
997                                                 disk_bytenr, num_bytes, 0,
998                                                 root->root_key.objectid,
999                                                 key.objectid, key.offset -
1000                                                 extent_offset);
1001                                 BUG_ON(ret); /* -ENOMEM */
1002                                 inode_sub_bytes(inode,
1003                                                 extent_end - key.offset);
1004                         }
1005
1006                         if (end == extent_end)
1007                                 break;
1008
1009                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1010                                 path->slots[0]++;
1011                                 goto next_slot;
1012                         }
1013
1014                         ret = btrfs_del_items(trans, root, path, del_slot,
1015                                               del_nr);
1016                         if (ret) {
1017                                 btrfs_abort_transaction(trans, ret);
1018                                 break;
1019                         }
1020
1021                         del_nr = 0;
1022                         del_slot = 0;
1023
1024                         btrfs_release_path(path);
1025                         continue;
1026                 }
1027
1028                 BUG_ON(1);
1029         }
1030
1031         if (!ret && del_nr > 0) {
1032                 /*
1033                  * Set path->slots[0] to first slot, so that after the delete
1034                  * if items are move off from our leaf to its immediate left or
1035                  * right neighbor leafs, we end up with a correct and adjusted
1036                  * path->slots[0] for our insertion (if replace_extent != 0).
1037                  */
1038                 path->slots[0] = del_slot;
1039                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1040                 if (ret)
1041                         btrfs_abort_transaction(trans, ret);
1042         }
1043
1044         leaf = path->nodes[0];
1045         /*
1046          * If btrfs_del_items() was called, it might have deleted a leaf, in
1047          * which case it unlocked our path, so check path->locks[0] matches a
1048          * write lock.
1049          */
1050         if (!ret && replace_extent && leafs_visited == 1 &&
1051             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1052              path->locks[0] == BTRFS_WRITE_LOCK) &&
1053             btrfs_leaf_free_space(fs_info, leaf) >=
1054             sizeof(struct btrfs_item) + extent_item_size) {
1055
1056                 key.objectid = ino;
1057                 key.type = BTRFS_EXTENT_DATA_KEY;
1058                 key.offset = start;
1059                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1060                         struct btrfs_key slot_key;
1061
1062                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1063                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1064                                 path->slots[0]++;
1065                 }
1066                 setup_items_for_insert(root, path, &key,
1067                                        &extent_item_size,
1068                                        extent_item_size,
1069                                        sizeof(struct btrfs_item) +
1070                                        extent_item_size, 1);
1071                 *key_inserted = 1;
1072         }
1073
1074         if (!replace_extent || !(*key_inserted))
1075                 btrfs_release_path(path);
1076         if (drop_end)
1077                 *drop_end = found ? min(end, last_end) : end;
1078         return ret;
1079 }
1080
1081 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1082                        struct btrfs_root *root, struct inode *inode, u64 start,
1083                        u64 end, int drop_cache)
1084 {
1085         struct btrfs_path *path;
1086         int ret;
1087
1088         path = btrfs_alloc_path();
1089         if (!path)
1090                 return -ENOMEM;
1091         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1092                                    drop_cache, 0, 0, NULL);
1093         btrfs_free_path(path);
1094         return ret;
1095 }
1096
1097 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1098                             u64 objectid, u64 bytenr, u64 orig_offset,
1099                             u64 *start, u64 *end)
1100 {
1101         struct btrfs_file_extent_item *fi;
1102         struct btrfs_key key;
1103         u64 extent_end;
1104
1105         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1106                 return 0;
1107
1108         btrfs_item_key_to_cpu(leaf, &key, slot);
1109         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1110                 return 0;
1111
1112         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1113         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1114             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1115             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1116             btrfs_file_extent_compression(leaf, fi) ||
1117             btrfs_file_extent_encryption(leaf, fi) ||
1118             btrfs_file_extent_other_encoding(leaf, fi))
1119                 return 0;
1120
1121         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1122         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1123                 return 0;
1124
1125         *start = key.offset;
1126         *end = extent_end;
1127         return 1;
1128 }
1129
1130 /*
1131  * Mark extent in the range start - end as written.
1132  *
1133  * This changes extent type from 'pre-allocated' to 'regular'. If only
1134  * part of extent is marked as written, the extent will be split into
1135  * two or three.
1136  */
1137 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1138                               struct btrfs_inode *inode, u64 start, u64 end)
1139 {
1140         struct btrfs_fs_info *fs_info = trans->fs_info;
1141         struct btrfs_root *root = inode->root;
1142         struct extent_buffer *leaf;
1143         struct btrfs_path *path;
1144         struct btrfs_file_extent_item *fi;
1145         struct btrfs_key key;
1146         struct btrfs_key new_key;
1147         u64 bytenr;
1148         u64 num_bytes;
1149         u64 extent_end;
1150         u64 orig_offset;
1151         u64 other_start;
1152         u64 other_end;
1153         u64 split;
1154         int del_nr = 0;
1155         int del_slot = 0;
1156         int recow;
1157         int ret;
1158         u64 ino = btrfs_ino(inode);
1159
1160         path = btrfs_alloc_path();
1161         if (!path)
1162                 return -ENOMEM;
1163 again:
1164         recow = 0;
1165         split = start;
1166         key.objectid = ino;
1167         key.type = BTRFS_EXTENT_DATA_KEY;
1168         key.offset = split;
1169
1170         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1171         if (ret < 0)
1172                 goto out;
1173         if (ret > 0 && path->slots[0] > 0)
1174                 path->slots[0]--;
1175
1176         leaf = path->nodes[0];
1177         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1178         if (key.objectid != ino ||
1179             key.type != BTRFS_EXTENT_DATA_KEY) {
1180                 ret = -EINVAL;
1181                 btrfs_abort_transaction(trans, ret);
1182                 goto out;
1183         }
1184         fi = btrfs_item_ptr(leaf, path->slots[0],
1185                             struct btrfs_file_extent_item);
1186         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1187                 ret = -EINVAL;
1188                 btrfs_abort_transaction(trans, ret);
1189                 goto out;
1190         }
1191         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1192         if (key.offset > start || extent_end < end) {
1193                 ret = -EINVAL;
1194                 btrfs_abort_transaction(trans, ret);
1195                 goto out;
1196         }
1197
1198         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1199         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1200         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1201         memcpy(&new_key, &key, sizeof(new_key));
1202
1203         if (start == key.offset && end < extent_end) {
1204                 other_start = 0;
1205                 other_end = start;
1206                 if (extent_mergeable(leaf, path->slots[0] - 1,
1207                                      ino, bytenr, orig_offset,
1208                                      &other_start, &other_end)) {
1209                         new_key.offset = end;
1210                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1211                         fi = btrfs_item_ptr(leaf, path->slots[0],
1212                                             struct btrfs_file_extent_item);
1213                         btrfs_set_file_extent_generation(leaf, fi,
1214                                                          trans->transid);
1215                         btrfs_set_file_extent_num_bytes(leaf, fi,
1216                                                         extent_end - end);
1217                         btrfs_set_file_extent_offset(leaf, fi,
1218                                                      end - orig_offset);
1219                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1220                                             struct btrfs_file_extent_item);
1221                         btrfs_set_file_extent_generation(leaf, fi,
1222                                                          trans->transid);
1223                         btrfs_set_file_extent_num_bytes(leaf, fi,
1224                                                         end - other_start);
1225                         btrfs_mark_buffer_dirty(leaf);
1226                         goto out;
1227                 }
1228         }
1229
1230         if (start > key.offset && end == extent_end) {
1231                 other_start = end;
1232                 other_end = 0;
1233                 if (extent_mergeable(leaf, path->slots[0] + 1,
1234                                      ino, bytenr, orig_offset,
1235                                      &other_start, &other_end)) {
1236                         fi = btrfs_item_ptr(leaf, path->slots[0],
1237                                             struct btrfs_file_extent_item);
1238                         btrfs_set_file_extent_num_bytes(leaf, fi,
1239                                                         start - key.offset);
1240                         btrfs_set_file_extent_generation(leaf, fi,
1241                                                          trans->transid);
1242                         path->slots[0]++;
1243                         new_key.offset = start;
1244                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1245
1246                         fi = btrfs_item_ptr(leaf, path->slots[0],
1247                                             struct btrfs_file_extent_item);
1248                         btrfs_set_file_extent_generation(leaf, fi,
1249                                                          trans->transid);
1250                         btrfs_set_file_extent_num_bytes(leaf, fi,
1251                                                         other_end - start);
1252                         btrfs_set_file_extent_offset(leaf, fi,
1253                                                      start - orig_offset);
1254                         btrfs_mark_buffer_dirty(leaf);
1255                         goto out;
1256                 }
1257         }
1258
1259         while (start > key.offset || end < extent_end) {
1260                 if (key.offset == start)
1261                         split = end;
1262
1263                 new_key.offset = split;
1264                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1265                 if (ret == -EAGAIN) {
1266                         btrfs_release_path(path);
1267                         goto again;
1268                 }
1269                 if (ret < 0) {
1270                         btrfs_abort_transaction(trans, ret);
1271                         goto out;
1272                 }
1273
1274                 leaf = path->nodes[0];
1275                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1276                                     struct btrfs_file_extent_item);
1277                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1278                 btrfs_set_file_extent_num_bytes(leaf, fi,
1279                                                 split - key.offset);
1280
1281                 fi = btrfs_item_ptr(leaf, path->slots[0],
1282                                     struct btrfs_file_extent_item);
1283
1284                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1285                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1286                 btrfs_set_file_extent_num_bytes(leaf, fi,
1287                                                 extent_end - split);
1288                 btrfs_mark_buffer_dirty(leaf);
1289
1290                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes,
1291                                            0, root->root_key.objectid,
1292                                            ino, orig_offset);
1293                 if (ret) {
1294                         btrfs_abort_transaction(trans, ret);
1295                         goto out;
1296                 }
1297
1298                 if (split == start) {
1299                         key.offset = start;
1300                 } else {
1301                         if (start != key.offset) {
1302                                 ret = -EINVAL;
1303                                 btrfs_abort_transaction(trans, ret);
1304                                 goto out;
1305                         }
1306                         path->slots[0]--;
1307                         extent_end = end;
1308                 }
1309                 recow = 1;
1310         }
1311
1312         other_start = end;
1313         other_end = 0;
1314         if (extent_mergeable(leaf, path->slots[0] + 1,
1315                              ino, bytenr, orig_offset,
1316                              &other_start, &other_end)) {
1317                 if (recow) {
1318                         btrfs_release_path(path);
1319                         goto again;
1320                 }
1321                 extent_end = other_end;
1322                 del_slot = path->slots[0] + 1;
1323                 del_nr++;
1324                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1325                                         0, root->root_key.objectid,
1326                                         ino, orig_offset);
1327                 if (ret) {
1328                         btrfs_abort_transaction(trans, ret);
1329                         goto out;
1330                 }
1331         }
1332         other_start = 0;
1333         other_end = start;
1334         if (extent_mergeable(leaf, path->slots[0] - 1,
1335                              ino, bytenr, orig_offset,
1336                              &other_start, &other_end)) {
1337                 if (recow) {
1338                         btrfs_release_path(path);
1339                         goto again;
1340                 }
1341                 key.offset = other_start;
1342                 del_slot = path->slots[0];
1343                 del_nr++;
1344                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1345                                         0, root->root_key.objectid,
1346                                         ino, orig_offset);
1347                 if (ret) {
1348                         btrfs_abort_transaction(trans, ret);
1349                         goto out;
1350                 }
1351         }
1352         if (del_nr == 0) {
1353                 fi = btrfs_item_ptr(leaf, path->slots[0],
1354                            struct btrfs_file_extent_item);
1355                 btrfs_set_file_extent_type(leaf, fi,
1356                                            BTRFS_FILE_EXTENT_REG);
1357                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1358                 btrfs_mark_buffer_dirty(leaf);
1359         } else {
1360                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1361                            struct btrfs_file_extent_item);
1362                 btrfs_set_file_extent_type(leaf, fi,
1363                                            BTRFS_FILE_EXTENT_REG);
1364                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1365                 btrfs_set_file_extent_num_bytes(leaf, fi,
1366                                                 extent_end - key.offset);
1367                 btrfs_mark_buffer_dirty(leaf);
1368
1369                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1370                 if (ret < 0) {
1371                         btrfs_abort_transaction(trans, ret);
1372                         goto out;
1373                 }
1374         }
1375 out:
1376         btrfs_free_path(path);
1377         return 0;
1378 }
1379
1380 /*
1381  * on error we return an unlocked page and the error value
1382  * on success we return a locked page and 0
1383  */
1384 static int prepare_uptodate_page(struct inode *inode,
1385                                  struct page *page, u64 pos,
1386                                  bool force_uptodate)
1387 {
1388         int ret = 0;
1389
1390         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1391             !PageUptodate(page)) {
1392                 ret = btrfs_readpage(NULL, page);
1393                 if (ret)
1394                         return ret;
1395                 lock_page(page);
1396                 if (!PageUptodate(page)) {
1397                         unlock_page(page);
1398                         return -EIO;
1399                 }
1400                 if (page->mapping != inode->i_mapping) {
1401                         unlock_page(page);
1402                         return -EAGAIN;
1403                 }
1404         }
1405         return 0;
1406 }
1407
1408 /*
1409  * this just gets pages into the page cache and locks them down.
1410  */
1411 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1412                                   size_t num_pages, loff_t pos,
1413                                   size_t write_bytes, bool force_uptodate)
1414 {
1415         int i;
1416         unsigned long index = pos >> PAGE_SHIFT;
1417         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1418         int err = 0;
1419         int faili;
1420
1421         for (i = 0; i < num_pages; i++) {
1422 again:
1423                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1424                                                mask | __GFP_WRITE);
1425                 if (!pages[i]) {
1426                         faili = i - 1;
1427                         err = -ENOMEM;
1428                         goto fail;
1429                 }
1430
1431                 if (i == 0)
1432                         err = prepare_uptodate_page(inode, pages[i], pos,
1433                                                     force_uptodate);
1434                 if (!err && i == num_pages - 1)
1435                         err = prepare_uptodate_page(inode, pages[i],
1436                                                     pos + write_bytes, false);
1437                 if (err) {
1438                         put_page(pages[i]);
1439                         if (err == -EAGAIN) {
1440                                 err = 0;
1441                                 goto again;
1442                         }
1443                         faili = i - 1;
1444                         goto fail;
1445                 }
1446                 wait_on_page_writeback(pages[i]);
1447         }
1448
1449         return 0;
1450 fail:
1451         while (faili >= 0) {
1452                 unlock_page(pages[faili]);
1453                 put_page(pages[faili]);
1454                 faili--;
1455         }
1456         return err;
1457
1458 }
1459
1460 /*
1461  * This function locks the extent and properly waits for data=ordered extents
1462  * to finish before allowing the pages to be modified if need.
1463  *
1464  * The return value:
1465  * 1 - the extent is locked
1466  * 0 - the extent is not locked, and everything is OK
1467  * -EAGAIN - need re-prepare the pages
1468  * the other < 0 number - Something wrong happens
1469  */
1470 static noinline int
1471 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1472                                 size_t num_pages, loff_t pos,
1473                                 size_t write_bytes,
1474                                 u64 *lockstart, u64 *lockend,
1475                                 struct extent_state **cached_state)
1476 {
1477         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1478         u64 start_pos;
1479         u64 last_pos;
1480         int i;
1481         int ret = 0;
1482
1483         start_pos = round_down(pos, fs_info->sectorsize);
1484         last_pos = start_pos
1485                 + round_up(pos + write_bytes - start_pos,
1486                            fs_info->sectorsize) - 1;
1487
1488         if (start_pos < inode->vfs_inode.i_size) {
1489                 struct btrfs_ordered_extent *ordered;
1490
1491                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1492                                 cached_state);
1493                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1494                                                      last_pos - start_pos + 1);
1495                 if (ordered &&
1496                     ordered->file_offset + ordered->len > start_pos &&
1497                     ordered->file_offset <= last_pos) {
1498                         unlock_extent_cached(&inode->io_tree, start_pos,
1499                                         last_pos, cached_state);
1500                         for (i = 0; i < num_pages; i++) {
1501                                 unlock_page(pages[i]);
1502                                 put_page(pages[i]);
1503                         }
1504                         btrfs_start_ordered_extent(&inode->vfs_inode,
1505                                         ordered, 1);
1506                         btrfs_put_ordered_extent(ordered);
1507                         return -EAGAIN;
1508                 }
1509                 if (ordered)
1510                         btrfs_put_ordered_extent(ordered);
1511
1512                 *lockstart = start_pos;
1513                 *lockend = last_pos;
1514                 ret = 1;
1515         }
1516
1517         /*
1518          * It's possible the pages are dirty right now, but we don't want
1519          * to clean them yet because copy_from_user may catch a page fault
1520          * and we might have to fall back to one page at a time.  If that
1521          * happens, we'll unlock these pages and we'd have a window where
1522          * reclaim could sneak in and drop the once-dirty page on the floor
1523          * without writing it.
1524          *
1525          * We have the pages locked and the extent range locked, so there's
1526          * no way someone can start IO on any dirty pages in this range.
1527          *
1528          * We'll call btrfs_dirty_pages() later on, and that will flip around
1529          * delalloc bits and dirty the pages as required.
1530          */
1531         for (i = 0; i < num_pages; i++) {
1532                 set_page_extent_mapped(pages[i]);
1533                 WARN_ON(!PageLocked(pages[i]));
1534         }
1535
1536         return ret;
1537 }
1538
1539 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1540                                     size_t *write_bytes)
1541 {
1542         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1543         struct btrfs_root *root = inode->root;
1544         struct btrfs_ordered_extent *ordered;
1545         u64 lockstart, lockend;
1546         u64 num_bytes;
1547         int ret;
1548
1549         ret = btrfs_start_write_no_snapshotting(root);
1550         if (!ret)
1551                 return -ENOSPC;
1552
1553         lockstart = round_down(pos, fs_info->sectorsize);
1554         lockend = round_up(pos + *write_bytes,
1555                            fs_info->sectorsize) - 1;
1556
1557         while (1) {
1558                 lock_extent(&inode->io_tree, lockstart, lockend);
1559                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1560                                                      lockend - lockstart + 1);
1561                 if (!ordered) {
1562                         break;
1563                 }
1564                 unlock_extent(&inode->io_tree, lockstart, lockend);
1565                 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1566                 btrfs_put_ordered_extent(ordered);
1567         }
1568
1569         num_bytes = lockend - lockstart + 1;
1570         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1571                         NULL, NULL, NULL);
1572         if (ret <= 0) {
1573                 ret = 0;
1574                 btrfs_end_write_no_snapshotting(root);
1575         } else {
1576                 *write_bytes = min_t(size_t, *write_bytes ,
1577                                      num_bytes - pos + lockstart);
1578         }
1579
1580         unlock_extent(&inode->io_tree, lockstart, lockend);
1581
1582         return ret;
1583 }
1584
1585 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1586                                                struct iov_iter *i)
1587 {
1588         struct file *file = iocb->ki_filp;
1589         loff_t pos = iocb->ki_pos;
1590         struct inode *inode = file_inode(file);
1591         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1592         struct btrfs_root *root = BTRFS_I(inode)->root;
1593         struct page **pages = NULL;
1594         struct extent_state *cached_state = NULL;
1595         struct extent_changeset *data_reserved = NULL;
1596         u64 release_bytes = 0;
1597         u64 lockstart;
1598         u64 lockend;
1599         size_t num_written = 0;
1600         int nrptrs;
1601         int ret = 0;
1602         bool only_release_metadata = false;
1603         bool force_page_uptodate = false;
1604
1605         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1606                         PAGE_SIZE / (sizeof(struct page *)));
1607         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1608         nrptrs = max(nrptrs, 8);
1609         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1610         if (!pages)
1611                 return -ENOMEM;
1612
1613         while (iov_iter_count(i) > 0) {
1614                 size_t offset = pos & (PAGE_SIZE - 1);
1615                 size_t sector_offset;
1616                 size_t write_bytes = min(iov_iter_count(i),
1617                                          nrptrs * (size_t)PAGE_SIZE -
1618                                          offset);
1619                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1620                                                 PAGE_SIZE);
1621                 size_t reserve_bytes;
1622                 size_t dirty_pages;
1623                 size_t copied;
1624                 size_t dirty_sectors;
1625                 size_t num_sectors;
1626                 int extents_locked;
1627
1628                 WARN_ON(num_pages > nrptrs);
1629
1630                 /*
1631                  * Fault pages before locking them in prepare_pages
1632                  * to avoid recursive lock
1633                  */
1634                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1635                         ret = -EFAULT;
1636                         break;
1637                 }
1638
1639                 sector_offset = pos & (fs_info->sectorsize - 1);
1640                 reserve_bytes = round_up(write_bytes + sector_offset,
1641                                 fs_info->sectorsize);
1642
1643                 extent_changeset_release(data_reserved);
1644                 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1645                                                   write_bytes);
1646                 if (ret < 0) {
1647                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1648                                                       BTRFS_INODE_PREALLOC)) &&
1649                             check_can_nocow(BTRFS_I(inode), pos,
1650                                         &write_bytes) > 0) {
1651                                 /*
1652                                  * For nodata cow case, no need to reserve
1653                                  * data space.
1654                                  */
1655                                 only_release_metadata = true;
1656                                 /*
1657                                  * our prealloc extent may be smaller than
1658                                  * write_bytes, so scale down.
1659                                  */
1660                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1661                                                          PAGE_SIZE);
1662                                 reserve_bytes = round_up(write_bytes +
1663                                                          sector_offset,
1664                                                          fs_info->sectorsize);
1665                         } else {
1666                                 break;
1667                         }
1668                 }
1669
1670                 WARN_ON(reserve_bytes == 0);
1671                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1672                                 reserve_bytes);
1673                 if (ret) {
1674                         if (!only_release_metadata)
1675                                 btrfs_free_reserved_data_space(inode,
1676                                                 data_reserved, pos,
1677                                                 write_bytes);
1678                         else
1679                                 btrfs_end_write_no_snapshotting(root);
1680                         break;
1681                 }
1682
1683                 release_bytes = reserve_bytes;
1684 again:
1685                 /*
1686                  * This is going to setup the pages array with the number of
1687                  * pages we want, so we don't really need to worry about the
1688                  * contents of pages from loop to loop
1689                  */
1690                 ret = prepare_pages(inode, pages, num_pages,
1691                                     pos, write_bytes,
1692                                     force_page_uptodate);
1693                 if (ret) {
1694                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1695                                                        reserve_bytes, true);
1696                         break;
1697                 }
1698
1699                 extents_locked = lock_and_cleanup_extent_if_need(
1700                                 BTRFS_I(inode), pages,
1701                                 num_pages, pos, write_bytes, &lockstart,
1702                                 &lockend, &cached_state);
1703                 if (extents_locked < 0) {
1704                         if (extents_locked == -EAGAIN)
1705                                 goto again;
1706                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1707                                                        reserve_bytes, true);
1708                         ret = extents_locked;
1709                         break;
1710                 }
1711
1712                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1713
1714                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1715                 dirty_sectors = round_up(copied + sector_offset,
1716                                         fs_info->sectorsize);
1717                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1718
1719                 /*
1720                  * if we have trouble faulting in the pages, fall
1721                  * back to one page at a time
1722                  */
1723                 if (copied < write_bytes)
1724                         nrptrs = 1;
1725
1726                 if (copied == 0) {
1727                         force_page_uptodate = true;
1728                         dirty_sectors = 0;
1729                         dirty_pages = 0;
1730                 } else {
1731                         force_page_uptodate = false;
1732                         dirty_pages = DIV_ROUND_UP(copied + offset,
1733                                                    PAGE_SIZE);
1734                 }
1735
1736                 if (num_sectors > dirty_sectors) {
1737                         /* release everything except the sectors we dirtied */
1738                         release_bytes -= dirty_sectors <<
1739                                                 fs_info->sb->s_blocksize_bits;
1740                         if (only_release_metadata) {
1741                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1742                                                         release_bytes, true);
1743                         } else {
1744                                 u64 __pos;
1745
1746                                 __pos = round_down(pos,
1747                                                    fs_info->sectorsize) +
1748                                         (dirty_pages << PAGE_SHIFT);
1749                                 btrfs_delalloc_release_space(inode,
1750                                                 data_reserved, __pos,
1751                                                 release_bytes, true);
1752                         }
1753                 }
1754
1755                 release_bytes = round_up(copied + sector_offset,
1756                                         fs_info->sectorsize);
1757
1758                 if (copied > 0)
1759                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1760                                                 pos, copied, &cached_state);
1761                 if (extents_locked)
1762                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1763                                              lockstart, lockend, &cached_state);
1764                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes,
1765                                                true);
1766                 if (ret) {
1767                         btrfs_drop_pages(pages, num_pages);
1768                         break;
1769                 }
1770
1771                 release_bytes = 0;
1772                 if (only_release_metadata)
1773                         btrfs_end_write_no_snapshotting(root);
1774
1775                 if (only_release_metadata && copied > 0) {
1776                         lockstart = round_down(pos,
1777                                                fs_info->sectorsize);
1778                         lockend = round_up(pos + copied,
1779                                            fs_info->sectorsize) - 1;
1780
1781                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1782                                        lockend, EXTENT_NORESERVE, NULL,
1783                                        NULL, GFP_NOFS);
1784                         only_release_metadata = false;
1785                 }
1786
1787                 btrfs_drop_pages(pages, num_pages);
1788
1789                 cond_resched();
1790
1791                 balance_dirty_pages_ratelimited(inode->i_mapping);
1792                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1793                         btrfs_btree_balance_dirty(fs_info);
1794
1795                 pos += copied;
1796                 num_written += copied;
1797         }
1798
1799         kfree(pages);
1800
1801         if (release_bytes) {
1802                 if (only_release_metadata) {
1803                         btrfs_end_write_no_snapshotting(root);
1804                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1805                                         release_bytes, true);
1806                 } else {
1807                         btrfs_delalloc_release_space(inode, data_reserved,
1808                                         round_down(pos, fs_info->sectorsize),
1809                                         release_bytes, true);
1810                 }
1811         }
1812
1813         extent_changeset_free(data_reserved);
1814         return num_written ? num_written : ret;
1815 }
1816
1817 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1818 {
1819         struct file *file = iocb->ki_filp;
1820         struct inode *inode = file_inode(file);
1821         loff_t pos;
1822         ssize_t written;
1823         ssize_t written_buffered;
1824         loff_t endbyte;
1825         int err;
1826
1827         written = generic_file_direct_write(iocb, from);
1828
1829         if (written < 0 || !iov_iter_count(from))
1830                 return written;
1831
1832         pos = iocb->ki_pos;
1833         written_buffered = btrfs_buffered_write(iocb, from);
1834         if (written_buffered < 0) {
1835                 err = written_buffered;
1836                 goto out;
1837         }
1838         /*
1839          * Ensure all data is persisted. We want the next direct IO read to be
1840          * able to read what was just written.
1841          */
1842         endbyte = pos + written_buffered - 1;
1843         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1844         if (err)
1845                 goto out;
1846         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1847         if (err)
1848                 goto out;
1849         written += written_buffered;
1850         iocb->ki_pos = pos + written_buffered;
1851         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1852                                  endbyte >> PAGE_SHIFT);
1853 out:
1854         return written ? written : err;
1855 }
1856
1857 static void update_time_for_write(struct inode *inode)
1858 {
1859         struct timespec64 now;
1860
1861         if (IS_NOCMTIME(inode))
1862                 return;
1863
1864         now = current_time(inode);
1865         if (!timespec64_equal(&inode->i_mtime, &now))
1866                 inode->i_mtime = now;
1867
1868         if (!timespec64_equal(&inode->i_ctime, &now))
1869                 inode->i_ctime = now;
1870
1871         if (IS_I_VERSION(inode))
1872                 inode_inc_iversion(inode);
1873 }
1874
1875 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1876                                     struct iov_iter *from)
1877 {
1878         struct file *file = iocb->ki_filp;
1879         struct inode *inode = file_inode(file);
1880         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1881         struct btrfs_root *root = BTRFS_I(inode)->root;
1882         u64 start_pos;
1883         u64 end_pos;
1884         ssize_t num_written = 0;
1885         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1886         ssize_t err;
1887         loff_t pos;
1888         size_t count = iov_iter_count(from);
1889         loff_t oldsize;
1890         int clean_page = 0;
1891
1892         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1893             (iocb->ki_flags & IOCB_NOWAIT))
1894                 return -EOPNOTSUPP;
1895
1896         if (!inode_trylock(inode)) {
1897                 if (iocb->ki_flags & IOCB_NOWAIT)
1898                         return -EAGAIN;
1899                 inode_lock(inode);
1900         }
1901
1902         err = generic_write_checks(iocb, from);
1903         if (err <= 0) {
1904                 inode_unlock(inode);
1905                 return err;
1906         }
1907
1908         pos = iocb->ki_pos;
1909         if (iocb->ki_flags & IOCB_NOWAIT) {
1910                 /*
1911                  * We will allocate space in case nodatacow is not set,
1912                  * so bail
1913                  */
1914                 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1915                                               BTRFS_INODE_PREALLOC)) ||
1916                     check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1917                         inode_unlock(inode);
1918                         return -EAGAIN;
1919                 }
1920         }
1921
1922         current->backing_dev_info = inode_to_bdi(inode);
1923         err = file_remove_privs(file);
1924         if (err) {
1925                 inode_unlock(inode);
1926                 goto out;
1927         }
1928
1929         /*
1930          * If BTRFS flips readonly due to some impossible error
1931          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1932          * although we have opened a file as writable, we have
1933          * to stop this write operation to ensure FS consistency.
1934          */
1935         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1936                 inode_unlock(inode);
1937                 err = -EROFS;
1938                 goto out;
1939         }
1940
1941         /*
1942          * We reserve space for updating the inode when we reserve space for the
1943          * extent we are going to write, so we will enospc out there.  We don't
1944          * need to start yet another transaction to update the inode as we will
1945          * update the inode when we finish writing whatever data we write.
1946          */
1947         update_time_for_write(inode);
1948
1949         start_pos = round_down(pos, fs_info->sectorsize);
1950         oldsize = i_size_read(inode);
1951         if (start_pos > oldsize) {
1952                 /* Expand hole size to cover write data, preventing empty gap */
1953                 end_pos = round_up(pos + count,
1954                                    fs_info->sectorsize);
1955                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1956                 if (err) {
1957                         inode_unlock(inode);
1958                         goto out;
1959                 }
1960                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1961                         clean_page = 1;
1962         }
1963
1964         if (sync)
1965                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1966
1967         if (iocb->ki_flags & IOCB_DIRECT) {
1968                 num_written = __btrfs_direct_write(iocb, from);
1969         } else {
1970                 num_written = btrfs_buffered_write(iocb, from);
1971                 if (num_written > 0)
1972                         iocb->ki_pos = pos + num_written;
1973                 if (clean_page)
1974                         pagecache_isize_extended(inode, oldsize,
1975                                                 i_size_read(inode));
1976         }
1977
1978         inode_unlock(inode);
1979
1980         /*
1981          * We also have to set last_sub_trans to the current log transid,
1982          * otherwise subsequent syncs to a file that's been synced in this
1983          * transaction will appear to have already occurred.
1984          */
1985         spin_lock(&BTRFS_I(inode)->lock);
1986         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1987         spin_unlock(&BTRFS_I(inode)->lock);
1988         if (num_written > 0)
1989                 num_written = generic_write_sync(iocb, num_written);
1990
1991         if (sync)
1992                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1993 out:
1994         current->backing_dev_info = NULL;
1995         return num_written ? num_written : err;
1996 }
1997
1998 int btrfs_release_file(struct inode *inode, struct file *filp)
1999 {
2000         struct btrfs_file_private *private = filp->private_data;
2001
2002         if (private && private->filldir_buf)
2003                 kfree(private->filldir_buf);
2004         kfree(private);
2005         filp->private_data = NULL;
2006
2007         /*
2008          * ordered_data_close is set by settattr when we are about to truncate
2009          * a file from a non-zero size to a zero size.  This tries to
2010          * flush down new bytes that may have been written if the
2011          * application were using truncate to replace a file in place.
2012          */
2013         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2014                                &BTRFS_I(inode)->runtime_flags))
2015                         filemap_flush(inode->i_mapping);
2016         return 0;
2017 }
2018
2019 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2020 {
2021         int ret;
2022         struct blk_plug plug;
2023
2024         /*
2025          * This is only called in fsync, which would do synchronous writes, so
2026          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2027          * multiple disks using raid profile, a large IO can be split to
2028          * several segments of stripe length (currently 64K).
2029          */
2030         blk_start_plug(&plug);
2031         atomic_inc(&BTRFS_I(inode)->sync_writers);
2032         ret = btrfs_fdatawrite_range(inode, start, end);
2033         atomic_dec(&BTRFS_I(inode)->sync_writers);
2034         blk_finish_plug(&plug);
2035
2036         return ret;
2037 }
2038
2039 /*
2040  * fsync call for both files and directories.  This logs the inode into
2041  * the tree log instead of forcing full commits whenever possible.
2042  *
2043  * It needs to call filemap_fdatawait so that all ordered extent updates are
2044  * in the metadata btree are up to date for copying to the log.
2045  *
2046  * It drops the inode mutex before doing the tree log commit.  This is an
2047  * important optimization for directories because holding the mutex prevents
2048  * new operations on the dir while we write to disk.
2049  */
2050 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2051 {
2052         struct dentry *dentry = file_dentry(file);
2053         struct inode *inode = d_inode(dentry);
2054         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2055         struct btrfs_root *root = BTRFS_I(inode)->root;
2056         struct btrfs_trans_handle *trans;
2057         struct btrfs_log_ctx ctx;
2058         int ret = 0, err;
2059         u64 len;
2060
2061         /*
2062          * The range length can be represented by u64, we have to do the typecasts
2063          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2064          */
2065         len = (u64)end - (u64)start + 1;
2066         trace_btrfs_sync_file(file, datasync);
2067
2068         btrfs_init_log_ctx(&ctx, inode);
2069
2070         /*
2071          * We write the dirty pages in the range and wait until they complete
2072          * out of the ->i_mutex. If so, we can flush the dirty pages by
2073          * multi-task, and make the performance up.  See
2074          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2075          */
2076         ret = start_ordered_ops(inode, start, end);
2077         if (ret)
2078                 goto out;
2079
2080         inode_lock(inode);
2081
2082         /*
2083          * We take the dio_sem here because the tree log stuff can race with
2084          * lockless dio writes and get an extent map logged for an extent we
2085          * never waited on.  We need it this high up for lockdep reasons.
2086          */
2087         down_write(&BTRFS_I(inode)->dio_sem);
2088
2089         atomic_inc(&root->log_batch);
2090
2091         /*
2092          * Before we acquired the inode's lock, someone may have dirtied more
2093          * pages in the target range. We need to make sure that writeback for
2094          * any such pages does not start while we are logging the inode, because
2095          * if it does, any of the following might happen when we are not doing a
2096          * full inode sync:
2097          *
2098          * 1) We log an extent after its writeback finishes but before its
2099          *    checksums are added to the csum tree, leading to -EIO errors
2100          *    when attempting to read the extent after a log replay.
2101          *
2102          * 2) We can end up logging an extent before its writeback finishes.
2103          *    Therefore after the log replay we will have a file extent item
2104          *    pointing to an unwritten extent (and no data checksums as well).
2105          *
2106          * So trigger writeback for any eventual new dirty pages and then we
2107          * wait for all ordered extents to complete below.
2108          */
2109         ret = start_ordered_ops(inode, start, end);
2110         if (ret) {
2111                 inode_unlock(inode);
2112                 goto out;
2113         }
2114
2115         /*
2116          * We have to do this here to avoid the priority inversion of waiting on
2117          * IO of a lower priority task while holding a transaciton open.
2118          */
2119         ret = btrfs_wait_ordered_range(inode, start, len);
2120         if (ret) {
2121                 up_write(&BTRFS_I(inode)->dio_sem);
2122                 inode_unlock(inode);
2123                 goto out;
2124         }
2125         atomic_inc(&root->log_batch);
2126
2127         smp_mb();
2128         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2129             BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2130                 /*
2131                  * We've had everything committed since the last time we were
2132                  * modified so clear this flag in case it was set for whatever
2133                  * reason, it's no longer relevant.
2134                  */
2135                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2136                           &BTRFS_I(inode)->runtime_flags);
2137                 /*
2138                  * An ordered extent might have started before and completed
2139                  * already with io errors, in which case the inode was not
2140                  * updated and we end up here. So check the inode's mapping
2141                  * for any errors that might have happened since we last
2142                  * checked called fsync.
2143                  */
2144                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2145                 up_write(&BTRFS_I(inode)->dio_sem);
2146                 inode_unlock(inode);
2147                 goto out;
2148         }
2149
2150         /*
2151          * We use start here because we will need to wait on the IO to complete
2152          * in btrfs_sync_log, which could require joining a transaction (for
2153          * example checking cross references in the nocow path).  If we use join
2154          * here we could get into a situation where we're waiting on IO to
2155          * happen that is blocked on a transaction trying to commit.  With start
2156          * we inc the extwriter counter, so we wait for all extwriters to exit
2157          * before we start blocking join'ers.  This comment is to keep somebody
2158          * from thinking they are super smart and changing this to
2159          * btrfs_join_transaction *cough*Josef*cough*.
2160          */
2161         trans = btrfs_start_transaction(root, 0);
2162         if (IS_ERR(trans)) {
2163                 ret = PTR_ERR(trans);
2164                 up_write(&BTRFS_I(inode)->dio_sem);
2165                 inode_unlock(inode);
2166                 goto out;
2167         }
2168         trans->sync = true;
2169
2170         ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2171         if (ret < 0) {
2172                 /* Fallthrough and commit/free transaction. */
2173                 ret = 1;
2174         }
2175
2176         /* we've logged all the items and now have a consistent
2177          * version of the file in the log.  It is possible that
2178          * someone will come in and modify the file, but that's
2179          * fine because the log is consistent on disk, and we
2180          * have references to all of the file's extents
2181          *
2182          * It is possible that someone will come in and log the
2183          * file again, but that will end up using the synchronization
2184          * inside btrfs_sync_log to keep things safe.
2185          */
2186         up_write(&BTRFS_I(inode)->dio_sem);
2187         inode_unlock(inode);
2188
2189         /*
2190          * If any of the ordered extents had an error, just return it to user
2191          * space, so that the application knows some writes didn't succeed and
2192          * can take proper action (retry for e.g.). Blindly committing the
2193          * transaction in this case, would fool userspace that everything was
2194          * successful. And we also want to make sure our log doesn't contain
2195          * file extent items pointing to extents that weren't fully written to -
2196          * just like in the non fast fsync path, where we check for the ordered
2197          * operation's error flag before writing to the log tree and return -EIO
2198          * if any of them had this flag set (btrfs_wait_ordered_range) -
2199          * therefore we need to check for errors in the ordered operations,
2200          * which are indicated by ctx.io_err.
2201          */
2202         if (ctx.io_err) {
2203                 btrfs_end_transaction(trans);
2204                 ret = ctx.io_err;
2205                 goto out;
2206         }
2207
2208         if (ret != BTRFS_NO_LOG_SYNC) {
2209                 if (!ret) {
2210                         ret = btrfs_sync_log(trans, root, &ctx);
2211                         if (!ret) {
2212                                 ret = btrfs_end_transaction(trans);
2213                                 goto out;
2214                         }
2215                 }
2216                 ret = btrfs_commit_transaction(trans);
2217         } else {
2218                 ret = btrfs_end_transaction(trans);
2219         }
2220 out:
2221         ASSERT(list_empty(&ctx.list));
2222         err = file_check_and_advance_wb_err(file);
2223         if (!ret)
2224                 ret = err;
2225         return ret > 0 ? -EIO : ret;
2226 }
2227
2228 static const struct vm_operations_struct btrfs_file_vm_ops = {
2229         .fault          = filemap_fault,
2230         .map_pages      = filemap_map_pages,
2231         .page_mkwrite   = btrfs_page_mkwrite,
2232 };
2233
2234 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2235 {
2236         struct address_space *mapping = filp->f_mapping;
2237
2238         if (!mapping->a_ops->readpage)
2239                 return -ENOEXEC;
2240
2241         file_accessed(filp);
2242         vma->vm_ops = &btrfs_file_vm_ops;
2243
2244         return 0;
2245 }
2246
2247 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2248                           int slot, u64 start, u64 end)
2249 {
2250         struct btrfs_file_extent_item *fi;
2251         struct btrfs_key key;
2252
2253         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2254                 return 0;
2255
2256         btrfs_item_key_to_cpu(leaf, &key, slot);
2257         if (key.objectid != btrfs_ino(inode) ||
2258             key.type != BTRFS_EXTENT_DATA_KEY)
2259                 return 0;
2260
2261         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2262
2263         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2264                 return 0;
2265
2266         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2267                 return 0;
2268
2269         if (key.offset == end)
2270                 return 1;
2271         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2272                 return 1;
2273         return 0;
2274 }
2275
2276 static int fill_holes(struct btrfs_trans_handle *trans,
2277                 struct btrfs_inode *inode,
2278                 struct btrfs_path *path, u64 offset, u64 end)
2279 {
2280         struct btrfs_fs_info *fs_info = trans->fs_info;
2281         struct btrfs_root *root = inode->root;
2282         struct extent_buffer *leaf;
2283         struct btrfs_file_extent_item *fi;
2284         struct extent_map *hole_em;
2285         struct extent_map_tree *em_tree = &inode->extent_tree;
2286         struct btrfs_key key;
2287         int ret;
2288
2289         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2290                 goto out;
2291
2292         key.objectid = btrfs_ino(inode);
2293         key.type = BTRFS_EXTENT_DATA_KEY;
2294         key.offset = offset;
2295
2296         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2297         if (ret <= 0) {
2298                 /*
2299                  * We should have dropped this offset, so if we find it then
2300                  * something has gone horribly wrong.
2301                  */
2302                 if (ret == 0)
2303                         ret = -EINVAL;
2304                 return ret;
2305         }
2306
2307         leaf = path->nodes[0];
2308         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2309                 u64 num_bytes;
2310
2311                 path->slots[0]--;
2312                 fi = btrfs_item_ptr(leaf, path->slots[0],
2313                                     struct btrfs_file_extent_item);
2314                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2315                         end - offset;
2316                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2317                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2318                 btrfs_set_file_extent_offset(leaf, fi, 0);
2319                 btrfs_mark_buffer_dirty(leaf);
2320                 goto out;
2321         }
2322
2323         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2324                 u64 num_bytes;
2325
2326                 key.offset = offset;
2327                 btrfs_set_item_key_safe(fs_info, path, &key);
2328                 fi = btrfs_item_ptr(leaf, path->slots[0],
2329                                     struct btrfs_file_extent_item);
2330                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2331                         offset;
2332                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2333                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2334                 btrfs_set_file_extent_offset(leaf, fi, 0);
2335                 btrfs_mark_buffer_dirty(leaf);
2336                 goto out;
2337         }
2338         btrfs_release_path(path);
2339
2340         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2341                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2342         if (ret)
2343                 return ret;
2344
2345 out:
2346         btrfs_release_path(path);
2347
2348         hole_em = alloc_extent_map();
2349         if (!hole_em) {
2350                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2351                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2352         } else {
2353                 hole_em->start = offset;
2354                 hole_em->len = end - offset;
2355                 hole_em->ram_bytes = hole_em->len;
2356                 hole_em->orig_start = offset;
2357
2358                 hole_em->block_start = EXTENT_MAP_HOLE;
2359                 hole_em->block_len = 0;
2360                 hole_em->orig_block_len = 0;
2361                 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2362                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2363                 hole_em->generation = trans->transid;
2364
2365                 do {
2366                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2367                         write_lock(&em_tree->lock);
2368                         ret = add_extent_mapping(em_tree, hole_em, 1);
2369                         write_unlock(&em_tree->lock);
2370                 } while (ret == -EEXIST);
2371                 free_extent_map(hole_em);
2372                 if (ret)
2373                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2374                                         &inode->runtime_flags);
2375         }
2376
2377         return 0;
2378 }
2379
2380 /*
2381  * Find a hole extent on given inode and change start/len to the end of hole
2382  * extent.(hole/vacuum extent whose em->start <= start &&
2383  *         em->start + em->len > start)
2384  * When a hole extent is found, return 1 and modify start/len.
2385  */
2386 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2387 {
2388         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2389         struct extent_map *em;
2390         int ret = 0;
2391
2392         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2393                               round_down(*start, fs_info->sectorsize),
2394                               round_up(*len, fs_info->sectorsize), 0);
2395         if (IS_ERR(em))
2396                 return PTR_ERR(em);
2397
2398         /* Hole or vacuum extent(only exists in no-hole mode) */
2399         if (em->block_start == EXTENT_MAP_HOLE) {
2400                 ret = 1;
2401                 *len = em->start + em->len > *start + *len ?
2402                        0 : *start + *len - em->start - em->len;
2403                 *start = em->start + em->len;
2404         }
2405         free_extent_map(em);
2406         return ret;
2407 }
2408
2409 static int btrfs_punch_hole_lock_range(struct inode *inode,
2410                                        const u64 lockstart,
2411                                        const u64 lockend,
2412                                        struct extent_state **cached_state)
2413 {
2414         while (1) {
2415                 struct btrfs_ordered_extent *ordered;
2416                 int ret;
2417
2418                 truncate_pagecache_range(inode, lockstart, lockend);
2419
2420                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2421                                  cached_state);
2422                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2423
2424                 /*
2425                  * We need to make sure we have no ordered extents in this range
2426                  * and nobody raced in and read a page in this range, if we did
2427                  * we need to try again.
2428                  */
2429                 if ((!ordered ||
2430                     (ordered->file_offset + ordered->len <= lockstart ||
2431                      ordered->file_offset > lockend)) &&
2432                      !filemap_range_has_page(inode->i_mapping,
2433                                              lockstart, lockend)) {
2434                         if (ordered)
2435                                 btrfs_put_ordered_extent(ordered);
2436                         break;
2437                 }
2438                 if (ordered)
2439                         btrfs_put_ordered_extent(ordered);
2440                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2441                                      lockend, cached_state);
2442                 ret = btrfs_wait_ordered_range(inode, lockstart,
2443                                                lockend - lockstart + 1);
2444                 if (ret)
2445                         return ret;
2446         }
2447         return 0;
2448 }
2449
2450 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2451 {
2452         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2453         struct btrfs_root *root = BTRFS_I(inode)->root;
2454         struct extent_state *cached_state = NULL;
2455         struct btrfs_path *path;
2456         struct btrfs_block_rsv *rsv;
2457         struct btrfs_trans_handle *trans;
2458         u64 lockstart;
2459         u64 lockend;
2460         u64 tail_start;
2461         u64 tail_len;
2462         u64 orig_start = offset;
2463         u64 cur_offset;
2464         u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2465         u64 drop_end;
2466         int ret = 0;
2467         int err = 0;
2468         unsigned int rsv_count;
2469         bool same_block;
2470         bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2471         u64 ino_size;
2472         bool truncated_block = false;
2473         bool updated_inode = false;
2474
2475         ret = btrfs_wait_ordered_range(inode, offset, len);
2476         if (ret)
2477                 return ret;
2478
2479         inode_lock(inode);
2480         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2481         ret = find_first_non_hole(inode, &offset, &len);
2482         if (ret < 0)
2483                 goto out_only_mutex;
2484         if (ret && !len) {
2485                 /* Already in a large hole */
2486                 ret = 0;
2487                 goto out_only_mutex;
2488         }
2489
2490         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2491         lockend = round_down(offset + len,
2492                              btrfs_inode_sectorsize(inode)) - 1;
2493         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2494                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2495         /*
2496          * We needn't truncate any block which is beyond the end of the file
2497          * because we are sure there is no data there.
2498          */
2499         /*
2500          * Only do this if we are in the same block and we aren't doing the
2501          * entire block.
2502          */
2503         if (same_block && len < fs_info->sectorsize) {
2504                 if (offset < ino_size) {
2505                         truncated_block = true;
2506                         ret = btrfs_truncate_block(inode, offset, len, 0);
2507                 } else {
2508                         ret = 0;
2509                 }
2510                 goto out_only_mutex;
2511         }
2512
2513         /* zero back part of the first block */
2514         if (offset < ino_size) {
2515                 truncated_block = true;
2516                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2517                 if (ret) {
2518                         inode_unlock(inode);
2519                         return ret;
2520                 }
2521         }
2522
2523         /* Check the aligned pages after the first unaligned page,
2524          * if offset != orig_start, which means the first unaligned page
2525          * including several following pages are already in holes,
2526          * the extra check can be skipped */
2527         if (offset == orig_start) {
2528                 /* after truncate page, check hole again */
2529                 len = offset + len - lockstart;
2530                 offset = lockstart;
2531                 ret = find_first_non_hole(inode, &offset, &len);
2532                 if (ret < 0)
2533                         goto out_only_mutex;
2534                 if (ret && !len) {
2535                         ret = 0;
2536                         goto out_only_mutex;
2537                 }
2538                 lockstart = offset;
2539         }
2540
2541         /* Check the tail unaligned part is in a hole */
2542         tail_start = lockend + 1;
2543         tail_len = offset + len - tail_start;
2544         if (tail_len) {
2545                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2546                 if (unlikely(ret < 0))
2547                         goto out_only_mutex;
2548                 if (!ret) {
2549                         /* zero the front end of the last page */
2550                         if (tail_start + tail_len < ino_size) {
2551                                 truncated_block = true;
2552                                 ret = btrfs_truncate_block(inode,
2553                                                         tail_start + tail_len,
2554                                                         0, 1);
2555                                 if (ret)
2556                                         goto out_only_mutex;
2557                         }
2558                 }
2559         }
2560
2561         if (lockend < lockstart) {
2562                 ret = 0;
2563                 goto out_only_mutex;
2564         }
2565
2566         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2567                                           &cached_state);
2568         if (ret) {
2569                 inode_unlock(inode);
2570                 goto out_only_mutex;
2571         }
2572
2573         path = btrfs_alloc_path();
2574         if (!path) {
2575                 ret = -ENOMEM;
2576                 goto out;
2577         }
2578
2579         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2580         if (!rsv) {
2581                 ret = -ENOMEM;
2582                 goto out_free;
2583         }
2584         rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2585         rsv->failfast = 1;
2586
2587         /*
2588          * 1 - update the inode
2589          * 1 - removing the extents in the range
2590          * 1 - adding the hole extent if no_holes isn't set
2591          */
2592         rsv_count = no_holes ? 2 : 3;
2593         trans = btrfs_start_transaction(root, rsv_count);
2594         if (IS_ERR(trans)) {
2595                 err = PTR_ERR(trans);
2596                 goto out_free;
2597         }
2598
2599         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2600                                       min_size, false);
2601         BUG_ON(ret);
2602         trans->block_rsv = rsv;
2603
2604         cur_offset = lockstart;
2605         len = lockend - cur_offset;
2606         while (cur_offset < lockend) {
2607                 ret = __btrfs_drop_extents(trans, root, inode, path,
2608                                            cur_offset, lockend + 1,
2609                                            &drop_end, 1, 0, 0, NULL);
2610                 if (ret != -ENOSPC)
2611                         break;
2612
2613                 trans->block_rsv = &fs_info->trans_block_rsv;
2614
2615                 if (cur_offset < drop_end && cur_offset < ino_size) {
2616                         ret = fill_holes(trans, BTRFS_I(inode), path,
2617                                         cur_offset, drop_end);
2618                         if (ret) {
2619                                 /*
2620                                  * If we failed then we didn't insert our hole
2621                                  * entries for the area we dropped, so now the
2622                                  * fs is corrupted, so we must abort the
2623                                  * transaction.
2624                                  */
2625                                 btrfs_abort_transaction(trans, ret);
2626                                 err = ret;
2627                                 break;
2628                         }
2629                 }
2630
2631                 cur_offset = drop_end;
2632
2633                 ret = btrfs_update_inode(trans, root, inode);
2634                 if (ret) {
2635                         err = ret;
2636                         break;
2637                 }
2638
2639                 btrfs_end_transaction(trans);
2640                 btrfs_btree_balance_dirty(fs_info);
2641
2642                 trans = btrfs_start_transaction(root, rsv_count);
2643                 if (IS_ERR(trans)) {
2644                         ret = PTR_ERR(trans);
2645                         trans = NULL;
2646                         break;
2647                 }
2648
2649                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2650                                               rsv, min_size, false);
2651                 BUG_ON(ret);    /* shouldn't happen */
2652                 trans->block_rsv = rsv;
2653
2654                 ret = find_first_non_hole(inode, &cur_offset, &len);
2655                 if (unlikely(ret < 0))
2656                         break;
2657                 if (ret && !len) {
2658                         ret = 0;
2659                         break;
2660                 }
2661         }
2662
2663         if (ret) {
2664                 err = ret;
2665                 goto out_trans;
2666         }
2667
2668         trans->block_rsv = &fs_info->trans_block_rsv;
2669         /*
2670          * If we are using the NO_HOLES feature we might have had already an
2671          * hole that overlaps a part of the region [lockstart, lockend] and
2672          * ends at (or beyond) lockend. Since we have no file extent items to
2673          * represent holes, drop_end can be less than lockend and so we must
2674          * make sure we have an extent map representing the existing hole (the
2675          * call to __btrfs_drop_extents() might have dropped the existing extent
2676          * map representing the existing hole), otherwise the fast fsync path
2677          * will not record the existence of the hole region
2678          * [existing_hole_start, lockend].
2679          */
2680         if (drop_end <= lockend)
2681                 drop_end = lockend + 1;
2682         /*
2683          * Don't insert file hole extent item if it's for a range beyond eof
2684          * (because it's useless) or if it represents a 0 bytes range (when
2685          * cur_offset == drop_end).
2686          */
2687         if (cur_offset < ino_size && cur_offset < drop_end) {
2688                 ret = fill_holes(trans, BTRFS_I(inode), path,
2689                                 cur_offset, drop_end);
2690                 if (ret) {
2691                         /* Same comment as above. */
2692                         btrfs_abort_transaction(trans, ret);
2693                         err = ret;
2694                         goto out_trans;
2695                 }
2696         }
2697
2698 out_trans:
2699         if (!trans)
2700                 goto out_free;
2701
2702         inode_inc_iversion(inode);
2703         inode->i_mtime = inode->i_ctime = current_time(inode);
2704
2705         trans->block_rsv = &fs_info->trans_block_rsv;
2706         ret = btrfs_update_inode(trans, root, inode);
2707         updated_inode = true;
2708         btrfs_end_transaction(trans);
2709         btrfs_btree_balance_dirty(fs_info);
2710 out_free:
2711         btrfs_free_path(path);
2712         btrfs_free_block_rsv(fs_info, rsv);
2713 out:
2714         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2715                              &cached_state);
2716 out_only_mutex:
2717         if (!updated_inode && truncated_block && !ret && !err) {
2718                 /*
2719                  * If we only end up zeroing part of a page, we still need to
2720                  * update the inode item, so that all the time fields are
2721                  * updated as well as the necessary btrfs inode in memory fields
2722                  * for detecting, at fsync time, if the inode isn't yet in the
2723                  * log tree or it's there but not up to date.
2724                  */
2725                 trans = btrfs_start_transaction(root, 1);
2726                 if (IS_ERR(trans)) {
2727                         err = PTR_ERR(trans);
2728                 } else {
2729                         err = btrfs_update_inode(trans, root, inode);
2730                         ret = btrfs_end_transaction(trans);
2731                 }
2732         }
2733         inode_unlock(inode);
2734         if (ret && !err)
2735                 err = ret;
2736         return err;
2737 }
2738
2739 /* Helper structure to record which range is already reserved */
2740 struct falloc_range {
2741         struct list_head list;
2742         u64 start;
2743         u64 len;
2744 };
2745
2746 /*
2747  * Helper function to add falloc range
2748  *
2749  * Caller should have locked the larger range of extent containing
2750  * [start, len)
2751  */
2752 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2753 {
2754         struct falloc_range *prev = NULL;
2755         struct falloc_range *range = NULL;
2756
2757         if (list_empty(head))
2758                 goto insert;
2759
2760         /*
2761          * As fallocate iterate by bytenr order, we only need to check
2762          * the last range.
2763          */
2764         prev = list_entry(head->prev, struct falloc_range, list);
2765         if (prev->start + prev->len == start) {
2766                 prev->len += len;
2767                 return 0;
2768         }
2769 insert:
2770         range = kmalloc(sizeof(*range), GFP_KERNEL);
2771         if (!range)
2772                 return -ENOMEM;
2773         range->start = start;
2774         range->len = len;
2775         list_add_tail(&range->list, head);
2776         return 0;
2777 }
2778
2779 static int btrfs_fallocate_update_isize(struct inode *inode,
2780                                         const u64 end,
2781                                         const int mode)
2782 {
2783         struct btrfs_trans_handle *trans;
2784         struct btrfs_root *root = BTRFS_I(inode)->root;
2785         int ret;
2786         int ret2;
2787
2788         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2789                 return 0;
2790
2791         trans = btrfs_start_transaction(root, 1);
2792         if (IS_ERR(trans))
2793                 return PTR_ERR(trans);
2794
2795         inode->i_ctime = current_time(inode);
2796         i_size_write(inode, end);
2797         btrfs_ordered_update_i_size(inode, end, NULL);
2798         ret = btrfs_update_inode(trans, root, inode);
2799         ret2 = btrfs_end_transaction(trans);
2800
2801         return ret ? ret : ret2;
2802 }
2803
2804 enum {
2805         RANGE_BOUNDARY_WRITTEN_EXTENT = 0,
2806         RANGE_BOUNDARY_PREALLOC_EXTENT = 1,
2807         RANGE_BOUNDARY_HOLE = 2,
2808 };
2809
2810 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2811                                                  u64 offset)
2812 {
2813         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2814         struct extent_map *em;
2815         int ret;
2816
2817         offset = round_down(offset, sectorsize);
2818         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2819         if (IS_ERR(em))
2820                 return PTR_ERR(em);
2821
2822         if (em->block_start == EXTENT_MAP_HOLE)
2823                 ret = RANGE_BOUNDARY_HOLE;
2824         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2825                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2826         else
2827                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2828
2829         free_extent_map(em);
2830         return ret;
2831 }
2832
2833 static int btrfs_zero_range(struct inode *inode,
2834                             loff_t offset,
2835                             loff_t len,
2836                             const int mode)
2837 {
2838         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2839         struct extent_map *em;
2840         struct extent_changeset *data_reserved = NULL;
2841         int ret;
2842         u64 alloc_hint = 0;
2843         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2844         u64 alloc_start = round_down(offset, sectorsize);
2845         u64 alloc_end = round_up(offset + len, sectorsize);
2846         u64 bytes_to_reserve = 0;
2847         bool space_reserved = false;
2848
2849         inode_dio_wait(inode);
2850
2851         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2852                               alloc_start, alloc_end - alloc_start, 0);
2853         if (IS_ERR(em)) {
2854                 ret = PTR_ERR(em);
2855                 goto out;
2856         }
2857
2858         /*
2859          * Avoid hole punching and extent allocation for some cases. More cases
2860          * could be considered, but these are unlikely common and we keep things
2861          * as simple as possible for now. Also, intentionally, if the target
2862          * range contains one or more prealloc extents together with regular
2863          * extents and holes, we drop all the existing extents and allocate a
2864          * new prealloc extent, so that we get a larger contiguous disk extent.
2865          */
2866         if (em->start <= alloc_start &&
2867             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2868                 const u64 em_end = em->start + em->len;
2869
2870                 if (em_end >= offset + len) {
2871                         /*
2872                          * The whole range is already a prealloc extent,
2873                          * do nothing except updating the inode's i_size if
2874                          * needed.
2875                          */
2876                         free_extent_map(em);
2877                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2878                                                            mode);
2879                         goto out;
2880                 }
2881                 /*
2882                  * Part of the range is already a prealloc extent, so operate
2883                  * only on the remaining part of the range.
2884                  */
2885                 alloc_start = em_end;
2886                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2887                 len = offset + len - alloc_start;
2888                 offset = alloc_start;
2889                 alloc_hint = em->block_start + em->len;
2890         }
2891         free_extent_map(em);
2892
2893         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2894             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2895                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2896                                       alloc_start, sectorsize, 0);
2897                 if (IS_ERR(em)) {
2898                         ret = PTR_ERR(em);
2899                         goto out;
2900                 }
2901
2902                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2903                         free_extent_map(em);
2904                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2905                                                            mode);
2906                         goto out;
2907                 }
2908                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2909                         free_extent_map(em);
2910                         ret = btrfs_truncate_block(inode, offset, len, 0);
2911                         if (!ret)
2912                                 ret = btrfs_fallocate_update_isize(inode,
2913                                                                    offset + len,
2914                                                                    mode);
2915                         return ret;
2916                 }
2917                 free_extent_map(em);
2918                 alloc_start = round_down(offset, sectorsize);
2919                 alloc_end = alloc_start + sectorsize;
2920                 goto reserve_space;
2921         }
2922
2923         alloc_start = round_up(offset, sectorsize);
2924         alloc_end = round_down(offset + len, sectorsize);
2925
2926         /*
2927          * For unaligned ranges, check the pages at the boundaries, they might
2928          * map to an extent, in which case we need to partially zero them, or
2929          * they might map to a hole, in which case we need our allocation range
2930          * to cover them.
2931          */
2932         if (!IS_ALIGNED(offset, sectorsize)) {
2933                 ret = btrfs_zero_range_check_range_boundary(inode, offset);
2934                 if (ret < 0)
2935                         goto out;
2936                 if (ret == RANGE_BOUNDARY_HOLE) {
2937                         alloc_start = round_down(offset, sectorsize);
2938                         ret = 0;
2939                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2940                         ret = btrfs_truncate_block(inode, offset, 0, 0);
2941                         if (ret)
2942                                 goto out;
2943                 } else {
2944                         ret = 0;
2945                 }
2946         }
2947
2948         if (!IS_ALIGNED(offset + len, sectorsize)) {
2949                 ret = btrfs_zero_range_check_range_boundary(inode,
2950                                                             offset + len);
2951                 if (ret < 0)
2952                         goto out;
2953                 if (ret == RANGE_BOUNDARY_HOLE) {
2954                         alloc_end = round_up(offset + len, sectorsize);
2955                         ret = 0;
2956                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2957                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
2958                         if (ret)
2959                                 goto out;
2960                 } else {
2961                         ret = 0;
2962                 }
2963         }
2964
2965 reserve_space:
2966         if (alloc_start < alloc_end) {
2967                 struct extent_state *cached_state = NULL;
2968                 const u64 lockstart = alloc_start;
2969                 const u64 lockend = alloc_end - 1;
2970
2971                 bytes_to_reserve = alloc_end - alloc_start;
2972                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2973                                                       bytes_to_reserve);
2974                 if (ret < 0)
2975                         goto out;
2976                 space_reserved = true;
2977                 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
2978                                                 alloc_start, bytes_to_reserve);
2979                 if (ret)
2980                         goto out;
2981                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2982                                                   &cached_state);
2983                 if (ret)
2984                         goto out;
2985                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
2986                                                 alloc_end - alloc_start,
2987                                                 i_blocksize(inode),
2988                                                 offset + len, &alloc_hint);
2989                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2990                                      lockend, &cached_state);
2991                 /* btrfs_prealloc_file_range releases reserved space on error */
2992                 if (ret) {
2993                         space_reserved = false;
2994                         goto out;
2995                 }
2996         }
2997         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
2998  out:
2999         if (ret && space_reserved)
3000                 btrfs_free_reserved_data_space(inode, data_reserved,
3001                                                alloc_start, bytes_to_reserve);
3002         extent_changeset_free(data_reserved);
3003
3004         return ret;
3005 }
3006
3007 static long btrfs_fallocate(struct file *file, int mode,
3008                             loff_t offset, loff_t len)
3009 {
3010         struct inode *inode = file_inode(file);
3011         struct extent_state *cached_state = NULL;
3012         struct extent_changeset *data_reserved = NULL;
3013         struct falloc_range *range;
3014         struct falloc_range *tmp;
3015         struct list_head reserve_list;
3016         u64 cur_offset;
3017         u64 last_byte;
3018         u64 alloc_start;
3019         u64 alloc_end;
3020         u64 alloc_hint = 0;
3021         u64 locked_end;
3022         u64 actual_end = 0;
3023         struct extent_map *em;
3024         int blocksize = btrfs_inode_sectorsize(inode);
3025         int ret;
3026
3027         alloc_start = round_down(offset, blocksize);
3028         alloc_end = round_up(offset + len, blocksize);
3029         cur_offset = alloc_start;
3030
3031         /* Make sure we aren't being give some crap mode */
3032         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3033                      FALLOC_FL_ZERO_RANGE))
3034                 return -EOPNOTSUPP;
3035
3036         if (mode & FALLOC_FL_PUNCH_HOLE)
3037                 return btrfs_punch_hole(inode, offset, len);
3038
3039         /*
3040          * Only trigger disk allocation, don't trigger qgroup reserve
3041          *
3042          * For qgroup space, it will be checked later.
3043          */
3044         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3045                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3046                                                       alloc_end - alloc_start);
3047                 if (ret < 0)
3048                         return ret;
3049         }
3050
3051         inode_lock(inode);
3052
3053         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3054                 ret = inode_newsize_ok(inode, offset + len);
3055                 if (ret)
3056                         goto out;
3057         }
3058
3059         /*
3060     &nb