Loading...
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#include "delalloc-space.h"
30
31static struct kmem_cache *btrfs_inode_defrag_cachep;
32/*
33 * when auto defrag is enabled we
34 * queue up these defrag structs to remember which
35 * inodes need defragging passes
36 */
37struct inode_defrag {
38 struct rb_node rb_node;
39 /* objectid */
40 u64 ino;
41 /*
42 * transid where the defrag was added, we search for
43 * extents newer than this
44 */
45 u64 transid;
46
47 /* root objectid */
48 u64 root;
49
50 /* last offset we were able to defrag */
51 u64 last_offset;
52
53 /* if we've wrapped around back to zero once already */
54 int cycled;
55};
56
57static int __compare_inode_defrag(struct inode_defrag *defrag1,
58 struct inode_defrag *defrag2)
59{
60 if (defrag1->root > defrag2->root)
61 return 1;
62 else if (defrag1->root < defrag2->root)
63 return -1;
64 else if (defrag1->ino > defrag2->ino)
65 return 1;
66 else if (defrag1->ino < defrag2->ino)
67 return -1;
68 else
69 return 0;
70}
71
72/* pop a record for an inode into the defrag tree. The lock
73 * must be held already
74 *
75 * If you're inserting a record for an older transid than an
76 * existing record, the transid already in the tree is lowered
77 *
78 * If an existing record is found the defrag item you
79 * pass in is freed
80 */
81static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
82 struct inode_defrag *defrag)
83{
84 struct btrfs_fs_info *fs_info = inode->root->fs_info;
85 struct inode_defrag *entry;
86 struct rb_node **p;
87 struct rb_node *parent = NULL;
88 int ret;
89
90 p = &fs_info->defrag_inodes.rb_node;
91 while (*p) {
92 parent = *p;
93 entry = rb_entry(parent, struct inode_defrag, rb_node);
94
95 ret = __compare_inode_defrag(defrag, entry);
96 if (ret < 0)
97 p = &parent->rb_left;
98 else if (ret > 0)
99 p = &parent->rb_right;
100 else {
101 /* if we're reinserting an entry for
102 * an old defrag run, make sure to
103 * lower the transid of our existing record
104 */
105 if (defrag->transid < entry->transid)
106 entry->transid = defrag->transid;
107 if (defrag->last_offset > entry->last_offset)
108 entry->last_offset = defrag->last_offset;
109 return -EEXIST;
110 }
111 }
112 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
113 rb_link_node(&defrag->rb_node, parent, p);
114 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
115 return 0;
116}
117
118static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
119{
120 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
121 return 0;
122
123 if (btrfs_fs_closing(fs_info))
124 return 0;
125
126 return 1;
127}
128
129/*
130 * insert a defrag record for this inode if auto defrag is
131 * enabled
132 */
133int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
134 struct btrfs_inode *inode)
135{
136 struct btrfs_root *root = inode->root;
137 struct btrfs_fs_info *fs_info = root->fs_info;
138 struct inode_defrag *defrag;
139 u64 transid;
140 int ret;
141
142 if (!__need_auto_defrag(fs_info))
143 return 0;
144
145 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
146 return 0;
147
148 if (trans)
149 transid = trans->transid;
150 else
151 transid = inode->root->last_trans;
152
153 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
154 if (!defrag)
155 return -ENOMEM;
156
157 defrag->ino = btrfs_ino(inode);
158 defrag->transid = transid;
159 defrag->root = root->root_key.objectid;
160
161 spin_lock(&fs_info->defrag_inodes_lock);
162 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
163 /*
164 * If we set IN_DEFRAG flag and evict the inode from memory,
165 * and then re-read this inode, this new inode doesn't have
166 * IN_DEFRAG flag. At the case, we may find the existed defrag.
167 */
168 ret = __btrfs_add_inode_defrag(inode, defrag);
169 if (ret)
170 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
171 } else {
172 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
173 }
174 spin_unlock(&fs_info->defrag_inodes_lock);
175 return 0;
176}
177
178/*
179 * Requeue the defrag object. If there is a defrag object that points to
180 * the same inode in the tree, we will merge them together (by
181 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
182 */
183static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
184 struct inode_defrag *defrag)
185{
186 struct btrfs_fs_info *fs_info = inode->root->fs_info;
187 int ret;
188
189 if (!__need_auto_defrag(fs_info))
190 goto out;
191
192 /*
193 * Here we don't check the IN_DEFRAG flag, because we need merge
194 * them together.
195 */
196 spin_lock(&fs_info->defrag_inodes_lock);
197 ret = __btrfs_add_inode_defrag(inode, defrag);
198 spin_unlock(&fs_info->defrag_inodes_lock);
199 if (ret)
200 goto out;
201 return;
202out:
203 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
204}
205
206/*
207 * pick the defragable inode that we want, if it doesn't exist, we will get
208 * the next one.
209 */
210static struct inode_defrag *
211btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
212{
213 struct inode_defrag *entry = NULL;
214 struct inode_defrag tmp;
215 struct rb_node *p;
216 struct rb_node *parent = NULL;
217 int ret;
218
219 tmp.ino = ino;
220 tmp.root = root;
221
222 spin_lock(&fs_info->defrag_inodes_lock);
223 p = fs_info->defrag_inodes.rb_node;
224 while (p) {
225 parent = p;
226 entry = rb_entry(parent, struct inode_defrag, rb_node);
227
228 ret = __compare_inode_defrag(&tmp, entry);
229 if (ret < 0)
230 p = parent->rb_left;
231 else if (ret > 0)
232 p = parent->rb_right;
233 else
234 goto out;
235 }
236
237 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
238 parent = rb_next(parent);
239 if (parent)
240 entry = rb_entry(parent, struct inode_defrag, rb_node);
241 else
242 entry = NULL;
243 }
244out:
245 if (entry)
246 rb_erase(parent, &fs_info->defrag_inodes);
247 spin_unlock(&fs_info->defrag_inodes_lock);
248 return entry;
249}
250
251void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
252{
253 struct inode_defrag *defrag;
254 struct rb_node *node;
255
256 spin_lock(&fs_info->defrag_inodes_lock);
257 node = rb_first(&fs_info->defrag_inodes);
258 while (node) {
259 rb_erase(node, &fs_info->defrag_inodes);
260 defrag = rb_entry(node, struct inode_defrag, rb_node);
261 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
262
263 cond_resched_lock(&fs_info->defrag_inodes_lock);
264
265 node = rb_first(&fs_info->defrag_inodes);
266 }
267 spin_unlock(&fs_info->defrag_inodes_lock);
268}
269
270#define BTRFS_DEFRAG_BATCH 1024
271
272static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
273 struct inode_defrag *defrag)
274{
275 struct btrfs_root *inode_root;
276 struct inode *inode;
277 struct btrfs_key key;
278 struct btrfs_ioctl_defrag_range_args range;
279 int num_defrag;
280 int index;
281 int ret;
282
283 /* get the inode */
284 key.objectid = defrag->root;
285 key.type = BTRFS_ROOT_ITEM_KEY;
286 key.offset = (u64)-1;
287
288 index = srcu_read_lock(&fs_info->subvol_srcu);
289
290 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
291 if (IS_ERR(inode_root)) {
292 ret = PTR_ERR(inode_root);
293 goto cleanup;
294 }
295
296 key.objectid = defrag->ino;
297 key.type = BTRFS_INODE_ITEM_KEY;
298 key.offset = 0;
299 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
300 if (IS_ERR(inode)) {
301 ret = PTR_ERR(inode);
302 goto cleanup;
303 }
304 srcu_read_unlock(&fs_info->subvol_srcu, index);
305
306 /* do a chunk of defrag */
307 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
308 memset(&range, 0, sizeof(range));
309 range.len = (u64)-1;
310 range.start = defrag->last_offset;
311
312 sb_start_write(fs_info->sb);
313 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
314 BTRFS_DEFRAG_BATCH);
315 sb_end_write(fs_info->sb);
316 /*
317 * if we filled the whole defrag batch, there
318 * must be more work to do. Queue this defrag
319 * again
320 */
321 if (num_defrag == BTRFS_DEFRAG_BATCH) {
322 defrag->last_offset = range.start;
323 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
324 } else if (defrag->last_offset && !defrag->cycled) {
325 /*
326 * we didn't fill our defrag batch, but
327 * we didn't start at zero. Make sure we loop
328 * around to the start of the file.
329 */
330 defrag->last_offset = 0;
331 defrag->cycled = 1;
332 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
333 } else {
334 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
335 }
336
337 iput(inode);
338 return 0;
339cleanup:
340 srcu_read_unlock(&fs_info->subvol_srcu, index);
341 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
342 return ret;
343}
344
345/*
346 * run through the list of inodes in the FS that need
347 * defragging
348 */
349int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
350{
351 struct inode_defrag *defrag;
352 u64 first_ino = 0;
353 u64 root_objectid = 0;
354
355 atomic_inc(&fs_info->defrag_running);
356 while (1) {
357 /* Pause the auto defragger. */
358 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
359 &fs_info->fs_state))
360 break;
361
362 if (!__need_auto_defrag(fs_info))
363 break;
364
365 /* find an inode to defrag */
366 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
367 first_ino);
368 if (!defrag) {
369 if (root_objectid || first_ino) {
370 root_objectid = 0;
371 first_ino = 0;
372 continue;
373 } else {
374 break;
375 }
376 }
377
378 first_ino = defrag->ino + 1;
379 root_objectid = defrag->root;
380
381 __btrfs_run_defrag_inode(fs_info, defrag);
382 }
383 atomic_dec(&fs_info->defrag_running);
384
385 /*
386 * during unmount, we use the transaction_wait queue to
387 * wait for the defragger to stop
388 */
389 wake_up(&fs_info->transaction_wait);
390 return 0;
391}
392
393/* simple helper to fault in pages and copy. This should go away
394 * and be replaced with calls into generic code.
395 */
396static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
397 struct page **prepared_pages,
398 struct iov_iter *i)
399{
400 size_t copied = 0;
401 size_t total_copied = 0;
402 int pg = 0;
403 int offset = offset_in_page(pos);
404
405 while (write_bytes > 0) {
406 size_t count = min_t(size_t,
407 PAGE_SIZE - offset, write_bytes);
408 struct page *page = prepared_pages[pg];
409 /*
410 * Copy data from userspace to the current page
411 */
412 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
413
414 /* Flush processor's dcache for this page */
415 flush_dcache_page(page);
416
417 /*
418 * if we get a partial write, we can end up with
419 * partially up to date pages. These add
420 * a lot of complexity, so make sure they don't
421 * happen by forcing this copy to be retried.
422 *
423 * The rest of the btrfs_file_write code will fall
424 * back to page at a time copies after we return 0.
425 */
426 if (!PageUptodate(page) && copied < count)
427 copied = 0;
428
429 iov_iter_advance(i, copied);
430 write_bytes -= copied;
431 total_copied += copied;
432
433 /* Return to btrfs_file_write_iter to fault page */
434 if (unlikely(copied == 0))
435 break;
436
437 if (copied < PAGE_SIZE - offset) {
438 offset += copied;
439 } else {
440 pg++;
441 offset = 0;
442 }
443 }
444 return total_copied;
445}
446
447/*
448 * unlocks pages after btrfs_file_write is done with them
449 */
450static void btrfs_drop_pages(struct page **pages, size_t num_pages)
451{
452 size_t i;
453 for (i = 0; i < num_pages; i++) {
454 /* page checked is some magic around finding pages that
455 * have been modified without going through btrfs_set_page_dirty
456 * clear it here. There should be no need to mark the pages
457 * accessed as prepare_pages should have marked them accessed
458 * in prepare_pages via find_or_create_page()
459 */
460 ClearPageChecked(pages[i]);
461 unlock_page(pages[i]);
462 put_page(pages[i]);
463 }
464}
465
466static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
467 const u64 start,
468 const u64 len,
469 struct extent_state **cached_state)
470{
471 u64 search_start = start;
472 const u64 end = start + len - 1;
473
474 while (search_start < end) {
475 const u64 search_len = end - search_start + 1;
476 struct extent_map *em;
477 u64 em_len;
478 int ret = 0;
479
480 em = btrfs_get_extent(inode, NULL, 0, search_start,
481 search_len, 0);
482 if (IS_ERR(em))
483 return PTR_ERR(em);
484
485 if (em->block_start != EXTENT_MAP_HOLE)
486 goto next;
487
488 em_len = em->len;
489 if (em->start < search_start)
490 em_len -= search_start - em->start;
491 if (em_len > search_len)
492 em_len = search_len;
493
494 ret = set_extent_bit(&inode->io_tree, search_start,
495 search_start + em_len - 1,
496 EXTENT_DELALLOC_NEW,
497 NULL, cached_state, GFP_NOFS);
498next:
499 search_start = extent_map_end(em);
500 free_extent_map(em);
501 if (ret)
502 return ret;
503 }
504 return 0;
505}
506
507/*
508 * after copy_from_user, pages need to be dirtied and we need to make
509 * sure holes are created between the current EOF and the start of
510 * any next extents (if required).
511 *
512 * this also makes the decision about creating an inline extent vs
513 * doing real data extents, marking pages dirty and delalloc as required.
514 */
515int btrfs_dirty_pages(struct inode *inode, struct page **pages,
516 size_t num_pages, loff_t pos, size_t write_bytes,
517 struct extent_state **cached)
518{
519 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
520 int err = 0;
521 int i;
522 u64 num_bytes;
523 u64 start_pos;
524 u64 end_of_last_block;
525 u64 end_pos = pos + write_bytes;
526 loff_t isize = i_size_read(inode);
527 unsigned int extra_bits = 0;
528
529 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
530 num_bytes = round_up(write_bytes + pos - start_pos,
531 fs_info->sectorsize);
532
533 end_of_last_block = start_pos + num_bytes - 1;
534
535 /*
536 * The pages may have already been dirty, clear out old accounting so
537 * we can set things up properly
538 */
539 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
540 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
541 0, 0, cached);
542
543 if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
544 if (start_pos >= isize &&
545 !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
546 /*
547 * There can't be any extents following eof in this case
548 * so just set the delalloc new bit for the range
549 * directly.
550 */
551 extra_bits |= EXTENT_DELALLOC_NEW;
552 } else {
553 err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
554 start_pos,
555 num_bytes, cached);
556 if (err)
557 return err;
558 }
559 }
560
561 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
562 extra_bits, cached);
563 if (err)
564 return err;
565
566 for (i = 0; i < num_pages; i++) {
567 struct page *p = pages[i];
568 SetPageUptodate(p);
569 ClearPageChecked(p);
570 set_page_dirty(p);
571 }
572
573 /*
574 * we've only changed i_size in ram, and we haven't updated
575 * the disk i_size. There is no need to log the inode
576 * at this time.
577 */
578 if (end_pos > isize)
579 i_size_write(inode, end_pos);
580 return 0;
581}
582
583/*
584 * this drops all the extents in the cache that intersect the range
585 * [start, end]. Existing extents are split as required.
586 */
587void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
588 int skip_pinned)
589{
590 struct extent_map *em;
591 struct extent_map *split = NULL;
592 struct extent_map *split2 = NULL;
593 struct extent_map_tree *em_tree = &inode->extent_tree;
594 u64 len = end - start + 1;
595 u64 gen;
596 int ret;
597 int testend = 1;
598 unsigned long flags;
599 int compressed = 0;
600 bool modified;
601
602 WARN_ON(end < start);
603 if (end == (u64)-1) {
604 len = (u64)-1;
605 testend = 0;
606 }
607 while (1) {
608 int no_splits = 0;
609
610 modified = false;
611 if (!split)
612 split = alloc_extent_map();
613 if (!split2)
614 split2 = alloc_extent_map();
615 if (!split || !split2)
616 no_splits = 1;
617
618 write_lock(&em_tree->lock);
619 em = lookup_extent_mapping(em_tree, start, len);
620 if (!em) {
621 write_unlock(&em_tree->lock);
622 break;
623 }
624 flags = em->flags;
625 gen = em->generation;
626 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
627 if (testend && em->start + em->len >= start + len) {
628 free_extent_map(em);
629 write_unlock(&em_tree->lock);
630 break;
631 }
632 start = em->start + em->len;
633 if (testend)
634 len = start + len - (em->start + em->len);
635 free_extent_map(em);
636 write_unlock(&em_tree->lock);
637 continue;
638 }
639 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
640 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
641 clear_bit(EXTENT_FLAG_LOGGING, &flags);
642 modified = !list_empty(&em->list);
643 if (no_splits)
644 goto next;
645
646 if (em->start < start) {
647 split->start = em->start;
648 split->len = start - em->start;
649
650 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
651 split->orig_start = em->orig_start;
652 split->block_start = em->block_start;
653
654 if (compressed)
655 split->block_len = em->block_len;
656 else
657 split->block_len = split->len;
658 split->orig_block_len = max(split->block_len,
659 em->orig_block_len);
660 split->ram_bytes = em->ram_bytes;
661 } else {
662 split->orig_start = split->start;
663 split->block_len = 0;
664 split->block_start = em->block_start;
665 split->orig_block_len = 0;
666 split->ram_bytes = split->len;
667 }
668
669 split->generation = gen;
670 split->bdev = em->bdev;
671 split->flags = flags;
672 split->compress_type = em->compress_type;
673 replace_extent_mapping(em_tree, em, split, modified);
674 free_extent_map(split);
675 split = split2;
676 split2 = NULL;
677 }
678 if (testend && em->start + em->len > start + len) {
679 u64 diff = start + len - em->start;
680
681 split->start = start + len;
682 split->len = em->start + em->len - (start + len);
683 split->bdev = em->bdev;
684 split->flags = flags;
685 split->compress_type = em->compress_type;
686 split->generation = gen;
687
688 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
689 split->orig_block_len = max(em->block_len,
690 em->orig_block_len);
691
692 split->ram_bytes = em->ram_bytes;
693 if (compressed) {
694 split->block_len = em->block_len;
695 split->block_start = em->block_start;
696 split->orig_start = em->orig_start;
697 } else {
698 split->block_len = split->len;
699 split->block_start = em->block_start
700 + diff;
701 split->orig_start = em->orig_start;
702 }
703 } else {
704 split->ram_bytes = split->len;
705 split->orig_start = split->start;
706 split->block_len = 0;
707 split->block_start = em->block_start;
708 split->orig_block_len = 0;
709 }
710
711 if (extent_map_in_tree(em)) {
712 replace_extent_mapping(em_tree, em, split,
713 modified);
714 } else {
715 ret = add_extent_mapping(em_tree, split,
716 modified);
717 ASSERT(ret == 0); /* Logic error */
718 }
719 free_extent_map(split);
720 split = NULL;
721 }
722next:
723 if (extent_map_in_tree(em))
724 remove_extent_mapping(em_tree, em);
725 write_unlock(&em_tree->lock);
726
727 /* once for us */
728 free_extent_map(em);
729 /* once for the tree*/
730 free_extent_map(em);
731 }
732 if (split)
733 free_extent_map(split);
734 if (split2)
735 free_extent_map(split2);
736}
737
738/*
739 * this is very complex, but the basic idea is to drop all extents
740 * in the range start - end. hint_block is filled in with a block number
741 * that would be a good hint to the block allocator for this file.
742 *
743 * If an extent intersects the range but is not entirely inside the range
744 * it is either truncated or split. Anything entirely inside the range
745 * is deleted from the tree.
746 */
747int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
748 struct btrfs_root *root, struct inode *inode,
749 struct btrfs_path *path, u64 start, u64 end,
750 u64 *drop_end, int drop_cache,
751 int replace_extent,
752 u32 extent_item_size,
753 int *key_inserted)
754{
755 struct btrfs_fs_info *fs_info = root->fs_info;
756 struct extent_buffer *leaf;
757 struct btrfs_file_extent_item *fi;
758 struct btrfs_ref ref = { 0 };
759 struct btrfs_key key;
760 struct btrfs_key new_key;
761 u64 ino = btrfs_ino(BTRFS_I(inode));
762 u64 search_start = start;
763 u64 disk_bytenr = 0;
764 u64 num_bytes = 0;
765 u64 extent_offset = 0;
766 u64 extent_end = 0;
767 u64 last_end = start;
768 int del_nr = 0;
769 int del_slot = 0;
770 int extent_type;
771 int recow;
772 int ret;
773 int modify_tree = -1;
774 int update_refs;
775 int found = 0;
776 int leafs_visited = 0;
777
778 if (drop_cache)
779 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
780
781 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
782 modify_tree = 0;
783
784 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
785 root == fs_info->tree_root);
786 while (1) {
787 recow = 0;
788 ret = btrfs_lookup_file_extent(trans, root, path, ino,
789 search_start, modify_tree);
790 if (ret < 0)
791 break;
792 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
793 leaf = path->nodes[0];
794 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
795 if (key.objectid == ino &&
796 key.type == BTRFS_EXTENT_DATA_KEY)
797 path->slots[0]--;
798 }
799 ret = 0;
800 leafs_visited++;
801next_slot:
802 leaf = path->nodes[0];
803 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
804 BUG_ON(del_nr > 0);
805 ret = btrfs_next_leaf(root, path);
806 if (ret < 0)
807 break;
808 if (ret > 0) {
809 ret = 0;
810 break;
811 }
812 leafs_visited++;
813 leaf = path->nodes[0];
814 recow = 1;
815 }
816
817 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
818
819 if (key.objectid > ino)
820 break;
821 if (WARN_ON_ONCE(key.objectid < ino) ||
822 key.type < BTRFS_EXTENT_DATA_KEY) {
823 ASSERT(del_nr == 0);
824 path->slots[0]++;
825 goto next_slot;
826 }
827 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
828 break;
829
830 fi = btrfs_item_ptr(leaf, path->slots[0],
831 struct btrfs_file_extent_item);
832 extent_type = btrfs_file_extent_type(leaf, fi);
833
834 if (extent_type == BTRFS_FILE_EXTENT_REG ||
835 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
836 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
837 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
838 extent_offset = btrfs_file_extent_offset(leaf, fi);
839 extent_end = key.offset +
840 btrfs_file_extent_num_bytes(leaf, fi);
841 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
842 extent_end = key.offset +
843 btrfs_file_extent_ram_bytes(leaf, fi);
844 } else {
845 /* can't happen */
846 BUG();
847 }
848
849 /*
850 * Don't skip extent items representing 0 byte lengths. They
851 * used to be created (bug) if while punching holes we hit
852 * -ENOSPC condition. So if we find one here, just ensure we
853 * delete it, otherwise we would insert a new file extent item
854 * with the same key (offset) as that 0 bytes length file
855 * extent item in the call to setup_items_for_insert() later
856 * in this function.
857 */
858 if (extent_end == key.offset && extent_end >= search_start) {
859 last_end = extent_end;
860 goto delete_extent_item;
861 }
862
863 if (extent_end <= search_start) {
864 path->slots[0]++;
865 goto next_slot;
866 }
867
868 found = 1;
869 search_start = max(key.offset, start);
870 if (recow || !modify_tree) {
871 modify_tree = -1;
872 btrfs_release_path(path);
873 continue;
874 }
875
876 /*
877 * | - range to drop - |
878 * | -------- extent -------- |
879 */
880 if (start > key.offset && end < extent_end) {
881 BUG_ON(del_nr > 0);
882 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
883 ret = -EOPNOTSUPP;
884 break;
885 }
886
887 memcpy(&new_key, &key, sizeof(new_key));
888 new_key.offset = start;
889 ret = btrfs_duplicate_item(trans, root, path,
890 &new_key);
891 if (ret == -EAGAIN) {
892 btrfs_release_path(path);
893 continue;
894 }
895 if (ret < 0)
896 break;
897
898 leaf = path->nodes[0];
899 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
900 struct btrfs_file_extent_item);
901 btrfs_set_file_extent_num_bytes(leaf, fi,
902 start - key.offset);
903
904 fi = btrfs_item_ptr(leaf, path->slots[0],
905 struct btrfs_file_extent_item);
906
907 extent_offset += start - key.offset;
908 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
909 btrfs_set_file_extent_num_bytes(leaf, fi,
910 extent_end - start);
911 btrfs_mark_buffer_dirty(leaf);
912
913 if (update_refs && disk_bytenr > 0) {
914 btrfs_init_generic_ref(&ref,
915 BTRFS_ADD_DELAYED_REF,
916 disk_bytenr, num_bytes, 0);
917 btrfs_init_data_ref(&ref,
918 root->root_key.objectid,
919 new_key.objectid,
920 start - extent_offset);
921 ret = btrfs_inc_extent_ref(trans, &ref);
922 BUG_ON(ret); /* -ENOMEM */
923 }
924 key.offset = start;
925 }
926 /*
927 * From here on out we will have actually dropped something, so
928 * last_end can be updated.
929 */
930 last_end = extent_end;
931
932 /*
933 * | ---- range to drop ----- |
934 * | -------- extent -------- |
935 */
936 if (start <= key.offset && end < extent_end) {
937 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
938 ret = -EOPNOTSUPP;
939 break;
940 }
941
942 memcpy(&new_key, &key, sizeof(new_key));
943 new_key.offset = end;
944 btrfs_set_item_key_safe(fs_info, path, &new_key);
945
946 extent_offset += end - key.offset;
947 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
948 btrfs_set_file_extent_num_bytes(leaf, fi,
949 extent_end - end);
950 btrfs_mark_buffer_dirty(leaf);
951 if (update_refs && disk_bytenr > 0)
952 inode_sub_bytes(inode, end - key.offset);
953 break;
954 }
955
956 search_start = extent_end;
957 /*
958 * | ---- range to drop ----- |
959 * | -------- extent -------- |
960 */
961 if (start > key.offset && end >= extent_end) {
962 BUG_ON(del_nr > 0);
963 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
964 ret = -EOPNOTSUPP;
965 break;
966 }
967
968 btrfs_set_file_extent_num_bytes(leaf, fi,
969 start - key.offset);
970 btrfs_mark_buffer_dirty(leaf);
971 if (update_refs && disk_bytenr > 0)
972 inode_sub_bytes(inode, extent_end - start);
973 if (end == extent_end)
974 break;
975
976 path->slots[0]++;
977 goto next_slot;
978 }
979
980 /*
981 * | ---- range to drop ----- |
982 * | ------ extent ------ |
983 */
984 if (start <= key.offset && end >= extent_end) {
985delete_extent_item:
986 if (del_nr == 0) {
987 del_slot = path->slots[0];
988 del_nr = 1;
989 } else {
990 BUG_ON(del_slot + del_nr != path->slots[0]);
991 del_nr++;
992 }
993
994 if (update_refs &&
995 extent_type == BTRFS_FILE_EXTENT_INLINE) {
996 inode_sub_bytes(inode,
997 extent_end - key.offset);
998 extent_end = ALIGN(extent_end,
999 fs_info->sectorsize);
1000 } else if (update_refs && disk_bytenr > 0) {
1001 btrfs_init_generic_ref(&ref,
1002 BTRFS_DROP_DELAYED_REF,
1003 disk_bytenr, num_bytes, 0);
1004 btrfs_init_data_ref(&ref,
1005 root->root_key.objectid,
1006 key.objectid,
1007 key.offset - extent_offset);
1008 ret = btrfs_free_extent(trans, &ref);
1009 BUG_ON(ret); /* -ENOMEM */
1010 inode_sub_bytes(inode,
1011 extent_end - key.offset);
1012 }
1013
1014 if (end == extent_end)
1015 break;
1016
1017 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1018 path->slots[0]++;
1019 goto next_slot;
1020 }
1021
1022 ret = btrfs_del_items(trans, root, path, del_slot,
1023 del_nr);
1024 if (ret) {
1025 btrfs_abort_transaction(trans, ret);
1026 break;
1027 }
1028
1029 del_nr = 0;
1030 del_slot = 0;
1031
1032 btrfs_release_path(path);
1033 continue;
1034 }
1035
1036 BUG();
1037 }
1038
1039 if (!ret && del_nr > 0) {
1040 /*
1041 * Set path->slots[0] to first slot, so that after the delete
1042 * if items are move off from our leaf to its immediate left or
1043 * right neighbor leafs, we end up with a correct and adjusted
1044 * path->slots[0] for our insertion (if replace_extent != 0).
1045 */
1046 path->slots[0] = del_slot;
1047 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1048 if (ret)
1049 btrfs_abort_transaction(trans, ret);
1050 }
1051
1052 leaf = path->nodes[0];
1053 /*
1054 * If btrfs_del_items() was called, it might have deleted a leaf, in
1055 * which case it unlocked our path, so check path->locks[0] matches a
1056 * write lock.
1057 */
1058 if (!ret && replace_extent && leafs_visited == 1 &&
1059 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1060 path->locks[0] == BTRFS_WRITE_LOCK) &&
1061 btrfs_leaf_free_space(leaf) >=
1062 sizeof(struct btrfs_item) + extent_item_size) {
1063
1064 key.objectid = ino;
1065 key.type = BTRFS_EXTENT_DATA_KEY;
1066 key.offset = start;
1067 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1068 struct btrfs_key slot_key;
1069
1070 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1071 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1072 path->slots[0]++;
1073 }
1074 setup_items_for_insert(root, path, &key,
1075 &extent_item_size,
1076 extent_item_size,
1077 sizeof(struct btrfs_item) +
1078 extent_item_size, 1);
1079 *key_inserted = 1;
1080 }
1081
1082 if (!replace_extent || !(*key_inserted))
1083 btrfs_release_path(path);
1084 if (drop_end)
1085 *drop_end = found ? min(end, last_end) : end;
1086 return ret;
1087}
1088
1089int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1090 struct btrfs_root *root, struct inode *inode, u64 start,
1091 u64 end, int drop_cache)
1092{
1093 struct btrfs_path *path;
1094 int ret;
1095
1096 path = btrfs_alloc_path();
1097 if (!path)
1098 return -ENOMEM;
1099 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1100 drop_cache, 0, 0, NULL);
1101 btrfs_free_path(path);
1102 return ret;
1103}
1104
1105static int extent_mergeable(struct extent_buffer *leaf, int slot,
1106 u64 objectid, u64 bytenr, u64 orig_offset,
1107 u64 *start, u64 *end)
1108{
1109 struct btrfs_file_extent_item *fi;
1110 struct btrfs_key key;
1111 u64 extent_end;
1112
1113 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1114 return 0;
1115
1116 btrfs_item_key_to_cpu(leaf, &key, slot);
1117 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1118 return 0;
1119
1120 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1121 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1122 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1123 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1124 btrfs_file_extent_compression(leaf, fi) ||
1125 btrfs_file_extent_encryption(leaf, fi) ||
1126 btrfs_file_extent_other_encoding(leaf, fi))
1127 return 0;
1128
1129 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1130 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1131 return 0;
1132
1133 *start = key.offset;
1134 *end = extent_end;
1135 return 1;
1136}
1137
1138/*
1139 * Mark extent in the range start - end as written.
1140 *
1141 * This changes extent type from 'pre-allocated' to 'regular'. If only
1142 * part of extent is marked as written, the extent will be split into
1143 * two or three.
1144 */
1145int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1146 struct btrfs_inode *inode, u64 start, u64 end)
1147{
1148 struct btrfs_fs_info *fs_info = trans->fs_info;
1149 struct btrfs_root *root = inode->root;
1150 struct extent_buffer *leaf;
1151 struct btrfs_path *path;
1152 struct btrfs_file_extent_item *fi;
1153 struct btrfs_ref ref = { 0 };
1154 struct btrfs_key key;
1155 struct btrfs_key new_key;
1156 u64 bytenr;
1157 u64 num_bytes;
1158 u64 extent_end;
1159 u64 orig_offset;
1160 u64 other_start;
1161 u64 other_end;
1162 u64 split;
1163 int del_nr = 0;
1164 int del_slot = 0;
1165 int recow;
1166 int ret;
1167 u64 ino = btrfs_ino(inode);
1168
1169 path = btrfs_alloc_path();
1170 if (!path)
1171 return -ENOMEM;
1172again:
1173 recow = 0;
1174 split = start;
1175 key.objectid = ino;
1176 key.type = BTRFS_EXTENT_DATA_KEY;
1177 key.offset = split;
1178
1179 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1180 if (ret < 0)
1181 goto out;
1182 if (ret > 0 && path->slots[0] > 0)
1183 path->slots[0]--;
1184
1185 leaf = path->nodes[0];
1186 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1187 if (key.objectid != ino ||
1188 key.type != BTRFS_EXTENT_DATA_KEY) {
1189 ret = -EINVAL;
1190 btrfs_abort_transaction(trans, ret);
1191 goto out;
1192 }
1193 fi = btrfs_item_ptr(leaf, path->slots[0],
1194 struct btrfs_file_extent_item);
1195 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1196 ret = -EINVAL;
1197 btrfs_abort_transaction(trans, ret);
1198 goto out;
1199 }
1200 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1201 if (key.offset > start || extent_end < end) {
1202 ret = -EINVAL;
1203 btrfs_abort_transaction(trans, ret);
1204 goto out;
1205 }
1206
1207 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1208 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1209 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1210 memcpy(&new_key, &key, sizeof(new_key));
1211
1212 if (start == key.offset && end < extent_end) {
1213 other_start = 0;
1214 other_end = start;
1215 if (extent_mergeable(leaf, path->slots[0] - 1,
1216 ino, bytenr, orig_offset,
1217 &other_start, &other_end)) {
1218 new_key.offset = end;
1219 btrfs_set_item_key_safe(fs_info, path, &new_key);
1220 fi = btrfs_item_ptr(leaf, path->slots[0],
1221 struct btrfs_file_extent_item);
1222 btrfs_set_file_extent_generation(leaf, fi,
1223 trans->transid);
1224 btrfs_set_file_extent_num_bytes(leaf, fi,
1225 extent_end - end);
1226 btrfs_set_file_extent_offset(leaf, fi,
1227 end - orig_offset);
1228 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1229 struct btrfs_file_extent_item);
1230 btrfs_set_file_extent_generation(leaf, fi,
1231 trans->transid);
1232 btrfs_set_file_extent_num_bytes(leaf, fi,
1233 end - other_start);
1234 btrfs_mark_buffer_dirty(leaf);
1235 goto out;
1236 }
1237 }
1238
1239 if (start > key.offset && end == extent_end) {
1240 other_start = end;
1241 other_end = 0;
1242 if (extent_mergeable(leaf, path->slots[0] + 1,
1243 ino, bytenr, orig_offset,
1244 &other_start, &other_end)) {
1245 fi = btrfs_item_ptr(leaf, path->slots[0],
1246 struct btrfs_file_extent_item);
1247 btrfs_set_file_extent_num_bytes(leaf, fi,
1248 start - key.offset);
1249 btrfs_set_file_extent_generation(leaf, fi,
1250 trans->transid);
1251 path->slots[0]++;
1252 new_key.offset = start;
1253 btrfs_set_item_key_safe(fs_info, path, &new_key);
1254
1255 fi = btrfs_item_ptr(leaf, path->slots[0],
1256 struct btrfs_file_extent_item);
1257 btrfs_set_file_extent_generation(leaf, fi,
1258 trans->transid);
1259 btrfs_set_file_extent_num_bytes(leaf, fi,
1260 other_end - start);
1261 btrfs_set_file_extent_offset(leaf, fi,
1262 start - orig_offset);
1263 btrfs_mark_buffer_dirty(leaf);
1264 goto out;
1265 }
1266 }
1267
1268 while (start > key.offset || end < extent_end) {
1269 if (key.offset == start)
1270 split = end;
1271
1272 new_key.offset = split;
1273 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1274 if (ret == -EAGAIN) {
1275 btrfs_release_path(path);
1276 goto again;
1277 }
1278 if (ret < 0) {
1279 btrfs_abort_transaction(trans, ret);
1280 goto out;
1281 }
1282
1283 leaf = path->nodes[0];
1284 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1285 struct btrfs_file_extent_item);
1286 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1287 btrfs_set_file_extent_num_bytes(leaf, fi,
1288 split - key.offset);
1289
1290 fi = btrfs_item_ptr(leaf, path->slots[0],
1291 struct btrfs_file_extent_item);
1292
1293 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1294 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1295 btrfs_set_file_extent_num_bytes(leaf, fi,
1296 extent_end - split);
1297 btrfs_mark_buffer_dirty(leaf);
1298
1299 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1300 num_bytes, 0);
1301 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1302 orig_offset);
1303 ret = btrfs_inc_extent_ref(trans, &ref);
1304 if (ret) {
1305 btrfs_abort_transaction(trans, ret);
1306 goto out;
1307 }
1308
1309 if (split == start) {
1310 key.offset = start;
1311 } else {
1312 if (start != key.offset) {
1313 ret = -EINVAL;
1314 btrfs_abort_transaction(trans, ret);
1315 goto out;
1316 }
1317 path->slots[0]--;
1318 extent_end = end;
1319 }
1320 recow = 1;
1321 }
1322
1323 other_start = end;
1324 other_end = 0;
1325 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1326 num_bytes, 0);
1327 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1328 if (extent_mergeable(leaf, path->slots[0] + 1,
1329 ino, bytenr, orig_offset,
1330 &other_start, &other_end)) {
1331 if (recow) {
1332 btrfs_release_path(path);
1333 goto again;
1334 }
1335 extent_end = other_end;
1336 del_slot = path->slots[0] + 1;
1337 del_nr++;
1338 ret = btrfs_free_extent(trans, &ref);
1339 if (ret) {
1340 btrfs_abort_transaction(trans, ret);
1341 goto out;
1342 }
1343 }
1344 other_start = 0;
1345 other_end = start;
1346 if (extent_mergeable(leaf, path->slots[0] - 1,
1347 ino, bytenr, orig_offset,
1348 &other_start, &other_end)) {
1349 if (recow) {
1350 btrfs_release_path(path);
1351 goto again;
1352 }
1353 key.offset = other_start;
1354 del_slot = path->slots[0];
1355 del_nr++;
1356 ret = btrfs_free_extent(trans, &ref);
1357 if (ret) {
1358 btrfs_abort_transaction(trans, ret);
1359 goto out;
1360 }
1361 }
1362 if (del_nr == 0) {
1363 fi = btrfs_item_ptr(leaf, path->slots[0],
1364 struct btrfs_file_extent_item);
1365 btrfs_set_file_extent_type(leaf, fi,
1366 BTRFS_FILE_EXTENT_REG);
1367 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1368 btrfs_mark_buffer_dirty(leaf);
1369 } else {
1370 fi = btrfs_item_ptr(leaf, del_slot - 1,
1371 struct btrfs_file_extent_item);
1372 btrfs_set_file_extent_type(leaf, fi,
1373 BTRFS_FILE_EXTENT_REG);
1374 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1375 btrfs_set_file_extent_num_bytes(leaf, fi,
1376 extent_end - key.offset);
1377 btrfs_mark_buffer_dirty(leaf);
1378
1379 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1380 if (ret < 0) {
1381 btrfs_abort_transaction(trans, ret);
1382 goto out;
1383 }
1384 }
1385out:
1386 btrfs_free_path(path);
1387 return 0;
1388}
1389
1390/*
1391 * on error we return an unlocked page and the error value
1392 * on success we return a locked page and 0
1393 */
1394static int prepare_uptodate_page(struct inode *inode,
1395 struct page *page, u64 pos,
1396 bool force_uptodate)
1397{
1398 int ret = 0;
1399
1400 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1401 !PageUptodate(page)) {
1402 ret = btrfs_readpage(NULL, page);
1403 if (ret)
1404 return ret;
1405 lock_page(page);
1406 if (!PageUptodate(page)) {
1407 unlock_page(page);
1408 return -EIO;
1409 }
1410 if (page->mapping != inode->i_mapping) {
1411 unlock_page(page);
1412 return -EAGAIN;
1413 }
1414 }
1415 return 0;
1416}
1417
1418/*
1419 * this just gets pages into the page cache and locks them down.
1420 */
1421static noinline int prepare_pages(struct inode *inode, struct page **pages,
1422 size_t num_pages, loff_t pos,
1423 size_t write_bytes, bool force_uptodate)
1424{
1425 int i;
1426 unsigned long index = pos >> PAGE_SHIFT;
1427 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1428 int err = 0;
1429 int faili;
1430
1431 for (i = 0; i < num_pages; i++) {
1432again:
1433 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1434 mask | __GFP_WRITE);
1435 if (!pages[i]) {
1436 faili = i - 1;
1437 err = -ENOMEM;
1438 goto fail;
1439 }
1440
1441 if (i == 0)
1442 err = prepare_uptodate_page(inode, pages[i], pos,
1443 force_uptodate);
1444 if (!err && i == num_pages - 1)
1445 err = prepare_uptodate_page(inode, pages[i],
1446 pos + write_bytes, false);
1447 if (err) {
1448 put_page(pages[i]);
1449 if (err == -EAGAIN) {
1450 err = 0;
1451 goto again;
1452 }
1453 faili = i - 1;
1454 goto fail;
1455 }
1456 wait_on_page_writeback(pages[i]);
1457 }
1458
1459 return 0;
1460fail:
1461 while (faili >= 0) {
1462 unlock_page(pages[faili]);
1463 put_page(pages[faili]);
1464 faili--;
1465 }
1466 return err;
1467
1468}
1469
1470/*
1471 * This function locks the extent and properly waits for data=ordered extents
1472 * to finish before allowing the pages to be modified if need.
1473 *
1474 * The return value:
1475 * 1 - the extent is locked
1476 * 0 - the extent is not locked, and everything is OK
1477 * -EAGAIN - need re-prepare the pages
1478 * the other < 0 number - Something wrong happens
1479 */
1480static noinline int
1481lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1482 size_t num_pages, loff_t pos,
1483 size_t write_bytes,
1484 u64 *lockstart, u64 *lockend,
1485 struct extent_state **cached_state)
1486{
1487 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1488 u64 start_pos;
1489 u64 last_pos;
1490 int i;
1491 int ret = 0;
1492
1493 start_pos = round_down(pos, fs_info->sectorsize);
1494 last_pos = start_pos
1495 + round_up(pos + write_bytes - start_pos,
1496 fs_info->sectorsize) - 1;
1497
1498 if (start_pos < inode->vfs_inode.i_size) {
1499 struct btrfs_ordered_extent *ordered;
1500
1501 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1502 cached_state);
1503 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1504 last_pos - start_pos + 1);
1505 if (ordered &&
1506 ordered->file_offset + ordered->len > start_pos &&
1507 ordered->file_offset <= last_pos) {
1508 unlock_extent_cached(&inode->io_tree, start_pos,
1509 last_pos, cached_state);
1510 for (i = 0; i < num_pages; i++) {
1511 unlock_page(pages[i]);
1512 put_page(pages[i]);
1513 }
1514 btrfs_start_ordered_extent(&inode->vfs_inode,
1515 ordered, 1);
1516 btrfs_put_ordered_extent(ordered);
1517 return -EAGAIN;
1518 }
1519 if (ordered)
1520 btrfs_put_ordered_extent(ordered);
1521
1522 *lockstart = start_pos;
1523 *lockend = last_pos;
1524 ret = 1;
1525 }
1526
1527 /*
1528 * It's possible the pages are dirty right now, but we don't want
1529 * to clean them yet because copy_from_user may catch a page fault
1530 * and we might have to fall back to one page at a time. If that
1531 * happens, we'll unlock these pages and we'd have a window where
1532 * reclaim could sneak in and drop the once-dirty page on the floor
1533 * without writing it.
1534 *
1535 * We have the pages locked and the extent range locked, so there's
1536 * no way someone can start IO on any dirty pages in this range.
1537 *
1538 * We'll call btrfs_dirty_pages() later on, and that will flip around
1539 * delalloc bits and dirty the pages as required.
1540 */
1541 for (i = 0; i < num_pages; i++) {
1542 set_page_extent_mapped(pages[i]);
1543 WARN_ON(!PageLocked(pages[i]));
1544 }
1545
1546 return ret;
1547}
1548
1549static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1550 size_t *write_bytes)
1551{
1552 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1553 struct btrfs_root *root = inode->root;
1554 u64 lockstart, lockend;
1555 u64 num_bytes;
1556 int ret;
1557
1558 ret = btrfs_start_write_no_snapshotting(root);
1559 if (!ret)
1560 return -EAGAIN;
1561
1562 lockstart = round_down(pos, fs_info->sectorsize);
1563 lockend = round_up(pos + *write_bytes,
1564 fs_info->sectorsize) - 1;
1565
1566 btrfs_lock_and_flush_ordered_range(&inode->io_tree, inode, lockstart,
1567 lockend, NULL);
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
1585static 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_changeset *data_reserved = NULL;
1595 u64 release_bytes = 0;
1596 u64 lockstart;
1597 u64 lockend;
1598 size_t num_written = 0;
1599 int nrptrs;
1600 int ret = 0;
1601 bool only_release_metadata = false;
1602 bool force_page_uptodate = false;
1603
1604 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1605 PAGE_SIZE / (sizeof(struct page *)));
1606 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1607 nrptrs = max(nrptrs, 8);
1608 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1609 if (!pages)
1610 return -ENOMEM;
1611
1612 while (iov_iter_count(i) > 0) {
1613 struct extent_state *cached_state = NULL;
1614 size_t offset = offset_in_page(pos);
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;
1684again:
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);
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);
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
1762 /*
1763 * If we have not locked the extent range, because the range's
1764 * start offset is >= i_size, we might still have a non-NULL
1765 * cached extent state, acquired while marking the extent range
1766 * as delalloc through btrfs_dirty_pages(). Therefore free any
1767 * possible cached extent state to avoid a memory leak.
1768 */
1769 if (extents_locked)
1770 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1771 lockstart, lockend, &cached_state);
1772 else
1773 free_extent_state(cached_state);
1774
1775 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1776 if (ret) {
1777 btrfs_drop_pages(pages, num_pages);
1778 break;
1779 }
1780
1781 release_bytes = 0;
1782 if (only_release_metadata)
1783 btrfs_end_write_no_snapshotting(root);
1784
1785 if (only_release_metadata && copied > 0) {
1786 lockstart = round_down(pos,
1787 fs_info->sectorsize);
1788 lockend = round_up(pos + copied,
1789 fs_info->sectorsize) - 1;
1790
1791 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1792 lockend, EXTENT_NORESERVE, NULL,
1793 NULL, GFP_NOFS);
1794 only_release_metadata = false;
1795 }
1796
1797 btrfs_drop_pages(pages, num_pages);
1798
1799 cond_resched();
1800
1801 balance_dirty_pages_ratelimited(inode->i_mapping);
1802 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1803 btrfs_btree_balance_dirty(fs_info);
1804
1805 pos += copied;
1806 num_written += copied;
1807 }
1808
1809 kfree(pages);
1810
1811 if (release_bytes) {
1812 if (only_release_metadata) {
1813 btrfs_end_write_no_snapshotting(root);
1814 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1815 release_bytes, true);
1816 } else {
1817 btrfs_delalloc_release_space(inode, data_reserved,
1818 round_down(pos, fs_info->sectorsize),
1819 release_bytes, true);
1820 }
1821 }
1822
1823 extent_changeset_free(data_reserved);
1824 return num_written ? num_written : ret;
1825}
1826
1827static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1828{
1829 struct file *file = iocb->ki_filp;
1830 struct inode *inode = file_inode(file);
1831 loff_t pos;
1832 ssize_t written;
1833 ssize_t written_buffered;
1834 loff_t endbyte;
1835 int err;
1836
1837 written = generic_file_direct_write(iocb, from);
1838
1839 if (written < 0 || !iov_iter_count(from))
1840 return written;
1841
1842 pos = iocb->ki_pos;
1843 written_buffered = btrfs_buffered_write(iocb, from);
1844 if (written_buffered < 0) {
1845 err = written_buffered;
1846 goto out;
1847 }
1848 /*
1849 * Ensure all data is persisted. We want the next direct IO read to be
1850 * able to read what was just written.
1851 */
1852 endbyte = pos + written_buffered - 1;
1853 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1854 if (err)
1855 goto out;
1856 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1857 if (err)
1858 goto out;
1859 written += written_buffered;
1860 iocb->ki_pos = pos + written_buffered;
1861 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1862 endbyte >> PAGE_SHIFT);
1863out:
1864 return written ? written : err;
1865}
1866
1867static void update_time_for_write(struct inode *inode)
1868{
1869 struct timespec64 now;
1870
1871 if (IS_NOCMTIME(inode))
1872 return;
1873
1874 now = current_time(inode);
1875 if (!timespec64_equal(&inode->i_mtime, &now))
1876 inode->i_mtime = now;
1877
1878 if (!timespec64_equal(&inode->i_ctime, &now))
1879 inode->i_ctime = now;
1880
1881 if (IS_I_VERSION(inode))
1882 inode_inc_iversion(inode);
1883}
1884
1885static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1886 struct iov_iter *from)
1887{
1888 struct file *file = iocb->ki_filp;
1889 struct inode *inode = file_inode(file);
1890 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1891 struct btrfs_root *root = BTRFS_I(inode)->root;
1892 u64 start_pos;
1893 u64 end_pos;
1894 ssize_t num_written = 0;
1895 const bool sync = iocb->ki_flags & IOCB_DSYNC;
1896 ssize_t err;
1897 loff_t pos;
1898 size_t count;
1899 loff_t oldsize;
1900 int clean_page = 0;
1901
1902 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1903 (iocb->ki_flags & IOCB_NOWAIT))
1904 return -EOPNOTSUPP;
1905
1906 if (!inode_trylock(inode)) {
1907 if (iocb->ki_flags & IOCB_NOWAIT)
1908 return -EAGAIN;
1909 inode_lock(inode);
1910 }
1911
1912 err = generic_write_checks(iocb, from);
1913 if (err <= 0) {
1914 inode_unlock(inode);
1915 return err;
1916 }
1917
1918 pos = iocb->ki_pos;
1919 count = iov_iter_count(from);
1920 if (iocb->ki_flags & IOCB_NOWAIT) {
1921 /*
1922 * We will allocate space in case nodatacow is not set,
1923 * so bail
1924 */
1925 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1926 BTRFS_INODE_PREALLOC)) ||
1927 check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1928 inode_unlock(inode);
1929 return -EAGAIN;
1930 }
1931 }
1932
1933 current->backing_dev_info = inode_to_bdi(inode);
1934 err = file_remove_privs(file);
1935 if (err) {
1936 inode_unlock(inode);
1937 goto out;
1938 }
1939
1940 /*
1941 * If BTRFS flips readonly due to some impossible error
1942 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1943 * although we have opened a file as writable, we have
1944 * to stop this write operation to ensure FS consistency.
1945 */
1946 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1947 inode_unlock(inode);
1948 err = -EROFS;
1949 goto out;
1950 }
1951
1952 /*
1953 * We reserve space for updating the inode when we reserve space for the
1954 * extent we are going to write, so we will enospc out there. We don't
1955 * need to start yet another transaction to update the inode as we will
1956 * update the inode when we finish writing whatever data we write.
1957 */
1958 update_time_for_write(inode);
1959
1960 start_pos = round_down(pos, fs_info->sectorsize);
1961 oldsize = i_size_read(inode);
1962 if (start_pos > oldsize) {
1963 /* Expand hole size to cover write data, preventing empty gap */
1964 end_pos = round_up(pos + count,
1965 fs_info->sectorsize);
1966 err = btrfs_cont_expand(inode, oldsize, end_pos);
1967 if (err) {
1968 inode_unlock(inode);
1969 goto out;
1970 }
1971 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1972 clean_page = 1;
1973 }
1974
1975 if (sync)
1976 atomic_inc(&BTRFS_I(inode)->sync_writers);
1977
1978 if (iocb->ki_flags & IOCB_DIRECT) {
1979 num_written = __btrfs_direct_write(iocb, from);
1980 } else {
1981 num_written = btrfs_buffered_write(iocb, from);
1982 if (num_written > 0)
1983 iocb->ki_pos = pos + num_written;
1984 if (clean_page)
1985 pagecache_isize_extended(inode, oldsize,
1986 i_size_read(inode));
1987 }
1988
1989 inode_unlock(inode);
1990
1991 /*
1992 * We also have to set last_sub_trans to the current log transid,
1993 * otherwise subsequent syncs to a file that's been synced in this
1994 * transaction will appear to have already occurred.
1995 */
1996 spin_lock(&BTRFS_I(inode)->lock);
1997 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1998 spin_unlock(&BTRFS_I(inode)->lock);
1999 if (num_written > 0)
2000 num_written = generic_write_sync(iocb, num_written);
2001
2002 if (sync)
2003 atomic_dec(&BTRFS_I(inode)->sync_writers);
2004out:
2005 current->backing_dev_info = NULL;
2006 return num_written ? num_written : err;
2007}
2008
2009int btrfs_release_file(struct inode *inode, struct file *filp)
2010{
2011 struct btrfs_file_private *private = filp->private_data;
2012
2013 if (private && private->filldir_buf)
2014 kfree(private->filldir_buf);
2015 kfree(private);
2016 filp->private_data = NULL;
2017
2018 /*
2019 * ordered_data_close is set by setattr when we are about to truncate
2020 * a file from a non-zero size to a zero size. This tries to
2021 * flush down new bytes that may have been written if the
2022 * application were using truncate to replace a file in place.
2023 */
2024 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2025 &BTRFS_I(inode)->runtime_flags))
2026 filemap_flush(inode->i_mapping);
2027 return 0;
2028}
2029
2030static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2031{
2032 int ret;
2033 struct blk_plug plug;
2034
2035 /*
2036 * This is only called in fsync, which would do synchronous writes, so
2037 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2038 * multiple disks using raid profile, a large IO can be split to
2039 * several segments of stripe length (currently 64K).
2040 */
2041 blk_start_plug(&plug);
2042 atomic_inc(&BTRFS_I(inode)->sync_writers);
2043 ret = btrfs_fdatawrite_range(inode, start, end);
2044 atomic_dec(&BTRFS_I(inode)->sync_writers);
2045 blk_finish_plug(&plug);
2046
2047 return ret;
2048}
2049
2050/*
2051 * fsync call for both files and directories. This logs the inode into
2052 * the tree log instead of forcing full commits whenever possible.
2053 *
2054 * It needs to call filemap_fdatawait so that all ordered extent updates are
2055 * in the metadata btree are up to date for copying to the log.
2056 *
2057 * It drops the inode mutex before doing the tree log commit. This is an
2058 * important optimization for directories because holding the mutex prevents
2059 * new operations on the dir while we write to disk.
2060 */
2061int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2062{
2063 struct dentry *dentry = file_dentry(file);
2064 struct inode *inode = d_inode(dentry);
2065 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2066 struct btrfs_root *root = BTRFS_I(inode)->root;
2067 struct btrfs_trans_handle *trans;
2068 struct btrfs_log_ctx ctx;
2069 int ret = 0, err;
2070
2071 trace_btrfs_sync_file(file, datasync);
2072
2073 btrfs_init_log_ctx(&ctx, inode);
2074
2075 /*
2076 * We write the dirty pages in the range and wait until they complete
2077 * out of the ->i_mutex. If so, we can flush the dirty pages by
2078 * multi-task, and make the performance up. See
2079 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2080 */
2081 ret = start_ordered_ops(inode, start, end);
2082 if (ret)
2083 goto out;
2084
2085 inode_lock(inode);
2086
2087 /*
2088 * We take the dio_sem here because the tree log stuff can race with
2089 * lockless dio writes and get an extent map logged for an extent we
2090 * never waited on. We need it this high up for lockdep reasons.
2091 */
2092 down_write(&BTRFS_I(inode)->dio_sem);
2093
2094 atomic_inc(&root->log_batch);
2095
2096 /*
2097 * If the inode needs a full sync, make sure we use a full range to
2098 * avoid log tree corruption, due to hole detection racing with ordered
2099 * extent completion for adjacent ranges, and assertion failures during
2100 * hole detection. Do this while holding the inode lock, to avoid races
2101 * with other tasks.
2102 */
2103 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2104 &BTRFS_I(inode)->runtime_flags)) {
2105 start = 0;
2106 end = LLONG_MAX;
2107 }
2108
2109 /*
2110 * Before we acquired the inode's lock, someone may have dirtied more
2111 * pages in the target range. We need to make sure that writeback for
2112 * any such pages does not start while we are logging the inode, because
2113 * if it does, any of the following might happen when we are not doing a
2114 * full inode sync:
2115 *
2116 * 1) We log an extent after its writeback finishes but before its
2117 * checksums are added to the csum tree, leading to -EIO errors
2118 * when attempting to read the extent after a log replay.
2119 *
2120 * 2) We can end up logging an extent before its writeback finishes.
2121 * Therefore after the log replay we will have a file extent item
2122 * pointing to an unwritten extent (and no data checksums as well).
2123 *
2124 * So trigger writeback for any eventual new dirty pages and then we
2125 * wait for all ordered extents to complete below.
2126 */
2127 ret = start_ordered_ops(inode, start, end);
2128 if (ret) {
2129 inode_unlock(inode);
2130 goto out;
2131 }
2132
2133 /*
2134 * We have to do this here to avoid the priority inversion of waiting on
2135 * IO of a lower priority task while holding a transaction open.
2136 *
2137 * Also, the range length can be represented by u64, we have to do the
2138 * typecasts to avoid signed overflow if it's [0, LLONG_MAX].
2139 */
2140 ret = btrfs_wait_ordered_range(inode, start, (u64)end - (u64)start + 1);
2141 if (ret) {
2142 up_write(&BTRFS_I(inode)->dio_sem);
2143 inode_unlock(inode);
2144 goto out;
2145 }
2146 atomic_inc(&root->log_batch);
2147
2148 smp_mb();
2149 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2150 BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2151 /*
2152 * We've had everything committed since the last time we were
2153 * modified so clear this flag in case it was set for whatever
2154 * reason, it's no longer relevant.
2155 */
2156 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2157 &BTRFS_I(inode)->runtime_flags);
2158 /*
2159 * An ordered extent might have started before and completed
2160 * already with io errors, in which case the inode was not
2161 * updated and we end up here. So check the inode's mapping
2162 * for any errors that might have happened since we last
2163 * checked called fsync.
2164 */
2165 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2166 up_write(&BTRFS_I(inode)->dio_sem);
2167 inode_unlock(inode);
2168 goto out;
2169 }
2170
2171 /*
2172 * We use start here because we will need to wait on the IO to complete
2173 * in btrfs_sync_log, which could require joining a transaction (for
2174 * example checking cross references in the nocow path). If we use join
2175 * here we could get into a situation where we're waiting on IO to
2176 * happen that is blocked on a transaction trying to commit. With start
2177 * we inc the extwriter counter, so we wait for all extwriters to exit
2178 * before we start blocking joiners. This comment is to keep somebody
2179 * from thinking they are super smart and changing this to
2180 * btrfs_join_transaction *cough*Josef*cough*.
2181 */
2182 trans = btrfs_start_transaction(root, 0);
2183 if (IS_ERR(trans)) {
2184 ret = PTR_ERR(trans);
2185 up_write(&BTRFS_I(inode)->dio_sem);
2186 inode_unlock(inode);
2187 goto out;
2188 }
2189
2190 ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2191 if (ret < 0) {
2192 /* Fallthrough and commit/free transaction. */
2193 ret = 1;
2194 }
2195
2196 /* we've logged all the items and now have a consistent
2197 * version of the file in the log. It is possible that
2198 * someone will come in and modify the file, but that's
2199 * fine because the log is consistent on disk, and we
2200 * have references to all of the file's extents
2201 *
2202 * It is possible that someone will come in and log the
2203 * file again, but that will end up using the synchronization
2204 * inside btrfs_sync_log to keep things safe.
2205 */
2206 up_write(&BTRFS_I(inode)->dio_sem);
2207 inode_unlock(inode);
2208
2209 if (ret != BTRFS_NO_LOG_SYNC) {
2210 if (!ret) {
2211 ret = btrfs_sync_log(trans, root, &ctx);
2212 if (!ret) {
2213 ret = btrfs_end_transaction(trans);
2214 goto out;
2215 }
2216 }
2217 ret = btrfs_commit_transaction(trans);
2218 } else {
2219 ret = btrfs_end_transaction(trans);
2220 }
2221out:
2222 ASSERT(list_empty(&ctx.list));
2223 err = file_check_and_advance_wb_err(file);
2224 if (!ret)
2225 ret = err;
2226 return ret > 0 ? -EIO : ret;
2227}
2228
2229static const struct vm_operations_struct btrfs_file_vm_ops = {
2230 .fault = filemap_fault,
2231 .map_pages = filemap_map_pages,
2232 .page_mkwrite = btrfs_page_mkwrite,
2233};
2234
2235static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2236{
2237 struct address_space *mapping = filp->f_mapping;
2238
2239 if (!mapping->a_ops->readpage)
2240 return -ENOEXEC;
2241
2242 file_accessed(filp);
2243 vma->vm_ops = &btrfs_file_vm_ops;
2244
2245 return 0;
2246}
2247
2248static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2249 int slot, u64 start, u64 end)
2250{
2251 struct btrfs_file_extent_item *fi;
2252 struct btrfs_key key;
2253
2254 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2255 return 0;
2256
2257 btrfs_item_key_to_cpu(leaf, &key, slot);
2258 if (key.objectid != btrfs_ino(inode) ||
2259 key.type != BTRFS_EXTENT_DATA_KEY)
2260 return 0;
2261
2262 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2263
2264 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2265 return 0;
2266
2267 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2268 return 0;
2269
2270 if (key.offset == end)
2271 return 1;
2272 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2273 return 1;
2274 return 0;
2275}
2276
2277static int fill_holes(struct btrfs_trans_handle *trans,
2278 struct btrfs_inode *inode,
2279 struct btrfs_path *path, u64 offset, u64 end)
2280{
2281 struct btrfs_fs_info *fs_info = trans->fs_info;
2282 struct btrfs_root *root = inode->root;
2283 struct extent_buffer *leaf;
2284 struct btrfs_file_extent_item *fi;
2285 struct extent_map *hole_em;
2286 struct extent_map_tree *em_tree = &inode->extent_tree;
2287 struct btrfs_key key;
2288 int ret;
2289
2290 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2291 goto out;
2292
2293 key.objectid = btrfs_ino(inode);
2294 key.type = BTRFS_EXTENT_DATA_KEY;
2295 key.offset = offset;
2296
2297 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2298 if (ret <= 0) {
2299 /*
2300 * We should have dropped this offset, so if we find it then
2301 * something has gone horribly wrong.
2302 */
2303 if (ret == 0)
2304 ret = -EINVAL;
2305 return ret;
2306 }
2307
2308 leaf = path->nodes[0];
2309 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2310 u64 num_bytes;
2311
2312 path->slots[0]--;
2313 fi = btrfs_item_ptr(leaf, path->slots[0],
2314 struct btrfs_file_extent_item);
2315 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2316 end - offset;
2317 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2318 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2319 btrfs_set_file_extent_offset(leaf, fi, 0);
2320 btrfs_mark_buffer_dirty(leaf);
2321 goto out;
2322 }
2323
2324 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2325 u64 num_bytes;
2326
2327 key.offset = offset;
2328 btrfs_set_item_key_safe(fs_info, path, &key);
2329 fi = btrfs_item_ptr(leaf, path->slots[0],
2330 struct btrfs_file_extent_item);
2331 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2332 offset;
2333 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2334 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2335 btrfs_set_file_extent_offset(leaf, fi, 0);
2336 btrfs_mark_buffer_dirty(leaf);
2337 goto out;
2338 }
2339 btrfs_release_path(path);
2340
2341 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2342 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2343 if (ret)
2344 return ret;
2345
2346out:
2347 btrfs_release_path(path);
2348
2349 hole_em = alloc_extent_map();
2350 if (!hole_em) {
2351 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2352 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2353 } else {
2354 hole_em->start = offset;
2355 hole_em->len = end - offset;
2356 hole_em->ram_bytes = hole_em->len;
2357 hole_em->orig_start = offset;
2358
2359 hole_em->block_start = EXTENT_MAP_HOLE;
2360 hole_em->block_len = 0;
2361 hole_em->orig_block_len = 0;
2362 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2363 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2364 hole_em->generation = trans->transid;
2365
2366 do {
2367 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2368 write_lock(&em_tree->lock);
2369 ret = add_extent_mapping(em_tree, hole_em, 1);
2370 write_unlock(&em_tree->lock);
2371 } while (ret == -EEXIST);
2372 free_extent_map(hole_em);
2373 if (ret)
2374 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2375 &inode->runtime_flags);
2376 }
2377
2378 return 0;
2379}
2380
2381/*
2382 * Find a hole extent on given inode and change start/len to the end of hole
2383 * extent.(hole/vacuum extent whose em->start <= start &&
2384 * em->start + em->len > start)
2385 * When a hole extent is found, return 1 and modify start/len.
2386 */
2387static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2388{
2389 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2390 struct extent_map *em;
2391 int ret = 0;
2392
2393 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2394 round_down(*start, fs_info->sectorsize),
2395 round_up(*len, fs_info->sectorsize), 0);
2396 if (IS_ERR(em))
2397 return PTR_ERR(em);
2398
2399 /* Hole or vacuum extent(only exists in no-hole mode) */
2400 if (em->block_start == EXTENT_MAP_HOLE) {
2401 ret = 1;
2402 *len = em->start + em->len > *start + *len ?
2403 0 : *start + *len - em->start - em->len;
2404 *start = em->start + em->len;
2405 }
2406 free_extent_map(em);
2407 return ret;
2408}
2409
2410static int btrfs_punch_hole_lock_range(struct inode *inode,
2411 const u64 lockstart,
2412 const u64 lockend,
2413 struct extent_state **cached_state)
2414{
2415 while (1) {
2416 struct btrfs_ordered_extent *ordered;
2417 int ret;
2418
2419 truncate_pagecache_range(inode, lockstart, lockend);
2420
2421 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2422 cached_state);
2423 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2424
2425 /*
2426 * We need to make sure we have no ordered extents in this range
2427 * and nobody raced in and read a page in this range, if we did
2428 * we need to try again.
2429 */
2430 if ((!ordered ||
2431 (ordered->file_offset + ordered->len <= lockstart ||
2432 ordered->file_offset > lockend)) &&
2433 !filemap_range_has_page(inode->i_mapping,
2434 lockstart, lockend)) {
2435 if (ordered)
2436 btrfs_put_ordered_extent(ordered);
2437 break;
2438 }
2439 if (ordered)
2440 btrfs_put_ordered_extent(ordered);
2441 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2442 lockend, cached_state);
2443 ret = btrfs_wait_ordered_range(inode, lockstart,
2444 lockend - lockstart + 1);
2445 if (ret)
2446 return ret;
2447 }
2448 return 0;
2449}
2450
2451static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2452 struct inode *inode,
2453 struct btrfs_path *path,
2454 struct btrfs_clone_extent_info *clone_info,
2455 const u64 clone_len)
2456{
2457 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2458 struct btrfs_root *root = BTRFS_I(inode)->root;
2459 struct btrfs_file_extent_item *extent;
2460 struct extent_buffer *leaf;
2461 struct btrfs_key key;
2462 int slot;
2463 struct btrfs_ref ref = { 0 };
2464 u64 ref_offset;
2465 int ret;
2466
2467 if (clone_len == 0)
2468 return 0;
2469
2470 if (clone_info->disk_offset == 0 &&
2471 btrfs_fs_incompat(fs_info, NO_HOLES))
2472 return 0;
2473
2474 key.objectid = btrfs_ino(BTRFS_I(inode));
2475 key.type = BTRFS_EXTENT_DATA_KEY;
2476 key.offset = clone_info->file_offset;
2477 ret = btrfs_insert_empty_item(trans, root, path, &key,
2478 clone_info->item_size);
2479 if (ret)
2480 return ret;
2481 leaf = path->nodes[0];
2482 slot = path->slots[0];
2483 write_extent_buffer(leaf, clone_info->extent_buf,
2484 btrfs_item_ptr_offset(leaf, slot),
2485 clone_info->item_size);
2486 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2487 btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2488 btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2489 btrfs_mark_buffer_dirty(leaf);
2490 btrfs_release_path(path);
2491
2492 /* If it's a hole, nothing more needs to be done. */
2493 if (clone_info->disk_offset == 0)
2494 return 0;
2495
2496 inode_add_bytes(inode, clone_len);
2497 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2498 clone_info->disk_offset,
2499 clone_info->disk_len, 0);
2500 ref_offset = clone_info->file_offset - clone_info->data_offset;
2501 btrfs_init_data_ref(&ref, root->root_key.objectid,
2502 btrfs_ino(BTRFS_I(inode)), ref_offset);
2503 ret = btrfs_inc_extent_ref(trans, &ref);
2504
2505 return ret;
2506}
2507
2508/*
2509 * The respective range must have been previously locked, as well as the inode.
2510 * The end offset is inclusive (last byte of the range).
2511 * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2512 * cloning.
2513 * When cloning, we don't want to end up in a state where we dropped extents
2514 * without inserting a new one, so we must abort the transaction to avoid a
2515 * corruption.
2516 */
2517int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2518 const u64 start, const u64 end,
2519 struct btrfs_clone_extent_info *clone_info,
2520 struct btrfs_trans_handle **trans_out)
2521{
2522 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2523 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2524 u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2525 struct btrfs_root *root = BTRFS_I(inode)->root;
2526 struct btrfs_trans_handle *trans = NULL;
2527 struct btrfs_block_rsv *rsv;
2528 unsigned int rsv_count;
2529 u64 cur_offset;
2530 u64 drop_end;
2531 u64 len = end - start;
2532 int ret = 0;
2533
2534 if (end <= start)
2535 return -EINVAL;
2536
2537 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2538 if (!rsv) {
2539 ret = -ENOMEM;
2540 goto out;
2541 }
2542 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2543 rsv->failfast = 1;
2544
2545 /*
2546 * 1 - update the inode
2547 * 1 - removing the extents in the range
2548 * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2549 * an extent
2550 */
2551 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2552 rsv_count = 3;
2553 else
2554 rsv_count = 2;
2555
2556 trans = btrfs_start_transaction(root, rsv_count);
2557 if (IS_ERR(trans)) {
2558 ret = PTR_ERR(trans);
2559 trans = NULL;
2560 goto out_free;
2561 }
2562
2563 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2564 min_size, false);
2565 BUG_ON(ret);
2566 trans->block_rsv = rsv;
2567
2568 cur_offset = start;
2569 while (cur_offset < end) {
2570 ret = __btrfs_drop_extents(trans, root, inode, path,
2571 cur_offset, end + 1, &drop_end,
2572 1, 0, 0, NULL);
2573 if (ret != -ENOSPC) {
2574 /*
2575 * When cloning we want to avoid transaction aborts when
2576 * nothing was done and we are attempting to clone parts
2577 * of inline extents, in such cases -EOPNOTSUPP is
2578 * returned by __btrfs_drop_extents() without having
2579 * changed anything in the file.
2580 */
2581 if (clone_info && ret && ret != -EOPNOTSUPP)
2582 btrfs_abort_transaction(trans, ret);
2583 break;
2584 }
2585
2586 trans->block_rsv = &fs_info->trans_block_rsv;
2587
2588 if (!clone_info && cur_offset < drop_end &&
2589 cur_offset < ino_size) {
2590 ret = fill_holes(trans, BTRFS_I(inode), path,
2591 cur_offset, drop_end);
2592 if (ret) {
2593 /*
2594 * If we failed then we didn't insert our hole
2595 * entries for the area we dropped, so now the
2596 * fs is corrupted, so we must abort the
2597 * transaction.
2598 */
2599 btrfs_abort_transaction(trans, ret);
2600 break;
2601 }
2602 }
2603
2604 if (clone_info) {
2605 u64 clone_len = drop_end - cur_offset;
2606
2607 ret = btrfs_insert_clone_extent(trans, inode, path,
2608 clone_info, clone_len);
2609 if (ret) {
2610 btrfs_abort_transaction(trans, ret);
2611 break;
2612 }
2613 clone_info->data_len -= clone_len;
2614 clone_info->data_offset += clone_len;
2615 clone_info->file_offset += clone_len;
2616 }
2617
2618 cur_offset = drop_end;
2619
2620 ret = btrfs_update_inode(trans, root, inode);
2621 if (ret)
2622 break;
2623
2624 btrfs_end_transaction(trans);
2625 btrfs_btree_balance_dirty(fs_info);
2626
2627 trans = btrfs_start_transaction(root, rsv_count);
2628 if (IS_ERR(trans)) {
2629 ret = PTR_ERR(trans);
2630 trans = NULL;
2631 break;
2632 }
2633
2634 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2635 rsv, min_size, false);
2636 BUG_ON(ret); /* shouldn't happen */
2637 trans->block_rsv = rsv;
2638
2639 if (!clone_info) {
2640 ret = find_first_non_hole(inode, &cur_offset, &len);
2641 if (unlikely(ret < 0))
2642 break;
2643 if (ret && !len) {
2644 ret = 0;
2645 break;
2646 }
2647 }
2648 }
2649
2650 /*
2651 * If we were cloning, force the next fsync to be a full one since we
2652 * we replaced (or just dropped in the case of cloning holes when
2653 * NO_HOLES is enabled) extents and extent maps.
2654 * This is for the sake of simplicity, and cloning into files larger
2655 * than 16Mb would force the full fsync any way (when
2656 * try_release_extent_mapping() is invoked during page cache truncation.
2657 */
2658 if (clone_info)
2659 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2660 &BTRFS_I(inode)->runtime_flags);
2661
2662 if (ret)
2663 goto out_trans;
2664
2665 trans->block_rsv = &fs_info->trans_block_rsv;
2666 /*
2667 * If we are using the NO_HOLES feature we might have had already an
2668 * hole that overlaps a part of the region [lockstart, lockend] and
2669 * ends at (or beyond) lockend. Since we have no file extent items to
2670 * represent holes, drop_end can be less than lockend and so we must
2671 * make sure we have an extent map representing the existing hole (the
2672 * call to __btrfs_drop_extents() might have dropped the existing extent
2673 * map representing the existing hole), otherwise the fast fsync path
2674 * will not record the existence of the hole region
2675 * [existing_hole_start, lockend].
2676 */
2677 if (drop_end <= end)
2678 drop_end = end + 1;
2679 /*
2680 * Don't insert file hole extent item if it's for a range beyond eof
2681 * (because it's useless) or if it represents a 0 bytes range (when
2682 * cur_offset == drop_end).
2683 */
2684 if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2685 ret = fill_holes(trans, BTRFS_I(inode), path,
2686 cur_offset, drop_end);
2687 if (ret) {
2688 /* Same comment as above. */
2689 btrfs_abort_transaction(trans, ret);
2690 goto out_trans;
2691 }
2692 }
2693 if (clone_info) {
2694 ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2695 clone_info->data_len);
2696 if (ret) {
2697 btrfs_abort_transaction(trans, ret);
2698 goto out_trans;
2699 }
2700 }
2701
2702out_trans:
2703 if (!trans)
2704 goto out_free;
2705
2706 trans->block_rsv = &fs_info->trans_block_rsv;
2707 if (ret)
2708 btrfs_end_transaction(trans);
2709 else
2710 *trans_out = trans;
2711out_free:
2712 btrfs_free_block_rsv(fs_info, rsv);
2713out:
2714 return ret;
2715}
2716
2717static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2718{
2719 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2720 struct btrfs_root *root = BTRFS_I(inode)->root;
2721 struct extent_state *cached_state = NULL;
2722 struct btrfs_path *path;
2723 struct btrfs_trans_handle *trans = NULL;
2724 u64 lockstart;
2725 u64 lockend;
2726 u64 tail_start;
2727 u64 tail_len;
2728 u64 orig_start = offset;
2729 int ret = 0;
2730 bool same_block;
2731 u64 ino_size;
2732 bool truncated_block = false;
2733 bool updated_inode = false;
2734
2735 ret = btrfs_wait_ordered_range(inode, offset, len);
2736 if (ret)
2737 return ret;
2738
2739 inode_lock(inode);
2740 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2741 ret = find_first_non_hole(inode, &offset, &len);
2742 if (ret < 0)
2743 goto out_only_mutex;
2744 if (ret && !len) {
2745 /* Already in a large hole */
2746 ret = 0;
2747 goto out_only_mutex;
2748 }
2749
2750 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2751 lockend = round_down(offset + len,
2752 btrfs_inode_sectorsize(inode)) - 1;
2753 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2754 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2755 /*
2756 * We needn't truncate any block which is beyond the end of the file
2757 * because we are sure there is no data there.
2758 */
2759 /*
2760 * Only do this if we are in the same block and we aren't doing the
2761 * entire block.
2762 */
2763 if (same_block && len < fs_info->sectorsize) {
2764 if (offset < ino_size) {
2765 truncated_block = true;
2766 ret = btrfs_truncate_block(inode, offset, len, 0);
2767 } else {
2768 ret = 0;
2769 }
2770 goto out_only_mutex;
2771 }
2772
2773 /* zero back part of the first block */
2774 if (offset < ino_size) {
2775 truncated_block = true;
2776 ret = btrfs_truncate_block(inode, offset, 0, 0);
2777 if (ret) {
2778 inode_unlock(inode);
2779 return ret;
2780 }
2781 }
2782
2783 /* Check the aligned pages after the first unaligned page,
2784 * if offset != orig_start, which means the first unaligned page
2785 * including several following pages are already in holes,
2786 * the extra check can be skipped */
2787 if (offset == orig_start) {
2788 /* after truncate page, check hole again */
2789 len = offset + len - lockstart;
2790 offset = lockstart;
2791 ret = find_first_non_hole(inode, &offset, &len);
2792 if (ret < 0)
2793 goto out_only_mutex;
2794 if (ret && !len) {
2795 ret = 0;
2796 goto out_only_mutex;
2797 }
2798 lockstart = offset;
2799 }
2800
2801 /* Check the tail unaligned part is in a hole */
2802 tail_start = lockend + 1;
2803 tail_len = offset + len - tail_start;
2804 if (tail_len) {
2805 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2806 if (unlikely(ret < 0))
2807 goto out_only_mutex;
2808 if (!ret) {
2809 /* zero the front end of the last page */
2810 if (tail_start + tail_len < ino_size) {
2811 truncated_block = true;
2812 ret = btrfs_truncate_block(inode,
2813 tail_start + tail_len,
2814 0, 1);
2815 if (ret)
2816 goto out_only_mutex;
2817 }
2818 }
2819 }
2820
2821 if (lockend < lockstart) {
2822 ret = 0;
2823 goto out_only_mutex;
2824 }
2825
2826 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2827 &cached_state);
2828 if (ret)
2829 goto out_only_mutex;
2830
2831 path = btrfs_alloc_path();
2832 if (!path) {
2833 ret = -ENOMEM;
2834 goto out;
2835 }
2836
2837 ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2838 &trans);
2839 btrfs_free_path(path);
2840 if (ret)
2841 goto out;
2842
2843 ASSERT(trans != NULL);
2844 inode_inc_iversion(inode);
2845 inode->i_mtime = inode->i_ctime = current_time(inode);
2846 ret = btrfs_update_inode(trans, root, inode);
2847 updated_inode = true;
2848 btrfs_end_transaction(trans);
2849 btrfs_btree_balance_dirty(fs_info);
2850out:
2851 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2852 &cached_state);
2853out_only_mutex:
2854 if (!updated_inode && truncated_block && !ret) {
2855 /*
2856 * If we only end up zeroing part of a page, we still need to
2857 * update the inode item, so that all the time fields are
2858 * updated as well as the necessary btrfs inode in memory fields
2859 * for detecting, at fsync time, if the inode isn't yet in the
2860 * log tree or it's there but not up to date.
2861 */
2862 struct timespec64 now = current_time(inode);
2863
2864 inode_inc_iversion(inode);
2865 inode->i_mtime = now;
2866 inode->i_ctime = now;
2867 trans = btrfs_start_transaction(root, 1);
2868 if (IS_ERR(trans)) {
2869 ret = PTR_ERR(trans);
2870 } else {
2871 int ret2;
2872
2873 ret = btrfs_update_inode(trans, root, inode);
2874 ret2 = btrfs_end_transaction(trans);
2875 if (!ret)
2876 ret = ret2;
2877 }
2878 }
2879 inode_unlock(inode);
2880 return ret;
2881}
2882
2883/* Helper structure to record which range is already reserved */
2884struct falloc_range {
2885 struct list_head list;
2886 u64 start;
2887 u64 len;
2888};
2889
2890/*
2891 * Helper function to add falloc range
2892 *
2893 * Caller should have locked the larger range of extent containing
2894 * [start, len)
2895 */
2896static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2897{
2898 struct falloc_range *prev = NULL;
2899 struct falloc_range *range = NULL;
2900
2901 if (list_empty(head))
2902 goto insert;
2903
2904 /*
2905 * As fallocate iterate by bytenr order, we only need to check
2906 * the last range.
2907 */
2908 prev = list_entry(head->prev, struct falloc_range, list);
2909 if (prev->start + prev->len == start) {
2910 prev->len += len;
2911 return 0;
2912 }
2913insert:
2914 range = kmalloc(sizeof(*range), GFP_KERNEL);
2915 if (!range)
2916 return -ENOMEM;
2917 range->start = start;
2918 range->len = len;
2919 list_add_tail(&range->list, head);
2920 return 0;
2921}
2922
2923static int btrfs_fallocate_update_isize(struct inode *inode,
2924 const u64 end,
2925 const int mode)
2926{
2927 struct btrfs_trans_handle *trans;
2928 struct btrfs_root *root = BTRFS_I(inode)->root;
2929 int ret;
2930 int ret2;
2931
2932 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2933 return 0;
2934
2935 trans = btrfs_start_transaction(root, 1);
2936 if (IS_ERR(trans))
2937 return PTR_ERR(trans);
2938
2939 inode->i_ctime = current_time(inode);
2940 i_size_write(inode, end);
2941 btrfs_ordered_update_i_size(inode, end, NULL);
2942 ret = btrfs_update_inode(trans, root, inode);
2943 ret2 = btrfs_end_transaction(trans);
2944
2945 return ret ? ret : ret2;
2946}
2947
2948enum {
2949 RANGE_BOUNDARY_WRITTEN_EXTENT,
2950 RANGE_BOUNDARY_PREALLOC_EXTENT,
2951 RANGE_BOUNDARY_HOLE,
2952};
2953
2954static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2955 u64 offset)
2956{
2957 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2958 struct extent_map *em;
2959 int ret;
2960
2961 offset = round_down(offset, sectorsize);
2962 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2963 if (IS_ERR(em))
2964 return PTR_ERR(em);
2965
2966 if (em->block_start == EXTENT_MAP_HOLE)
2967 ret = RANGE_BOUNDARY_HOLE;
2968 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2969 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2970 else
2971 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2972
2973 free_extent_map(em);
2974 return ret;
2975}
2976
2977static int btrfs_zero_range(struct inode *inode,
2978 loff_t offset,
2979 loff_t len,
2980 const int mode)
2981{
2982 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2983 struct extent_map *em;
2984 struct extent_changeset *data_reserved = NULL;
2985 int ret;
2986 u64 alloc_hint = 0;
2987 const u64 sectorsize = btrfs_inode_sectorsize(inode);
2988 u64 alloc_start = round_down(offset, sectorsize);
2989 u64 alloc_end = round_up(offset + len, sectorsize);
2990 u64 bytes_to_reserve = 0;
2991 bool space_reserved = false;
2992
2993 inode_dio_wait(inode);
2994
2995 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2996 alloc_start, alloc_end - alloc_start, 0);
2997 if (IS_ERR(em)) {
2998 ret = PTR_ERR(em);
2999 goto out;
3000 }
3001
3002 /*
3003 * Avoid hole punching and extent allocation for some cases. More cases
3004 * could be considered, but these are unlikely common and we keep things
3005 * as simple as possible for now. Also, intentionally, if the target
3006 * range contains one or more prealloc extents together with regular
3007 * extents and holes, we drop all the existing extents and allocate a
3008 * new prealloc extent, so that we get a larger contiguous disk extent.
3009 */
3010 if (em->start <= alloc_start &&
3011 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3012 const u64 em_end = em->start + em->len;
3013
3014 if (em_end >= offset + len) {
3015 /*
3016 * The whole range is already a prealloc extent,
3017 * do nothing except updating the inode's i_size if
3018 * needed.
3019 */
3020 free_extent_map(em);
3021 ret = btrfs_fallocate_update_isize(inode, offset + len,
3022 mode);
3023 goto out;
3024 }
3025 /*
3026 * Part of the range is already a prealloc extent, so operate
3027 * only on the remaining part of the range.
3028 */
3029 alloc_start = em_end;
3030 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3031 len = offset + len - alloc_start;
3032 offset = alloc_start;
3033 alloc_hint = em->block_start + em->len;
3034 }
3035 free_extent_map(em);
3036
3037 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3038 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3039 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
3040 alloc_start, sectorsize, 0);
3041 if (IS_ERR(em)) {
3042 ret = PTR_ERR(em);
3043 goto out;
3044 }
3045
3046 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3047 free_extent_map(em);
3048 ret = btrfs_fallocate_update_isize(inode, offset + len,
3049 mode);
3050 goto out;
3051 }
3052 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3053 free_extent_map(em);
3054 ret = btrfs_truncate_block(inode, offset, len, 0);
3055 if (!ret)
3056 ret = btrfs_fallocate_update_isize(inode,
3057 offset + len,
3058 mode);
3059 return ret;
3060 }
3061 free_extent_map(em);
3062 alloc_start = round_down(offset, sectorsize);
3063 alloc_end = alloc_start + sectorsize;
3064 goto reserve_space;
3065 }
3066
3067 alloc_start = round_up(offset, sectorsize);
3068 alloc_end = round_down(offset + len, sectorsize);
3069
3070 /*
3071 * For unaligned ranges, check the pages at the boundaries, they might
3072 * map to an extent, in which case we need to partially zero them, or
3073 * they might map to a hole, in which case we need our allocation range
3074 * to cover them.
3075 */
3076 if (!IS_ALIGNED(offset, sectorsize)) {
3077 ret = btrfs_zero_range_check_range_boundary(inode, offset);
3078 if (ret < 0)
3079 goto out;
3080 if (ret == RANGE_BOUNDARY_HOLE) {
3081 alloc_start = round_down(offset, sectorsize);
3082 ret = 0;
3083 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3084 ret = btrfs_truncate_block(inode, offset, 0, 0);
3085 if (ret)
3086 goto out;
3087 } else {
3088 ret = 0;
3089 }
3090 }
3091
3092 if (!IS_ALIGNED(offset + len, sectorsize)) {
3093 ret = btrfs_zero_range_check_range_boundary(inode,
3094 offset + len);
3095 if (ret < 0)
3096 goto out;
3097 if (ret == RANGE_BOUNDARY_HOLE) {
3098 alloc_end = round_up(offset + len, sectorsize);
3099 ret = 0;
3100 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3101 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3102 if (ret)
3103 goto out;
3104 } else {
3105 ret = 0;
3106 }
3107 }
3108
3109reserve_space:
3110 if (alloc_start < alloc_end) {
3111 struct extent_state *cached_state = NULL;
3112 const u64 lockstart = alloc_start;
3113 const u64 lockend = alloc_end - 1;
3114
3115 bytes_to_reserve = alloc_end - alloc_start;
3116 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3117 bytes_to_reserve);
3118 if (ret < 0)
3119 goto out;
3120 space_reserved = true;
3121 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3122 alloc_start, bytes_to_reserve);
3123 if (ret)
3124 goto out;
3125 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3126 &cached_state);
3127 if (ret)
3128 goto out;
3129 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3130 alloc_end - alloc_start,
3131 i_blocksize(inode),
3132 offset + len, &alloc_hint);
3133 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3134 lockend, &cached_state);
3135 /* btrfs_prealloc_file_range releases reserved space on error */
3136 if (ret) {
3137 space_reserved = false;
3138 goto out;
3139 }
3140 }
3141 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3142 out:
3143 if (ret && space_reserved)
3144 btrfs_free_reserved_data_space(inode, data_reserved,
3145 alloc_start, bytes_to_reserve);
3146 extent_changeset_free(data_reserved);
3147
3148 return ret;
3149}
3150
3151static long btrfs_fallocate(struct file *file, int mode,
3152 loff_t offset, loff_t len)
3153{
3154 struct inode *inode = file_inode(file);
3155 struct extent_state *cached_state = NULL;
3156 struct extent_changeset *data_reserved = NULL;
3157 struct falloc_range *range;
3158 struct falloc_range *tmp;
3159 struct list_head reserve_list;
3160 u64 cur_offset;
3161 u64 last_byte;
3162 u64 alloc_start;
3163 u64 alloc_end;
3164 u64 alloc_hint = 0;
3165 u64 locked_end;
3166 u64 actual_end = 0;
3167 struct extent_map *em;
3168 int blocksize = btrfs_inode_sectorsize(inode);
3169 int ret;
3170
3171 alloc_start = round_down(offset, blocksize);
3172 alloc_end = round_up(offset + len, blocksize);
3173 cur_offset = alloc_start;
3174
3175 /* Make sure we aren't being give some crap mode */
3176 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3177 FALLOC_FL_ZERO_RANGE))
3178 return -EOPNOTSUPP;
3179
3180 if (mode & FALLOC_FL_PUNCH_HOLE)
3181 return btrfs_punch_hole(inode, offset, len);
3182
3183 /*
3184 * Only trigger disk allocation, don't trigger qgroup reserve
3185 *
3186 * For qgroup space, it will be checked later.
3187 */
3188 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3189 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3190 alloc_end - alloc_start);
3191 if (ret < 0)
3192 return ret;
3193 }
3194
3195 inode_lock(inode);
3196
3197 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3198 ret = inode_newsize_ok(inode, offset + len);
3199 if (ret)
3200 goto out;
3201 }
3202
3203 /*
3204 * TODO: Move these two operations after we have checked
3205 * accurate reserved space, or fallocate can still fail but
3206 * with page truncated or size expanded.
3207 *
3208 * But that's a minor problem and won't do much harm BTW.
3209 */
3210 if (alloc_start > inode->i_size) {
3211 ret = btrfs_cont_expand(inode, i_size_read(inode),
3212 alloc_start);
3213 if (ret)
3214 goto out;
3215 } else if (offset + len > inode->i_size) {
3216 /*
3217 * If we are fallocating from the end of the file onward we
3218 * need to zero out the end of the block if i_size lands in the
3219 * middle of a block.
3220 */
3221 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3222 if (ret)
3223 goto out;
3224 }
3225
3226 /*
3227 * wait for ordered IO before we have any locks. We'll loop again
3228 * below with the locks held.
3229 */
3230 ret = btrfs_wait_ordered_range(inode, alloc_start,
3231 alloc_end - alloc_start);
3232 if (ret)
3233 goto out;
3234
3235 if (mode & FALLOC_FL_ZERO_RANGE) {
3236 ret = btrfs_zero_range(inode, offset, len, mode);
3237 inode_unlock(inode);
3238 return ret;
3239 }
3240
3241 locked_end = alloc_end - 1;
3242 while (1) {
3243 struct btrfs_ordered_extent *ordered;
3244
3245 /* the extent lock is ordered inside the running
3246 * transaction
3247 */
3248 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3249 locked_end, &cached_state);
3250 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3251
3252 if (ordered &&
3253 ordered->file_offset + ordered->len > alloc_start &&
3254 ordered->file_offset < alloc_end) {
3255 btrfs_put_ordered_extent(ordered);
3256 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3257 alloc_start, locked_end,
3258 &cached_state);
3259 /*
3260 * we can't wait on the range with the transaction
3261 * running or with the extent lock held
3262 */
3263 ret = btrfs_wait_ordered_range(inode, alloc_start,
3264 alloc_end - alloc_start);
3265 if (ret)
3266 goto out;
3267 } else {
3268 if (ordered)
3269 btrfs_put_ordered_extent(ordered);
3270 break;
3271 }
3272 }
3273
3274 /* First, check if we exceed the qgroup limit */
3275 INIT_LIST_HEAD(&reserve_list);
3276 while (cur_offset < alloc_end) {
3277 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3278 alloc_end - cur_offset, 0);
3279 if (IS_ERR(em)) {
3280 ret = PTR_ERR(em);
3281 break;
3282 }
3283 last_byte = min(extent_map_end(em), alloc_end);
3284 actual_end = min_t(u64, extent_map_end(em), offset + len);
3285 last_byte = ALIGN(last_byte, blocksize);
3286 if (em->block_start == EXTENT_MAP_HOLE ||
3287 (cur_offset >= inode->i_size &&
3288 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3289 ret = add_falloc_range(&reserve_list, cur_offset,
3290 last_byte - cur_offset);
3291 if (ret < 0) {
3292 free_extent_map(em);
3293 break;
3294 }
3295 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3296 cur_offset, last_byte - cur_offset);
3297 if (ret < 0) {
3298 cur_offset = last_byte;
3299 free_extent_map(em);
3300 break;
3301 }
3302 } else {
3303 /*
3304 * Do not need to reserve unwritten extent for this
3305 * range, free reserved data space first, otherwise
3306 * it'll result in false ENOSPC error.
3307 */
3308 btrfs_free_reserved_data_space(inode, data_reserved,
3309 cur_offset, last_byte - cur_offset);
3310 }
3311 free_extent_map(em);
3312 cur_offset = last_byte;
3313 }
3314
3315 /*
3316 * If ret is still 0, means we're OK to fallocate.
3317 * Or just cleanup the list and exit.
3318 */
3319 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3320 if (!ret)
3321 ret = btrfs_prealloc_file_range(inode, mode,
3322 range->start,
3323 range->len, i_blocksize(inode),
3324 offset + len, &alloc_hint);
3325 else
3326 btrfs_free_reserved_data_space(inode,
3327 data_reserved, range->start,
3328 range->len);
3329 list_del(&range->list);
3330 kfree(range);
3331 }
3332 if (ret < 0)
3333 goto out_unlock;
3334
3335 /*
3336 * We didn't need to allocate any more space, but we still extended the
3337 * size of the file so we need to update i_size and the inode item.
3338 */
3339 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3340out_unlock:
3341 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3342 &cached_state);
3343out:
3344 inode_unlock(inode);
3345 /* Let go of our reservation. */
3346 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3347 btrfs_free_reserved_data_space(inode, data_reserved,
3348 cur_offset, alloc_end - cur_offset);
3349 extent_changeset_free(data_reserved);
3350 return ret;
3351}
3352
3353static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3354{
3355 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3356 struct extent_map *em = NULL;
3357 struct extent_state *cached_state = NULL;
3358 u64 lockstart;
3359 u64 lockend;
3360 u64 start;
3361 u64 len;
3362 int ret = 0;
3363
3364 if (inode->i_size == 0)
3365 return -ENXIO;
3366
3367 /*
3368 * *offset can be negative, in this case we start finding DATA/HOLE from
3369 * the very start of the file.
3370 */
3371 start = max_t(loff_t, 0, *offset);
3372
3373 lockstart = round_down(start, fs_info->sectorsize);
3374 lockend = round_up(i_size_read(inode),
3375 fs_info->sectorsize);
3376 if (lockend <= lockstart)
3377 lockend = lockstart + fs_info->sectorsize;
3378 lockend--;
3379 len = lockend - lockstart + 1;
3380
3381 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3382 &cached_state);
3383
3384 while (start < inode->i_size) {
3385 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3386 if (IS_ERR(em)) {
3387 ret = PTR_ERR(em);
3388 em = NULL;
3389 break;
3390 }
3391
3392 if (whence == SEEK_HOLE &&
3393 (em->block_start == EXTENT_MAP_HOLE ||
3394 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3395 break;
3396 else if (whence == SEEK_DATA &&
3397 (em->block_start != EXTENT_MAP_HOLE &&
3398 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3399 break;
3400
3401 start = em->start + em->len;
3402 free_extent_map(em);
3403 em = NULL;
3404 cond_resched();
3405 }
3406 free_extent_map(em);
3407 if (!ret) {
3408 if (whence == SEEK_DATA && start >= inode->i_size)
3409 ret = -ENXIO;
3410 else
3411 *offset = min_t(loff_t, start, inode->i_size);
3412 }
3413 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3414 &cached_state);
3415 return ret;
3416}
3417
3418static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3419{
3420 struct inode *inode = file->f_mapping->host;
3421 int ret;
3422
3423 inode_lock(inode);
3424 switch (whence) {
3425 case SEEK_END:
3426 case SEEK_CUR:
3427 offset = generic_file_llseek(file, offset, whence);
3428 goto out;
3429 case SEEK_DATA:
3430 case SEEK_HOLE:
3431 if (offset >= i_size_read(inode)) {
3432 inode_unlock(inode);
3433 return -ENXIO;
3434 }
3435
3436 ret = find_desired_extent(inode, &offset, whence);
3437 if (ret) {
3438 inode_unlock(inode);
3439 return ret;
3440 }
3441 }
3442
3443 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3444out:
3445 inode_unlock(inode);
3446 return offset;
3447}
3448
3449static int btrfs_file_open(struct inode *inode, struct file *filp)
3450{
3451 filp->f_mode |= FMODE_NOWAIT;
3452 return generic_file_open(inode, filp);
3453}
3454
3455const struct file_operations btrfs_file_operations = {
3456 .llseek = btrfs_file_llseek,
3457 .read_iter = generic_file_read_iter,
3458 .splice_read = generic_file_splice_read,
3459 .write_iter = btrfs_file_write_iter,
3460 .mmap = btrfs_file_mmap,
3461 .open = btrfs_file_open,
3462 .release = btrfs_release_file,
3463 .fsync = btrfs_sync_file,
3464 .fallocate = btrfs_fallocate,
3465 .unlocked_ioctl = btrfs_ioctl,
3466#ifdef CONFIG_COMPAT
3467 .compat_ioctl = btrfs_compat_ioctl,
3468#endif
3469 .remap_file_range = btrfs_remap_file_range,
3470};
3471
3472void __cold btrfs_auto_defrag_exit(void)
3473{
3474 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3475}
3476
3477int __init btrfs_auto_defrag_init(void)
3478{
3479 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3480 sizeof(struct inode_defrag), 0,
3481 SLAB_MEM_SPREAD,
3482 NULL);
3483 if (!btrfs_inode_defrag_cachep)
3484 return -ENOMEM;
3485
3486 return 0;
3487}
3488
3489int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3490{
3491 int ret;
3492
3493 /*
3494 * So with compression we will find and lock a dirty page and clear the
3495 * first one as dirty, setup an async extent, and immediately return
3496 * with the entire range locked but with nobody actually marked with
3497 * writeback. So we can't just filemap_write_and_wait_range() and
3498 * expect it to work since it will just kick off a thread to do the
3499 * actual work. So we need to call filemap_fdatawrite_range _again_
3500 * since it will wait on the page lock, which won't be unlocked until
3501 * after the pages have been marked as writeback and so we're good to go
3502 * from there. We have to do this otherwise we'll miss the ordered
3503 * extents and that results in badness. Please Josef, do not think you
3504 * know better and pull this out at some point in the future, it is
3505 * right and you are wrong.
3506 */
3507 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3508 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3509 &BTRFS_I(inode)->runtime_flags))
3510 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3511
3512 return ret;
3513}
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 <linux/fsverity.h>
20#include "ctree.h"
21#include "disk-io.h"
22#include "transaction.h"
23#include "btrfs_inode.h"
24#include "print-tree.h"
25#include "tree-log.h"
26#include "locking.h"
27#include "volumes.h"
28#include "qgroup.h"
29#include "compression.h"
30#include "delalloc-space.h"
31#include "reflink.h"
32#include "subpage.h"
33#include "fs.h"
34#include "accessors.h"
35#include "extent-tree.h"
36#include "file-item.h"
37#include "ioctl.h"
38#include "file.h"
39#include "super.h"
40
41/* simple helper to fault in pages and copy. This should go away
42 * and be replaced with calls into generic code.
43 */
44static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
45 struct page **prepared_pages,
46 struct iov_iter *i)
47{
48 size_t copied = 0;
49 size_t total_copied = 0;
50 int pg = 0;
51 int offset = offset_in_page(pos);
52
53 while (write_bytes > 0) {
54 size_t count = min_t(size_t,
55 PAGE_SIZE - offset, write_bytes);
56 struct page *page = prepared_pages[pg];
57 /*
58 * Copy data from userspace to the current page
59 */
60 copied = copy_page_from_iter_atomic(page, offset, count, i);
61
62 /* Flush processor's dcache for this page */
63 flush_dcache_page(page);
64
65 /*
66 * if we get a partial write, we can end up with
67 * partially up to date pages. These add
68 * a lot of complexity, so make sure they don't
69 * happen by forcing this copy to be retried.
70 *
71 * The rest of the btrfs_file_write code will fall
72 * back to page at a time copies after we return 0.
73 */
74 if (unlikely(copied < count)) {
75 if (!PageUptodate(page)) {
76 iov_iter_revert(i, copied);
77 copied = 0;
78 }
79 if (!copied)
80 break;
81 }
82
83 write_bytes -= copied;
84 total_copied += copied;
85 offset += copied;
86 if (offset == PAGE_SIZE) {
87 pg++;
88 offset = 0;
89 }
90 }
91 return total_copied;
92}
93
94/*
95 * unlocks pages after btrfs_file_write is done with them
96 */
97static void btrfs_drop_pages(struct btrfs_fs_info *fs_info,
98 struct page **pages, size_t num_pages,
99 u64 pos, u64 copied)
100{
101 size_t i;
102 u64 block_start = round_down(pos, fs_info->sectorsize);
103 u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
104
105 ASSERT(block_len <= U32_MAX);
106 for (i = 0; i < num_pages; i++) {
107 /* page checked is some magic around finding pages that
108 * have been modified without going through btrfs_set_page_dirty
109 * clear it here. There should be no need to mark the pages
110 * accessed as prepare_pages should have marked them accessed
111 * in prepare_pages via find_or_create_page()
112 */
113 btrfs_page_clamp_clear_checked(fs_info, pages[i], block_start,
114 block_len);
115 unlock_page(pages[i]);
116 put_page(pages[i]);
117 }
118}
119
120/*
121 * After btrfs_copy_from_user(), update the following things for delalloc:
122 * - Mark newly dirtied pages as DELALLOC in the io tree.
123 * Used to advise which range is to be written back.
124 * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
125 * - Update inode size for past EOF write
126 */
127int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
128 size_t num_pages, loff_t pos, size_t write_bytes,
129 struct extent_state **cached, bool noreserve)
130{
131 struct btrfs_fs_info *fs_info = inode->root->fs_info;
132 int err = 0;
133 int i;
134 u64 num_bytes;
135 u64 start_pos;
136 u64 end_of_last_block;
137 u64 end_pos = pos + write_bytes;
138 loff_t isize = i_size_read(&inode->vfs_inode);
139 unsigned int extra_bits = 0;
140
141 if (write_bytes == 0)
142 return 0;
143
144 if (noreserve)
145 extra_bits |= EXTENT_NORESERVE;
146
147 start_pos = round_down(pos, fs_info->sectorsize);
148 num_bytes = round_up(write_bytes + pos - start_pos,
149 fs_info->sectorsize);
150 ASSERT(num_bytes <= U32_MAX);
151
152 end_of_last_block = start_pos + num_bytes - 1;
153
154 /*
155 * The pages may have already been dirty, clear out old accounting so
156 * we can set things up properly
157 */
158 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
159 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
160 cached);
161
162 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
163 extra_bits, cached);
164 if (err)
165 return err;
166
167 for (i = 0; i < num_pages; i++) {
168 struct page *p = pages[i];
169
170 btrfs_page_clamp_set_uptodate(fs_info, p, start_pos, num_bytes);
171 btrfs_page_clamp_clear_checked(fs_info, p, start_pos, num_bytes);
172 btrfs_page_clamp_set_dirty(fs_info, p, start_pos, num_bytes);
173 }
174
175 /*
176 * we've only changed i_size in ram, and we haven't updated
177 * the disk i_size. There is no need to log the inode
178 * at this time.
179 */
180 if (end_pos > isize)
181 i_size_write(&inode->vfs_inode, end_pos);
182 return 0;
183}
184
185/*
186 * this is very complex, but the basic idea is to drop all extents
187 * in the range start - end. hint_block is filled in with a block number
188 * that would be a good hint to the block allocator for this file.
189 *
190 * If an extent intersects the range but is not entirely inside the range
191 * it is either truncated or split. Anything entirely inside the range
192 * is deleted from the tree.
193 *
194 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
195 * to deal with that. We set the field 'bytes_found' of the arguments structure
196 * with the number of allocated bytes found in the target range, so that the
197 * caller can update the inode's number of bytes in an atomic way when
198 * replacing extents in a range to avoid races with stat(2).
199 */
200int btrfs_drop_extents(struct btrfs_trans_handle *trans,
201 struct btrfs_root *root, struct btrfs_inode *inode,
202 struct btrfs_drop_extents_args *args)
203{
204 struct btrfs_fs_info *fs_info = root->fs_info;
205 struct extent_buffer *leaf;
206 struct btrfs_file_extent_item *fi;
207 struct btrfs_ref ref = { 0 };
208 struct btrfs_key key;
209 struct btrfs_key new_key;
210 u64 ino = btrfs_ino(inode);
211 u64 search_start = args->start;
212 u64 disk_bytenr = 0;
213 u64 num_bytes = 0;
214 u64 extent_offset = 0;
215 u64 extent_end = 0;
216 u64 last_end = args->start;
217 int del_nr = 0;
218 int del_slot = 0;
219 int extent_type;
220 int recow;
221 int ret;
222 int modify_tree = -1;
223 int update_refs;
224 int found = 0;
225 struct btrfs_path *path = args->path;
226
227 args->bytes_found = 0;
228 args->extent_inserted = false;
229
230 /* Must always have a path if ->replace_extent is true */
231 ASSERT(!(args->replace_extent && !args->path));
232
233 if (!path) {
234 path = btrfs_alloc_path();
235 if (!path) {
236 ret = -ENOMEM;
237 goto out;
238 }
239 }
240
241 if (args->drop_cache)
242 btrfs_drop_extent_map_range(inode, args->start, args->end - 1, false);
243
244 if (args->start >= inode->disk_i_size && !args->replace_extent)
245 modify_tree = 0;
246
247 update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
248 while (1) {
249 recow = 0;
250 ret = btrfs_lookup_file_extent(trans, root, path, ino,
251 search_start, modify_tree);
252 if (ret < 0)
253 break;
254 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
255 leaf = path->nodes[0];
256 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
257 if (key.objectid == ino &&
258 key.type == BTRFS_EXTENT_DATA_KEY)
259 path->slots[0]--;
260 }
261 ret = 0;
262next_slot:
263 leaf = path->nodes[0];
264 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
265 BUG_ON(del_nr > 0);
266 ret = btrfs_next_leaf(root, path);
267 if (ret < 0)
268 break;
269 if (ret > 0) {
270 ret = 0;
271 break;
272 }
273 leaf = path->nodes[0];
274 recow = 1;
275 }
276
277 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
278
279 if (key.objectid > ino)
280 break;
281 if (WARN_ON_ONCE(key.objectid < ino) ||
282 key.type < BTRFS_EXTENT_DATA_KEY) {
283 ASSERT(del_nr == 0);
284 path->slots[0]++;
285 goto next_slot;
286 }
287 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
288 break;
289
290 fi = btrfs_item_ptr(leaf, path->slots[0],
291 struct btrfs_file_extent_item);
292 extent_type = btrfs_file_extent_type(leaf, fi);
293
294 if (extent_type == BTRFS_FILE_EXTENT_REG ||
295 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
296 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
297 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
298 extent_offset = btrfs_file_extent_offset(leaf, fi);
299 extent_end = key.offset +
300 btrfs_file_extent_num_bytes(leaf, fi);
301 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
302 extent_end = key.offset +
303 btrfs_file_extent_ram_bytes(leaf, fi);
304 } else {
305 /* can't happen */
306 BUG();
307 }
308
309 /*
310 * Don't skip extent items representing 0 byte lengths. They
311 * used to be created (bug) if while punching holes we hit
312 * -ENOSPC condition. So if we find one here, just ensure we
313 * delete it, otherwise we would insert a new file extent item
314 * with the same key (offset) as that 0 bytes length file
315 * extent item in the call to setup_items_for_insert() later
316 * in this function.
317 */
318 if (extent_end == key.offset && extent_end >= search_start) {
319 last_end = extent_end;
320 goto delete_extent_item;
321 }
322
323 if (extent_end <= search_start) {
324 path->slots[0]++;
325 goto next_slot;
326 }
327
328 found = 1;
329 search_start = max(key.offset, args->start);
330 if (recow || !modify_tree) {
331 modify_tree = -1;
332 btrfs_release_path(path);
333 continue;
334 }
335
336 /*
337 * | - range to drop - |
338 * | -------- extent -------- |
339 */
340 if (args->start > key.offset && args->end < extent_end) {
341 BUG_ON(del_nr > 0);
342 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
343 ret = -EOPNOTSUPP;
344 break;
345 }
346
347 memcpy(&new_key, &key, sizeof(new_key));
348 new_key.offset = args->start;
349 ret = btrfs_duplicate_item(trans, root, path,
350 &new_key);
351 if (ret == -EAGAIN) {
352 btrfs_release_path(path);
353 continue;
354 }
355 if (ret < 0)
356 break;
357
358 leaf = path->nodes[0];
359 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
360 struct btrfs_file_extent_item);
361 btrfs_set_file_extent_num_bytes(leaf, fi,
362 args->start - key.offset);
363
364 fi = btrfs_item_ptr(leaf, path->slots[0],
365 struct btrfs_file_extent_item);
366
367 extent_offset += args->start - key.offset;
368 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
369 btrfs_set_file_extent_num_bytes(leaf, fi,
370 extent_end - args->start);
371 btrfs_mark_buffer_dirty(leaf);
372
373 if (update_refs && disk_bytenr > 0) {
374 btrfs_init_generic_ref(&ref,
375 BTRFS_ADD_DELAYED_REF,
376 disk_bytenr, num_bytes, 0);
377 btrfs_init_data_ref(&ref,
378 root->root_key.objectid,
379 new_key.objectid,
380 args->start - extent_offset,
381 0, false);
382 ret = btrfs_inc_extent_ref(trans, &ref);
383 if (ret) {
384 btrfs_abort_transaction(trans, ret);
385 break;
386 }
387 }
388 key.offset = args->start;
389 }
390 /*
391 * From here on out we will have actually dropped something, so
392 * last_end can be updated.
393 */
394 last_end = extent_end;
395
396 /*
397 * | ---- range to drop ----- |
398 * | -------- extent -------- |
399 */
400 if (args->start <= key.offset && args->end < extent_end) {
401 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
402 ret = -EOPNOTSUPP;
403 break;
404 }
405
406 memcpy(&new_key, &key, sizeof(new_key));
407 new_key.offset = args->end;
408 btrfs_set_item_key_safe(fs_info, path, &new_key);
409
410 extent_offset += args->end - key.offset;
411 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
412 btrfs_set_file_extent_num_bytes(leaf, fi,
413 extent_end - args->end);
414 btrfs_mark_buffer_dirty(leaf);
415 if (update_refs && disk_bytenr > 0)
416 args->bytes_found += args->end - key.offset;
417 break;
418 }
419
420 search_start = extent_end;
421 /*
422 * | ---- range to drop ----- |
423 * | -------- extent -------- |
424 */
425 if (args->start > key.offset && args->end >= extent_end) {
426 BUG_ON(del_nr > 0);
427 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
428 ret = -EOPNOTSUPP;
429 break;
430 }
431
432 btrfs_set_file_extent_num_bytes(leaf, fi,
433 args->start - key.offset);
434 btrfs_mark_buffer_dirty(leaf);
435 if (update_refs && disk_bytenr > 0)
436 args->bytes_found += extent_end - args->start;
437 if (args->end == extent_end)
438 break;
439
440 path->slots[0]++;
441 goto next_slot;
442 }
443
444 /*
445 * | ---- range to drop ----- |
446 * | ------ extent ------ |
447 */
448 if (args->start <= key.offset && args->end >= extent_end) {
449delete_extent_item:
450 if (del_nr == 0) {
451 del_slot = path->slots[0];
452 del_nr = 1;
453 } else {
454 BUG_ON(del_slot + del_nr != path->slots[0]);
455 del_nr++;
456 }
457
458 if (update_refs &&
459 extent_type == BTRFS_FILE_EXTENT_INLINE) {
460 args->bytes_found += extent_end - key.offset;
461 extent_end = ALIGN(extent_end,
462 fs_info->sectorsize);
463 } else if (update_refs && disk_bytenr > 0) {
464 btrfs_init_generic_ref(&ref,
465 BTRFS_DROP_DELAYED_REF,
466 disk_bytenr, num_bytes, 0);
467 btrfs_init_data_ref(&ref,
468 root->root_key.objectid,
469 key.objectid,
470 key.offset - extent_offset, 0,
471 false);
472 ret = btrfs_free_extent(trans, &ref);
473 if (ret) {
474 btrfs_abort_transaction(trans, ret);
475 break;
476 }
477 args->bytes_found += extent_end - key.offset;
478 }
479
480 if (args->end == extent_end)
481 break;
482
483 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
484 path->slots[0]++;
485 goto next_slot;
486 }
487
488 ret = btrfs_del_items(trans, root, path, del_slot,
489 del_nr);
490 if (ret) {
491 btrfs_abort_transaction(trans, ret);
492 break;
493 }
494
495 del_nr = 0;
496 del_slot = 0;
497
498 btrfs_release_path(path);
499 continue;
500 }
501
502 BUG();
503 }
504
505 if (!ret && del_nr > 0) {
506 /*
507 * Set path->slots[0] to first slot, so that after the delete
508 * if items are move off from our leaf to its immediate left or
509 * right neighbor leafs, we end up with a correct and adjusted
510 * path->slots[0] for our insertion (if args->replace_extent).
511 */
512 path->slots[0] = del_slot;
513 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
514 if (ret)
515 btrfs_abort_transaction(trans, ret);
516 }
517
518 leaf = path->nodes[0];
519 /*
520 * If btrfs_del_items() was called, it might have deleted a leaf, in
521 * which case it unlocked our path, so check path->locks[0] matches a
522 * write lock.
523 */
524 if (!ret && args->replace_extent &&
525 path->locks[0] == BTRFS_WRITE_LOCK &&
526 btrfs_leaf_free_space(leaf) >=
527 sizeof(struct btrfs_item) + args->extent_item_size) {
528
529 key.objectid = ino;
530 key.type = BTRFS_EXTENT_DATA_KEY;
531 key.offset = args->start;
532 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
533 struct btrfs_key slot_key;
534
535 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
536 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
537 path->slots[0]++;
538 }
539 btrfs_setup_item_for_insert(root, path, &key, args->extent_item_size);
540 args->extent_inserted = true;
541 }
542
543 if (!args->path)
544 btrfs_free_path(path);
545 else if (!args->extent_inserted)
546 btrfs_release_path(path);
547out:
548 args->drop_end = found ? min(args->end, last_end) : args->end;
549
550 return ret;
551}
552
553static int extent_mergeable(struct extent_buffer *leaf, int slot,
554 u64 objectid, u64 bytenr, u64 orig_offset,
555 u64 *start, u64 *end)
556{
557 struct btrfs_file_extent_item *fi;
558 struct btrfs_key key;
559 u64 extent_end;
560
561 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
562 return 0;
563
564 btrfs_item_key_to_cpu(leaf, &key, slot);
565 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
566 return 0;
567
568 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
569 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
570 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
571 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
572 btrfs_file_extent_compression(leaf, fi) ||
573 btrfs_file_extent_encryption(leaf, fi) ||
574 btrfs_file_extent_other_encoding(leaf, fi))
575 return 0;
576
577 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
578 if ((*start && *start != key.offset) || (*end && *end != extent_end))
579 return 0;
580
581 *start = key.offset;
582 *end = extent_end;
583 return 1;
584}
585
586/*
587 * Mark extent in the range start - end as written.
588 *
589 * This changes extent type from 'pre-allocated' to 'regular'. If only
590 * part of extent is marked as written, the extent will be split into
591 * two or three.
592 */
593int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
594 struct btrfs_inode *inode, u64 start, u64 end)
595{
596 struct btrfs_fs_info *fs_info = trans->fs_info;
597 struct btrfs_root *root = inode->root;
598 struct extent_buffer *leaf;
599 struct btrfs_path *path;
600 struct btrfs_file_extent_item *fi;
601 struct btrfs_ref ref = { 0 };
602 struct btrfs_key key;
603 struct btrfs_key new_key;
604 u64 bytenr;
605 u64 num_bytes;
606 u64 extent_end;
607 u64 orig_offset;
608 u64 other_start;
609 u64 other_end;
610 u64 split;
611 int del_nr = 0;
612 int del_slot = 0;
613 int recow;
614 int ret = 0;
615 u64 ino = btrfs_ino(inode);
616
617 path = btrfs_alloc_path();
618 if (!path)
619 return -ENOMEM;
620again:
621 recow = 0;
622 split = start;
623 key.objectid = ino;
624 key.type = BTRFS_EXTENT_DATA_KEY;
625 key.offset = split;
626
627 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
628 if (ret < 0)
629 goto out;
630 if (ret > 0 && path->slots[0] > 0)
631 path->slots[0]--;
632
633 leaf = path->nodes[0];
634 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
635 if (key.objectid != ino ||
636 key.type != BTRFS_EXTENT_DATA_KEY) {
637 ret = -EINVAL;
638 btrfs_abort_transaction(trans, ret);
639 goto out;
640 }
641 fi = btrfs_item_ptr(leaf, path->slots[0],
642 struct btrfs_file_extent_item);
643 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
644 ret = -EINVAL;
645 btrfs_abort_transaction(trans, ret);
646 goto out;
647 }
648 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
649 if (key.offset > start || extent_end < end) {
650 ret = -EINVAL;
651 btrfs_abort_transaction(trans, ret);
652 goto out;
653 }
654
655 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
656 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
657 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
658 memcpy(&new_key, &key, sizeof(new_key));
659
660 if (start == key.offset && end < extent_end) {
661 other_start = 0;
662 other_end = start;
663 if (extent_mergeable(leaf, path->slots[0] - 1,
664 ino, bytenr, orig_offset,
665 &other_start, &other_end)) {
666 new_key.offset = end;
667 btrfs_set_item_key_safe(fs_info, path, &new_key);
668 fi = btrfs_item_ptr(leaf, path->slots[0],
669 struct btrfs_file_extent_item);
670 btrfs_set_file_extent_generation(leaf, fi,
671 trans->transid);
672 btrfs_set_file_extent_num_bytes(leaf, fi,
673 extent_end - end);
674 btrfs_set_file_extent_offset(leaf, fi,
675 end - orig_offset);
676 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
677 struct btrfs_file_extent_item);
678 btrfs_set_file_extent_generation(leaf, fi,
679 trans->transid);
680 btrfs_set_file_extent_num_bytes(leaf, fi,
681 end - other_start);
682 btrfs_mark_buffer_dirty(leaf);
683 goto out;
684 }
685 }
686
687 if (start > key.offset && end == extent_end) {
688 other_start = end;
689 other_end = 0;
690 if (extent_mergeable(leaf, path->slots[0] + 1,
691 ino, bytenr, orig_offset,
692 &other_start, &other_end)) {
693 fi = btrfs_item_ptr(leaf, path->slots[0],
694 struct btrfs_file_extent_item);
695 btrfs_set_file_extent_num_bytes(leaf, fi,
696 start - key.offset);
697 btrfs_set_file_extent_generation(leaf, fi,
698 trans->transid);
699 path->slots[0]++;
700 new_key.offset = start;
701 btrfs_set_item_key_safe(fs_info, path, &new_key);
702
703 fi = btrfs_item_ptr(leaf, path->slots[0],
704 struct btrfs_file_extent_item);
705 btrfs_set_file_extent_generation(leaf, fi,
706 trans->transid);
707 btrfs_set_file_extent_num_bytes(leaf, fi,
708 other_end - start);
709 btrfs_set_file_extent_offset(leaf, fi,
710 start - orig_offset);
711 btrfs_mark_buffer_dirty(leaf);
712 goto out;
713 }
714 }
715
716 while (start > key.offset || end < extent_end) {
717 if (key.offset == start)
718 split = end;
719
720 new_key.offset = split;
721 ret = btrfs_duplicate_item(trans, root, path, &new_key);
722 if (ret == -EAGAIN) {
723 btrfs_release_path(path);
724 goto again;
725 }
726 if (ret < 0) {
727 btrfs_abort_transaction(trans, ret);
728 goto out;
729 }
730
731 leaf = path->nodes[0];
732 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
733 struct btrfs_file_extent_item);
734 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
735 btrfs_set_file_extent_num_bytes(leaf, fi,
736 split - key.offset);
737
738 fi = btrfs_item_ptr(leaf, path->slots[0],
739 struct btrfs_file_extent_item);
740
741 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
742 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
743 btrfs_set_file_extent_num_bytes(leaf, fi,
744 extent_end - split);
745 btrfs_mark_buffer_dirty(leaf);
746
747 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
748 num_bytes, 0);
749 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
750 orig_offset, 0, false);
751 ret = btrfs_inc_extent_ref(trans, &ref);
752 if (ret) {
753 btrfs_abort_transaction(trans, ret);
754 goto out;
755 }
756
757 if (split == start) {
758 key.offset = start;
759 } else {
760 if (start != key.offset) {
761 ret = -EINVAL;
762 btrfs_abort_transaction(trans, ret);
763 goto out;
764 }
765 path->slots[0]--;
766 extent_end = end;
767 }
768 recow = 1;
769 }
770
771 other_start = end;
772 other_end = 0;
773 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
774 num_bytes, 0);
775 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset,
776 0, false);
777 if (extent_mergeable(leaf, path->slots[0] + 1,
778 ino, bytenr, orig_offset,
779 &other_start, &other_end)) {
780 if (recow) {
781 btrfs_release_path(path);
782 goto again;
783 }
784 extent_end = other_end;
785 del_slot = path->slots[0] + 1;
786 del_nr++;
787 ret = btrfs_free_extent(trans, &ref);
788 if (ret) {
789 btrfs_abort_transaction(trans, ret);
790 goto out;
791 }
792 }
793 other_start = 0;
794 other_end = start;
795 if (extent_mergeable(leaf, path->slots[0] - 1,
796 ino, bytenr, orig_offset,
797 &other_start, &other_end)) {
798 if (recow) {
799 btrfs_release_path(path);
800 goto again;
801 }
802 key.offset = other_start;
803 del_slot = path->slots[0];
804 del_nr++;
805 ret = btrfs_free_extent(trans, &ref);
806 if (ret) {
807 btrfs_abort_transaction(trans, ret);
808 goto out;
809 }
810 }
811 if (del_nr == 0) {
812 fi = btrfs_item_ptr(leaf, path->slots[0],
813 struct btrfs_file_extent_item);
814 btrfs_set_file_extent_type(leaf, fi,
815 BTRFS_FILE_EXTENT_REG);
816 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
817 btrfs_mark_buffer_dirty(leaf);
818 } else {
819 fi = btrfs_item_ptr(leaf, del_slot - 1,
820 struct btrfs_file_extent_item);
821 btrfs_set_file_extent_type(leaf, fi,
822 BTRFS_FILE_EXTENT_REG);
823 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
824 btrfs_set_file_extent_num_bytes(leaf, fi,
825 extent_end - key.offset);
826 btrfs_mark_buffer_dirty(leaf);
827
828 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
829 if (ret < 0) {
830 btrfs_abort_transaction(trans, ret);
831 goto out;
832 }
833 }
834out:
835 btrfs_free_path(path);
836 return ret;
837}
838
839/*
840 * on error we return an unlocked page and the error value
841 * on success we return a locked page and 0
842 */
843static int prepare_uptodate_page(struct inode *inode,
844 struct page *page, u64 pos,
845 bool force_uptodate)
846{
847 struct folio *folio = page_folio(page);
848 int ret = 0;
849
850 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
851 !PageUptodate(page)) {
852 ret = btrfs_read_folio(NULL, folio);
853 if (ret)
854 return ret;
855 lock_page(page);
856 if (!PageUptodate(page)) {
857 unlock_page(page);
858 return -EIO;
859 }
860
861 /*
862 * Since btrfs_read_folio() will unlock the folio before it
863 * returns, there is a window where btrfs_release_folio() can be
864 * called to release the page. Here we check both inode
865 * mapping and PagePrivate() to make sure the page was not
866 * released.
867 *
868 * The private flag check is essential for subpage as we need
869 * to store extra bitmap using page->private.
870 */
871 if (page->mapping != inode->i_mapping || !PagePrivate(page)) {
872 unlock_page(page);
873 return -EAGAIN;
874 }
875 }
876 return 0;
877}
878
879static unsigned int get_prepare_fgp_flags(bool nowait)
880{
881 unsigned int fgp_flags = FGP_LOCK | FGP_ACCESSED | FGP_CREAT;
882
883 if (nowait)
884 fgp_flags |= FGP_NOWAIT;
885
886 return fgp_flags;
887}
888
889static gfp_t get_prepare_gfp_flags(struct inode *inode, bool nowait)
890{
891 gfp_t gfp;
892
893 gfp = btrfs_alloc_write_mask(inode->i_mapping);
894 if (nowait) {
895 gfp &= ~__GFP_DIRECT_RECLAIM;
896 gfp |= GFP_NOWAIT;
897 }
898
899 return gfp;
900}
901
902/*
903 * this just gets pages into the page cache and locks them down.
904 */
905static noinline int prepare_pages(struct inode *inode, struct page **pages,
906 size_t num_pages, loff_t pos,
907 size_t write_bytes, bool force_uptodate,
908 bool nowait)
909{
910 int i;
911 unsigned long index = pos >> PAGE_SHIFT;
912 gfp_t mask = get_prepare_gfp_flags(inode, nowait);
913 unsigned int fgp_flags = get_prepare_fgp_flags(nowait);
914 int err = 0;
915 int faili;
916
917 for (i = 0; i < num_pages; i++) {
918again:
919 pages[i] = pagecache_get_page(inode->i_mapping, index + i,
920 fgp_flags, mask | __GFP_WRITE);
921 if (!pages[i]) {
922 faili = i - 1;
923 if (nowait)
924 err = -EAGAIN;
925 else
926 err = -ENOMEM;
927 goto fail;
928 }
929
930 err = set_page_extent_mapped(pages[i]);
931 if (err < 0) {
932 faili = i;
933 goto fail;
934 }
935
936 if (i == 0)
937 err = prepare_uptodate_page(inode, pages[i], pos,
938 force_uptodate);
939 if (!err && i == num_pages - 1)
940 err = prepare_uptodate_page(inode, pages[i],
941 pos + write_bytes, false);
942 if (err) {
943 put_page(pages[i]);
944 if (!nowait && err == -EAGAIN) {
945 err = 0;
946 goto again;
947 }
948 faili = i - 1;
949 goto fail;
950 }
951 wait_on_page_writeback(pages[i]);
952 }
953
954 return 0;
955fail:
956 while (faili >= 0) {
957 unlock_page(pages[faili]);
958 put_page(pages[faili]);
959 faili--;
960 }
961 return err;
962
963}
964
965/*
966 * This function locks the extent and properly waits for data=ordered extents
967 * to finish before allowing the pages to be modified if need.
968 *
969 * The return value:
970 * 1 - the extent is locked
971 * 0 - the extent is not locked, and everything is OK
972 * -EAGAIN - need re-prepare the pages
973 * the other < 0 number - Something wrong happens
974 */
975static noinline int
976lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
977 size_t num_pages, loff_t pos,
978 size_t write_bytes,
979 u64 *lockstart, u64 *lockend, bool nowait,
980 struct extent_state **cached_state)
981{
982 struct btrfs_fs_info *fs_info = inode->root->fs_info;
983 u64 start_pos;
984 u64 last_pos;
985 int i;
986 int ret = 0;
987
988 start_pos = round_down(pos, fs_info->sectorsize);
989 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
990
991 if (start_pos < inode->vfs_inode.i_size) {
992 struct btrfs_ordered_extent *ordered;
993
994 if (nowait) {
995 if (!try_lock_extent(&inode->io_tree, start_pos, last_pos,
996 cached_state)) {
997 for (i = 0; i < num_pages; i++) {
998 unlock_page(pages[i]);
999 put_page(pages[i]);
1000 pages[i] = NULL;
1001 }
1002
1003 return -EAGAIN;
1004 }
1005 } else {
1006 lock_extent(&inode->io_tree, start_pos, last_pos, cached_state);
1007 }
1008
1009 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1010 last_pos - start_pos + 1);
1011 if (ordered &&
1012 ordered->file_offset + ordered->num_bytes > start_pos &&
1013 ordered->file_offset <= last_pos) {
1014 unlock_extent(&inode->io_tree, start_pos, last_pos,
1015 cached_state);
1016 for (i = 0; i < num_pages; i++) {
1017 unlock_page(pages[i]);
1018 put_page(pages[i]);
1019 }
1020 btrfs_start_ordered_extent(ordered, 1);
1021 btrfs_put_ordered_extent(ordered);
1022 return -EAGAIN;
1023 }
1024 if (ordered)
1025 btrfs_put_ordered_extent(ordered);
1026
1027 *lockstart = start_pos;
1028 *lockend = last_pos;
1029 ret = 1;
1030 }
1031
1032 /*
1033 * We should be called after prepare_pages() which should have locked
1034 * all pages in the range.
1035 */
1036 for (i = 0; i < num_pages; i++)
1037 WARN_ON(!PageLocked(pages[i]));
1038
1039 return ret;
1040}
1041
1042/*
1043 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1044 *
1045 * @pos: File offset.
1046 * @write_bytes: The length to write, will be updated to the nocow writeable
1047 * range.
1048 *
1049 * This function will flush ordered extents in the range to ensure proper
1050 * nocow checks.
1051 *
1052 * Return:
1053 * > 0 If we can nocow, and updates @write_bytes.
1054 * 0 If we can't do a nocow write.
1055 * -EAGAIN If we can't do a nocow write because snapshoting of the inode's
1056 * root is in progress.
1057 * < 0 If an error happened.
1058 *
1059 * NOTE: Callers need to call btrfs_check_nocow_unlock() if we return > 0.
1060 */
1061int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1062 size_t *write_bytes, bool nowait)
1063{
1064 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1065 struct btrfs_root *root = inode->root;
1066 struct extent_state *cached_state = NULL;
1067 u64 lockstart, lockend;
1068 u64 num_bytes;
1069 int ret;
1070
1071 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1072 return 0;
1073
1074 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1075 return -EAGAIN;
1076
1077 lockstart = round_down(pos, fs_info->sectorsize);
1078 lockend = round_up(pos + *write_bytes,
1079 fs_info->sectorsize) - 1;
1080 num_bytes = lockend - lockstart + 1;
1081
1082 if (nowait) {
1083 if (!btrfs_try_lock_ordered_range(inode, lockstart, lockend,
1084 &cached_state)) {
1085 btrfs_drew_write_unlock(&root->snapshot_lock);
1086 return -EAGAIN;
1087 }
1088 } else {
1089 btrfs_lock_and_flush_ordered_range(inode, lockstart, lockend,
1090 &cached_state);
1091 }
1092 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1093 NULL, NULL, NULL, nowait, false);
1094 if (ret <= 0)
1095 btrfs_drew_write_unlock(&root->snapshot_lock);
1096 else
1097 *write_bytes = min_t(size_t, *write_bytes ,
1098 num_bytes - pos + lockstart);
1099 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
1100
1101 return ret;
1102}
1103
1104void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1105{
1106 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1107}
1108
1109static void update_time_for_write(struct inode *inode)
1110{
1111 struct timespec64 now;
1112
1113 if (IS_NOCMTIME(inode))
1114 return;
1115
1116 now = current_time(inode);
1117 if (!timespec64_equal(&inode->i_mtime, &now))
1118 inode->i_mtime = now;
1119
1120 if (!timespec64_equal(&inode->i_ctime, &now))
1121 inode->i_ctime = now;
1122
1123 if (IS_I_VERSION(inode))
1124 inode_inc_iversion(inode);
1125}
1126
1127static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1128 size_t count)
1129{
1130 struct file *file = iocb->ki_filp;
1131 struct inode *inode = file_inode(file);
1132 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1133 loff_t pos = iocb->ki_pos;
1134 int ret;
1135 loff_t oldsize;
1136 loff_t start_pos;
1137
1138 /*
1139 * Quickly bail out on NOWAIT writes if we don't have the nodatacow or
1140 * prealloc flags, as without those flags we always have to COW. We will
1141 * later check if we can really COW into the target range (using
1142 * can_nocow_extent() at btrfs_get_blocks_direct_write()).
1143 */
1144 if ((iocb->ki_flags & IOCB_NOWAIT) &&
1145 !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1146 return -EAGAIN;
1147
1148 current->backing_dev_info = inode_to_bdi(inode);
1149 ret = file_remove_privs(file);
1150 if (ret)
1151 return ret;
1152
1153 /*
1154 * We reserve space for updating the inode when we reserve space for the
1155 * extent we are going to write, so we will enospc out there. We don't
1156 * need to start yet another transaction to update the inode as we will
1157 * update the inode when we finish writing whatever data we write.
1158 */
1159 update_time_for_write(inode);
1160
1161 start_pos = round_down(pos, fs_info->sectorsize);
1162 oldsize = i_size_read(inode);
1163 if (start_pos > oldsize) {
1164 /* Expand hole size to cover write data, preventing empty gap */
1165 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1166
1167 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1168 if (ret) {
1169 current->backing_dev_info = NULL;
1170 return ret;
1171 }
1172 }
1173
1174 return 0;
1175}
1176
1177static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1178 struct iov_iter *i)
1179{
1180 struct file *file = iocb->ki_filp;
1181 loff_t pos;
1182 struct inode *inode = file_inode(file);
1183 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1184 struct page **pages = NULL;
1185 struct extent_changeset *data_reserved = NULL;
1186 u64 release_bytes = 0;
1187 u64 lockstart;
1188 u64 lockend;
1189 size_t num_written = 0;
1190 int nrptrs;
1191 ssize_t ret;
1192 bool only_release_metadata = false;
1193 bool force_page_uptodate = false;
1194 loff_t old_isize = i_size_read(inode);
1195 unsigned int ilock_flags = 0;
1196 const bool nowait = (iocb->ki_flags & IOCB_NOWAIT);
1197 unsigned int bdp_flags = (nowait ? BDP_ASYNC : 0);
1198
1199 if (nowait)
1200 ilock_flags |= BTRFS_ILOCK_TRY;
1201
1202 ret = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1203 if (ret < 0)
1204 return ret;
1205
1206 ret = generic_write_checks(iocb, i);
1207 if (ret <= 0)
1208 goto out;
1209
1210 ret = btrfs_write_check(iocb, i, ret);
1211 if (ret < 0)
1212 goto out;
1213
1214 pos = iocb->ki_pos;
1215 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1216 PAGE_SIZE / (sizeof(struct page *)));
1217 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1218 nrptrs = max(nrptrs, 8);
1219 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1220 if (!pages) {
1221 ret = -ENOMEM;
1222 goto out;
1223 }
1224
1225 while (iov_iter_count(i) > 0) {
1226 struct extent_state *cached_state = NULL;
1227 size_t offset = offset_in_page(pos);
1228 size_t sector_offset;
1229 size_t write_bytes = min(iov_iter_count(i),
1230 nrptrs * (size_t)PAGE_SIZE -
1231 offset);
1232 size_t num_pages;
1233 size_t reserve_bytes;
1234 size_t dirty_pages;
1235 size_t copied;
1236 size_t dirty_sectors;
1237 size_t num_sectors;
1238 int extents_locked;
1239
1240 /*
1241 * Fault pages before locking them in prepare_pages
1242 * to avoid recursive lock
1243 */
1244 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1245 ret = -EFAULT;
1246 break;
1247 }
1248
1249 only_release_metadata = false;
1250 sector_offset = pos & (fs_info->sectorsize - 1);
1251
1252 extent_changeset_release(data_reserved);
1253 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1254 &data_reserved, pos,
1255 write_bytes, nowait);
1256 if (ret < 0) {
1257 int can_nocow;
1258
1259 if (nowait && (ret == -ENOSPC || ret == -EAGAIN)) {
1260 ret = -EAGAIN;
1261 break;
1262 }
1263
1264 /*
1265 * If we don't have to COW at the offset, reserve
1266 * metadata only. write_bytes may get smaller than
1267 * requested here.
1268 */
1269 can_nocow = btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1270 &write_bytes, nowait);
1271 if (can_nocow < 0)
1272 ret = can_nocow;
1273 if (can_nocow > 0)
1274 ret = 0;
1275 if (ret)
1276 break;
1277 only_release_metadata = true;
1278 }
1279
1280 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1281 WARN_ON(num_pages > nrptrs);
1282 reserve_bytes = round_up(write_bytes + sector_offset,
1283 fs_info->sectorsize);
1284 WARN_ON(reserve_bytes == 0);
1285 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1286 reserve_bytes,
1287 reserve_bytes, nowait);
1288 if (ret) {
1289 if (!only_release_metadata)
1290 btrfs_free_reserved_data_space(BTRFS_I(inode),
1291 data_reserved, pos,
1292 write_bytes);
1293 else
1294 btrfs_check_nocow_unlock(BTRFS_I(inode));
1295
1296 if (nowait && ret == -ENOSPC)
1297 ret = -EAGAIN;
1298 break;
1299 }
1300
1301 release_bytes = reserve_bytes;
1302again:
1303 ret = balance_dirty_pages_ratelimited_flags(inode->i_mapping, bdp_flags);
1304 if (ret) {
1305 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1306 break;
1307 }
1308
1309 /*
1310 * This is going to setup the pages array with the number of
1311 * pages we want, so we don't really need to worry about the
1312 * contents of pages from loop to loop
1313 */
1314 ret = prepare_pages(inode, pages, num_pages,
1315 pos, write_bytes, force_page_uptodate, false);
1316 if (ret) {
1317 btrfs_delalloc_release_extents(BTRFS_I(inode),
1318 reserve_bytes);
1319 break;
1320 }
1321
1322 extents_locked = lock_and_cleanup_extent_if_need(
1323 BTRFS_I(inode), pages,
1324 num_pages, pos, write_bytes, &lockstart,
1325 &lockend, nowait, &cached_state);
1326 if (extents_locked < 0) {
1327 if (!nowait && extents_locked == -EAGAIN)
1328 goto again;
1329
1330 btrfs_delalloc_release_extents(BTRFS_I(inode),
1331 reserve_bytes);
1332 ret = extents_locked;
1333 break;
1334 }
1335
1336 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1337
1338 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1339 dirty_sectors = round_up(copied + sector_offset,
1340 fs_info->sectorsize);
1341 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1342
1343 /*
1344 * if we have trouble faulting in the pages, fall
1345 * back to one page at a time
1346 */
1347 if (copied < write_bytes)
1348 nrptrs = 1;
1349
1350 if (copied == 0) {
1351 force_page_uptodate = true;
1352 dirty_sectors = 0;
1353 dirty_pages = 0;
1354 } else {
1355 force_page_uptodate = false;
1356 dirty_pages = DIV_ROUND_UP(copied + offset,
1357 PAGE_SIZE);
1358 }
1359
1360 if (num_sectors > dirty_sectors) {
1361 /* release everything except the sectors we dirtied */
1362 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1363 if (only_release_metadata) {
1364 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1365 release_bytes, true);
1366 } else {
1367 u64 __pos;
1368
1369 __pos = round_down(pos,
1370 fs_info->sectorsize) +
1371 (dirty_pages << PAGE_SHIFT);
1372 btrfs_delalloc_release_space(BTRFS_I(inode),
1373 data_reserved, __pos,
1374 release_bytes, true);
1375 }
1376 }
1377
1378 release_bytes = round_up(copied + sector_offset,
1379 fs_info->sectorsize);
1380
1381 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1382 dirty_pages, pos, copied,
1383 &cached_state, only_release_metadata);
1384
1385 /*
1386 * If we have not locked the extent range, because the range's
1387 * start offset is >= i_size, we might still have a non-NULL
1388 * cached extent state, acquired while marking the extent range
1389 * as delalloc through btrfs_dirty_pages(). Therefore free any
1390 * possible cached extent state to avoid a memory leak.
1391 */
1392 if (extents_locked)
1393 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
1394 lockend, &cached_state);
1395 else
1396 free_extent_state(cached_state);
1397
1398 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1399 if (ret) {
1400 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1401 break;
1402 }
1403
1404 release_bytes = 0;
1405 if (only_release_metadata)
1406 btrfs_check_nocow_unlock(BTRFS_I(inode));
1407
1408 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1409
1410 cond_resched();
1411
1412 pos += copied;
1413 num_written += copied;
1414 }
1415
1416 kfree(pages);
1417
1418 if (release_bytes) {
1419 if (only_release_metadata) {
1420 btrfs_check_nocow_unlock(BTRFS_I(inode));
1421 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1422 release_bytes, true);
1423 } else {
1424 btrfs_delalloc_release_space(BTRFS_I(inode),
1425 data_reserved,
1426 round_down(pos, fs_info->sectorsize),
1427 release_bytes, true);
1428 }
1429 }
1430
1431 extent_changeset_free(data_reserved);
1432 if (num_written > 0) {
1433 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1434 iocb->ki_pos += num_written;
1435 }
1436out:
1437 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1438 return num_written ? num_written : ret;
1439}
1440
1441static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1442 const struct iov_iter *iter, loff_t offset)
1443{
1444 const u32 blocksize_mask = fs_info->sectorsize - 1;
1445
1446 if (offset & blocksize_mask)
1447 return -EINVAL;
1448
1449 if (iov_iter_alignment(iter) & blocksize_mask)
1450 return -EINVAL;
1451
1452 return 0;
1453}
1454
1455static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1456{
1457 struct file *file = iocb->ki_filp;
1458 struct inode *inode = file_inode(file);
1459 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1460 loff_t pos;
1461 ssize_t written = 0;
1462 ssize_t written_buffered;
1463 size_t prev_left = 0;
1464 loff_t endbyte;
1465 ssize_t err;
1466 unsigned int ilock_flags = 0;
1467 struct iomap_dio *dio;
1468
1469 if (iocb->ki_flags & IOCB_NOWAIT)
1470 ilock_flags |= BTRFS_ILOCK_TRY;
1471
1472 /* If the write DIO is within EOF, use a shared lock */
1473 if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1474 ilock_flags |= BTRFS_ILOCK_SHARED;
1475
1476relock:
1477 err = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1478 if (err < 0)
1479 return err;
1480
1481 err = generic_write_checks(iocb, from);
1482 if (err <= 0) {
1483 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1484 return err;
1485 }
1486
1487 err = btrfs_write_check(iocb, from, err);
1488 if (err < 0) {
1489 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1490 goto out;
1491 }
1492
1493 pos = iocb->ki_pos;
1494 /*
1495 * Re-check since file size may have changed just before taking the
1496 * lock or pos may have changed because of O_APPEND in generic_write_check()
1497 */
1498 if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1499 pos + iov_iter_count(from) > i_size_read(inode)) {
1500 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1501 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1502 goto relock;
1503 }
1504
1505 if (check_direct_IO(fs_info, from, pos)) {
1506 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1507 goto buffered;
1508 }
1509
1510 /*
1511 * The iov_iter can be mapped to the same file range we are writing to.
1512 * If that's the case, then we will deadlock in the iomap code, because
1513 * it first calls our callback btrfs_dio_iomap_begin(), which will create
1514 * an ordered extent, and after that it will fault in the pages that the
1515 * iov_iter refers to. During the fault in we end up in the readahead
1516 * pages code (starting at btrfs_readahead()), which will lock the range,
1517 * find that ordered extent and then wait for it to complete (at
1518 * btrfs_lock_and_flush_ordered_range()), resulting in a deadlock since
1519 * obviously the ordered extent can never complete as we didn't submit
1520 * yet the respective bio(s). This always happens when the buffer is
1521 * memory mapped to the same file range, since the iomap DIO code always
1522 * invalidates pages in the target file range (after starting and waiting
1523 * for any writeback).
1524 *
1525 * So here we disable page faults in the iov_iter and then retry if we
1526 * got -EFAULT, faulting in the pages before the retry.
1527 */
1528 from->nofault = true;
1529 dio = btrfs_dio_write(iocb, from, written);
1530 from->nofault = false;
1531
1532 /*
1533 * iomap_dio_complete() will call btrfs_sync_file() if we have a dsync
1534 * iocb, and that needs to lock the inode. So unlock it before calling
1535 * iomap_dio_complete() to avoid a deadlock.
1536 */
1537 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1538
1539 if (IS_ERR_OR_NULL(dio))
1540 err = PTR_ERR_OR_ZERO(dio);
1541 else
1542 err = iomap_dio_complete(dio);
1543
1544 /* No increment (+=) because iomap returns a cumulative value. */
1545 if (err > 0)
1546 written = err;
1547
1548 if (iov_iter_count(from) > 0 && (err == -EFAULT || err > 0)) {
1549 const size_t left = iov_iter_count(from);
1550 /*
1551 * We have more data left to write. Try to fault in as many as
1552 * possible of the remainder pages and retry. We do this without
1553 * releasing and locking again the inode, to prevent races with
1554 * truncate.
1555 *
1556 * Also, in case the iov refers to pages in the file range of the
1557 * file we want to write to (due to a mmap), we could enter an
1558 * infinite loop if we retry after faulting the pages in, since
1559 * iomap will invalidate any pages in the range early on, before
1560 * it tries to fault in the pages of the iov. So we keep track of
1561 * how much was left of iov in the previous EFAULT and fallback
1562 * to buffered IO in case we haven't made any progress.
1563 */
1564 if (left == prev_left) {
1565 err = -ENOTBLK;
1566 } else {
1567 fault_in_iov_iter_readable(from, left);
1568 prev_left = left;
1569 goto relock;
1570 }
1571 }
1572
1573 /*
1574 * If 'err' is -ENOTBLK or we have not written all data, then it means
1575 * we must fallback to buffered IO.
1576 */
1577 if ((err < 0 && err != -ENOTBLK) || !iov_iter_count(from))
1578 goto out;
1579
1580buffered:
1581 /*
1582 * If we are in a NOWAIT context, then return -EAGAIN to signal the caller
1583 * it must retry the operation in a context where blocking is acceptable,
1584 * because even if we end up not blocking during the buffered IO attempt
1585 * below, we will block when flushing and waiting for the IO.
1586 */
1587 if (iocb->ki_flags & IOCB_NOWAIT) {
1588 err = -EAGAIN;
1589 goto out;
1590 }
1591
1592 pos = iocb->ki_pos;
1593 written_buffered = btrfs_buffered_write(iocb, from);
1594 if (written_buffered < 0) {
1595 err = written_buffered;
1596 goto out;
1597 }
1598 /*
1599 * Ensure all data is persisted. We want the next direct IO read to be
1600 * able to read what was just written.
1601 */
1602 endbyte = pos + written_buffered - 1;
1603 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1604 if (err)
1605 goto out;
1606 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1607 if (err)
1608 goto out;
1609 written += written_buffered;
1610 iocb->ki_pos = pos + written_buffered;
1611 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1612 endbyte >> PAGE_SHIFT);
1613out:
1614 return err < 0 ? err : written;
1615}
1616
1617static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
1618 const struct btrfs_ioctl_encoded_io_args *encoded)
1619{
1620 struct file *file = iocb->ki_filp;
1621 struct inode *inode = file_inode(file);
1622 loff_t count;
1623 ssize_t ret;
1624
1625 btrfs_inode_lock(BTRFS_I(inode), 0);
1626 count = encoded->len;
1627 ret = generic_write_checks_count(iocb, &count);
1628 if (ret == 0 && count != encoded->len) {
1629 /*
1630 * The write got truncated by generic_write_checks_count(). We
1631 * can't do a partial encoded write.
1632 */
1633 ret = -EFBIG;
1634 }
1635 if (ret || encoded->len == 0)
1636 goto out;
1637
1638 ret = btrfs_write_check(iocb, from, encoded->len);
1639 if (ret < 0)
1640 goto out;
1641
1642 ret = btrfs_do_encoded_write(iocb, from, encoded);
1643out:
1644 btrfs_inode_unlock(BTRFS_I(inode), 0);
1645 return ret;
1646}
1647
1648ssize_t btrfs_do_write_iter(struct kiocb *iocb, struct iov_iter *from,
1649 const struct btrfs_ioctl_encoded_io_args *encoded)
1650{
1651 struct file *file = iocb->ki_filp;
1652 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1653 ssize_t num_written, num_sync;
1654 const bool sync = iocb_is_dsync(iocb);
1655
1656 /*
1657 * If the fs flips readonly due to some impossible error, although we
1658 * have opened a file as writable, we have to stop this write operation
1659 * to ensure consistency.
1660 */
1661 if (BTRFS_FS_ERROR(inode->root->fs_info))
1662 return -EROFS;
1663
1664 if (encoded && (iocb->ki_flags & IOCB_NOWAIT))
1665 return -EOPNOTSUPP;
1666
1667 if (sync)
1668 atomic_inc(&inode->sync_writers);
1669
1670 if (encoded) {
1671 num_written = btrfs_encoded_write(iocb, from, encoded);
1672 num_sync = encoded->len;
1673 } else if (iocb->ki_flags & IOCB_DIRECT) {
1674 num_written = btrfs_direct_write(iocb, from);
1675 num_sync = num_written;
1676 } else {
1677 num_written = btrfs_buffered_write(iocb, from);
1678 num_sync = num_written;
1679 }
1680
1681 btrfs_set_inode_last_sub_trans(inode);
1682
1683 if (num_sync > 0) {
1684 num_sync = generic_write_sync(iocb, num_sync);
1685 if (num_sync < 0)
1686 num_written = num_sync;
1687 }
1688
1689 if (sync)
1690 atomic_dec(&inode->sync_writers);
1691
1692 current->backing_dev_info = NULL;
1693 return num_written;
1694}
1695
1696static ssize_t btrfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1697{
1698 return btrfs_do_write_iter(iocb, from, NULL);
1699}
1700
1701int btrfs_release_file(struct inode *inode, struct file *filp)
1702{
1703 struct btrfs_file_private *private = filp->private_data;
1704
1705 if (private) {
1706 kfree(private->filldir_buf);
1707 free_extent_state(private->llseek_cached_state);
1708 kfree(private);
1709 filp->private_data = NULL;
1710 }
1711
1712 /*
1713 * Set by setattr when we are about to truncate a file from a non-zero
1714 * size to a zero size. This tries to flush down new bytes that may
1715 * have been written if the application were using truncate to replace
1716 * a file in place.
1717 */
1718 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
1719 &BTRFS_I(inode)->runtime_flags))
1720 filemap_flush(inode->i_mapping);
1721 return 0;
1722}
1723
1724static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1725{
1726 int ret;
1727 struct blk_plug plug;
1728
1729 /*
1730 * This is only called in fsync, which would do synchronous writes, so
1731 * a plug can merge adjacent IOs as much as possible. Esp. in case of
1732 * multiple disks using raid profile, a large IO can be split to
1733 * several segments of stripe length (currently 64K).
1734 */
1735 blk_start_plug(&plug);
1736 atomic_inc(&BTRFS_I(inode)->sync_writers);
1737 ret = btrfs_fdatawrite_range(inode, start, end);
1738 atomic_dec(&BTRFS_I(inode)->sync_writers);
1739 blk_finish_plug(&plug);
1740
1741 return ret;
1742}
1743
1744static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
1745{
1746 struct btrfs_inode *inode = BTRFS_I(ctx->inode);
1747 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1748
1749 if (btrfs_inode_in_log(inode, fs_info->generation) &&
1750 list_empty(&ctx->ordered_extents))
1751 return true;
1752
1753 /*
1754 * If we are doing a fast fsync we can not bail out if the inode's
1755 * last_trans is <= then the last committed transaction, because we only
1756 * update the last_trans of the inode during ordered extent completion,
1757 * and for a fast fsync we don't wait for that, we only wait for the
1758 * writeback to complete.
1759 */
1760 if (inode->last_trans <= fs_info->last_trans_committed &&
1761 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
1762 list_empty(&ctx->ordered_extents)))
1763 return true;
1764
1765 return false;
1766}
1767
1768/*
1769 * fsync call for both files and directories. This logs the inode into
1770 * the tree log instead of forcing full commits whenever possible.
1771 *
1772 * It needs to call filemap_fdatawait so that all ordered extent updates are
1773 * in the metadata btree are up to date for copying to the log.
1774 *
1775 * It drops the inode mutex before doing the tree log commit. This is an
1776 * important optimization for directories because holding the mutex prevents
1777 * new operations on the dir while we write to disk.
1778 */
1779int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1780{
1781 struct dentry *dentry = file_dentry(file);
1782 struct inode *inode = d_inode(dentry);
1783 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1784 struct btrfs_root *root = BTRFS_I(inode)->root;
1785 struct btrfs_trans_handle *trans;
1786 struct btrfs_log_ctx ctx;
1787 int ret = 0, err;
1788 u64 len;
1789 bool full_sync;
1790
1791 trace_btrfs_sync_file(file, datasync);
1792
1793 btrfs_init_log_ctx(&ctx, inode);
1794
1795 /*
1796 * Always set the range to a full range, otherwise we can get into
1797 * several problems, from missing file extent items to represent holes
1798 * when not using the NO_HOLES feature, to log tree corruption due to
1799 * races between hole detection during logging and completion of ordered
1800 * extents outside the range, to missing checksums due to ordered extents
1801 * for which we flushed only a subset of their pages.
1802 */
1803 start = 0;
1804 end = LLONG_MAX;
1805 len = (u64)LLONG_MAX + 1;
1806
1807 /*
1808 * We write the dirty pages in the range and wait until they complete
1809 * out of the ->i_mutex. If so, we can flush the dirty pages by
1810 * multi-task, and make the performance up. See
1811 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1812 */
1813 ret = start_ordered_ops(inode, start, end);
1814 if (ret)
1815 goto out;
1816
1817 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1818
1819 atomic_inc(&root->log_batch);
1820
1821 /*
1822 * Before we acquired the inode's lock and the mmap lock, someone may
1823 * have dirtied more pages in the target range. We need to make sure
1824 * that writeback for any such pages does not start while we are logging
1825 * the inode, because if it does, any of the following might happen when
1826 * we are not doing a full inode sync:
1827 *
1828 * 1) We log an extent after its writeback finishes but before its
1829 * checksums are added to the csum tree, leading to -EIO errors
1830 * when attempting to read the extent after a log replay.
1831 *
1832 * 2) We can end up logging an extent before its writeback finishes.
1833 * Therefore after the log replay we will have a file extent item
1834 * pointing to an unwritten extent (and no data checksums as well).
1835 *
1836 * So trigger writeback for any eventual new dirty pages and then we
1837 * wait for all ordered extents to complete below.
1838 */
1839 ret = start_ordered_ops(inode, start, end);
1840 if (ret) {
1841 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1842 goto out;
1843 }
1844
1845 /*
1846 * Always check for the full sync flag while holding the inode's lock,
1847 * to avoid races with other tasks. The flag must be either set all the
1848 * time during logging or always off all the time while logging.
1849 * We check the flag here after starting delalloc above, because when
1850 * running delalloc the full sync flag may be set if we need to drop
1851 * extra extent map ranges due to temporary memory allocation failures.
1852 */
1853 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1854 &BTRFS_I(inode)->runtime_flags);
1855
1856 /*
1857 * We have to do this here to avoid the priority inversion of waiting on
1858 * IO of a lower priority task while holding a transaction open.
1859 *
1860 * For a full fsync we wait for the ordered extents to complete while
1861 * for a fast fsync we wait just for writeback to complete, and then
1862 * attach the ordered extents to the transaction so that a transaction
1863 * commit waits for their completion, to avoid data loss if we fsync,
1864 * the current transaction commits before the ordered extents complete
1865 * and a power failure happens right after that.
1866 *
1867 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
1868 * logical address recorded in the ordered extent may change. We need
1869 * to wait for the IO to stabilize the logical address.
1870 */
1871 if (full_sync || btrfs_is_zoned(fs_info)) {
1872 ret = btrfs_wait_ordered_range(inode, start, len);
1873 } else {
1874 /*
1875 * Get our ordered extents as soon as possible to avoid doing
1876 * checksum lookups in the csum tree, and use instead the
1877 * checksums attached to the ordered extents.
1878 */
1879 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
1880 &ctx.ordered_extents);
1881 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
1882 }
1883
1884 if (ret)
1885 goto out_release_extents;
1886
1887 atomic_inc(&root->log_batch);
1888
1889 smp_mb();
1890 if (skip_inode_logging(&ctx)) {
1891 /*
1892 * We've had everything committed since the last time we were
1893 * modified so clear this flag in case it was set for whatever
1894 * reason, it's no longer relevant.
1895 */
1896 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1897 &BTRFS_I(inode)->runtime_flags);
1898 /*
1899 * An ordered extent might have started before and completed
1900 * already with io errors, in which case the inode was not
1901 * updated and we end up here. So check the inode's mapping
1902 * for any errors that might have happened since we last
1903 * checked called fsync.
1904 */
1905 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
1906 goto out_release_extents;
1907 }
1908
1909 /*
1910 * We use start here because we will need to wait on the IO to complete
1911 * in btrfs_sync_log, which could require joining a transaction (for
1912 * example checking cross references in the nocow path). If we use join
1913 * here we could get into a situation where we're waiting on IO to
1914 * happen that is blocked on a transaction trying to commit. With start
1915 * we inc the extwriter counter, so we wait for all extwriters to exit
1916 * before we start blocking joiners. This comment is to keep somebody
1917 * from thinking they are super smart and changing this to
1918 * btrfs_join_transaction *cough*Josef*cough*.
1919 */
1920 trans = btrfs_start_transaction(root, 0);
1921 if (IS_ERR(trans)) {
1922 ret = PTR_ERR(trans);
1923 goto out_release_extents;
1924 }
1925 trans->in_fsync = true;
1926
1927 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
1928 btrfs_release_log_ctx_extents(&ctx);
1929 if (ret < 0) {
1930 /* Fallthrough and commit/free transaction. */
1931 ret = BTRFS_LOG_FORCE_COMMIT;
1932 }
1933
1934 /* we've logged all the items and now have a consistent
1935 * version of the file in the log. It is possible that
1936 * someone will come in and modify the file, but that's
1937 * fine because the log is consistent on disk, and we
1938 * have references to all of the file's extents
1939 *
1940 * It is possible that someone will come in and log the
1941 * file again, but that will end up using the synchronization
1942 * inside btrfs_sync_log to keep things safe.
1943 */
1944 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1945
1946 if (ret == BTRFS_NO_LOG_SYNC) {
1947 ret = btrfs_end_transaction(trans);
1948 goto out;
1949 }
1950
1951 /* We successfully logged the inode, attempt to sync the log. */
1952 if (!ret) {
1953 ret = btrfs_sync_log(trans, root, &ctx);
1954 if (!ret) {
1955 ret = btrfs_end_transaction(trans);
1956 goto out;
1957 }
1958 }
1959
1960 /*
1961 * At this point we need to commit the transaction because we had
1962 * btrfs_need_log_full_commit() or some other error.
1963 *
1964 * If we didn't do a full sync we have to stop the trans handle, wait on
1965 * the ordered extents, start it again and commit the transaction. If
1966 * we attempt to wait on the ordered extents here we could deadlock with
1967 * something like fallocate() that is holding the extent lock trying to
1968 * start a transaction while some other thread is trying to commit the
1969 * transaction while we (fsync) are currently holding the transaction
1970 * open.
1971 */
1972 if (!full_sync) {
1973 ret = btrfs_end_transaction(trans);
1974 if (ret)
1975 goto out;
1976 ret = btrfs_wait_ordered_range(inode, start, len);
1977 if (ret)
1978 goto out;
1979
1980 /*
1981 * This is safe to use here because we're only interested in
1982 * making sure the transaction that had the ordered extents is
1983 * committed. We aren't waiting on anything past this point,
1984 * we're purely getting the transaction and committing it.
1985 */
1986 trans = btrfs_attach_transaction_barrier(root);
1987 if (IS_ERR(trans)) {
1988 ret = PTR_ERR(trans);
1989
1990 /*
1991 * We committed the transaction and there's no currently
1992 * running transaction, this means everything we care
1993 * about made it to disk and we are done.
1994 */
1995 if (ret == -ENOENT)
1996 ret = 0;
1997 goto out;
1998 }
1999 }
2000
2001 ret = btrfs_commit_transaction(trans);
2002out:
2003 ASSERT(list_empty(&ctx.list));
2004 ASSERT(list_empty(&ctx.conflict_inodes));
2005 err = file_check_and_advance_wb_err(file);
2006 if (!ret)
2007 ret = err;
2008 return ret > 0 ? -EIO : ret;
2009
2010out_release_extents:
2011 btrfs_release_log_ctx_extents(&ctx);
2012 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2013 goto out;
2014}
2015
2016static const struct vm_operations_struct btrfs_file_vm_ops = {
2017 .fault = filemap_fault,
2018 .map_pages = filemap_map_pages,
2019 .page_mkwrite = btrfs_page_mkwrite,
2020};
2021
2022static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2023{
2024 struct address_space *mapping = filp->f_mapping;
2025
2026 if (!mapping->a_ops->read_folio)
2027 return -ENOEXEC;
2028
2029 file_accessed(filp);
2030 vma->vm_ops = &btrfs_file_vm_ops;
2031
2032 return 0;
2033}
2034
2035static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2036 int slot, u64 start, u64 end)
2037{
2038 struct btrfs_file_extent_item *fi;
2039 struct btrfs_key key;
2040
2041 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2042 return 0;
2043
2044 btrfs_item_key_to_cpu(leaf, &key, slot);
2045 if (key.objectid != btrfs_ino(inode) ||
2046 key.type != BTRFS_EXTENT_DATA_KEY)
2047 return 0;
2048
2049 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2050
2051 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2052 return 0;
2053
2054 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2055 return 0;
2056
2057 if (key.offset == end)
2058 return 1;
2059 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2060 return 1;
2061 return 0;
2062}
2063
2064static int fill_holes(struct btrfs_trans_handle *trans,
2065 struct btrfs_inode *inode,
2066 struct btrfs_path *path, u64 offset, u64 end)
2067{
2068 struct btrfs_fs_info *fs_info = trans->fs_info;
2069 struct btrfs_root *root = inode->root;
2070 struct extent_buffer *leaf;
2071 struct btrfs_file_extent_item *fi;
2072 struct extent_map *hole_em;
2073 struct btrfs_key key;
2074 int ret;
2075
2076 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2077 goto out;
2078
2079 key.objectid = btrfs_ino(inode);
2080 key.type = BTRFS_EXTENT_DATA_KEY;
2081 key.offset = offset;
2082
2083 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2084 if (ret <= 0) {
2085 /*
2086 * We should have dropped this offset, so if we find it then
2087 * something has gone horribly wrong.
2088 */
2089 if (ret == 0)
2090 ret = -EINVAL;
2091 return ret;
2092 }
2093
2094 leaf = path->nodes[0];
2095 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2096 u64 num_bytes;
2097
2098 path->slots[0]--;
2099 fi = btrfs_item_ptr(leaf, path->slots[0],
2100 struct btrfs_file_extent_item);
2101 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2102 end - offset;
2103 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2104 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2105 btrfs_set_file_extent_offset(leaf, fi, 0);
2106 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2107 btrfs_mark_buffer_dirty(leaf);
2108 goto out;
2109 }
2110
2111 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2112 u64 num_bytes;
2113
2114 key.offset = offset;
2115 btrfs_set_item_key_safe(fs_info, path, &key);
2116 fi = btrfs_item_ptr(leaf, path->slots[0],
2117 struct btrfs_file_extent_item);
2118 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2119 offset;
2120 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2121 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2122 btrfs_set_file_extent_offset(leaf, fi, 0);
2123 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2124 btrfs_mark_buffer_dirty(leaf);
2125 goto out;
2126 }
2127 btrfs_release_path(path);
2128
2129 ret = btrfs_insert_hole_extent(trans, root, btrfs_ino(inode), offset,
2130 end - offset);
2131 if (ret)
2132 return ret;
2133
2134out:
2135 btrfs_release_path(path);
2136
2137 hole_em = alloc_extent_map();
2138 if (!hole_em) {
2139 btrfs_drop_extent_map_range(inode, offset, end - 1, false);
2140 btrfs_set_inode_full_sync(inode);
2141 } else {
2142 hole_em->start = offset;
2143 hole_em->len = end - offset;
2144 hole_em->ram_bytes = hole_em->len;
2145 hole_em->orig_start = offset;
2146
2147 hole_em->block_start = EXTENT_MAP_HOLE;
2148 hole_em->block_len = 0;
2149 hole_em->orig_block_len = 0;
2150 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2151 hole_em->generation = trans->transid;
2152
2153 ret = btrfs_replace_extent_map_range(inode, hole_em, true);
2154 free_extent_map(hole_em);
2155 if (ret)
2156 btrfs_set_inode_full_sync(inode);
2157 }
2158
2159 return 0;
2160}
2161
2162/*
2163 * Find a hole extent on given inode and change start/len to the end of hole
2164 * extent.(hole/vacuum extent whose em->start <= start &&
2165 * em->start + em->len > start)
2166 * When a hole extent is found, return 1 and modify start/len.
2167 */
2168static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2169{
2170 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2171 struct extent_map *em;
2172 int ret = 0;
2173
2174 em = btrfs_get_extent(inode, NULL, 0,
2175 round_down(*start, fs_info->sectorsize),
2176 round_up(*len, fs_info->sectorsize));
2177 if (IS_ERR(em))
2178 return PTR_ERR(em);
2179
2180 /* Hole or vacuum extent(only exists in no-hole mode) */
2181 if (em->block_start == EXTENT_MAP_HOLE) {
2182 ret = 1;
2183 *len = em->start + em->len > *start + *len ?
2184 0 : *start + *len - em->start - em->len;
2185 *start = em->start + em->len;
2186 }
2187 free_extent_map(em);
2188 return ret;
2189}
2190
2191static void btrfs_punch_hole_lock_range(struct inode *inode,
2192 const u64 lockstart,
2193 const u64 lockend,
2194 struct extent_state **cached_state)
2195{
2196 /*
2197 * For subpage case, if the range is not at page boundary, we could
2198 * have pages at the leading/tailing part of the range.
2199 * This could lead to dead loop since filemap_range_has_page()
2200 * will always return true.
2201 * So here we need to do extra page alignment for
2202 * filemap_range_has_page().
2203 */
2204 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2205 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2206
2207 while (1) {
2208 truncate_pagecache_range(inode, lockstart, lockend);
2209
2210 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2211 cached_state);
2212 /*
2213 * We can't have ordered extents in the range, nor dirty/writeback
2214 * pages, because we have locked the inode's VFS lock in exclusive
2215 * mode, we have locked the inode's i_mmap_lock in exclusive mode,
2216 * we have flushed all delalloc in the range and we have waited
2217 * for any ordered extents in the range to complete.
2218 * We can race with anyone reading pages from this range, so after
2219 * locking the range check if we have pages in the range, and if
2220 * we do, unlock the range and retry.
2221 */
2222 if (!filemap_range_has_page(inode->i_mapping, page_lockstart,
2223 page_lockend))
2224 break;
2225
2226 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2227 cached_state);
2228 }
2229
2230 btrfs_assert_inode_range_clean(BTRFS_I(inode), lockstart, lockend);
2231}
2232
2233static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2234 struct btrfs_inode *inode,
2235 struct btrfs_path *path,
2236 struct btrfs_replace_extent_info *extent_info,
2237 const u64 replace_len,
2238 const u64 bytes_to_drop)
2239{
2240 struct btrfs_fs_info *fs_info = trans->fs_info;
2241 struct btrfs_root *root = inode->root;
2242 struct btrfs_file_extent_item *extent;
2243 struct extent_buffer *leaf;
2244 struct btrfs_key key;
2245 int slot;
2246 struct btrfs_ref ref = { 0 };
2247 int ret;
2248
2249 if (replace_len == 0)
2250 return 0;
2251
2252 if (extent_info->disk_offset == 0 &&
2253 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2254 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2255 return 0;
2256 }
2257
2258 key.objectid = btrfs_ino(inode);
2259 key.type = BTRFS_EXTENT_DATA_KEY;
2260 key.offset = extent_info->file_offset;
2261 ret = btrfs_insert_empty_item(trans, root, path, &key,
2262 sizeof(struct btrfs_file_extent_item));
2263 if (ret)
2264 return ret;
2265 leaf = path->nodes[0];
2266 slot = path->slots[0];
2267 write_extent_buffer(leaf, extent_info->extent_buf,
2268 btrfs_item_ptr_offset(leaf, slot),
2269 sizeof(struct btrfs_file_extent_item));
2270 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2271 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2272 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2273 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2274 if (extent_info->is_new_extent)
2275 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2276 btrfs_mark_buffer_dirty(leaf);
2277 btrfs_release_path(path);
2278
2279 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2280 replace_len);
2281 if (ret)
2282 return ret;
2283
2284 /* If it's a hole, nothing more needs to be done. */
2285 if (extent_info->disk_offset == 0) {
2286 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2287 return 0;
2288 }
2289
2290 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2291
2292 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2293 key.objectid = extent_info->disk_offset;
2294 key.type = BTRFS_EXTENT_ITEM_KEY;
2295 key.offset = extent_info->disk_len;
2296 ret = btrfs_alloc_reserved_file_extent(trans, root,
2297 btrfs_ino(inode),
2298 extent_info->file_offset,
2299 extent_info->qgroup_reserved,
2300 &key);
2301 } else {
2302 u64 ref_offset;
2303
2304 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2305 extent_info->disk_offset,
2306 extent_info->disk_len, 0);
2307 ref_offset = extent_info->file_offset - extent_info->data_offset;
2308 btrfs_init_data_ref(&ref, root->root_key.objectid,
2309 btrfs_ino(inode), ref_offset, 0, false);
2310 ret = btrfs_inc_extent_ref(trans, &ref);
2311 }
2312
2313 extent_info->insertions++;
2314
2315 return ret;
2316}
2317
2318/*
2319 * The respective range must have been previously locked, as well as the inode.
2320 * The end offset is inclusive (last byte of the range).
2321 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2322 * the file range with an extent.
2323 * When not punching a hole, we don't want to end up in a state where we dropped
2324 * extents without inserting a new one, so we must abort the transaction to avoid
2325 * a corruption.
2326 */
2327int btrfs_replace_file_extents(struct btrfs_inode *inode,
2328 struct btrfs_path *path, const u64 start,
2329 const u64 end,
2330 struct btrfs_replace_extent_info *extent_info,
2331 struct btrfs_trans_handle **trans_out)
2332{
2333 struct btrfs_drop_extents_args drop_args = { 0 };
2334 struct btrfs_root *root = inode->root;
2335 struct btrfs_fs_info *fs_info = root->fs_info;
2336 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2337 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2338 struct btrfs_trans_handle *trans = NULL;
2339 struct btrfs_block_rsv *rsv;
2340 unsigned int rsv_count;
2341 u64 cur_offset;
2342 u64 len = end - start;
2343 int ret = 0;
2344
2345 if (end <= start)
2346 return -EINVAL;
2347
2348 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2349 if (!rsv) {
2350 ret = -ENOMEM;
2351 goto out;
2352 }
2353 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2354 rsv->failfast = true;
2355
2356 /*
2357 * 1 - update the inode
2358 * 1 - removing the extents in the range
2359 * 1 - adding the hole extent if no_holes isn't set or if we are
2360 * replacing the range with a new extent
2361 */
2362 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2363 rsv_count = 3;
2364 else
2365 rsv_count = 2;
2366
2367 trans = btrfs_start_transaction(root, rsv_count);
2368 if (IS_ERR(trans)) {
2369 ret = PTR_ERR(trans);
2370 trans = NULL;
2371 goto out_free;
2372 }
2373
2374 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2375 min_size, false);
2376 if (WARN_ON(ret))
2377 goto out_trans;
2378 trans->block_rsv = rsv;
2379
2380 cur_offset = start;
2381 drop_args.path = path;
2382 drop_args.end = end + 1;
2383 drop_args.drop_cache = true;
2384 while (cur_offset < end) {
2385 drop_args.start = cur_offset;
2386 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2387 /* If we are punching a hole decrement the inode's byte count */
2388 if (!extent_info)
2389 btrfs_update_inode_bytes(inode, 0,
2390 drop_args.bytes_found);
2391 if (ret != -ENOSPC) {
2392 /*
2393 * The only time we don't want to abort is if we are
2394 * attempting to clone a partial inline extent, in which
2395 * case we'll get EOPNOTSUPP. However if we aren't
2396 * clone we need to abort no matter what, because if we
2397 * got EOPNOTSUPP via prealloc then we messed up and
2398 * need to abort.
2399 */
2400 if (ret &&
2401 (ret != -EOPNOTSUPP ||
2402 (extent_info && extent_info->is_new_extent)))
2403 btrfs_abort_transaction(trans, ret);
2404 break;
2405 }
2406
2407 trans->block_rsv = &fs_info->trans_block_rsv;
2408
2409 if (!extent_info && cur_offset < drop_args.drop_end &&
2410 cur_offset < ino_size) {
2411 ret = fill_holes(trans, inode, path, cur_offset,
2412 drop_args.drop_end);
2413 if (ret) {
2414 /*
2415 * If we failed then we didn't insert our hole
2416 * entries for the area we dropped, so now the
2417 * fs is corrupted, so we must abort the
2418 * transaction.
2419 */
2420 btrfs_abort_transaction(trans, ret);
2421 break;
2422 }
2423 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2424 /*
2425 * We are past the i_size here, but since we didn't
2426 * insert holes we need to clear the mapped area so we
2427 * know to not set disk_i_size in this area until a new
2428 * file extent is inserted here.
2429 */
2430 ret = btrfs_inode_clear_file_extent_range(inode,
2431 cur_offset,
2432 drop_args.drop_end - cur_offset);
2433 if (ret) {
2434 /*
2435 * We couldn't clear our area, so we could
2436 * presumably adjust up and corrupt the fs, so
2437 * we need to abort.
2438 */
2439 btrfs_abort_transaction(trans, ret);
2440 break;
2441 }
2442 }
2443
2444 if (extent_info &&
2445 drop_args.drop_end > extent_info->file_offset) {
2446 u64 replace_len = drop_args.drop_end -
2447 extent_info->file_offset;
2448
2449 ret = btrfs_insert_replace_extent(trans, inode, path,
2450 extent_info, replace_len,
2451 drop_args.bytes_found);
2452 if (ret) {
2453 btrfs_abort_transaction(trans, ret);
2454 break;
2455 }
2456 extent_info->data_len -= replace_len;
2457 extent_info->data_offset += replace_len;
2458 extent_info->file_offset += replace_len;
2459 }
2460
2461 /*
2462 * We are releasing our handle on the transaction, balance the
2463 * dirty pages of the btree inode and flush delayed items, and
2464 * then get a new transaction handle, which may now point to a
2465 * new transaction in case someone else may have committed the
2466 * transaction we used to replace/drop file extent items. So
2467 * bump the inode's iversion and update mtime and ctime except
2468 * if we are called from a dedupe context. This is because a
2469 * power failure/crash may happen after the transaction is
2470 * committed and before we finish replacing/dropping all the
2471 * file extent items we need.
2472 */
2473 inode_inc_iversion(&inode->vfs_inode);
2474
2475 if (!extent_info || extent_info->update_times) {
2476 inode->vfs_inode.i_mtime = current_time(&inode->vfs_inode);
2477 inode->vfs_inode.i_ctime = inode->vfs_inode.i_mtime;
2478 }
2479
2480 ret = btrfs_update_inode(trans, root, inode);
2481 if (ret)
2482 break;
2483
2484 btrfs_end_transaction(trans);
2485 btrfs_btree_balance_dirty(fs_info);
2486
2487 trans = btrfs_start_transaction(root, rsv_count);
2488 if (IS_ERR(trans)) {
2489 ret = PTR_ERR(trans);
2490 trans = NULL;
2491 break;
2492 }
2493
2494 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2495 rsv, min_size, false);
2496 if (WARN_ON(ret))
2497 break;
2498 trans->block_rsv = rsv;
2499
2500 cur_offset = drop_args.drop_end;
2501 len = end - cur_offset;
2502 if (!extent_info && len) {
2503 ret = find_first_non_hole(inode, &cur_offset, &len);
2504 if (unlikely(ret < 0))
2505 break;
2506 if (ret && !len) {
2507 ret = 0;
2508 break;
2509 }
2510 }
2511 }
2512
2513 /*
2514 * If we were cloning, force the next fsync to be a full one since we
2515 * we replaced (or just dropped in the case of cloning holes when
2516 * NO_HOLES is enabled) file extent items and did not setup new extent
2517 * maps for the replacement extents (or holes).
2518 */
2519 if (extent_info && !extent_info->is_new_extent)
2520 btrfs_set_inode_full_sync(inode);
2521
2522 if (ret)
2523 goto out_trans;
2524
2525 trans->block_rsv = &fs_info->trans_block_rsv;
2526 /*
2527 * If we are using the NO_HOLES feature we might have had already an
2528 * hole that overlaps a part of the region [lockstart, lockend] and
2529 * ends at (or beyond) lockend. Since we have no file extent items to
2530 * represent holes, drop_end can be less than lockend and so we must
2531 * make sure we have an extent map representing the existing hole (the
2532 * call to __btrfs_drop_extents() might have dropped the existing extent
2533 * map representing the existing hole), otherwise the fast fsync path
2534 * will not record the existence of the hole region
2535 * [existing_hole_start, lockend].
2536 */
2537 if (drop_args.drop_end <= end)
2538 drop_args.drop_end = end + 1;
2539 /*
2540 * Don't insert file hole extent item if it's for a range beyond eof
2541 * (because it's useless) or if it represents a 0 bytes range (when
2542 * cur_offset == drop_end).
2543 */
2544 if (!extent_info && cur_offset < ino_size &&
2545 cur_offset < drop_args.drop_end) {
2546 ret = fill_holes(trans, inode, path, cur_offset,
2547 drop_args.drop_end);
2548 if (ret) {
2549 /* Same comment as above. */
2550 btrfs_abort_transaction(trans, ret);
2551 goto out_trans;
2552 }
2553 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2554 /* See the comment in the loop above for the reasoning here. */
2555 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2556 drop_args.drop_end - cur_offset);
2557 if (ret) {
2558 btrfs_abort_transaction(trans, ret);
2559 goto out_trans;
2560 }
2561
2562 }
2563 if (extent_info) {
2564 ret = btrfs_insert_replace_extent(trans, inode, path,
2565 extent_info, extent_info->data_len,
2566 drop_args.bytes_found);
2567 if (ret) {
2568 btrfs_abort_transaction(trans, ret);
2569 goto out_trans;
2570 }
2571 }
2572
2573out_trans:
2574 if (!trans)
2575 goto out_free;
2576
2577 trans->block_rsv = &fs_info->trans_block_rsv;
2578 if (ret)
2579 btrfs_end_transaction(trans);
2580 else
2581 *trans_out = trans;
2582out_free:
2583 btrfs_free_block_rsv(fs_info, rsv);
2584out:
2585 return ret;
2586}
2587
2588static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2589{
2590 struct inode *inode = file_inode(file);
2591 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2592 struct btrfs_root *root = BTRFS_I(inode)->root;
2593 struct extent_state *cached_state = NULL;
2594 struct btrfs_path *path;
2595 struct btrfs_trans_handle *trans = NULL;
2596 u64 lockstart;
2597 u64 lockend;
2598 u64 tail_start;
2599 u64 tail_len;
2600 u64 orig_start = offset;
2601 int ret = 0;
2602 bool same_block;
2603 u64 ino_size;
2604 bool truncated_block = false;
2605 bool updated_inode = false;
2606
2607 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2608
2609 ret = btrfs_wait_ordered_range(inode, offset, len);
2610 if (ret)
2611 goto out_only_mutex;
2612
2613 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2614 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2615 if (ret < 0)
2616 goto out_only_mutex;
2617 if (ret && !len) {
2618 /* Already in a large hole */
2619 ret = 0;
2620 goto out_only_mutex;
2621 }
2622
2623 ret = file_modified(file);
2624 if (ret)
2625 goto out_only_mutex;
2626
2627 lockstart = round_up(offset, fs_info->sectorsize);
2628 lockend = round_down(offset + len, fs_info->sectorsize) - 1;
2629 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2630 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2631 /*
2632 * We needn't truncate any block which is beyond the end of the file
2633 * because we are sure there is no data there.
2634 */
2635 /*
2636 * Only do this if we are in the same block and we aren't doing the
2637 * entire block.
2638 */
2639 if (same_block && len < fs_info->sectorsize) {
2640 if (offset < ino_size) {
2641 truncated_block = true;
2642 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2643 0);
2644 } else {
2645 ret = 0;
2646 }
2647 goto out_only_mutex;
2648 }
2649
2650 /* zero back part of the first block */
2651 if (offset < ino_size) {
2652 truncated_block = true;
2653 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2654 if (ret) {
2655 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2656 return ret;
2657 }
2658 }
2659
2660 /* Check the aligned pages after the first unaligned page,
2661 * if offset != orig_start, which means the first unaligned page
2662 * including several following pages are already in holes,
2663 * the extra check can be skipped */
2664 if (offset == orig_start) {
2665 /* after truncate page, check hole again */
2666 len = offset + len - lockstart;
2667 offset = lockstart;
2668 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2669 if (ret < 0)
2670 goto out_only_mutex;
2671 if (ret && !len) {
2672 ret = 0;
2673 goto out_only_mutex;
2674 }
2675 lockstart = offset;
2676 }
2677
2678 /* Check the tail unaligned part is in a hole */
2679 tail_start = lockend + 1;
2680 tail_len = offset + len - tail_start;
2681 if (tail_len) {
2682 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2683 if (unlikely(ret < 0))
2684 goto out_only_mutex;
2685 if (!ret) {
2686 /* zero the front end of the last page */
2687 if (tail_start + tail_len < ino_size) {
2688 truncated_block = true;
2689 ret = btrfs_truncate_block(BTRFS_I(inode),
2690 tail_start + tail_len,
2691 0, 1);
2692 if (ret)
2693 goto out_only_mutex;
2694 }
2695 }
2696 }
2697
2698 if (lockend < lockstart) {
2699 ret = 0;
2700 goto out_only_mutex;
2701 }
2702
2703 btrfs_punch_hole_lock_range(inode, lockstart, lockend, &cached_state);
2704
2705 path = btrfs_alloc_path();
2706 if (!path) {
2707 ret = -ENOMEM;
2708 goto out;
2709 }
2710
2711 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
2712 lockend, NULL, &trans);
2713 btrfs_free_path(path);
2714 if (ret)
2715 goto out;
2716
2717 ASSERT(trans != NULL);
2718 inode_inc_iversion(inode);
2719 inode->i_mtime = current_time(inode);
2720 inode->i_ctime = inode->i_mtime;
2721 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2722 updated_inode = true;
2723 btrfs_end_transaction(trans);
2724 btrfs_btree_balance_dirty(fs_info);
2725out:
2726 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2727 &cached_state);
2728out_only_mutex:
2729 if (!updated_inode && truncated_block && !ret) {
2730 /*
2731 * If we only end up zeroing part of a page, we still need to
2732 * update the inode item, so that all the time fields are
2733 * updated as well as the necessary btrfs inode in memory fields
2734 * for detecting, at fsync time, if the inode isn't yet in the
2735 * log tree or it's there but not up to date.
2736 */
2737 struct timespec64 now = current_time(inode);
2738
2739 inode_inc_iversion(inode);
2740 inode->i_mtime = now;
2741 inode->i_ctime = now;
2742 trans = btrfs_start_transaction(root, 1);
2743 if (IS_ERR(trans)) {
2744 ret = PTR_ERR(trans);
2745 } else {
2746 int ret2;
2747
2748 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2749 ret2 = btrfs_end_transaction(trans);
2750 if (!ret)
2751 ret = ret2;
2752 }
2753 }
2754 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2755 return ret;
2756}
2757
2758/* Helper structure to record which range is already reserved */
2759struct falloc_range {
2760 struct list_head list;
2761 u64 start;
2762 u64 len;
2763};
2764
2765/*
2766 * Helper function to add falloc range
2767 *
2768 * Caller should have locked the larger range of extent containing
2769 * [start, len)
2770 */
2771static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2772{
2773 struct falloc_range *range = NULL;
2774
2775 if (!list_empty(head)) {
2776 /*
2777 * As fallocate iterates by bytenr order, we only need to check
2778 * the last range.
2779 */
2780 range = list_last_entry(head, struct falloc_range, list);
2781 if (range->start + range->len == start) {
2782 range->len += len;
2783 return 0;
2784 }
2785 }
2786
2787 range = kmalloc(sizeof(*range), GFP_KERNEL);
2788 if (!range)
2789 return -ENOMEM;
2790 range->start = start;
2791 range->len = len;
2792 list_add_tail(&range->list, head);
2793 return 0;
2794}
2795
2796static int btrfs_fallocate_update_isize(struct inode *inode,
2797 const u64 end,
2798 const int mode)
2799{
2800 struct btrfs_trans_handle *trans;
2801 struct btrfs_root *root = BTRFS_I(inode)->root;
2802 int ret;
2803 int ret2;
2804
2805 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2806 return 0;
2807
2808 trans = btrfs_start_transaction(root, 1);
2809 if (IS_ERR(trans))
2810 return PTR_ERR(trans);
2811
2812 inode->i_ctime = current_time(inode);
2813 i_size_write(inode, end);
2814 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
2815 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2816 ret2 = btrfs_end_transaction(trans);
2817
2818 return ret ? ret : ret2;
2819}
2820
2821enum {
2822 RANGE_BOUNDARY_WRITTEN_EXTENT,
2823 RANGE_BOUNDARY_PREALLOC_EXTENT,
2824 RANGE_BOUNDARY_HOLE,
2825};
2826
2827static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
2828 u64 offset)
2829{
2830 const u64 sectorsize = inode->root->fs_info->sectorsize;
2831 struct extent_map *em;
2832 int ret;
2833
2834 offset = round_down(offset, sectorsize);
2835 em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
2836 if (IS_ERR(em))
2837 return PTR_ERR(em);
2838
2839 if (em->block_start == EXTENT_MAP_HOLE)
2840 ret = RANGE_BOUNDARY_HOLE;
2841 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2842 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2843 else
2844 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2845
2846 free_extent_map(em);
2847 return ret;
2848}
2849
2850static int btrfs_zero_range(struct inode *inode,
2851 loff_t offset,
2852 loff_t len,
2853 const int mode)
2854{
2855 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2856 struct extent_map *em;
2857 struct extent_changeset *data_reserved = NULL;
2858 int ret;
2859 u64 alloc_hint = 0;
2860 const u64 sectorsize = fs_info->sectorsize;
2861 u64 alloc_start = round_down(offset, sectorsize);
2862 u64 alloc_end = round_up(offset + len, sectorsize);
2863 u64 bytes_to_reserve = 0;
2864 bool space_reserved = false;
2865
2866 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
2867 alloc_end - alloc_start);
2868 if (IS_ERR(em)) {
2869 ret = PTR_ERR(em);
2870 goto out;
2871 }
2872
2873 /*
2874 * Avoid hole punching and extent allocation for some cases. More cases
2875 * could be considered, but these are unlikely common and we keep things
2876 * as simple as possible for now. Also, intentionally, if the target
2877 * range contains one or more prealloc extents together with regular
2878 * extents and holes, we drop all the existing extents and allocate a
2879 * new prealloc extent, so that we get a larger contiguous disk extent.
2880 */
2881 if (em->start <= alloc_start &&
2882 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2883 const u64 em_end = em->start + em->len;
2884
2885 if (em_end >= offset + len) {
2886 /*
2887 * The whole range is already a prealloc extent,
2888 * do nothing except updating the inode's i_size if
2889 * needed.
2890 */
2891 free_extent_map(em);
2892 ret = btrfs_fallocate_update_isize(inode, offset + len,
2893 mode);
2894 goto out;
2895 }
2896 /*
2897 * Part of the range is already a prealloc extent, so operate
2898 * only on the remaining part of the range.
2899 */
2900 alloc_start = em_end;
2901 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2902 len = offset + len - alloc_start;
2903 offset = alloc_start;
2904 alloc_hint = em->block_start + em->len;
2905 }
2906 free_extent_map(em);
2907
2908 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2909 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2910 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
2911 sectorsize);
2912 if (IS_ERR(em)) {
2913 ret = PTR_ERR(em);
2914 goto out;
2915 }
2916
2917 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2918 free_extent_map(em);
2919 ret = btrfs_fallocate_update_isize(inode, offset + len,
2920 mode);
2921 goto out;
2922 }
2923 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2924 free_extent_map(em);
2925 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2926 0);
2927 if (!ret)
2928 ret = btrfs_fallocate_update_isize(inode,
2929 offset + len,
2930 mode);
2931 return ret;
2932 }
2933 free_extent_map(em);
2934 alloc_start = round_down(offset, sectorsize);
2935 alloc_end = alloc_start + sectorsize;
2936 goto reserve_space;
2937 }
2938
2939 alloc_start = round_up(offset, sectorsize);
2940 alloc_end = round_down(offset + len, sectorsize);
2941
2942 /*
2943 * For unaligned ranges, check the pages at the boundaries, they might
2944 * map to an extent, in which case we need to partially zero them, or
2945 * they might map to a hole, in which case we need our allocation range
2946 * to cover them.
2947 */
2948 if (!IS_ALIGNED(offset, sectorsize)) {
2949 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2950 offset);
2951 if (ret < 0)
2952 goto out;
2953 if (ret == RANGE_BOUNDARY_HOLE) {
2954 alloc_start = round_down(offset, sectorsize);
2955 ret = 0;
2956 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2957 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2958 if (ret)
2959 goto out;
2960 } else {
2961 ret = 0;
2962 }
2963 }
2964
2965 if (!IS_ALIGNED(offset + len, sectorsize)) {
2966 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2967 offset + len);
2968 if (ret < 0)
2969 goto out;
2970 if (ret == RANGE_BOUNDARY_HOLE) {
2971 alloc_end = round_up(offset + len, sectorsize);
2972 ret = 0;
2973 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2974 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
2975 0, 1);
2976 if (ret)
2977 goto out;
2978 } else {
2979 ret = 0;
2980 }
2981 }
2982
2983reserve_space:
2984 if (alloc_start < alloc_end) {
2985 struct extent_state *cached_state = NULL;
2986 const u64 lockstart = alloc_start;
2987 const u64 lockend = alloc_end - 1;
2988
2989 bytes_to_reserve = alloc_end - alloc_start;
2990 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2991 bytes_to_reserve);
2992 if (ret < 0)
2993 goto out;
2994 space_reserved = true;
2995 btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2996 &cached_state);
2997 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
2998 alloc_start, bytes_to_reserve);
2999 if (ret) {
3000 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
3001 lockend, &cached_state);
3002 goto out;
3003 }
3004 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3005 alloc_end - alloc_start,
3006 i_blocksize(inode),
3007 offset + len, &alloc_hint);
3008 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3009 &cached_state);
3010 /* btrfs_prealloc_file_range releases reserved space on error */
3011 if (ret) {
3012 space_reserved = false;
3013 goto out;
3014 }
3015 }
3016 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3017 out:
3018 if (ret && space_reserved)
3019 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3020 alloc_start, bytes_to_reserve);
3021 extent_changeset_free(data_reserved);
3022
3023 return ret;
3024}
3025
3026static long btrfs_fallocate(struct file *file, int mode,
3027 loff_t offset, loff_t len)
3028{
3029 struct inode *inode = file_inode(file);
3030 struct extent_state *cached_state = NULL;
3031 struct extent_changeset *data_reserved = NULL;
3032 struct falloc_range *range;
3033 struct falloc_range *tmp;
3034 struct list_head reserve_list;
3035 u64 cur_offset;
3036 u64 last_byte;
3037 u64 alloc_start;
3038 u64 alloc_end;
3039 u64 alloc_hint = 0;
3040 u64 locked_end;
3041 u64 actual_end = 0;
3042 u64 data_space_needed = 0;
3043 u64 data_space_reserved = 0;
3044 u64 qgroup_reserved = 0;
3045 struct extent_map *em;
3046 int blocksize = BTRFS_I(inode)->root->fs_info->sectorsize;
3047 int ret;
3048
3049 /* Do not allow fallocate in ZONED mode */
3050 if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3051 return -EOPNOTSUPP;
3052
3053 alloc_start = round_down(offset, blocksize);
3054 alloc_end = round_up(offset + len, blocksize);
3055 cur_offset = alloc_start;
3056
3057 /* Make sure we aren't being give some crap mode */
3058 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3059 FALLOC_FL_ZERO_RANGE))
3060 return -EOPNOTSUPP;
3061
3062 if (mode & FALLOC_FL_PUNCH_HOLE)
3063 return btrfs_punch_hole(file, offset, len);
3064
3065 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3066
3067 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3068 ret = inode_newsize_ok(inode, offset + len);
3069 if (ret)
3070 goto out;
3071 }
3072
3073 ret = file_modified(file);
3074 if (ret)
3075 goto out;
3076
3077 /*
3078 * TODO: Move these two operations after we have checked
3079 * accurate reserved space, or fallocate can still fail but
3080 * with page truncated or size expanded.
3081 *
3082 * But that's a minor problem and won't do much harm BTW.
3083 */
3084 if (alloc_start > inode->i_size) {
3085 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3086 alloc_start);
3087 if (ret)
3088 goto out;
3089 } else if (offset + len > inode->i_size) {
3090 /*
3091 * If we are fallocating from the end of the file onward we
3092 * need to zero out the end of the block if i_size lands in the
3093 * middle of a block.
3094 */
3095 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3096 if (ret)
3097 goto out;
3098 }
3099
3100 /*
3101 * We have locked the inode at the VFS level (in exclusive mode) and we
3102 * have locked the i_mmap_lock lock (in exclusive mode). Now before
3103 * locking the file range, flush all dealloc in the range and wait for
3104 * all ordered extents in the range to complete. After this we can lock
3105 * the file range and, due to the previous locking we did, we know there
3106 * can't be more delalloc or ordered extents in the range.
3107 */
3108 ret = btrfs_wait_ordered_range(inode, alloc_start,
3109 alloc_end - alloc_start);
3110 if (ret)
3111 goto out;
3112
3113 if (mode & FALLOC_FL_ZERO_RANGE) {
3114 ret = btrfs_zero_range(inode, offset, len, mode);
3115 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3116 return ret;
3117 }
3118
3119 locked_end = alloc_end - 1;
3120 lock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3121 &cached_state);
3122
3123 btrfs_assert_inode_range_clean(BTRFS_I(inode), alloc_start, locked_end);
3124
3125 /* First, check if we exceed the qgroup limit */
3126 INIT_LIST_HEAD(&reserve_list);
3127 while (cur_offset < alloc_end) {
3128 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3129 alloc_end - cur_offset);
3130 if (IS_ERR(em)) {
3131 ret = PTR_ERR(em);
3132 break;
3133 }
3134 last_byte = min(extent_map_end(em), alloc_end);
3135 actual_end = min_t(u64, extent_map_end(em), offset + len);
3136 last_byte = ALIGN(last_byte, blocksize);
3137 if (em->block_start == EXTENT_MAP_HOLE ||
3138 (cur_offset >= inode->i_size &&
3139 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3140 const u64 range_len = last_byte - cur_offset;
3141
3142 ret = add_falloc_range(&reserve_list, cur_offset, range_len);
3143 if (ret < 0) {
3144 free_extent_map(em);
3145 break;
3146 }
3147 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3148 &data_reserved, cur_offset, range_len);
3149 if (ret < 0) {
3150 free_extent_map(em);
3151 break;
3152 }
3153 qgroup_reserved += range_len;
3154 data_space_needed += range_len;
3155 }
3156 free_extent_map(em);
3157 cur_offset = last_byte;
3158 }
3159
3160 if (!ret && data_space_needed > 0) {
3161 /*
3162 * We are safe to reserve space here as we can't have delalloc
3163 * in the range, see above.
3164 */
3165 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3166 data_space_needed);
3167 if (!ret)
3168 data_space_reserved = data_space_needed;
3169 }
3170
3171 /*
3172 * If ret is still 0, means we're OK to fallocate.
3173 * Or just cleanup the list and exit.
3174 */
3175 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3176 if (!ret) {
3177 ret = btrfs_prealloc_file_range(inode, mode,
3178 range->start,
3179 range->len, i_blocksize(inode),
3180 offset + len, &alloc_hint);
3181 /*
3182 * btrfs_prealloc_file_range() releases space even
3183 * if it returns an error.
3184 */
3185 data_space_reserved -= range->len;
3186 qgroup_reserved -= range->len;
3187 } else if (data_space_reserved > 0) {
3188 btrfs_free_reserved_data_space(BTRFS_I(inode),
3189 data_reserved, range->start,
3190 range->len);
3191 data_space_reserved -= range->len;
3192 qgroup_reserved -= range->len;
3193 } else if (qgroup_reserved > 0) {
3194 btrfs_qgroup_free_data(BTRFS_I(inode), data_reserved,
3195 range->start, range->len);
3196 qgroup_reserved -= range->len;
3197 }
3198 list_del(&range->list);
3199 kfree(range);
3200 }
3201 if (ret < 0)
3202 goto out_unlock;
3203
3204 /*
3205 * We didn't need to allocate any more space, but we still extended the
3206 * size of the file so we need to update i_size and the inode item.
3207 */
3208 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3209out_unlock:
3210 unlock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3211 &cached_state);
3212out:
3213 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3214 extent_changeset_free(data_reserved);
3215 return ret;
3216}
3217
3218/*
3219 * Helper for btrfs_find_delalloc_in_range(). Find a subrange in a given range
3220 * that has unflushed and/or flushing delalloc. There might be other adjacent
3221 * subranges after the one it found, so btrfs_find_delalloc_in_range() keeps
3222 * looping while it gets adjacent subranges, and merging them together.
3223 */
3224static bool find_delalloc_subrange(struct btrfs_inode *inode, u64 start, u64 end,
3225 struct extent_state **cached_state,
3226 bool *search_io_tree,
3227 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3228{
3229 u64 len = end + 1 - start;
3230 u64 delalloc_len = 0;
3231 struct btrfs_ordered_extent *oe;
3232 u64 oe_start;
3233 u64 oe_end;
3234
3235 /*
3236 * Search the io tree first for EXTENT_DELALLOC. If we find any, it
3237 * means we have delalloc (dirty pages) for which writeback has not
3238 * started yet.
3239 */
3240 if (*search_io_tree) {
3241 spin_lock(&inode->lock);
3242 if (inode->delalloc_bytes > 0) {
3243 spin_unlock(&inode->lock);
3244 *delalloc_start_ret = start;
3245 delalloc_len = count_range_bits(&inode->io_tree,
3246 delalloc_start_ret, end,
3247 len, EXTENT_DELALLOC, 1,
3248 cached_state);
3249 } else {
3250 spin_unlock(&inode->lock);
3251 }
3252 }
3253
3254 if (delalloc_len > 0) {
3255 /*
3256 * If delalloc was found then *delalloc_start_ret has a sector size
3257 * aligned value (rounded down).
3258 */
3259 *delalloc_end_ret = *delalloc_start_ret + delalloc_len - 1;
3260
3261 if (*delalloc_start_ret == start) {
3262 /* Delalloc for the whole range, nothing more to do. */
3263 if (*delalloc_end_ret == end)
3264 return true;
3265 /* Else trim our search range for ordered extents. */
3266 start = *delalloc_end_ret + 1;
3267 len = end + 1 - start;
3268 }
3269 } else {
3270 /* No delalloc, future calls don't need to search again. */
3271 *search_io_tree = false;
3272 }
3273
3274 /*
3275 * Now also check if there's any ordered extent in the range.
3276 * We do this because:
3277 *
3278 * 1) When delalloc is flushed, the file range is locked, we clear the
3279 * EXTENT_DELALLOC bit from the io tree and create an extent map and
3280 * an ordered extent for the write. So we might just have been called
3281 * after delalloc is flushed and before the ordered extent completes
3282 * and inserts the new file extent item in the subvolume's btree;
3283 *
3284 * 2) We may have an ordered extent created by flushing delalloc for a
3285 * subrange that starts before the subrange we found marked with
3286 * EXTENT_DELALLOC in the io tree.
3287 *
3288 * We could also use the extent map tree to find such delalloc that is
3289 * being flushed, but using the ordered extents tree is more efficient
3290 * because it's usually much smaller as ordered extents are removed from
3291 * the tree once they complete. With the extent maps, we mau have them
3292 * in the extent map tree for a very long time, and they were either
3293 * created by previous writes or loaded by read operations.
3294 */
3295 oe = btrfs_lookup_first_ordered_range(inode, start, len);
3296 if (!oe)
3297 return (delalloc_len > 0);
3298
3299 /* The ordered extent may span beyond our search range. */
3300 oe_start = max(oe->file_offset, start);
3301 oe_end = min(oe->file_offset + oe->num_bytes - 1, end);
3302
3303 btrfs_put_ordered_extent(oe);
3304
3305 /* Don't have unflushed delalloc, return the ordered extent range. */
3306 if (delalloc_len == 0) {
3307 *delalloc_start_ret = oe_start;
3308 *delalloc_end_ret = oe_end;
3309 return true;
3310 }
3311
3312 /*
3313 * We have both unflushed delalloc (io_tree) and an ordered extent.
3314 * If the ranges are adjacent returned a combined range, otherwise
3315 * return the leftmost range.
3316 */
3317 if (oe_start < *delalloc_start_ret) {
3318 if (oe_end < *delalloc_start_ret)
3319 *delalloc_end_ret = oe_end;
3320 *delalloc_start_ret = oe_start;
3321 } else if (*delalloc_end_ret + 1 == oe_start) {
3322 *delalloc_end_ret = oe_end;
3323 }
3324
3325 return true;
3326}
3327
3328/*
3329 * Check if there's delalloc in a given range.
3330 *
3331 * @inode: The inode.
3332 * @start: The start offset of the range. It does not need to be
3333 * sector size aligned.
3334 * @end: The end offset (inclusive value) of the search range.
3335 * It does not need to be sector size aligned.
3336 * @cached_state: Extent state record used for speeding up delalloc
3337 * searches in the inode's io_tree. Can be NULL.
3338 * @delalloc_start_ret: Output argument, set to the start offset of the
3339 * subrange found with delalloc (may not be sector size
3340 * aligned).
3341 * @delalloc_end_ret: Output argument, set to he end offset (inclusive value)
3342 * of the subrange found with delalloc.
3343 *
3344 * Returns true if a subrange with delalloc is found within the given range, and
3345 * if so it sets @delalloc_start_ret and @delalloc_end_ret with the start and
3346 * end offsets of the subrange.
3347 */
3348bool btrfs_find_delalloc_in_range(struct btrfs_inode *inode, u64 start, u64 end,
3349 struct extent_state **cached_state,
3350 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3351{
3352 u64 cur_offset = round_down(start, inode->root->fs_info->sectorsize);
3353 u64 prev_delalloc_end = 0;
3354 bool search_io_tree = true;
3355 bool ret = false;
3356
3357 while (cur_offset <= end) {
3358 u64 delalloc_start;
3359 u64 delalloc_end;
3360 bool delalloc;
3361
3362 delalloc = find_delalloc_subrange(inode, cur_offset, end,
3363 cached_state, &search_io_tree,
3364 &delalloc_start,
3365 &delalloc_end);
3366 if (!delalloc)
3367 break;
3368
3369 if (prev_delalloc_end == 0) {
3370 /* First subrange found. */
3371 *delalloc_start_ret = max(delalloc_start, start);
3372 *delalloc_end_ret = delalloc_end;
3373 ret = true;
3374 } else if (delalloc_start == prev_delalloc_end + 1) {
3375 /* Subrange adjacent to the previous one, merge them. */
3376 *delalloc_end_ret = delalloc_end;
3377 } else {
3378 /* Subrange not adjacent to the previous one, exit. */
3379 break;
3380 }
3381
3382 prev_delalloc_end = delalloc_end;
3383 cur_offset = delalloc_end + 1;
3384 cond_resched();
3385 }
3386
3387 return ret;
3388}
3389
3390/*
3391 * Check if there's a hole or delalloc range in a range representing a hole (or
3392 * prealloc extent) found in the inode's subvolume btree.
3393 *
3394 * @inode: The inode.
3395 * @whence: Seek mode (SEEK_DATA or SEEK_HOLE).
3396 * @start: Start offset of the hole region. It does not need to be sector
3397 * size aligned.
3398 * @end: End offset (inclusive value) of the hole region. It does not
3399 * need to be sector size aligned.
3400 * @start_ret: Return parameter, used to set the start of the subrange in the
3401 * hole that matches the search criteria (seek mode), if such
3402 * subrange is found (return value of the function is true).
3403 * The value returned here may not be sector size aligned.
3404 *
3405 * Returns true if a subrange matching the given seek mode is found, and if one
3406 * is found, it updates @start_ret with the start of the subrange.
3407 */
3408static bool find_desired_extent_in_hole(struct btrfs_inode *inode, int whence,
3409 struct extent_state **cached_state,
3410 u64 start, u64 end, u64 *start_ret)
3411{
3412 u64 delalloc_start;
3413 u64 delalloc_end;
3414 bool delalloc;
3415
3416 delalloc = btrfs_find_delalloc_in_range(inode, start, end, cached_state,
3417 &delalloc_start, &delalloc_end);
3418 if (delalloc && whence == SEEK_DATA) {
3419 *start_ret = delalloc_start;
3420 return true;
3421 }
3422
3423 if (delalloc && whence == SEEK_HOLE) {
3424 /*
3425 * We found delalloc but it starts after out start offset. So we
3426 * have a hole between our start offset and the delalloc start.
3427 */
3428 if (start < delalloc_start) {
3429 *start_ret = start;
3430 return true;
3431 }
3432 /*
3433 * Delalloc range starts at our start offset.
3434 * If the delalloc range's length is smaller than our range,
3435 * then it means we have a hole that starts where the delalloc
3436 * subrange ends.
3437 */
3438 if (delalloc_end < end) {
3439 *start_ret = delalloc_end + 1;
3440 return true;
3441 }
3442
3443 /* There's delalloc for the whole range. */
3444 return false;
3445 }
3446
3447 if (!delalloc && whence == SEEK_HOLE) {
3448 *start_ret = start;
3449 return true;
3450 }
3451
3452 /*
3453 * No delalloc in the range and we are seeking for data. The caller has
3454 * to iterate to the next extent item in the subvolume btree.
3455 */
3456 return false;
3457}
3458
3459static loff_t find_desired_extent(struct file *file, loff_t offset, int whence)
3460{
3461 struct btrfs_inode *inode = BTRFS_I(file->f_mapping->host);
3462 struct btrfs_file_private *private = file->private_data;
3463 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3464 struct extent_state *cached_state = NULL;
3465 struct extent_state **delalloc_cached_state;
3466 const loff_t i_size = i_size_read(&inode->vfs_inode);
3467 const u64 ino = btrfs_ino(inode);
3468 struct btrfs_root *root = inode->root;
3469 struct btrfs_path *path;
3470 struct btrfs_key key;
3471 u64 last_extent_end;
3472 u64 lockstart;
3473 u64 lockend;
3474 u64 start;
3475 int ret;
3476 bool found = false;
3477
3478 if (i_size == 0 || offset >= i_size)
3479 return -ENXIO;
3480
3481 /*
3482 * Quick path. If the inode has no prealloc extents and its number of
3483 * bytes used matches its i_size, then it can not have holes.
3484 */
3485 if (whence == SEEK_HOLE &&
3486 !(inode->flags & BTRFS_INODE_PREALLOC) &&
3487 inode_get_bytes(&inode->vfs_inode) == i_size)
3488 return i_size;
3489
3490 if (!private) {
3491 private = kzalloc(sizeof(*private), GFP_KERNEL);
3492 /*
3493 * No worries if memory allocation failed.
3494 * The private structure is used only for speeding up multiple
3495 * lseek SEEK_HOLE/DATA calls to a file when there's delalloc,
3496 * so everything will still be correct.
3497 */
3498 file->private_data = private;
3499 }
3500
3501 if (private)
3502 delalloc_cached_state = &private->llseek_cached_state;
3503 else
3504 delalloc_cached_state = NULL;
3505
3506 /*
3507 * offset can be negative, in this case we start finding DATA/HOLE from
3508 * the very start of the file.
3509 */
3510 start = max_t(loff_t, 0, offset);
3511
3512 lockstart = round_down(start, fs_info->sectorsize);
3513 lockend = round_up(i_size, fs_info->sectorsize);
3514 if (lockend <= lockstart)
3515 lockend = lockstart + fs_info->sectorsize;
3516 lockend--;
3517
3518 path = btrfs_alloc_path();
3519 if (!path)
3520 return -ENOMEM;
3521 path->reada = READA_FORWARD;
3522
3523 key.objectid = ino;
3524 key.type = BTRFS_EXTENT_DATA_KEY;
3525 key.offset = start;
3526
3527 last_extent_end = lockstart;
3528
3529 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3530
3531 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3532 if (ret < 0) {
3533 goto out;
3534 } else if (ret > 0 && path->slots[0] > 0) {
3535 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
3536 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
3537 path->slots[0]--;
3538 }
3539
3540 while (start < i_size) {
3541 struct extent_buffer *leaf = path->nodes[0];
3542 struct btrfs_file_extent_item *extent;
3543 u64 extent_end;
3544 u8 type;
3545
3546 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3547 ret = btrfs_next_leaf(root, path);
3548 if (ret < 0)
3549 goto out;
3550 else if (ret > 0)
3551 break;
3552
3553 leaf = path->nodes[0];
3554 }
3555
3556 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3557 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3558 break;
3559
3560 extent_end = btrfs_file_extent_end(path);
3561
3562 /*
3563 * In the first iteration we may have a slot that points to an
3564 * extent that ends before our start offset, so skip it.
3565 */
3566 if (extent_end <= start) {
3567 path->slots[0]++;
3568 continue;
3569 }
3570
3571 /* We have an implicit hole, NO_HOLES feature is likely set. */
3572 if (last_extent_end < key.offset) {
3573 u64 search_start = last_extent_end;
3574 u64 found_start;
3575
3576 /*
3577 * First iteration, @start matches @offset and it's
3578 * within the hole.
3579 */
3580 if (start == offset)
3581 search_start = offset;
3582
3583 found = find_desired_extent_in_hole(inode, whence,
3584 delalloc_cached_state,
3585 search_start,
3586 key.offset - 1,
3587 &found_start);
3588 if (found) {
3589 start = found_start;
3590 break;
3591 }
3592 /*
3593 * Didn't find data or a hole (due to delalloc) in the
3594 * implicit hole range, so need to analyze the extent.
3595 */
3596 }
3597
3598 extent = btrfs_item_ptr(leaf, path->slots[0],
3599 struct btrfs_file_extent_item);
3600 type = btrfs_file_extent_type(leaf, extent);
3601
3602 /*
3603 * Can't access the extent's disk_bytenr field if this is an
3604 * inline extent, since at that offset, it's where the extent
3605 * data starts.
3606 */
3607 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
3608 (type == BTRFS_FILE_EXTENT_REG &&
3609 btrfs_file_extent_disk_bytenr(leaf, extent) == 0)) {
3610 /*
3611 * Explicit hole or prealloc extent, search for delalloc.
3612 * A prealloc extent is treated like a hole.
3613 */
3614 u64 search_start = key.offset;
3615 u64 found_start;
3616
3617 /*
3618 * First iteration, @start matches @offset and it's
3619 * within the hole.
3620 */
3621 if (start == offset)
3622 search_start = offset;
3623
3624 found = find_desired_extent_in_hole(inode, whence,
3625 delalloc_cached_state,
3626 search_start,
3627 extent_end - 1,
3628 &found_start);
3629 if (found) {
3630 start = found_start;
3631 break;
3632 }
3633 /*
3634 * Didn't find data or a hole (due to delalloc) in the
3635 * implicit hole range, so need to analyze the next
3636 * extent item.
3637 */
3638 } else {
3639 /*
3640 * Found a regular or inline extent.
3641 * If we are seeking for data, adjust the start offset
3642 * and stop, we're done.
3643 */
3644 if (whence == SEEK_DATA) {
3645 start = max_t(u64, key.offset, offset);
3646 found = true;
3647 break;
3648 }
3649 /*
3650 * Else, we are seeking for a hole, check the next file
3651 * extent item.
3652 */
3653 }
3654
3655 start = extent_end;
3656 last_extent_end = extent_end;
3657 path->slots[0]++;
3658 if (fatal_signal_pending(current)) {
3659 ret = -EINTR;
3660 goto out;
3661 }
3662 cond_resched();
3663 }
3664
3665 /* We have an implicit hole from the last extent found up to i_size. */
3666 if (!found && start < i_size) {
3667 found = find_desired_extent_in_hole(inode, whence,
3668 delalloc_cached_state, start,
3669 i_size - 1, &start);
3670 if (!found)
3671 start = i_size;
3672 }
3673
3674out:
3675 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3676 btrfs_free_path(path);
3677
3678 if (ret < 0)
3679 return ret;
3680
3681 if (whence == SEEK_DATA && start >= i_size)
3682 return -ENXIO;
3683
3684 return min_t(loff_t, start, i_size);
3685}
3686
3687static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3688{
3689 struct inode *inode = file->f_mapping->host;
3690
3691 switch (whence) {
3692 default:
3693 return generic_file_llseek(file, offset, whence);
3694 case SEEK_DATA:
3695 case SEEK_HOLE:
3696 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3697 offset = find_desired_extent(file, offset, whence);
3698 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3699 break;
3700 }
3701
3702 if (offset < 0)
3703 return offset;
3704
3705 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3706}
3707
3708static int btrfs_file_open(struct inode *inode, struct file *filp)
3709{
3710 int ret;
3711
3712 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC;
3713
3714 ret = fsverity_file_open(inode, filp);
3715 if (ret)
3716 return ret;
3717 return generic_file_open(inode, filp);
3718}
3719
3720static int check_direct_read(struct btrfs_fs_info *fs_info,
3721 const struct iov_iter *iter, loff_t offset)
3722{
3723 int ret;
3724 int i, seg;
3725
3726 ret = check_direct_IO(fs_info, iter, offset);
3727 if (ret < 0)
3728 return ret;
3729
3730 if (!iter_is_iovec(iter))
3731 return 0;
3732
3733 for (seg = 0; seg < iter->nr_segs; seg++)
3734 for (i = seg + 1; i < iter->nr_segs; i++)
3735 if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3736 return -EINVAL;
3737 return 0;
3738}
3739
3740static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3741{
3742 struct inode *inode = file_inode(iocb->ki_filp);
3743 size_t prev_left = 0;
3744 ssize_t read = 0;
3745 ssize_t ret;
3746
3747 if (fsverity_active(inode))
3748 return 0;
3749
3750 if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3751 return 0;
3752
3753 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3754again:
3755 /*
3756 * This is similar to what we do for direct IO writes, see the comment
3757 * at btrfs_direct_write(), but we also disable page faults in addition
3758 * to disabling them only at the iov_iter level. This is because when
3759 * reading from a hole or prealloc extent, iomap calls iov_iter_zero(),
3760 * which can still trigger page fault ins despite having set ->nofault
3761 * to true of our 'to' iov_iter.
3762 *
3763 * The difference to direct IO writes is that we deadlock when trying
3764 * to lock the extent range in the inode's tree during he page reads
3765 * triggered by the fault in (while for writes it is due to waiting for
3766 * our own ordered extent). This is because for direct IO reads,
3767 * btrfs_dio_iomap_begin() returns with the extent range locked, which
3768 * is only unlocked in the endio callback (end_bio_extent_readpage()).
3769 */
3770 pagefault_disable();
3771 to->nofault = true;
3772 ret = btrfs_dio_read(iocb, to, read);
3773 to->nofault = false;
3774 pagefault_enable();
3775
3776 /* No increment (+=) because iomap returns a cumulative value. */
3777 if (ret > 0)
3778 read = ret;
3779
3780 if (iov_iter_count(to) > 0 && (ret == -EFAULT || ret > 0)) {
3781 const size_t left = iov_iter_count(to);
3782
3783 if (left == prev_left) {
3784 /*
3785 * We didn't make any progress since the last attempt,
3786 * fallback to a buffered read for the remainder of the
3787 * range. This is just to avoid any possibility of looping
3788 * for too long.
3789 */
3790 ret = read;
3791 } else {
3792 /*
3793 * We made some progress since the last retry or this is
3794 * the first time we are retrying. Fault in as many pages
3795 * as possible and retry.
3796 */
3797 fault_in_iov_iter_writeable(to, left);
3798 prev_left = left;
3799 goto again;
3800 }
3801 }
3802 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3803 return ret < 0 ? ret : read;
3804}
3805
3806static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3807{
3808 ssize_t ret = 0;
3809
3810 if (iocb->ki_flags & IOCB_DIRECT) {
3811 ret = btrfs_direct_read(iocb, to);
3812 if (ret < 0 || !iov_iter_count(to) ||
3813 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3814 return ret;
3815 }
3816
3817 return filemap_read(iocb, to, ret);
3818}
3819
3820const struct file_operations btrfs_file_operations = {
3821 .llseek = btrfs_file_llseek,
3822 .read_iter = btrfs_file_read_iter,
3823 .splice_read = generic_file_splice_read,
3824 .write_iter = btrfs_file_write_iter,
3825 .splice_write = iter_file_splice_write,
3826 .mmap = btrfs_file_mmap,
3827 .open = btrfs_file_open,
3828 .release = btrfs_release_file,
3829 .get_unmapped_area = thp_get_unmapped_area,
3830 .fsync = btrfs_sync_file,
3831 .fallocate = btrfs_fallocate,
3832 .unlocked_ioctl = btrfs_ioctl,
3833#ifdef CONFIG_COMPAT
3834 .compat_ioctl = btrfs_compat_ioctl,
3835#endif
3836 .remap_file_range = btrfs_remap_file_range,
3837};
3838
3839int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3840{
3841 int ret;
3842
3843 /*
3844 * So with compression we will find and lock a dirty page and clear the
3845 * first one as dirty, setup an async extent, and immediately return
3846 * with the entire range locked but with nobody actually marked with
3847 * writeback. So we can't just filemap_write_and_wait_range() and
3848 * expect it to work since it will just kick off a thread to do the
3849 * actual work. So we need to call filemap_fdatawrite_range _again_
3850 * since it will wait on the page lock, which won't be unlocked until
3851 * after the pages have been marked as writeback and so we're good to go
3852 * from there. We have to do this otherwise we'll miss the ordered
3853 * extents and that results in badness. Please Josef, do not think you
3854 * know better and pull this out at some point in the future, it is
3855 * right and you are wrong.
3856 */
3857 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3858 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3859 &BTRFS_I(inode)->runtime_flags))
3860 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3861
3862 return ret;
3863}