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
34 static struct kmem_cache *btrfs_inode_defrag_cachep;
35 /*
36 * when auto defrag is enabled we
37 * queue up these defrag structs to remember which
38 * inodes need defragging passes
39 */
40 struct inode_defrag {
41 struct rb_node rb_node;
42 /* objectid */
43 u64 ino;
44 /*
45 * transid where the defrag was added, we search for
46 * extents newer than this
47 */
48 u64 transid;
49
50 /* root objectid */
51 u64 root;
52
53 /*
54 * The extent size threshold for autodefrag.
55 *
56 * This value is different for compressed/non-compressed extents,
57 * thus needs to be passed from higher layer.
58 * (aka, inode_should_defrag())
59 */
60 u32 extent_thresh;
61 };
62
63 static int __compare_inode_defrag(struct inode_defrag *defrag1,
64 struct inode_defrag *defrag2)
65 {
66 if (defrag1->root > defrag2->root)
67 return 1;
68 else if (defrag1->root < defrag2->root)
69 return -1;
70 else if (defrag1->ino > defrag2->ino)
71 return 1;
72 else if (defrag1->ino < defrag2->ino)
73 return -1;
74 else
75 return 0;
76 }
77
78 /* pop a record for an inode into the defrag tree. The lock
79 * must be held already
80 *
81 * If you're inserting a record for an older transid than an
82 * existing record, the transid already in the tree is lowered
83 *
84 * If an existing record is found the defrag item you
85 * pass in is freed
86 */
87 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
88 struct inode_defrag *defrag)
89 {
90 struct btrfs_fs_info *fs_info = inode->root->fs_info;
91 struct inode_defrag *entry;
92 struct rb_node **p;
93 struct rb_node *parent = NULL;
94 int ret;
95
96 p = &fs_info->defrag_inodes.rb_node;
97 while (*p) {
98 parent = *p;
99 entry = rb_entry(parent, struct inode_defrag, rb_node);
100
101 ret = __compare_inode_defrag(defrag, entry);
102 if (ret < 0)
103 p = &parent->rb_left;
104 else if (ret > 0)
105 p = &parent->rb_right;
106 else {
107 /* if we're reinserting an entry for
108 * an old defrag run, make sure to
109 * lower the transid of our existing record
110 */
111 if (defrag->transid < entry->transid)
112 entry->transid = defrag->transid;
113 entry->extent_thresh = min(defrag->extent_thresh,
114 entry->extent_thresh);
115 return -EEXIST;
116 }
117 }
118 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
119 rb_link_node(&defrag->rb_node, parent, p);
120 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
121 return 0;
122 }
123
124 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
125 {
126 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
127 return 0;
128
129 if (btrfs_fs_closing(fs_info))
130 return 0;
131
132 return 1;
133 }
134
135 /*
136 * insert a defrag record for this inode if auto defrag is
137 * enabled
138 */
139 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
140 struct btrfs_inode *inode, u32 extent_thresh)
141 {
142 struct btrfs_root *root = inode->root;
143 struct btrfs_fs_info *fs_info = root->fs_info;
144 struct inode_defrag *defrag;
145 u64 transid;
146 int ret;
147
148 if (!__need_auto_defrag(fs_info))
149 return 0;
150
151 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
152 return 0;
153
154 if (trans)
155 transid = trans->transid;
156 else
157 transid = inode->root->last_trans;
158
159 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
160 if (!defrag)
161 return -ENOMEM;
162
163 defrag->ino = btrfs_ino(inode);
164 defrag->transid = transid;
165 defrag->root = root->root_key.objectid;
166 defrag->extent_thresh = extent_thresh;
167
168 spin_lock(&fs_info->defrag_inodes_lock);
169 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
170 /*
171 * If we set IN_DEFRAG flag and evict the inode from memory,
172 * and then re-read this inode, this new inode doesn't have
173 * IN_DEFRAG flag. At the case, we may find the existed defrag.
174 */
175 ret = __btrfs_add_inode_defrag(inode, defrag);
176 if (ret)
177 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
178 } else {
179 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
180 }
181 spin_unlock(&fs_info->defrag_inodes_lock);
182 return 0;
183 }
184
185 /*
186 * pick the defragable inode that we want, if it doesn't exist, we will get
187 * the next one.
188 */
189 static struct inode_defrag *
190 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
191 {
192 struct inode_defrag *entry = NULL;
193 struct inode_defrag tmp;
194 struct rb_node *p;
195 struct rb_node *parent = NULL;
196 int ret;
197
198 tmp.ino = ino;
199 tmp.root = root;
200
201 spin_lock(&fs_info->defrag_inodes_lock);
202 p = fs_info->defrag_inodes.rb_node;
203 while (p) {
204 parent = p;
205 entry = rb_entry(parent, struct inode_defrag, rb_node);
206
207 ret = __compare_inode_defrag(&tmp, entry);
208 if (ret < 0)
209 p = parent->rb_left;
210 else if (ret > 0)
211 p = parent->rb_right;
212 else
213 goto out;
214 }
215
216 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
217 parent = rb_next(parent);
218 if (parent)
219 entry = rb_entry(parent, struct inode_defrag, rb_node);
220 else
221 entry = NULL;
222 }
223 out:
224 if (entry)
225 rb_erase(parent, &fs_info->defrag_inodes);
226 spin_unlock(&fs_info->defrag_inodes_lock);
227 return entry;
228 }
229
230 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
231 {
232 struct inode_defrag *defrag;
233 struct rb_node *node;
234
235 spin_lock(&fs_info->defrag_inodes_lock);
236 node = rb_first(&fs_info->defrag_inodes);
237 while (node) {
238 rb_erase(node, &fs_info->defrag_inodes);
239 defrag = rb_entry(node, struct inode_defrag, rb_node);
240 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
241
242 cond_resched_lock(&fs_info->defrag_inodes_lock);
243
244 node = rb_first(&fs_info->defrag_inodes);
245 }
246 spin_unlock(&fs_info->defrag_inodes_lock);
247 }
248
249 #define BTRFS_DEFRAG_BATCH 1024
250
251 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
252 struct inode_defrag *defrag)
253 {
254 struct btrfs_root *inode_root;
255 struct inode *inode;
256 struct btrfs_ioctl_defrag_range_args range;
257 int ret = 0;
258 u64 cur = 0;
259
260 again:
261 if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
262 goto cleanup;
263 if (!__need_auto_defrag(fs_info))
264 goto cleanup;
265
266 /* get the inode */
267 inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
268 if (IS_ERR(inode_root)) {
269 ret = PTR_ERR(inode_root);
270 goto cleanup;
271 }
272
273 inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
274 btrfs_put_root(inode_root);
275 if (IS_ERR(inode)) {
276 ret = PTR_ERR(inode);
277 goto cleanup;
278 }
279
280 if (cur >= i_size_read(inode)) {
281 iput(inode);
282 goto cleanup;
283 }
284
285 /* do a chunk of defrag */
286 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
287 memset(&range, 0, sizeof(range));
288 range.len = (u64)-1;
289 range.start = cur;
290 range.extent_thresh = defrag->extent_thresh;
291
292 sb_start_write(fs_info->sb);
293 ret = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
294 BTRFS_DEFRAG_BATCH);
295 sb_end_write(fs_info->sb);
296 iput(inode);
297
298 if (ret < 0)
299 goto cleanup;
300
301 cur = max(cur + fs_info->sectorsize, range.start);
302 goto again;
303
304 cleanup:
305 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
306 return ret;
307 }
308
309 /*
310 * run through the list of inodes in the FS that need
311 * defragging
312 */
313 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
314 {
315 struct inode_defrag *defrag;
316 u64 first_ino = 0;
317 u64 root_objectid = 0;
318
319 atomic_inc(&fs_info->defrag_running);
320 while (1) {
321 /* Pause the auto defragger. */
322 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
323 &fs_info->fs_state))
324 break;
325
326 if (!__need_auto_defrag(fs_info))
327 break;
328
329 /* find an inode to defrag */
330 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
331 first_ino);
332 if (!defrag) {
333 if (root_objectid || first_ino) {
334 root_objectid = 0;
335 first_ino = 0;
336 continue;
337 } else {
338 break;
339 }
340 }
341
342 first_ino = defrag->ino + 1;
343 root_objectid = defrag->root;
344
345 __btrfs_run_defrag_inode(fs_info, defrag);
346 }
347 atomic_dec(&fs_info->defrag_running);
348
349 /*
350 * during unmount, we use the transaction_wait queue to
351 * wait for the defragger to stop
352 */
353 wake_up(&fs_info->transaction_wait);
354 return 0;
355 }
356
357 /* simple helper to fault in pages and copy. This should go away
358 * and be replaced with calls into generic code.
359 */
360 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
361 struct page **prepared_pages,
362 struct iov_iter *i)
363 {
364 size_t copied = 0;
365 size_t total_copied = 0;
366 int pg = 0;
367 int offset = offset_in_page(pos);
368
369 while (write_bytes > 0) {
370 size_t count = min_t(size_t,
371 PAGE_SIZE - offset, write_bytes);
372 struct page *page = prepared_pages[pg];
373 /*
374 * Copy data from userspace to the current page
375 */
376 copied = copy_page_from_iter_atomic(page, offset, count, i);
377
378 /* Flush processor's dcache for this page */
379 flush_dcache_page(page);
380
381 /*
382 * if we get a partial write, we can end up with
383 * partially up to date pages. These add
384 * a lot of complexity, so make sure they don't
385 * happen by forcing this copy to be retried.
386 *
387 * The rest of the btrfs_file_write code will fall
388 * back to page at a time copies after we return 0.
389 */
390 if (unlikely(copied < count)) {
391 if (!PageUptodate(page)) {
392 iov_iter_revert(i, copied);
393 copied = 0;
394 }
395 if (!copied)
396 break;
397 }
398
399 write_bytes -= copied;
400 total_copied += copied;
401 offset += copied;
402 if (offset == PAGE_SIZE) {
403 pg++;
404 offset = 0;
405 }
406 }
407 return total_copied;
408 }
409
410 /*
411 * unlocks pages after btrfs_file_write is done with them
412 */
413 static void btrfs_drop_pages(struct btrfs_fs_info *fs_info,
414 struct page **pages, size_t num_pages,
415 u64 pos, u64 copied)
416 {
417 size_t i;
418 u64 block_start = round_down(pos, fs_info->sectorsize);
419 u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
420
421 ASSERT(block_len <= U32_MAX);
422 for (i = 0; i < num_pages; i++) {
423 /* page checked is some magic around finding pages that
424 * have been modified without going through btrfs_set_page_dirty
425 * clear it here. There should be no need to mark the pages
426 * accessed as prepare_pages should have marked them accessed
427 * in prepare_pages via find_or_create_page()
428 */
429 btrfs_page_clamp_clear_checked(fs_info, pages[i], block_start,
430 block_len);
431 unlock_page(pages[i]);
432 put_page(pages[i]);
433 }
434 }
435
436 /*
437 * After btrfs_copy_from_user(), update the following things for delalloc:
438 * - Mark newly dirtied pages as DELALLOC in the io tree.
439 * Used to advise which range is to be written back.
440 * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
441 * - Update inode size for past EOF write
442 */
443 int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
444 size_t num_pages, loff_t pos, size_t write_bytes,
445 struct extent_state **cached, bool noreserve)
446 {
447 struct btrfs_fs_info *fs_info = inode->root->fs_info;
448 int err = 0;
449 int i;
450 u64 num_bytes;
451 u64 start_pos;
452 u64 end_of_last_block;
453 u64 end_pos = pos + write_bytes;
454 loff_t isize = i_size_read(&inode->vfs_inode);
455 unsigned int extra_bits = 0;
456
457 if (write_bytes == 0)
458 return 0;
459
460 if (noreserve)
461 extra_bits |= EXTENT_NORESERVE;
462
463 start_pos = round_down(pos, fs_info->sectorsize);
464 num_bytes = round_up(write_bytes + pos - start_pos,
465 fs_info->sectorsize);
466 ASSERT(num_bytes <= U32_MAX);
467
468 end_of_last_block = start_pos + num_bytes - 1;
469
470 /*
471 * The pages may have already been dirty, clear out old accounting so
472 * we can set things up properly
473 */
474 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
475 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
476 0, 0, cached);
477
478 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
479 extra_bits, cached);
480 if (err)
481 return err;
482
483 for (i = 0; i < num_pages; i++) {
484 struct page *p = pages[i];
485
486 btrfs_page_clamp_set_uptodate(fs_info, p, start_pos, num_bytes);
487 btrfs_page_clamp_clear_checked(fs_info, p, start_pos, num_bytes);
488 btrfs_page_clamp_set_dirty(fs_info, p, start_pos, num_bytes);
489 }
490
491 /*
492 * we've only changed i_size in ram, and we haven't updated
493 * the disk i_size. There is no need to log the inode
494 * at this time.
495 */
496 if (end_pos > isize)
497 i_size_write(&inode->vfs_inode, end_pos);
498 return 0;
499 }
500
501 /*
502 * this drops all the extents in the cache that intersect the range
503 * [start, end]. Existing extents are split as required.
504 */
505 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
506 int skip_pinned)
507 {
508 struct extent_map *em;
509 struct extent_map *split = NULL;
510 struct extent_map *split2 = NULL;
511 struct extent_map_tree *em_tree = &inode->extent_tree;
512 u64 len = end - start + 1;
513 u64 gen;
514 int ret;
515 int testend = 1;
516 unsigned long flags;
517 int compressed = 0;
518 bool modified;
519
520 WARN_ON(end < start);
521 if (end == (u64)-1) {
522 len = (u64)-1;
523 testend = 0;
524 }
525 while (1) {
526 bool ends_after_range = false;
527 int no_splits = 0;
528
529 modified = false;
530 if (!split)
531 split = alloc_extent_map();
532 if (!split2)
533 split2 = alloc_extent_map();
534 if (!split || !split2)
535 no_splits = 1;
536
537 write_lock(&em_tree->lock);
538 em = lookup_extent_mapping(em_tree, start, len);
539 if (!em) {
540 write_unlock(&em_tree->lock);
541 break;
542 }
543 if (testend && em->start + em->len > start + len)
544 ends_after_range = true;
545 flags = em->flags;
546 gen = em->generation;
547 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
548 if (ends_after_range) {
549 free_extent_map(em);
550 write_unlock(&em_tree->lock);
551 break;
552 }
553 start = em->start + em->len;
554 if (testend)
555 len = start + len - (em->start + em->len);
556 free_extent_map(em);
557 write_unlock(&em_tree->lock);
558 continue;
559 }
560 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
561 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
562 clear_bit(EXTENT_FLAG_LOGGING, &flags);
563 modified = !list_empty(&em->list);
564 if (no_splits)
565 goto next;
566
567 if (em->start < start) {
568 split->start = em->start;
569 split->len = start - em->start;
570
571 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
572 split->orig_start = em->orig_start;
573 split->block_start = em->block_start;
574
575 if (compressed)
576 split->block_len = em->block_len;
577 else
578 split->block_len = split->len;
579 split->orig_block_len = max(split->block_len,
580 em->orig_block_len);
581 split->ram_bytes = em->ram_bytes;
582 } else {
583 split->orig_start = split->start;
584 split->block_len = 0;
585 split->block_start = em->block_start;
586 split->orig_block_len = 0;
587 split->ram_bytes = split->len;
588 }
589
590 split->generation = gen;
591 split->flags = flags;
592 split->compress_type = em->compress_type;
593 replace_extent_mapping(em_tree, em, split, modified);
594 free_extent_map(split);
595 split = split2;
596 split2 = NULL;
597 }
598 if (ends_after_range) {
599 u64 diff = start + len - em->start;
600
601 split->start = start + len;
602 split->len = em->start + em->len - (start + len);
603 split->flags = flags;
604 split->compress_type = em->compress_type;
605 split->generation = gen;
606
607 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
608 split->orig_block_len = max(em->block_len,
609 em->orig_block_len);
610
611 split->ram_bytes = em->ram_bytes;
612 if (compressed) {
613 split->block_len = em->block_len;
614 split->block_start = em->block_start;
615 split->orig_start = em->orig_start;
616 } else {
617 split->block_len = split->len;
618 split->block_start = em->block_start
619 + diff;
620 split->orig_start = em->orig_start;
621 }
622 } else {
623 split->ram_bytes = split->len;
624 split->orig_start = split->start;
625 split->block_len = 0;
626 split->block_start = em->block_start;
627 split->orig_block_len = 0;
628 }
629
630 if (extent_map_in_tree(em)) {
631 replace_extent_mapping(em_tree, em, split,
632 modified);
633 } else {
634 ret = add_extent_mapping(em_tree, split,
635 modified);
636 /* Logic error, shouldn't happen. */
637 ASSERT(ret == 0);
638 if (WARN_ON(ret != 0) && modified)
639 btrfs_set_inode_full_sync(inode);
640 }
641 free_extent_map(split);
642 split = NULL;
643 }
644 next:
645 if (extent_map_in_tree(em)) {
646 /*
647 * If the extent map is still in the tree it means that
648 * either of the following is true:
649 *
650 * 1) It fits entirely in our range (doesn't end beyond
651 * it or starts before it);
652 *
653 * 2) It starts before our range and/or ends after our
654 * range, and we were not able to allocate the extent
655 * maps for split operations, @split and @split2.
656 *
657 * If we are at case 2) then we just remove the entire
658 * extent map - this is fine since if anyone needs it to
659 * access the subranges outside our range, will just
660 * load it again from the subvolume tree's file extent
661 * item. However if the extent map was in the list of
662 * modified extents, then we must mark the inode for a
663 * full fsync, otherwise a fast fsync will miss this
664 * extent if it's new and needs to be logged.
665 */
666 if ((em->start < start || ends_after_range) && modified) {
667 ASSERT(no_splits);
668 btrfs_set_inode_full_sync(inode);
669 }
670 remove_extent_mapping(em_tree, em);
671 }
672 write_unlock(&em_tree->lock);
673
674 /* once for us */
675 free_extent_map(em);
676 /* once for the tree*/
677 free_extent_map(em);
678 }
679 if (split)
680 free_extent_map(split);
681 if (split2)
682 free_extent_map(split2);
683 }
684
685 /*
686 * this is very complex, but the basic idea is to drop all extents
687 * in the range start - end. hint_block is filled in with a block number
688 * that would be a good hint to the block allocator for this file.
689 *
690 * If an extent intersects the range but is not entirely inside the range
691 * it is either truncated or split. Anything entirely inside the range
692 * is deleted from the tree.
693 *
694 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
695 * to deal with that. We set the field 'bytes_found' of the arguments structure
696 * with the number of allocated bytes found in the target range, so that the
697 * caller can update the inode's number of bytes in an atomic way when
698 * replacing extents in a range to avoid races with stat(2).
699 */
700 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
701 struct btrfs_root *root, struct btrfs_inode *inode,
702 struct btrfs_drop_extents_args *args)
703 {
704 struct btrfs_fs_info *fs_info = root->fs_info;
705 struct extent_buffer *leaf;
706 struct btrfs_file_extent_item *fi;
707 struct btrfs_ref ref = { 0 };
708 struct btrfs_key key;
709 struct btrfs_key new_key;
710 u64 ino = btrfs_ino(inode);
711 u64 search_start = args->start;
712 u64 disk_bytenr = 0;
713 u64 num_bytes = 0;
714 u64 extent_offset = 0;
715 u64 extent_end = 0;
716 u64 last_end = args->start;
717 int del_nr = 0;
718 int del_slot = 0;
719 int extent_type;
720 int recow;
721 int ret;
722 int modify_tree = -1;
723 int update_refs;
724 int found = 0;
725 struct btrfs_path *path = args->path;
726
727 args->bytes_found = 0;
728 args->extent_inserted = false;
729
730 /* Must always have a path if ->replace_extent is true */
731 ASSERT(!(args->replace_extent && !args->path));
732
733 if (!path) {
734 path = btrfs_alloc_path();
735 if (!path) {
736 ret = -ENOMEM;
737 goto out;
738 }
739 }
740
741 if (args->drop_cache)
742 btrfs_drop_extent_cache(inode, args->start, args->end - 1, 0);
743
744 if (args->start >= inode->disk_i_size && !args->replace_extent)
745 modify_tree = 0;
746
747 update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
748 while (1) {
749 recow = 0;
750 ret = btrfs_lookup_file_extent(trans, root, path, ino,
751 search_start, modify_tree);
752 if (ret < 0)
753 break;
754 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
755 leaf = path->nodes[0];
756 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
757 if (key.objectid == ino &&
758 key.type == BTRFS_EXTENT_DATA_KEY)
759 path->slots[0]--;
760 }
761 ret = 0;
762 next_slot:
763 leaf = path->nodes[0];
764 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
765 BUG_ON(del_nr > 0);
766 ret = btrfs_next_leaf(root, path);
767 if (ret < 0)
768 break;
769 if (ret > 0) {
770 ret = 0;
771 break;
772 }
773 leaf = path->nodes[0];
774 recow = 1;
775 }
776
777 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
778
779 if (key.objectid > ino)
780 break;
781 if (WARN_ON_ONCE(key.objectid < ino) ||
782 key.type < BTRFS_EXTENT_DATA_KEY) {
783 ASSERT(del_nr == 0);
784 path->slots[0]++;
785 goto next_slot;
786 }
787 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
788 break;
789
790 fi = btrfs_item_ptr(leaf, path->slots[0],
791 struct btrfs_file_extent_item);
792 extent_type = btrfs_file_extent_type(leaf, fi);
793
794 if (extent_type == BTRFS_FILE_EXTENT_REG ||
795 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
796 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
797 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
798 extent_offset = btrfs_file_extent_offset(leaf, fi);
799 extent_end = key.offset +
800 btrfs_file_extent_num_bytes(leaf, fi);
801 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
802 extent_end = key.offset +
803 btrfs_file_extent_ram_bytes(leaf, fi);
804 } else {
805 /* can't happen */
806 BUG();
807 }
808
809 /*
810 * Don't skip extent items representing 0 byte lengths. They
811 * used to be created (bug) if while punching holes we hit
812 * -ENOSPC condition. So if we find one here, just ensure we
813 * delete it, otherwise we would insert a new file extent item
814 * with the same key (offset) as that 0 bytes length file
815 * extent item in the call to setup_items_for_insert() later
816 * in this function.
817 */
818 if (extent_end == key.offset && extent_end >= search_start) {
819 last_end = extent_end;
820 goto delete_extent_item;
821 }
822
823 if (extent_end <= search_start) {
824 path->slots[0]++;
825 goto next_slot;
826 }
827
828 found = 1;
829 search_start = max(key.offset, args->start);
830 if (recow || !modify_tree) {
831 modify_tree = -1;
832 btrfs_release_path(path);
833 continue;
834 }
835
836 /*
837 * | - range to drop - |
838 * | -------- extent -------- |
839 */
840 if (args->start > key.offset && args->end < extent_end) {
841 BUG_ON(del_nr > 0);
842 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
843 ret = -EOPNOTSUPP;
844 break;
845 }
846
847 memcpy(&new_key, &key, sizeof(new_key));
848 new_key.offset = args->start;
849 ret = btrfs_duplicate_item(trans, root, path,
850 &new_key);
851 if (ret == -EAGAIN) {
852 btrfs_release_path(path);
853 continue;
854 }
855 if (ret < 0)
856 break;
857
858 leaf = path->nodes[0];
859 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
860 struct btrfs_file_extent_item);
861 btrfs_set_file_extent_num_bytes(leaf, fi,
862 args->start - key.offset);
863
864 fi = btrfs_item_ptr(leaf, path->slots[0],
865 struct btrfs_file_extent_item);
866
867 extent_offset += args->start - key.offset;
868 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
869 btrfs_set_file_extent_num_bytes(leaf, fi,
870 extent_end - args->start);
871 btrfs_mark_buffer_dirty(leaf);
872
873 if (update_refs && disk_bytenr > 0) {
874 btrfs_init_generic_ref(&ref,
875 BTRFS_ADD_DELAYED_REF,
876 disk_bytenr, num_bytes, 0);
877 btrfs_init_data_ref(&ref,
878 root->root_key.objectid,
879 new_key.objectid,
880 args->start - extent_offset,
881 0, false);
882 ret = btrfs_inc_extent_ref(trans, &ref);
883 BUG_ON(ret); /* -ENOMEM */
884 }
885 key.offset = args->start;
886 }
887 /*
888 * From here on out we will have actually dropped something, so
889 * last_end can be updated.
890 */
891 last_end = extent_end;
892
893 /*
894 * | ---- range to drop ----- |
895 * | -------- extent -------- |
896 */
897 if (args->start <= key.offset && args->end < extent_end) {
898 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
899 ret = -EOPNOTSUPP;
900 break;
901 }
902
903 memcpy(&new_key, &key, sizeof(new_key));
904 new_key.offset = args->end;
905 btrfs_set_item_key_safe(fs_info, path, &new_key);
906
907 extent_offset += args->end - key.offset;
908 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
909 btrfs_set_file_extent_num_bytes(leaf, fi,
910 extent_end - args->end);
911 btrfs_mark_buffer_dirty(leaf);
912 if (update_refs && disk_bytenr > 0)
913 args->bytes_found += args->end - key.offset;
914 break;
915 }
916
917 search_start = extent_end;
918 /*
919 * | ---- range to drop ----- |
920 * | -------- extent -------- |
921 */
922 if (args->start > key.offset && args->end >= extent_end) {
923 BUG_ON(del_nr > 0);
924 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
925 ret = -EOPNOTSUPP;
926 break;
927 }
928
929 btrfs_set_file_extent_num_bytes(leaf, fi,
930 args->start - key.offset);
931 btrfs_mark_buffer_dirty(leaf);
932 if (update_refs && disk_bytenr > 0)
933 args->bytes_found += extent_end - args->start;
934 if (args->end == extent_end)
935 break;
936
937 path->slots[0]++;
938 goto next_slot;
939 }
940
941 /*
942 * | ---- range to drop ----- |
943 * | ------ extent ------ |
944 */
945 if (args->start <= key.offset && args->end >= extent_end) {
946 delete_extent_item:
947 if (del_nr == 0) {
948 del_slot = path->slots[0];
949 del_nr = 1;
950 } else {
951 BUG_ON(del_slot + del_nr != path->slots[0]);
952 del_nr++;
953 }
954
955 if (update_refs &&
956 extent_type == BTRFS_FILE_EXTENT_INLINE) {
957 args->bytes_found += extent_end - key.offset;
958 extent_end = ALIGN(extent_end,
959 fs_info->sectorsize);
960 } else if (update_refs && disk_bytenr > 0) {
961 btrfs_init_generic_ref(&ref,
962 BTRFS_DROP_DELAYED_REF,
963 disk_bytenr, num_bytes, 0);
964 btrfs_init_data_ref(&ref,
965 root->root_key.objectid,
966 key.objectid,
967 key.offset - extent_offset, 0,
968 false);
969 ret = btrfs_free_extent(trans, &ref);
970 BUG_ON(ret); /* -ENOMEM */
971 args->bytes_found += extent_end - key.offset;
972 }
973
974 if (args->end == extent_end)
975 break;
976
977 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
978 path->slots[0]++;
979 goto next_slot;
980 }
981
982 ret = btrfs_del_items(trans, root, path, del_slot,
983 del_nr);
984 if (ret) {
985 btrfs_abort_transaction(trans, ret);
986 break;
987 }
988
989 del_nr = 0;
990 del_slot = 0;
991
992 btrfs_release_path(path);
993 continue;
994 }
995
996 BUG();
997 }
998
999 if (!ret && del_nr > 0) {
1000 /*
1001 * Set path->slots[0] to first slot, so that after the delete
1002 * if items are move off from our leaf to its immediate left or
1003 * right neighbor leafs, we end up with a correct and adjusted
1004 * path->slots[0] for our insertion (if args->replace_extent).
1005 */
1006 path->slots[0] = del_slot;
1007 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1008 if (ret)
1009 btrfs_abort_transaction(trans, ret);
1010 }
1011
1012 leaf = path->nodes[0];
1013 /*
1014 * If btrfs_del_items() was called, it might have deleted a leaf, in
1015 * which case it unlocked our path, so check path->locks[0] matches a
1016 * write lock.
1017 */
1018 if (!ret && args->replace_extent &&
1019 path->locks[0] == BTRFS_WRITE_LOCK &&
1020 btrfs_leaf_free_space(leaf) >=
1021 sizeof(struct btrfs_item) + args->extent_item_size) {
1022
1023 key.objectid = ino;
1024 key.type = BTRFS_EXTENT_DATA_KEY;
1025 key.offset = args->start;
1026 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1027 struct btrfs_key slot_key;
1028
1029 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1030 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1031 path->slots[0]++;
1032 }
1033 btrfs_setup_item_for_insert(root, path, &key, args->extent_item_size);
1034 args->extent_inserted = true;
1035 }
1036
1037 if (!args->path)
1038 btrfs_free_path(path);
1039 else if (!args->extent_inserted)
1040 btrfs_release_path(path);
1041 out:
1042 args->drop_end = found ? min(args->end, last_end) : args->end;
1043
1044 return ret;
1045 }
1046
1047 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1048 u64 objectid, u64 bytenr, u64 orig_offset,
1049 u64 *start, u64 *end)
1050 {
1051 struct btrfs_file_extent_item *fi;
1052 struct btrfs_key key;
1053 u64 extent_end;
1054
1055 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1056 return 0;
1057
1058 btrfs_item_key_to_cpu(leaf, &key, slot);
1059 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1060 return 0;
1061
1062 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1063 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1064 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1065 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1066 btrfs_file_extent_compression(leaf, fi) ||
1067 btrfs_file_extent_encryption(leaf, fi) ||
1068 btrfs_file_extent_other_encoding(leaf, fi))
1069 return 0;
1070
1071 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1072 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1073 return 0;
1074
1075 *start = key.offset;
1076 *end = extent_end;
1077 return 1;
1078 }
1079
1080 /*
1081 * Mark extent in the range start - end as written.
1082 *
1083 * This changes extent type from 'pre-allocated' to 'regular'. If only
1084 * part of extent is marked as written, the extent will be split into
1085 * two or three.
1086 */
1087 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1088 struct btrfs_inode *inode, u64 start, u64 end)
1089 {
1090 struct btrfs_fs_info *fs_info = trans->fs_info;
1091 struct btrfs_root *root = inode->root;
1092 struct extent_buffer *leaf;
1093 struct btrfs_path *path;
1094 struct btrfs_file_extent_item *fi;
1095 struct btrfs_ref ref = { 0 };
1096 struct btrfs_key key;
1097 struct btrfs_key new_key;
1098 u64 bytenr;
1099 u64 num_bytes;
1100 u64 extent_end;
1101 u64 orig_offset;
1102 u64 other_start;
1103 u64 other_end;
1104 u64 split;
1105 int del_nr = 0;
1106 int del_slot = 0;
1107 int recow;
1108 int ret = 0;
1109 u64 ino = btrfs_ino(inode);
1110
1111 path = btrfs_alloc_path();
1112 if (!path)
1113 return -ENOMEM;
1114 again:
1115 recow = 0;
1116 split = start;
1117 key.objectid = ino;
1118 key.type = BTRFS_EXTENT_DATA_KEY;
1119 key.offset = split;
1120
1121 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1122 if (ret < 0)
1123 goto out;
1124 if (ret > 0 && path->slots[0] > 0)
1125 path->slots[0]--;
1126
1127 leaf = path->nodes[0];
1128 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1129 if (key.objectid != ino ||
1130 key.type != BTRFS_EXTENT_DATA_KEY) {
1131 ret = -EINVAL;
1132 btrfs_abort_transaction(trans, ret);
1133 goto out;
1134 }
1135 fi = btrfs_item_ptr(leaf, path->slots[0],
1136 struct btrfs_file_extent_item);
1137 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1138 ret = -EINVAL;
1139 btrfs_abort_transaction(trans, ret);
1140 goto out;
1141 }
1142 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1143 if (key.offset > start || extent_end < end) {
1144 ret = -EINVAL;
1145 btrfs_abort_transaction(trans, ret);
1146 goto out;
1147 }
1148
1149 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1150 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1151 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1152 memcpy(&new_key, &key, sizeof(new_key));
1153
1154 if (start == key.offset && end < extent_end) {
1155 other_start = 0;
1156 other_end = start;
1157 if (extent_mergeable(leaf, path->slots[0] - 1,
1158 ino, bytenr, orig_offset,
1159 &other_start, &other_end)) {
1160 new_key.offset = end;
1161 btrfs_set_item_key_safe(fs_info, path, &new_key);
1162 fi = btrfs_item_ptr(leaf, path->slots[0],
1163 struct btrfs_file_extent_item);
1164 btrfs_set_file_extent_generation(leaf, fi,
1165 trans->transid);
1166 btrfs_set_file_extent_num_bytes(leaf, fi,
1167 extent_end - end);
1168 btrfs_set_file_extent_offset(leaf, fi,
1169 end - orig_offset);
1170 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1171 struct btrfs_file_extent_item);
1172 btrfs_set_file_extent_generation(leaf, fi,
1173 trans->transid);
1174 btrfs_set_file_extent_num_bytes(leaf, fi,
1175 end - other_start);
1176 btrfs_mark_buffer_dirty(leaf);
1177 goto out;
1178 }
1179 }
1180
1181 if (start > key.offset && end == extent_end) {
1182 other_start = end;
1183 other_end = 0;
1184 if (extent_mergeable(leaf, path->slots[0] + 1,
1185 ino, bytenr, orig_offset,
1186 &other_start, &other_end)) {
1187 fi = btrfs_item_ptr(leaf, path->slots[0],
1188 struct btrfs_file_extent_item);
1189 btrfs_set_file_extent_num_bytes(leaf, fi,
1190 start - key.offset);
1191 btrfs_set_file_extent_generation(leaf, fi,
1192 trans->transid);
1193 path->slots[0]++;
1194 new_key.offset = start;
1195 btrfs_set_item_key_safe(fs_info, path, &new_key);
1196
1197 fi = btrfs_item_ptr(leaf, path->slots[0],
1198 struct btrfs_file_extent_item);
1199 btrfs_set_file_extent_generation(leaf, fi,
1200 trans->transid);
1201 btrfs_set_file_extent_num_bytes(leaf, fi,
1202 other_end - start);
1203 btrfs_set_file_extent_offset(leaf, fi,
1204 start - orig_offset);
1205 btrfs_mark_buffer_dirty(leaf);
1206 goto out;
1207 }
1208 }
1209
1210 while (start > key.offset || end < extent_end) {
1211 if (key.offset == start)
1212 split = end;
1213
1214 new_key.offset = split;
1215 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1216 if (ret == -EAGAIN) {
1217 btrfs_release_path(path);
1218 goto again;
1219 }
1220 if (ret < 0) {
1221 btrfs_abort_transaction(trans, ret);
1222 goto out;
1223 }
1224
1225 leaf = path->nodes[0];
1226 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1227 struct btrfs_file_extent_item);
1228 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1229 btrfs_set_file_extent_num_bytes(leaf, fi,
1230 split - key.offset);
1231
1232 fi = btrfs_item_ptr(leaf, path->slots[0],
1233 struct btrfs_file_extent_item);
1234
1235 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1236 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1237 btrfs_set_file_extent_num_bytes(leaf, fi,
1238 extent_end - split);
1239 btrfs_mark_buffer_dirty(leaf);
1240
1241 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1242 num_bytes, 0);
1243 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1244 orig_offset, 0, false);
1245 ret = btrfs_inc_extent_ref(trans, &ref);
1246 if (ret) {
1247 btrfs_abort_transaction(trans, ret);
1248 goto out;
1249 }
1250
1251 if (split == start) {
1252 key.offset = start;
1253 } else {
1254 if (start != key.offset) {
1255 ret = -EINVAL;
1256 btrfs_abort_transaction(trans, ret);
1257 goto out;
1258 }
1259 path->slots[0]--;
1260 extent_end = end;
1261 }
1262 recow = 1;
1263 }
1264
1265 other_start = end;
1266 other_end = 0;
1267 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1268 num_bytes, 0);
1269 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset,
1270 0, false);
1271 if (extent_mergeable(leaf, path->slots[0] + 1,
1272 ino, bytenr, orig_offset,
1273 &other_start, &other_end)) {
1274 if (recow) {
1275 btrfs_release_path(path);
1276 goto again;
1277 }
1278 extent_end = other_end;
1279 del_slot = path->slots[0] + 1;
1280 del_nr++;
1281 ret = btrfs_free_extent(trans, &ref);
1282 if (ret) {
1283 btrfs_abort_transaction(trans, ret);
1284 goto out;
1285 }
1286 }
1287 other_start = 0;
1288 other_end = start;
1289 if (extent_mergeable(leaf, path->slots[0] - 1,
1290 ino, bytenr, orig_offset,
1291 &other_start, &other_end)) {
1292 if (recow) {
1293 btrfs_release_path(path);
1294 goto again;
1295 }
1296 key.offset = other_start;
1297 del_slot = path->slots[0];
1298 del_nr++;
1299 ret = btrfs_free_extent(trans, &ref);
1300 if (ret) {
1301 btrfs_abort_transaction(trans, ret);
1302 goto out;
1303 }
1304 }
1305 if (del_nr == 0) {
1306 fi = btrfs_item_ptr(leaf, path->slots[0],
1307 struct btrfs_file_extent_item);
1308 btrfs_set_file_extent_type(leaf, fi,
1309 BTRFS_FILE_EXTENT_REG);
1310 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1311 btrfs_mark_buffer_dirty(leaf);
1312 } else {
1313 fi = btrfs_item_ptr(leaf, del_slot - 1,
1314 struct btrfs_file_extent_item);
1315 btrfs_set_file_extent_type(leaf, fi,
1316 BTRFS_FILE_EXTENT_REG);
1317 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1318 btrfs_set_file_extent_num_bytes(leaf, fi,
1319 extent_end - key.offset);
1320 btrfs_mark_buffer_dirty(leaf);
1321
1322 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1323 if (ret < 0) {
1324 btrfs_abort_transaction(trans, ret);
1325 goto out;
1326 }
1327 }
1328 out:
1329 btrfs_free_path(path);
1330 return ret;
1331 }
1332
1333 /*
1334 * on error we return an unlocked page and the error value
1335 * on success we return a locked page and 0
1336 */
1337 static int prepare_uptodate_page(struct inode *inode,
1338 struct page *page, u64 pos,
1339 bool force_uptodate)
1340 {
1341 struct folio *folio = page_folio(page);
1342 int ret = 0;
1343
1344 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1345 !PageUptodate(page)) {
1346 ret = btrfs_read_folio(NULL, folio);
1347 if (ret)
1348 return ret;
1349 lock_page(page);
1350 if (!PageUptodate(page)) {
1351 unlock_page(page);
1352 return -EIO;
1353 }
1354
1355 /*
1356 * Since btrfs_read_folio() will unlock the folio before it
1357 * returns, there is a window where btrfs_release_folio() can be
1358 * called to release the page. Here we check both inode
1359 * mapping and PagePrivate() to make sure the page was not
1360 * released.
1361 *
1362 * The private flag check is essential for subpage as we need
1363 * to store extra bitmap using page->private.
1364 */
1365 if (page->mapping != inode->i_mapping || !PagePrivate(page)) {
1366 unlock_page(page);
1367 return -EAGAIN;
1368 }
1369 }
1370 return 0;
1371 }
1372
1373 /*
1374 * this just gets pages into the page cache and locks them down.
1375 */
1376 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1377 size_t num_pages, loff_t pos,
1378 size_t write_bytes, bool force_uptodate)
1379 {
1380 int i;
1381 unsigned long index = pos >> PAGE_SHIFT;
1382 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1383 int err = 0;
1384 int faili;
1385
1386 for (i = 0; i < num_pages; i++) {
1387 again:
1388 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1389 mask | __GFP_WRITE);
1390 if (!pages[i]) {
1391 faili = i - 1;
1392 err = -ENOMEM;
1393 goto fail;
1394 }
1395
1396 err = set_page_extent_mapped(pages[i]);
1397 if (err < 0) {
1398 faili = i;
1399 goto fail;
1400 }
1401
1402 if (i == 0)
1403 err = prepare_uptodate_page(inode, pages[i], pos,
1404 force_uptodate);
1405 if (!err && i == num_pages - 1)
1406 err = prepare_uptodate_page(inode, pages[i],
1407 pos + write_bytes, false);
1408 if (err) {
1409 put_page(pages[i]);
1410 if (err == -EAGAIN) {
1411 err = 0;
1412 goto again;
1413 }
1414 faili = i - 1;
1415 goto fail;
1416 }
1417 wait_on_page_writeback(pages[i]);
1418 }
1419
1420 return 0;
1421 fail:
1422 while (faili >= 0) {
1423 unlock_page(pages[faili]);
1424 put_page(pages[faili]);
1425 faili--;
1426 }
1427 return err;
1428
1429 }
1430
1431 /*
1432 * This function locks the extent and properly waits for data=ordered extents
1433 * to finish before allowing the pages to be modified if need.
1434 *
1435 * The return value:
1436 * 1 - the extent is locked
1437 * 0 - the extent is not locked, and everything is OK
1438 * -EAGAIN - need re-prepare the pages
1439 * the other < 0 number - Something wrong happens
1440 */
1441 static noinline int
1442 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1443 size_t num_pages, loff_t pos,
1444 size_t write_bytes,
1445 u64 *lockstart, u64 *lockend,
1446 struct extent_state **cached_state)
1447 {
1448 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1449 u64 start_pos;
1450 u64 last_pos;
1451 int i;
1452 int ret = 0;
1453
1454 start_pos = round_down(pos, fs_info->sectorsize);
1455 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1456
1457 if (start_pos < inode->vfs_inode.i_size) {
1458 struct btrfs_ordered_extent *ordered;
1459
1460 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1461 cached_state);
1462 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1463 last_pos - start_pos + 1);
1464 if (ordered &&
1465 ordered->file_offset + ordered->num_bytes > start_pos &&
1466 ordered->file_offset <= last_pos) {
1467 unlock_extent_cached(&inode->io_tree, start_pos,
1468 last_pos, cached_state);
1469 for (i = 0; i < num_pages; i++) {
1470 unlock_page(pages[i]);
1471 put_page(pages[i]);
1472 }
1473 btrfs_start_ordered_extent(ordered, 1);
1474 btrfs_put_ordered_extent(ordered);
1475 return -EAGAIN;
1476 }
1477 if (ordered)
1478 btrfs_put_ordered_extent(ordered);
1479
1480 *lockstart = start_pos;
1481 *lockend = last_pos;
1482 ret = 1;
1483 }
1484
1485 /*
1486 * We should be called after prepare_pages() which should have locked
1487 * all pages in the range.
1488 */
1489 for (i = 0; i < num_pages; i++)
1490 WARN_ON(!PageLocked(pages[i]));
1491
1492 return ret;
1493 }
1494
1495 /*
1496 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1497 *
1498 * @pos: File offset.
1499 * @write_bytes: The length to write, will be updated to the nocow writeable
1500 * range.
1501 *
1502 * This function will flush ordered extents in the range to ensure proper
1503 * nocow checks.
1504 *
1505 * Return:
1506 * > 0 If we can nocow, and updates @write_bytes.
1507 * 0 If we can't do a nocow write.
1508 * -EAGAIN If we can't do a nocow write because snapshoting of the inode's
1509 * root is in progress.
1510 * < 0 If an error happened.
1511 *
1512 * NOTE: Callers need to call btrfs_check_nocow_unlock() if we return > 0.
1513 */
1514 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1515 size_t *write_bytes)
1516 {
1517 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1518 struct btrfs_root *root = inode->root;
1519 u64 lockstart, lockend;
1520 u64 num_bytes;
1521 int ret;
1522
1523 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1524 return 0;
1525
1526 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1527 return -EAGAIN;
1528
1529 lockstart = round_down(pos, fs_info->sectorsize);
1530 lockend = round_up(pos + *write_bytes,
1531 fs_info->sectorsize) - 1;
1532 num_bytes = lockend - lockstart + 1;
1533
1534 btrfs_lock_and_flush_ordered_range(inode, lockstart, lockend, NULL);
1535 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1536 NULL, NULL, NULL, false);
1537 if (ret <= 0) {
1538 ret = 0;
1539 btrfs_drew_write_unlock(&root->snapshot_lock);
1540 } else {
1541 *write_bytes = min_t(size_t, *write_bytes ,
1542 num_bytes - pos + lockstart);
1543 }
1544 unlock_extent(&inode->io_tree, lockstart, lockend);
1545
1546 return ret;
1547 }
1548
1549 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1550 {
1551 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1552 }
1553
1554 static void update_time_for_write(struct inode *inode)
1555 {
1556 struct timespec64 now;
1557
1558 if (IS_NOCMTIME(inode))
1559 return;
1560
1561 now = current_time(inode);
1562 if (!timespec64_equal(&inode->i_mtime, &now))
1563 inode->i_mtime = now;
1564
1565 if (!timespec64_equal(&inode->i_ctime, &now))
1566 inode->i_ctime = now;
1567
1568 if (IS_I_VERSION(inode))
1569 inode_inc_iversion(inode);
1570 }
1571
1572 static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1573 size_t count)
1574 {
1575 struct file *file = iocb->ki_filp;
1576 struct inode *inode = file_inode(file);
1577 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1578 loff_t pos = iocb->ki_pos;
1579 int ret;
1580 loff_t oldsize;
1581 loff_t start_pos;
1582
1583 /*
1584 * Quickly bail out on NOWAIT writes if we don't have the nodatacow or
1585 * prealloc flags, as without those flags we always have to COW. We will
1586 * later check if we can really COW into the target range (using
1587 * can_nocow_extent() at btrfs_get_blocks_direct_write()).
1588 */
1589 if ((iocb->ki_flags & IOCB_NOWAIT) &&
1590 !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1591 return -EAGAIN;
1592
1593 current->backing_dev_info = inode_to_bdi(inode);
1594 ret = file_remove_privs(file);
1595 if (ret)
1596 return ret;
1597
1598 /*
1599 * We reserve space for updating the inode when we reserve space for the
1600 * extent we are going to write, so we will enospc out there. We don't
1601 * need to start yet another transaction to update the inode as we will
1602 * update the inode when we finish writing whatever data we write.
1603 */
1604 update_time_for_write(inode);
1605
1606 start_pos = round_down(pos, fs_info->sectorsize);
1607 oldsize = i_size_read(inode);
1608 if (start_pos > oldsize) {
1609 /* Expand hole size to cover write data, preventing empty gap */
1610 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1611
1612 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1613 if (ret) {
1614 current->backing_dev_info = NULL;
1615 return ret;
1616 }
1617 }
1618
1619 return 0;
1620 }
1621
1622 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1623 struct iov_iter *i)
1624 {
1625 struct file *file = iocb->ki_filp;
1626 loff_t pos;
1627 struct inode *inode = file_inode(file);
1628 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1629 struct page **pages = NULL;
1630 struct extent_changeset *data_reserved = NULL;
1631 u64 release_bytes = 0;
1632 u64 lockstart;
1633 u64 lockend;
1634 size_t num_written = 0;
1635 int nrptrs;
1636 ssize_t ret;
1637 bool only_release_metadata = false;
1638 bool force_page_uptodate = false;
1639 loff_t old_isize = i_size_read(inode);
1640 unsigned int ilock_flags = 0;
1641
1642 if (iocb->ki_flags & IOCB_NOWAIT)
1643 ilock_flags |= BTRFS_ILOCK_TRY;
1644
1645 ret = btrfs_inode_lock(inode, ilock_flags);
1646 if (ret < 0)
1647 return ret;
1648
1649 ret = generic_write_checks(iocb, i);
1650 if (ret <= 0)
1651 goto out;
1652
1653 ret = btrfs_write_check(iocb, i, ret);
1654 if (ret < 0)
1655 goto out;
1656
1657 pos = iocb->ki_pos;
1658 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1659 PAGE_SIZE / (sizeof(struct page *)));
1660 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1661 nrptrs = max(nrptrs, 8);
1662 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1663 if (!pages) {
1664 ret = -ENOMEM;
1665 goto out;
1666 }
1667
1668 while (iov_iter_count(i) > 0) {
1669 struct extent_state *cached_state = NULL;
1670 size_t offset = offset_in_page(pos);
1671 size_t sector_offset;
1672 size_t write_bytes = min(iov_iter_count(i),
1673 nrptrs * (size_t)PAGE_SIZE -
1674 offset);
1675 size_t num_pages;
1676 size_t reserve_bytes;
1677 size_t dirty_pages;
1678 size_t copied;
1679 size_t dirty_sectors;
1680 size_t num_sectors;
1681 int extents_locked;
1682
1683 /*
1684 * Fault pages before locking them in prepare_pages
1685 * to avoid recursive lock
1686 */
1687 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1688 ret = -EFAULT;
1689 break;
1690 }
1691
1692 only_release_metadata = false;
1693 sector_offset = pos & (fs_info->sectorsize - 1);
1694
1695 extent_changeset_release(data_reserved);
1696 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1697 &data_reserved, pos,
1698 write_bytes);
1699 if (ret < 0) {
1700 /*
1701 * If we don't have to COW at the offset, reserve
1702 * metadata only. write_bytes may get smaller than
1703 * requested here.
1704 */
1705 if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1706 &write_bytes) > 0)
1707 only_release_metadata = true;
1708 else
1709 break;
1710 }
1711
1712 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1713 WARN_ON(num_pages > nrptrs);
1714 reserve_bytes = round_up(write_bytes + sector_offset,
1715 fs_info->sectorsize);
1716 WARN_ON(reserve_bytes == 0);
1717 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1718 reserve_bytes,
1719 reserve_bytes, false);
1720 if (ret) {
1721 if (!only_release_metadata)
1722 btrfs_free_reserved_data_space(BTRFS_I(inode),
1723 data_reserved, pos,
1724 write_bytes);
1725 else
1726 btrfs_check_nocow_unlock(BTRFS_I(inode));
1727 break;
1728 }
1729
1730 release_bytes = reserve_bytes;
1731 again:
1732 /*
1733 * This is going to setup the pages array with the number of
1734 * pages we want, so we don't really need to worry about the
1735 * contents of pages from loop to loop
1736 */
1737 ret = prepare_pages(inode, pages, num_pages,
1738 pos, write_bytes,
1739 force_page_uptodate);
1740 if (ret) {
1741 btrfs_delalloc_release_extents(BTRFS_I(inode),
1742 reserve_bytes);
1743 break;
1744 }
1745
1746 extents_locked = lock_and_cleanup_extent_if_need(
1747 BTRFS_I(inode), pages,
1748 num_pages, pos, write_bytes, &lockstart,
1749 &lockend, &cached_state);
1750 if (extents_locked < 0) {
1751 if (extents_locked == -EAGAIN)
1752 goto again;
1753 btrfs_delalloc_release_extents(BTRFS_I(inode),
1754 reserve_bytes);
1755 ret = extents_locked;
1756 break;
1757 }
1758
1759 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1760
1761 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1762 dirty_sectors = round_up(copied + sector_offset,
1763 fs_info->sectorsize);
1764 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1765
1766 /*
1767 * if we have trouble faulting in the pages, fall
1768 * back to one page at a time
1769 */
1770 if (copied < write_bytes)
1771 nrptrs = 1;
1772
1773 if (copied == 0) {
1774 force_page_uptodate = true;
1775 dirty_sectors = 0;
1776 dirty_pages = 0;
1777 } else {
1778 force_page_uptodate = false;
1779 dirty_pages = DIV_ROUND_UP(copied + offset,
1780 PAGE_SIZE);
1781 }
1782
1783 if (num_sectors > dirty_sectors) {
1784 /* release everything except the sectors we dirtied */
1785 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1786 if (only_release_metadata) {
1787 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1788 release_bytes, true);
1789 } else {
1790 u64 __pos;
1791
1792 __pos = round_down(pos,
1793 fs_info->sectorsize) +
1794 (dirty_pages << PAGE_SHIFT);
1795 btrfs_delalloc_release_space(BTRFS_I(inode),
1796 data_reserved, __pos,
1797 release_bytes, true);
1798 }
1799 }
1800
1801 release_bytes = round_up(copied + sector_offset,
1802 fs_info->sectorsize);
1803
1804 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1805 dirty_pages, pos, copied,
1806 &cached_state, only_release_metadata);
1807
1808 /*
1809 * If we have not locked the extent range, because the range's
1810 * start offset is >= i_size, we might still have a non-NULL
1811 * cached extent state, acquired while marking the extent range
1812 * as delalloc through btrfs_dirty_pages(). Therefore free any
1813 * possible cached extent state to avoid a memory leak.
1814 */
1815 if (extents_locked)
1816 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1817 lockstart, lockend, &cached_state);
1818 else
1819 free_extent_state(cached_state);
1820
1821 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1822 if (ret) {
1823 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1824 break;
1825 }
1826
1827 release_bytes = 0;
1828 if (only_release_metadata)
1829 btrfs_check_nocow_unlock(BTRFS_I(inode));
1830
1831 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1832
1833 cond_resched();
1834
1835 balance_dirty_pages_ratelimited(inode->i_mapping);
1836
1837 pos += copied;
1838 num_written += copied;
1839 }
1840
1841 kfree(pages);
1842
1843 if (release_bytes) {
1844 if (only_release_metadata) {
1845 btrfs_check_nocow_unlock(BTRFS_I(inode));
1846 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1847 release_bytes, true);
1848 } else {
1849 btrfs_delalloc_release_space(BTRFS_I(inode),
1850 data_reserved,
1851 round_down(pos, fs_info->sectorsize),
1852 release_bytes, true);
1853 }
1854 }
1855
1856 extent_changeset_free(data_reserved);
1857 if (num_written > 0) {
1858 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1859 iocb->ki_pos += num_written;
1860 }
1861 out:
1862 btrfs_inode_unlock(inode, ilock_flags);
1863 return num_written ? num_written : ret;
1864 }
1865
1866 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1867 const struct iov_iter *iter, loff_t offset)
1868 {
1869 const u32 blocksize_mask = fs_info->sectorsize - 1;
1870
1871 if (offset & blocksize_mask)
1872 return -EINVAL;
1873
1874 if (iov_iter_alignment(iter) & blocksize_mask)
1875 return -EINVAL;
1876
1877 return 0;
1878 }
1879
1880 static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1881 {
1882 const bool is_sync_write = (iocb->ki_flags & IOCB_DSYNC);
1883 struct file *file = iocb->ki_filp;
1884 struct inode *inode = file_inode(file);
1885 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1886 loff_t pos;
1887 ssize_t written = 0;
1888 ssize_t written_buffered;
1889 size_t prev_left = 0;
1890 loff_t endbyte;
1891 ssize_t err;
1892 unsigned int ilock_flags = 0;
1893
1894 if (iocb->ki_flags & IOCB_NOWAIT)
1895 ilock_flags |= BTRFS_ILOCK_TRY;
1896
1897 /* If the write DIO is within EOF, use a shared lock */
1898 if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1899 ilock_flags |= BTRFS_ILOCK_SHARED;
1900
1901 relock:
1902 err = btrfs_inode_lock(inode, ilock_flags);
1903 if (err < 0)
1904 return err;
1905
1906 err = generic_write_checks(iocb, from);
1907 if (err <= 0) {
1908 btrfs_inode_unlock(inode, ilock_flags);
1909 return err;
1910 }
1911
1912 err = btrfs_write_check(iocb, from, err);
1913 if (err < 0) {
1914 btrfs_inode_unlock(inode, ilock_flags);
1915 goto out;
1916 }
1917
1918 pos = iocb->ki_pos;
1919 /*
1920 * Re-check since file size may have changed just before taking the
1921 * lock or pos may have changed because of O_APPEND in generic_write_check()
1922 */
1923 if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1924 pos + iov_iter_count(from) > i_size_read(inode)) {
1925 btrfs_inode_unlock(inode, ilock_flags);
1926 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1927 goto relock;
1928 }
1929
1930 if (check_direct_IO(fs_info, from, pos)) {
1931 btrfs_inode_unlock(inode, ilock_flags);
1932 goto buffered;
1933 }
1934
1935 /*
1936 * We remove IOCB_DSYNC so that we don't deadlock when iomap_dio_rw()
1937 * calls generic_write_sync() (through iomap_dio_complete()), because
1938 * that results in calling fsync (btrfs_sync_file()) which will try to
1939 * lock the inode in exclusive/write mode.
1940 */
1941 if (is_sync_write)
1942 iocb->ki_flags &= ~IOCB_DSYNC;
1943
1944 /*
1945 * The iov_iter can be mapped to the same file range we are writing to.
1946 * If that's the case, then we will deadlock in the iomap code, because
1947 * it first calls our callback btrfs_dio_iomap_begin(), which will create
1948 * an ordered extent, and after that it will fault in the pages that the
1949 * iov_iter refers to. During the fault in we end up in the readahead
1950 * pages code (starting at btrfs_readahead()), which will lock the range,
1951 * find that ordered extent and then wait for it to complete (at
1952 * btrfs_lock_and_flush_ordered_range()), resulting in a deadlock since
1953 * obviously the ordered extent can never complete as we didn't submit
1954 * yet the respective bio(s). This always happens when the buffer is
1955 * memory mapped to the same file range, since the iomap DIO code always
1956 * invalidates pages in the target file range (after starting and waiting
1957 * for any writeback).
1958 *
1959 * So here we disable page faults in the iov_iter and then retry if we
1960 * got -EFAULT, faulting in the pages before the retry.
1961 */
1962 again:
1963 from->nofault = true;
1964 err = btrfs_dio_rw(iocb, from, written);
1965 from->nofault = false;
1966
1967 /* No increment (+=) because iomap returns a cumulative value. */
1968 if (err > 0)
1969 written = err;
1970
1971 if (iov_iter_count(from) > 0 && (err == -EFAULT || err > 0)) {
1972 const size_t left = iov_iter_count(from);
1973 /*
1974 * We have more data left to write. Try to fault in as many as
1975 * possible of the remainder pages and retry. We do this without
1976 * releasing and locking again the inode, to prevent races with
1977 * truncate.
1978 *
1979 * Also, in case the iov refers to pages in the file range of the
1980 * file we want to write to (due to a mmap), we could enter an
1981 * infinite loop if we retry after faulting the pages in, since
1982 * iomap will invalidate any pages in the range early on, before
1983 * it tries to fault in the pages of the iov. So we keep track of
1984 * how much was left of iov in the previous EFAULT and fallback
1985 * to buffered IO in case we haven't made any progress.
1986 */
1987 if (left == prev_left) {
1988 err = -ENOTBLK;
1989 } else {
1990 fault_in_iov_iter_readable(from, left);
1991 prev_left = left;
1992 goto again;
1993 }
1994 }
1995
1996 btrfs_inode_unlock(inode, ilock_flags);
1997
1998 /*
1999 * Add back IOCB_DSYNC. Our caller, btrfs_file_write_iter(), will do
2000 * the fsync (call generic_write_sync()).
2001 */
2002 if (is_sync_write)
2003 iocb->ki_flags |= IOCB_DSYNC;
2004
2005 /* If 'err' is -ENOTBLK then it means we must fallback to buffered IO. */
2006 if ((err < 0 && err != -ENOTBLK) || !iov_iter_count(from))
2007 goto out;
2008
2009 buffered:
2010 pos = iocb->ki_pos;
2011 written_buffered = btrfs_buffered_write(iocb, from);
2012 if (written_buffered < 0) {
2013 err = written_buffered;
2014 goto out;
2015 }
2016 /*
2017 * Ensure all data is persisted. We want the next direct IO read to be
2018 * able to read what was just written.
2019 */
2020 endbyte = pos + written_buffered - 1;
2021 err = btrfs_fdatawrite_range(inode, pos, endbyte);
2022 if (err)
2023 goto out;
2024 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
2025 if (err)
2026 goto out;
2027 written += written_buffered;
2028 iocb->ki_pos = pos + written_buffered;
2029 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
2030 endbyte >> PAGE_SHIFT);
2031 out:
2032 return err < 0 ? err : written;
2033 }
2034
2035 static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
2036 const struct btrfs_ioctl_encoded_io_args *encoded)
2037 {
2038 struct file *file = iocb->ki_filp;
2039 struct inode *inode = file_inode(file);
2040 loff_t count;
2041 ssize_t ret;
2042
2043 btrfs_inode_lock(inode, 0);
2044 count = encoded->len;
2045 ret = generic_write_checks_count(iocb, &count);
2046 if (ret == 0 && count != encoded->len) {
2047 /*
2048 * The write got truncated by generic_write_checks_count(). We
2049 * can't do a partial encoded write.
2050 */
2051 ret = -EFBIG;
2052 }
2053 if (ret || encoded->len == 0)
2054 goto out;
2055
2056 ret = btrfs_write_check(iocb, from, encoded->len);
2057 if (ret < 0)
2058 goto out;
2059
2060 ret = btrfs_do_encoded_write(iocb, from, encoded);
2061 out:
2062 btrfs_inode_unlock(inode, 0);
2063 return ret;
2064 }
2065
2066 ssize_t btrfs_do_write_iter(struct kiocb *iocb, struct iov_iter *from,
2067 const struct btrfs_ioctl_encoded_io_args *encoded)
2068 {
2069 struct file *file = iocb->ki_filp;
2070 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
2071 ssize_t num_written, num_sync;
2072 const bool sync = iocb->ki_flags & IOCB_DSYNC;
2073
2074 /*
2075 * If the fs flips readonly due to some impossible error, although we
2076 * have opened a file as writable, we have to stop this write operation
2077 * to ensure consistency.
2078 */
2079 if (BTRFS_FS_ERROR(inode->root->fs_info))
2080 return -EROFS;
2081
2082 if ((iocb->ki_flags & IOCB_NOWAIT) && !(iocb->ki_flags & IOCB_DIRECT))
2083 return -EOPNOTSUPP;
2084
2085 if (sync)
2086 atomic_inc(&inode->sync_writers);
2087
2088 if (encoded) {
2089 num_written = btrfs_encoded_write(iocb, from, encoded);
2090 num_sync = encoded->len;
2091 } else if (iocb->ki_flags & IOCB_DIRECT) {
2092 num_written = num_sync = btrfs_direct_write(iocb, from);
2093 } else {
2094 num_written = num_sync = btrfs_buffered_write(iocb, from);
2095 }
2096
2097 btrfs_set_inode_last_sub_trans(inode);
2098
2099 if (num_sync > 0) {
2100 num_sync = generic_write_sync(iocb, num_sync);
2101 if (num_sync < 0)
2102 num_written = num_sync;
2103 }
2104
2105 if (sync)
2106 atomic_dec(&inode->sync_writers);
2107
2108 current->backing_dev_info = NULL;
2109 return num_written;
2110 }
2111
2112 static ssize_t btrfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
2113 {
2114 return btrfs_do_write_iter(iocb, from, NULL);
2115 }
2116
2117 int btrfs_release_file(struct inode *inode, struct file *filp)
2118 {
2119 struct btrfs_file_private *private = filp->private_data;
2120
2121 if (private && private->filldir_buf)
2122 kfree(private->filldir_buf);
2123 kfree(private);
2124 filp->private_data = NULL;
2125
2126 /*
2127 * Set by setattr when we are about to truncate a file from a non-zero
2128 * size to a zero size. This tries to flush down new bytes that may
2129 * have been written if the application were using truncate to replace
2130 * a file in place.
2131 */
2132 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2133 &BTRFS_I(inode)->runtime_flags))
2134 filemap_flush(inode->i_mapping);
2135 return 0;
2136 }
2137
2138 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2139 {
2140 int ret;
2141 struct blk_plug plug;
2142
2143 /*
2144 * This is only called in fsync, which would do synchronous writes, so
2145 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2146 * multiple disks using raid profile, a large IO can be split to
2147 * several segments of stripe length (currently 64K).
2148 */
2149 blk_start_plug(&plug);
2150 atomic_inc(&BTRFS_I(inode)->sync_writers);
2151 ret = btrfs_fdatawrite_range(inode, start, end);
2152 atomic_dec(&BTRFS_I(inode)->sync_writers);
2153 blk_finish_plug(&plug);
2154
2155 return ret;
2156 }
2157
2158 static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
2159 {
2160 struct btrfs_inode *inode = BTRFS_I(ctx->inode);
2161 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2162
2163 if (btrfs_inode_in_log(inode, fs_info->generation) &&
2164 list_empty(&ctx->ordered_extents))
2165 return true;
2166
2167 /*
2168 * If we are doing a fast fsync we can not bail out if the inode's
2169 * last_trans is <= then the last committed transaction, because we only
2170 * update the last_trans of the inode during ordered extent completion,
2171 * and for a fast fsync we don't wait for that, we only wait for the
2172 * writeback to complete.
2173 */
2174 if (inode->last_trans <= fs_info->last_trans_committed &&
2175 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
2176 list_empty(&ctx->ordered_extents)))
2177 return true;
2178
2179 return false;
2180 }
2181
2182 /*
2183 * fsync call for both files and directories. This logs the inode into
2184 * the tree log instead of forcing full commits whenever possible.
2185 *
2186 * It needs to call filemap_fdatawait so that all ordered extent updates are
2187 * in the metadata btree are up to date for copying to the log.
2188 *
2189 * It drops the inode mutex before doing the tree log commit. This is an
2190 * important optimization for directories because holding the mutex prevents
2191 * new operations on the dir while we write to disk.
2192 */
2193 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2194 {
2195 struct dentry *dentry = file_dentry(file);
2196 struct inode *inode = d_inode(dentry);
2197 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2198 struct btrfs_root *root = BTRFS_I(inode)->root;
2199 struct btrfs_trans_handle *trans;
2200 struct btrfs_log_ctx ctx;
2201 int ret = 0, err;
2202 u64 len;
2203 bool full_sync;
2204
2205 trace_btrfs_sync_file(file, datasync);
2206
2207 btrfs_init_log_ctx(&ctx, inode);
2208
2209 /*
2210 * Always set the range to a full range, otherwise we can get into
2211 * several problems, from missing file extent items to represent holes
2212 * when not using the NO_HOLES feature, to log tree corruption due to
2213 * races between hole detection during logging and completion of ordered
2214 * extents outside the range, to missing checksums due to ordered extents
2215 * for which we flushed only a subset of their pages.
2216 */
2217 start = 0;
2218 end = LLONG_MAX;
2219 len = (u64)LLONG_MAX + 1;
2220
2221 /*
2222 * We write the dirty pages in the range and wait until they complete
2223 * out of the ->i_mutex. If so, we can flush the dirty pages by
2224 * multi-task, and make the performance up. See
2225 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2226 */
2227 ret = start_ordered_ops(inode, start, end);
2228 if (ret)
2229 goto out;
2230
2231 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2232
2233 atomic_inc(&root->log_batch);
2234
2235 /*
2236 * Before we acquired the inode's lock and the mmap lock, someone may
2237 * have dirtied more pages in the target range. We need to make sure
2238 * that writeback for any such pages does not start while we are logging
2239 * the inode, because if it does, any of the following might happen when
2240 * we are not doing a full inode sync:
2241 *
2242 * 1) We log an extent after its writeback finishes but before its
2243 * checksums are added to the csum tree, leading to -EIO errors
2244 * when attempting to read the extent after a log replay.
2245 *
2246 * 2) We can end up logging an extent before its writeback finishes.
2247 * Therefore after the log replay we will have a file extent item
2248 * pointing to an unwritten extent (and no data checksums as well).
2249 *
2250 * So trigger writeback for any eventual new dirty pages and then we
2251 * wait for all ordered extents to complete below.
2252 */
2253 ret = start_ordered_ops(inode, start, end);
2254 if (ret) {
2255 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2256 goto out;
2257 }
2258
2259 /*
2260 * Always check for the full sync flag while holding the inode's lock,
2261 * to avoid races with other tasks. The flag must be either set all the
2262 * time during logging or always off all the time while logging.
2263 * We check the flag here after starting delalloc above, because when
2264 * running delalloc the full sync flag may be set if we need to drop
2265 * extra extent map ranges due to temporary memory allocation failures.
2266 */
2267 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2268 &BTRFS_I(inode)->runtime_flags);
2269
2270 /*
2271 * We have to do this here to avoid the priority inversion of waiting on
2272 * IO of a lower priority task while holding a transaction open.
2273 *
2274 * For a full fsync we wait for the ordered extents to complete while
2275 * for a fast fsync we wait just for writeback to complete, and then
2276 * attach the ordered extents to the transaction so that a transaction
2277 * commit waits for their completion, to avoid data loss if we fsync,
2278 * the current transaction commits before the ordered extents complete
2279 * and a power failure happens right after that.
2280 *
2281 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
2282 * logical address recorded in the ordered extent may change. We need
2283 * to wait for the IO to stabilize the logical address.
2284 */
2285 if (full_sync || btrfs_is_zoned(fs_info)) {
2286 ret = btrfs_wait_ordered_range(inode, start, len);
2287 } else {
2288 /*
2289 * Get our ordered extents as soon as possible to avoid doing
2290 * checksum lookups in the csum tree, and use instead the
2291 * checksums attached to the ordered extents.
2292 */
2293 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2294 &ctx.ordered_extents);
2295 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2296 }
2297
2298 if (ret)
2299 goto out_release_extents;
2300
2301 atomic_inc(&root->log_batch);
2302
2303 smp_mb();
2304 if (skip_inode_logging(&ctx)) {
2305 /*
2306 * We've had everything committed since the last time we were
2307 * modified so clear this flag in case it was set for whatever
2308 * reason, it's no longer relevant.
2309 */
2310 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2311 &BTRFS_I(inode)->runtime_flags);
2312 /*
2313 * An ordered extent might have started before and completed
2314 * already with io errors, in which case the inode was not
2315 * updated and we end up here. So check the inode's mapping
2316 * for any errors that might have happened since we last
2317 * checked called fsync.
2318 */
2319 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2320 goto out_release_extents;
2321 }
2322
2323 /*
2324 * We use start here because we will need to wait on the IO to complete
2325 * in btrfs_sync_log, which could require joining a transaction (for
2326 * example checking cross references in the nocow path). If we use join
2327 * here we could get into a situation where we're waiting on IO to
2328 * happen that is blocked on a transaction trying to commit. With start
2329 * we inc the extwriter counter, so we wait for all extwriters to exit
2330 * before we start blocking joiners. This comment is to keep somebody
2331 * from thinking they are super smart and changing this to
2332 * btrfs_join_transaction *cough*Josef*cough*.
2333 */
2334 trans = btrfs_start_transaction(root, 0);
2335 if (IS_ERR(trans)) {
2336 ret = PTR_ERR(trans);
2337 goto out_release_extents;
2338 }
2339 trans->in_fsync = true;
2340
2341 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2342 btrfs_release_log_ctx_extents(&ctx);
2343 if (ret < 0) {
2344 /* Fallthrough and commit/free transaction. */
2345 ret = BTRFS_LOG_FORCE_COMMIT;
2346 }
2347
2348 /* we've logged all the items and now have a consistent
2349 * version of the file in the log. It is possible that
2350 * someone will come in and modify the file, but that's
2351 * fine because the log is consistent on disk, and we
2352 * have references to all of the file's extents
2353 *
2354 * It is possible that someone will come in and log the
2355 * file again, but that will end up using the synchronization
2356 * inside btrfs_sync_log to keep things safe.
2357 */
2358 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2359
2360 if (ret == BTRFS_NO_LOG_SYNC) {
2361 ret = btrfs_end_transaction(trans);
2362 goto out;
2363 }
2364
2365 /* We successfully logged the inode, attempt to sync the log. */
2366 if (!ret) {
2367 ret = btrfs_sync_log(trans, root, &ctx);
2368 if (!ret) {
2369 ret = btrfs_end_transaction(trans);
2370 goto out;
2371 }
2372 }
2373
2374 /*
2375 * At this point we need to commit the transaction because we had
2376 * btrfs_need_log_full_commit() or some other error.
2377 *
2378 * If we didn't do a full sync we have to stop the trans handle, wait on
2379 * the ordered extents, start it again and commit the transaction. If
2380 * we attempt to wait on the ordered extents here we could deadlock with
2381 * something like fallocate() that is holding the extent lock trying to
2382 * start a transaction while some other thread is trying to commit the
2383 * transaction while we (fsync) are currently holding the transaction
2384 * open.
2385 */
2386 if (!full_sync) {
2387 ret = btrfs_end_transaction(trans);
2388 if (ret)
2389 goto out;
2390 ret = btrfs_wait_ordered_range(inode, start, len);
2391 if (ret)
2392 goto out;
2393
2394 /*
2395 * This is safe to use here because we're only interested in
2396 * making sure the transaction that had the ordered extents is
2397 * committed. We aren't waiting on anything past this point,
2398 * we're purely getting the transaction and committing it.
2399 */
2400 trans = btrfs_attach_transaction_barrier(root);
2401 if (IS_ERR(trans)) {
2402 ret = PTR_ERR(trans);
2403
2404 /*
2405 * We committed the transaction and there's no currently
2406 * running transaction, this means everything we care
2407 * about made it to disk and we are done.
2408 */
2409 if (ret == -ENOENT)
2410 ret = 0;
2411 goto out;
2412 }
2413 }
2414
2415 ret = btrfs_commit_transaction(trans);
2416 out:
2417 ASSERT(list_empty(&ctx.list));
2418 err = file_check_and_advance_wb_err(file);
2419 if (!ret)
2420 ret = err;
2421 return ret > 0 ? -EIO : ret;
2422
2423 out_release_extents:
2424 btrfs_release_log_ctx_extents(&ctx);
2425 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2426 goto out;
2427 }
2428
2429 static const struct vm_operations_struct btrfs_file_vm_ops = {
2430 .fault = filemap_fault,
2431 .map_pages = filemap_map_pages,
2432 .page_mkwrite = btrfs_page_mkwrite,
2433 };
2434
2435 static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2436 {
2437 struct address_space *mapping = filp->f_mapping;
2438
2439 if (!mapping->a_ops->read_folio)
2440 return -ENOEXEC;
2441
2442 file_accessed(filp);
2443 vma->vm_ops = &btrfs_file_vm_ops;
2444
2445 return 0;
2446 }
2447
2448 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2449 int slot, u64 start, u64 end)
2450 {
2451 struct btrfs_file_extent_item *fi;
2452 struct btrfs_key key;
2453
2454 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2455 return 0;
2456
2457 btrfs_item_key_to_cpu(leaf, &key, slot);
2458 if (key.objectid != btrfs_ino(inode) ||
2459 key.type != BTRFS_EXTENT_DATA_KEY)
2460 return 0;
2461
2462 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2463
2464 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2465 return 0;
2466
2467 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2468 return 0;
2469
2470 if (key.offset == end)
2471 return 1;
2472 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2473 return 1;
2474 return 0;
2475 }
2476
2477 static int fill_holes(struct btrfs_trans_handle *trans,
2478 struct btrfs_inode *inode,
2479 struct btrfs_path *path, u64 offset, u64 end)
2480 {
2481 struct btrfs_fs_info *fs_info = trans->fs_info;
2482 struct btrfs_root *root = inode->root;
2483 struct extent_buffer *leaf;
2484 struct btrfs_file_extent_item *fi;
2485 struct extent_map *hole_em;
2486 struct extent_map_tree *em_tree = &inode->extent_tree;
2487 struct btrfs_key key;
2488 int ret;
2489
2490 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2491 goto out;
2492
2493 key.objectid = btrfs_ino(inode);
2494 key.type = BTRFS_EXTENT_DATA_KEY;
2495 key.offset = offset;
2496
2497 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2498 if (ret <= 0) {
2499 /*
2500 * We should have dropped this offset, so if we find it then
2501 * something has gone horribly wrong.
2502 */
2503 if (ret == 0)
2504 ret = -EINVAL;
2505 return ret;
2506 }
2507
2508 leaf = path->nodes[0];
2509 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2510 u64 num_bytes;
2511
2512 path->slots[0]--;
2513 fi = btrfs_item_ptr(leaf, path->slots[0],
2514 struct btrfs_file_extent_item);
2515 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2516 end - offset;
2517 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2518 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2519 btrfs_set_file_extent_offset(leaf, fi, 0);
2520 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2521 btrfs_mark_buffer_dirty(leaf);
2522 goto out;
2523 }
2524
2525 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2526 u64 num_bytes;
2527
2528 key.offset = offset;
2529 btrfs_set_item_key_safe(fs_info, path, &key);
2530 fi = btrfs_item_ptr(leaf, path->slots[0],
2531 struct btrfs_file_extent_item);
2532 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2533 offset;
2534 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2535 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2536 btrfs_set_file_extent_offset(leaf, fi, 0);
2537 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2538 btrfs_mark_buffer_dirty(leaf);
2539 goto out;
2540 }
2541 btrfs_release_path(path);
2542
2543 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2544 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2545 if (ret)
2546 return ret;
2547
2548 out:
2549 btrfs_release_path(path);
2550
2551 hole_em = alloc_extent_map();
2552 if (!hole_em) {
2553 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2554 btrfs_set_inode_full_sync(inode);
2555 } else {
2556 hole_em->start = offset;
2557 hole_em->len = end - offset;
2558 hole_em->ram_bytes = hole_em->len;
2559 hole_em->orig_start = offset;
2560
2561 hole_em->block_start = EXTENT_MAP_HOLE;
2562 hole_em->block_len = 0;
2563 hole_em->orig_block_len = 0;
2564 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2565 hole_em->generation = trans->transid;
2566
2567 do {
2568 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2569 write_lock(&em_tree->lock);
2570 ret = add_extent_mapping(em_tree, hole_em, 1);
2571 write_unlock(&em_tree->lock);
2572 } while (ret == -EEXIST);
2573 free_extent_map(hole_em);
2574 if (ret)
2575 btrfs_set_inode_full_sync(inode);
2576 }
2577
2578 return 0;
2579 }
2580
2581 /*
2582 * Find a hole extent on given inode and change start/len to the end of hole
2583 * extent.(hole/vacuum extent whose em->start <= start &&
2584 * em->start + em->len > start)
2585 * When a hole extent is found, return 1 and modify start/len.
2586 */
2587 static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2588 {
2589 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2590 struct extent_map *em;
2591 int ret = 0;
2592
2593 em = btrfs_get_extent(inode, NULL, 0,
2594 round_down(*start, fs_info->sectorsize),
2595 round_up(*len, fs_info->sectorsize));
2596 if (IS_ERR(em))
2597 return PTR_ERR(em);
2598
2599 /* Hole or vacuum extent(only exists in no-hole mode) */
2600 if (em->block_start == EXTENT_MAP_HOLE) {
2601 ret = 1;
2602 *len = em->start + em->len > *start + *len ?
2603 0 : *start + *len - em->start - em->len;
2604 *start = em->start + em->len;
2605 }
2606 free_extent_map(em);
2607 return ret;
2608 }
2609
2610 static void btrfs_punch_hole_lock_range(struct inode *inode,
2611 const u64 lockstart,
2612 const u64 lockend,
2613 struct extent_state **cached_state)
2614 {
2615 /*
2616 * For subpage case, if the range is not at page boundary, we could
2617 * have pages at the leading/tailing part of the range.
2618 * This could lead to dead loop since filemap_range_has_page()
2619 * will always return true.
2620 * So here we need to do extra page alignment for
2621 * filemap_range_has_page().
2622 */
2623 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2624 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2625
2626 while (1) {
2627 truncate_pagecache_range(inode, lockstart, lockend);
2628
2629 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2630 cached_state);
2631 /*
2632 * We can't have ordered extents in the range, nor dirty/writeback
2633 * pages, because we have locked the inode's VFS lock in exclusive
2634 * mode, we have locked the inode's i_mmap_lock in exclusive mode,
2635 * we have flushed all delalloc in the range and we have waited
2636 * for any ordered extents in the range to complete.
2637 * We can race with anyone reading pages from this range, so after
2638 * locking the range check if we have pages in the range, and if
2639 * we do, unlock the range and retry.
2640 */
2641 if (!filemap_range_has_page(inode->i_mapping, page_lockstart,
2642 page_lockend))
2643 break;
2644
2645 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2646 lockend, cached_state);
2647 }
2648
2649 btrfs_assert_inode_range_clean(BTRFS_I(inode), lockstart, lockend);
2650 }
2651
2652 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2653 struct btrfs_inode *inode,
2654 struct btrfs_path *path,
2655 struct btrfs_replace_extent_info *extent_info,
2656 const u64 replace_len,
2657 const u64 bytes_to_drop)
2658 {
2659 struct btrfs_fs_info *fs_info = trans->fs_info;
2660 struct btrfs_root *root = inode->root;
2661 struct btrfs_file_extent_item *extent;
2662 struct extent_buffer *leaf;
2663 struct btrfs_key key;
2664 int slot;
2665 struct btrfs_ref ref = { 0 };
2666 int ret;
2667
2668 if (replace_len == 0)
2669 return 0;
2670
2671 if (extent_info->disk_offset == 0 &&
2672 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2673 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2674 return 0;
2675 }
2676
2677 key.objectid = btrfs_ino(inode);
2678 key.type = BTRFS_EXTENT_DATA_KEY;
2679 key.offset = extent_info->file_offset;
2680 ret = btrfs_insert_empty_item(trans, root, path, &key,
2681 sizeof(struct btrfs_file_extent_item));
2682 if (ret)
2683 return ret;
2684 leaf = path->nodes[0];
2685 slot = path->slots[0];
2686 write_extent_buffer(leaf, extent_info->extent_buf,
2687 btrfs_item_ptr_offset(leaf, slot),
2688 sizeof(struct btrfs_file_extent_item));
2689 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2690 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2691 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2692 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2693 if (extent_info->is_new_extent)
2694 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2695 btrfs_mark_buffer_dirty(leaf);
2696 btrfs_release_path(path);
2697
2698 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2699 replace_len);
2700 if (ret)
2701 return ret;
2702
2703 /* If it's a hole, nothing more needs to be done. */
2704 if (extent_info->disk_offset == 0) {
2705 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2706 return 0;
2707 }
2708
2709 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2710
2711 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2712 key.objectid = extent_info->disk_offset;
2713 key.type = BTRFS_EXTENT_ITEM_KEY;
2714 key.offset = extent_info->disk_len;
2715 ret = btrfs_alloc_reserved_file_extent(trans, root,
2716 btrfs_ino(inode),
2717 extent_info->file_offset,
2718 extent_info->qgroup_reserved,
2719 &key);
2720 } else {
2721 u64 ref_offset;
2722
2723 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2724 extent_info->disk_offset,
2725 extent_info->disk_len, 0);
2726 ref_offset = extent_info->file_offset - extent_info->data_offset;
2727 btrfs_init_data_ref(&ref, root->root_key.objectid,
2728 btrfs_ino(inode), ref_offset, 0, false);
2729 ret = btrfs_inc_extent_ref(trans, &ref);
2730 }
2731
2732 extent_info->insertions++;
2733
2734 return ret;
2735 }
2736
2737 /*
2738 * The respective range must have been previously locked, as well as the inode.
2739 * The end offset is inclusive (last byte of the range).
2740 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2741 * the file range with an extent.
2742 * When not punching a hole, we don't want to end up in a state where we dropped
2743 * extents without inserting a new one, so we must abort the transaction to avoid
2744 * a corruption.
2745 */
2746 int btrfs_replace_file_extents(struct btrfs_inode *inode,
2747 struct btrfs_path *path, const u64 start,
2748 const u64 end,
2749 struct btrfs_replace_extent_info *extent_info,
2750 struct btrfs_trans_handle **trans_out)
2751 {
2752 struct btrfs_drop_extents_args drop_args = { 0 };
2753 struct btrfs_root *root = inode->root;
2754 struct btrfs_fs_info *fs_info = root->fs_info;
2755 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2756 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2757 struct btrfs_trans_handle *trans = NULL;
2758 struct btrfs_block_rsv *rsv;
2759 unsigned int rsv_count;
2760 u64 cur_offset;
2761 u64 len = end - start;
2762 int ret = 0;
2763
2764 if (end <= start)
2765 return -EINVAL;
2766
2767 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2768 if (!rsv) {
2769 ret = -ENOMEM;
2770 goto out;
2771 }
2772 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2773 rsv->failfast = 1;
2774
2775 /*
2776 * 1 - update the inode
2777 * 1 - removing the extents in the range
2778 * 1 - adding the hole extent if no_holes isn't set or if we are
2779 * replacing the range with a new extent
2780 */
2781 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2782 rsv_count = 3;
2783 else
2784 rsv_count = 2;
2785
2786 trans = btrfs_start_transaction(root, rsv_count);
2787 if (IS_ERR(trans)) {
2788 ret = PTR_ERR(trans);
2789 trans = NULL;
2790 goto out_free;
2791 }
2792
2793 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2794 min_size, false);
2795 if (WARN_ON(ret))
2796 goto out_trans;
2797 trans->block_rsv = rsv;
2798
2799 cur_offset = start;
2800 drop_args.path = path;
2801 drop_args.end = end + 1;
2802 drop_args.drop_cache = true;
2803 while (cur_offset < end) {
2804 drop_args.start = cur_offset;
2805 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2806 /* If we are punching a hole decrement the inode's byte count */
2807 if (!extent_info)
2808 btrfs_update_inode_bytes(inode, 0,
2809 drop_args.bytes_found);
2810 if (ret != -ENOSPC) {
2811 /*
2812 * The only time we don't want to abort is if we are
2813 * attempting to clone a partial inline extent, in which
2814 * case we'll get EOPNOTSUPP. However if we aren't
2815 * clone we need to abort no matter what, because if we
2816 * got EOPNOTSUPP via prealloc then we messed up and
2817 * need to abort.
2818 */
2819 if (ret &&
2820 (ret != -EOPNOTSUPP ||
2821 (extent_info && extent_info->is_new_extent)))
2822 btrfs_abort_transaction(trans, ret);
2823 break;
2824 }
2825
2826 trans->block_rsv = &fs_info->trans_block_rsv;
2827
2828 if (!extent_info && cur_offset < drop_args.drop_end &&
2829 cur_offset < ino_size) {
2830 ret = fill_holes(trans, inode, path, cur_offset,
2831 drop_args.drop_end);
2832 if (ret) {
2833 /*
2834 * If we failed then we didn't insert our hole
2835 * entries for the area we dropped, so now the
2836 * fs is corrupted, so we must abort the
2837 * transaction.
2838 */
2839 btrfs_abort_transaction(trans, ret);
2840 break;
2841 }
2842 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2843 /*
2844 * We are past the i_size here, but since we didn't
2845 * insert holes we need to clear the mapped area so we
2846 * know to not set disk_i_size in this area until a new
2847 * file extent is inserted here.
2848 */
2849 ret = btrfs_inode_clear_file_extent_range(inode,
2850 cur_offset,
2851 drop_args.drop_end - cur_offset);
2852 if (ret) {
2853 /*
2854 * We couldn't clear our area, so we could
2855 * presumably adjust up and corrupt the fs, so
2856 * we need to abort.
2857 */
2858 btrfs_abort_transaction(trans, ret);
2859 break;
2860 }
2861 }
2862
2863 if (extent_info &&
2864 drop_args.drop_end > extent_info->file_offset) {
2865 u64 replace_len = drop_args.drop_end -
2866 extent_info->file_offset;
2867
2868 ret = btrfs_insert_replace_extent(trans, inode, path,
2869 extent_info, replace_len,
2870 drop_args.bytes_found);
2871 if (ret) {
2872 btrfs_abort_transaction(trans, ret);
2873 break;
2874 }
2875 extent_info->data_len -= replace_len;
2876 extent_info->data_offset += replace_len;
2877 extent_info->file_offset += replace_len;
2878 }
2879
2880 /*
2881 * We are releasing our handle on the transaction, balance the
2882 * dirty pages of the btree inode and flush delayed items, and
2883 * then get a new transaction handle, which may now point to a
2884 * new transaction in case someone else may have committed the
2885 * transaction we used to replace/drop file extent items. So
2886 * bump the inode's iversion and update mtime and ctime except
2887 * if we are called from a dedupe context. This is because a
2888 * power failure/crash may happen after the transaction is
2889 * committed and before we finish replacing/dropping all the
2890 * file extent items we need.
2891 */
2892 inode_inc_iversion(&inode->vfs_inode);
2893
2894 if (!extent_info || extent_info->update_times) {
2895 inode->vfs_inode.i_mtime = current_time(&inode->vfs_inode);
2896 inode->vfs_inode.i_ctime = inode->vfs_inode.i_mtime;
2897 }
2898
2899 ret = btrfs_update_inode(trans, root, inode);
2900 if (ret)
2901 break;
2902
2903 btrfs_end_transaction(trans);
2904 btrfs_btree_balance_dirty(fs_info);
2905
2906 trans = btrfs_start_transaction(root, rsv_count);
2907 if (IS_ERR(trans)) {
2908 ret = PTR_ERR(trans);
2909 trans = NULL;
2910 break;
2911 }
2912
2913 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2914 rsv, min_size, false);
2915 if (WARN_ON(ret))
2916 break;
2917 trans->block_rsv = rsv;
2918
2919 cur_offset = drop_args.drop_end;
2920 len = end - cur_offset;
2921 if (!extent_info && len) {
2922 ret = find_first_non_hole(inode, &cur_offset, &len);
2923 if (unlikely(ret < 0))
2924 break;
2925 if (ret && !len) {
2926 ret = 0;
2927 break;
2928 }
2929 }
2930 }
2931
2932 /*
2933 * If we were cloning, force the next fsync to be a full one since we
2934 * we replaced (or just dropped in the case of cloning holes when
2935 * NO_HOLES is enabled) file extent items and did not setup new extent
2936 * maps for the replacement extents (or holes).
2937 */
2938 if (extent_info && !extent_info->is_new_extent)
2939 btrfs_set_inode_full_sync(inode);
2940
2941 if (ret)
2942 goto out_trans;
2943
2944 trans->block_rsv = &fs_info->trans_block_rsv;
2945 /*
2946 * If we are using the NO_HOLES feature we might have had already an
2947 * hole that overlaps a part of the region [lockstart, lockend] and
2948 * ends at (or beyond) lockend. Since we have no file extent items to
2949 * represent holes, drop_end can be less than lockend and so we must
2950 * make sure we have an extent map representing the existing hole (the
2951 * call to __btrfs_drop_extents() might have dropped the existing extent
2952 * map representing the existing hole), otherwise the fast fsync path
2953 * will not record the existence of the hole region
2954 * [existing_hole_start, lockend].
2955 */
2956 if (drop_args.drop_end <= end)
2957 drop_args.drop_end = end + 1;
2958 /*
2959 * Don't insert file hole extent item if it's for a range beyond eof
2960 * (because it's useless) or if it represents a 0 bytes range (when
2961 * cur_offset == drop_end).
2962 */
2963 if (!extent_info && cur_offset < ino_size &&
2964 cur_offset < drop_args.drop_end) {
2965 ret = fill_holes(trans, inode, path, cur_offset,
2966 drop_args.drop_end);
2967 if (ret) {
2968 /* Same comment as above. */
2969 btrfs_abort_transaction(trans, ret);
2970 goto out_trans;
2971 }
2972 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2973 /* See the comment in the loop above for the reasoning here. */
2974 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2975 drop_args.drop_end - cur_offset);
2976 if (ret) {
2977 btrfs_abort_transaction(trans, ret);
2978 goto out_trans;
2979 }
2980
2981 }
2982 if (extent_info) {
2983 ret = btrfs_insert_replace_extent(trans, inode, path,
2984 extent_info, extent_info->data_len,
2985 drop_args.bytes_found);
2986 if (ret) {
2987 btrfs_abort_transaction(trans, ret);
2988 goto out_trans;
2989 }
2990 }
2991
2992 out_trans:
2993 if (!trans)
2994 goto out_free;
2995
2996 trans->block_rsv = &fs_info->trans_block_rsv;
2997 if (ret)
2998 btrfs_end_transaction(trans);
2999 else
3000 *trans_out = trans;
3001 out_free:
3002 btrfs_free_block_rsv(fs_info, rsv);
3003 out:
3004 return ret;
3005 }
3006
3007 static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
3008 {
3009 struct inode *inode = file_inode(file);
3010 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3011 struct btrfs_root *root = BTRFS_I(inode)->root;
3012 struct extent_state *cached_state = NULL;
3013 struct btrfs_path *path;
3014 struct btrfs_trans_handle *trans = NULL;
3015 u64 lockstart;
3016 u64 lockend;
3017 u64 tail_start;
3018 u64 tail_len;
3019 u64 orig_start = offset;
3020 int ret = 0;
3021 bool same_block;
3022 u64 ino_size;
3023 bool truncated_block = false;
3024 bool updated_inode = false;
3025
3026 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
3027
3028 ret = btrfs_wait_ordered_range(inode, offset, len);
3029 if (ret)
3030 goto out_only_mutex;
3031
3032 ino_size = round_up(inode->i_size, fs_info->sectorsize);
3033 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
3034 if (ret < 0)
3035 goto out_only_mutex;
3036 if (ret && !len) {
3037 /* Already in a large hole */
3038 ret = 0;
3039 goto out_only_mutex;
3040 }
3041
3042 ret = file_modified(file);
3043 if (ret)
3044 goto out_only_mutex;
3045
3046 lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
3047 lockend = round_down(offset + len,
3048 btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
3049 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
3050 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
3051 /*
3052 * We needn't truncate any block which is beyond the end of the file
3053 * because we are sure there is no data there.
3054 */
3055 /*
3056 * Only do this if we are in the same block and we aren't doing the
3057 * entire block.
3058 */
3059 if (same_block && len < fs_info->sectorsize) {
3060 if (offset < ino_size) {
3061 truncated_block = true;
3062 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3063 0);
3064 } else {
3065 ret = 0;
3066 }
3067 goto out_only_mutex;
3068 }
3069
3070 /* zero back part of the first block */
3071 if (offset < ino_size) {
3072 truncated_block = true;
3073 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3074 if (ret) {
3075 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3076 return ret;
3077 }
3078 }
3079
3080 /* Check the aligned pages after the first unaligned page,
3081 * if offset != orig_start, which means the first unaligned page
3082 * including several following pages are already in holes,
3083 * the extra check can be skipped */
3084 if (offset == orig_start) {
3085 /* after truncate page, check hole again */
3086 len = offset + len - lockstart;
3087 offset = lockstart;
3088 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
3089 if (ret < 0)
3090 goto out_only_mutex;
3091 if (ret && !len) {
3092 ret = 0;
3093 goto out_only_mutex;
3094 }
3095 lockstart = offset;
3096 }
3097
3098 /* Check the tail unaligned part is in a hole */
3099 tail_start = lockend + 1;
3100 tail_len = offset + len - tail_start;
3101 if (tail_len) {
3102 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
3103 if (unlikely(ret < 0))
3104 goto out_only_mutex;
3105 if (!ret) {
3106 /* zero the front end of the last page */
3107 if (tail_start + tail_len < ino_size) {
3108 truncated_block = true;
3109 ret = btrfs_truncate_block(BTRFS_I(inode),
3110 tail_start + tail_len,
3111 0, 1);
3112 if (ret)
3113 goto out_only_mutex;
3114 }
3115 }
3116 }
3117
3118 if (lockend < lockstart) {
3119 ret = 0;
3120 goto out_only_mutex;
3121 }
3122
3123 btrfs_punch_hole_lock_range(inode, lockstart, lockend, &cached_state);
3124
3125 path = btrfs_alloc_path();
3126 if (!path) {
3127 ret = -ENOMEM;
3128 goto out;
3129 }
3130
3131 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
3132 lockend, NULL, &trans);
3133 btrfs_free_path(path);
3134 if (ret)
3135 goto out;
3136
3137 ASSERT(trans != NULL);
3138 inode_inc_iversion(inode);
3139 inode->i_mtime = inode->i_ctime = current_time(inode);
3140 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3141 updated_inode = true;
3142 btrfs_end_transaction(trans);
3143 btrfs_btree_balance_dirty(fs_info);
3144 out:
3145 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3146 &cached_state);
3147 out_only_mutex:
3148 if (!updated_inode && truncated_block && !ret) {
3149 /*
3150 * If we only end up zeroing part of a page, we still need to
3151 * update the inode item, so that all the time fields are
3152 * updated as well as the necessary btrfs inode in memory fields
3153 * for detecting, at fsync time, if the inode isn't yet in the
3154 * log tree or it's there but not up to date.
3155 */
3156 struct timespec64 now = current_time(inode);
3157
3158 inode_inc_iversion(inode);
3159 inode->i_mtime = now;
3160 inode->i_ctime = now;
3161 trans = btrfs_start_transaction(root, 1);
3162 if (IS_ERR(trans)) {
3163 ret = PTR_ERR(trans);
3164 } else {
3165 int ret2;
3166
3167 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3168 ret2 = btrfs_end_transaction(trans);
3169 if (!ret)
3170 ret = ret2;
3171 }
3172 }
3173 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3174 return ret;
3175 }
3176
3177 /* Helper structure to record which range is already reserved */
3178 struct falloc_range {
3179 struct list_head list;
3180 u64 start;
3181 u64 len;
3182 };
3183
3184 /*
3185 * Helper function to add falloc range
3186 *
3187 * Caller should have locked the larger range of extent containing
3188 * [start, len)
3189 */
3190 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3191 {
3192 struct falloc_range *range = NULL;
3193
3194 if (!list_empty(head)) {
3195 /*
3196 * As fallocate iterates by bytenr order, we only need to check
3197 * the last range.
3198 */
3199 range = list_last_entry(head, struct falloc_range, list);
3200 if (range->start + range->len == start) {
3201 range->len += len;
3202 return 0;
3203 }
3204 }
3205
3206 range = kmalloc(sizeof(*range), GFP_KERNEL);
3207 if (!range)
3208 return -ENOMEM;
3209 range->start = start;
3210 range->len = len;
3211 list_add_tail(&range->list, head);
3212 return 0;
3213 }
3214
3215 static int btrfs_fallocate_update_isize(struct inode *inode,
3216 const u64 end,
3217 const int mode)
3218 {
3219 struct btrfs_trans_handle *trans;
3220 struct btrfs_root *root = BTRFS_I(inode)->root;
3221 int ret;
3222 int ret2;
3223
3224 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3225 return 0;
3226
3227 trans = btrfs_start_transaction(root, 1);
3228 if (IS_ERR(trans))
3229 return PTR_ERR(trans);
3230
3231 inode->i_ctime = current_time(inode);
3232 i_size_write(inode, end);
3233 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
3234 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3235 ret2 = btrfs_end_transaction(trans);
3236
3237 return ret ? ret : ret2;
3238 }
3239
3240 enum {
3241 RANGE_BOUNDARY_WRITTEN_EXTENT,
3242 RANGE_BOUNDARY_PREALLOC_EXTENT,
3243 RANGE_BOUNDARY_HOLE,
3244 };
3245
3246 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3247 u64 offset)
3248 {
3249 const u64 sectorsize = btrfs_inode_sectorsize(inode);
3250 struct extent_map *em;
3251 int ret;
3252
3253 offset = round_down(offset, sectorsize);
3254 em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3255 if (IS_ERR(em))
3256 return PTR_ERR(em);
3257
3258 if (em->block_start == EXTENT_MAP_HOLE)
3259 ret = RANGE_BOUNDARY_HOLE;
3260 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3261 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3262 else
3263 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3264
3265 free_extent_map(em);
3266 return ret;
3267 }
3268
3269 static int btrfs_zero_range(struct inode *inode,
3270 loff_t offset,
3271 loff_t len,
3272 const int mode)
3273 {
3274 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3275 struct extent_map *em;
3276 struct extent_changeset *data_reserved = NULL;
3277 int ret;
3278 u64 alloc_hint = 0;
3279 const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3280 u64 alloc_start = round_down(offset, sectorsize);
3281 u64 alloc_end = round_up(offset + len, sectorsize);
3282 u64 bytes_to_reserve = 0;
3283 bool space_reserved = false;
3284
3285 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3286 alloc_end - alloc_start);
3287 if (IS_ERR(em)) {
3288 ret = PTR_ERR(em);
3289 goto out;
3290 }
3291
3292 /*
3293 * Avoid hole punching and extent allocation for some cases. More cases
3294 * could be considered, but these are unlikely common and we keep things
3295 * as simple as possible for now. Also, intentionally, if the target
3296 * range contains one or more prealloc extents together with regular
3297 * extents and holes, we drop all the existing extents and allocate a
3298 * new prealloc extent, so that we get a larger contiguous disk extent.
3299 */
3300 if (em->start <= alloc_start &&
3301 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3302 const u64 em_end = em->start + em->len;
3303
3304 if (em_end >= offset + len) {
3305 /*
3306 * The whole range is already a prealloc extent,
3307 * do nothing except updating the inode's i_size if
3308 * needed.
3309 */
3310 free_extent_map(em);
3311 ret = btrfs_fallocate_update_isize(inode, offset + len,
3312 mode);
3313 goto out;
3314 }
3315 /*
3316 * Part of the range is already a prealloc extent, so operate
3317 * only on the remaining part of the range.
3318 */
3319 alloc_start = em_end;
3320 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3321 len = offset + len - alloc_start;
3322 offset = alloc_start;
3323 alloc_hint = em->block_start + em->len;
3324 }
3325 free_extent_map(em);
3326
3327 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3328 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3329 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3330 sectorsize);
3331 if (IS_ERR(em)) {
3332 ret = PTR_ERR(em);
3333 goto out;
3334 }
3335
3336 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3337 free_extent_map(em);
3338 ret = btrfs_fallocate_update_isize(inode, offset + len,
3339 mode);
3340 goto out;
3341 }
3342 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3343 free_extent_map(em);
3344 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3345 0);
3346 if (!ret)
3347 ret = btrfs_fallocate_update_isize(inode,
3348 offset + len,
3349 mode);
3350 return ret;
3351 }
3352 free_extent_map(em);
3353 alloc_start = round_down(offset, sectorsize);
3354 alloc_end = alloc_start + sectorsize;
3355 goto reserve_space;
3356 }
3357
3358 alloc_start = round_up(offset, sectorsize);
3359 alloc_end = round_down(offset + len, sectorsize);
3360
3361 /*
3362 * For unaligned ranges, check the pages at the boundaries, they might
3363 * map to an extent, in which case we need to partially zero them, or
3364 * they might map to a hole, in which case we need our allocation range
3365 * to cover them.
3366 */
3367 if (!IS_ALIGNED(offset, sectorsize)) {
3368 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3369 offset);
3370 if (ret < 0)
3371 goto out;
3372 if (ret == RANGE_BOUNDARY_HOLE) {
3373 alloc_start = round_down(offset, sectorsize);
3374 ret = 0;
3375 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3376 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3377 if (ret)
3378 goto out;
3379 } else {
3380 ret = 0;
3381 }
3382 }
3383
3384 if (!IS_ALIGNED(offset + len, sectorsize)) {
3385 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3386 offset + len);
3387 if (ret < 0)
3388 goto out;
3389 if (ret == RANGE_BOUNDARY_HOLE) {
3390 alloc_end = round_up(offset + len, sectorsize);
3391 ret = 0;
3392 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3393 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
3394 0, 1);
3395 if (ret)
3396 goto out;
3397 } else {
3398 ret = 0;
3399 }
3400 }
3401
3402 reserve_space:
3403 if (alloc_start < alloc_end) {
3404 struct extent_state *cached_state = NULL;
3405 const u64 lockstart = alloc_start;
3406 const u64 lockend = alloc_end - 1;
3407
3408 bytes_to_reserve = alloc_end - alloc_start;
3409 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3410 bytes_to_reserve);
3411 if (ret < 0)
3412 goto out;
3413 space_reserved = true;
3414 btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3415 &cached_state);
3416 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3417 alloc_start, bytes_to_reserve);
3418 if (ret) {
3419 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3420 lockend, &cached_state);
3421 goto out;
3422 }
3423 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3424 alloc_end - alloc_start,
3425 i_blocksize(inode),
3426 offset + len, &alloc_hint);
3427 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3428 lockend, &cached_state);
3429 /* btrfs_prealloc_file_range releases reserved space on error */
3430 if (ret) {
3431 space_reserved = false;
3432 goto out;
3433 }
3434 }
3435 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3436 out:
3437 if (ret && space_reserved)
3438 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3439 alloc_start, bytes_to_reserve);
3440 extent_changeset_free(data_reserved);
3441
3442 return ret;
3443 }
3444
3445 static long btrfs_fallocate(struct file *file, int mode,
3446 loff_t offset, loff_t len)
3447 {
3448 struct inode *inode = file_inode(file);
3449 struct extent_state *cached_state = NULL;
3450 struct extent_changeset *data_reserved = NULL;
3451 struct falloc_range *range;
3452 struct falloc_range *tmp;
3453 struct list_head reserve_list;
3454 u64 cur_offset;
3455 u64 last_byte;
3456 u64 alloc_start;
3457 u64 alloc_end;
3458 u64 alloc_hint = 0;
3459 u64 locked_end;
3460 u64 actual_end = 0;
3461 u64 data_space_needed = 0;
3462 u64 data_space_reserved = 0;
3463 u64 qgroup_reserved = 0;
3464 struct extent_map *em;
3465 int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3466 int ret;
3467
3468 /* Do not allow fallocate in ZONED mode */
3469 if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3470 return -EOPNOTSUPP;
3471
3472 alloc_start = round_down(offset, blocksize);
3473 alloc_end = round_up(offset + len, blocksize);
3474 cur_offset = alloc_start;
3475
3476 /* Make sure we aren't being give some crap mode */
3477 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3478 FALLOC_FL_ZERO_RANGE))
3479 return -EOPNOTSUPP;
3480
3481 if (mode & FALLOC_FL_PUNCH_HOLE)
3482 return btrfs_punch_hole(file, offset, len);
3483
3484 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
3485
3486 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3487 ret = inode_newsize_ok(inode, offset + len);
3488 if (ret)
3489 goto out;
3490 }
3491
3492 ret = file_modified(file);
3493 if (ret)
3494 goto out;
3495
3496 /*
3497 * TODO: Move these two operations after we have checked
3498 * accurate reserved space, or fallocate can still fail but
3499 * with page truncated or size expanded.
3500 *
3501 * But that's a minor problem and won't do much harm BTW.
3502 */
3503 if (alloc_start > inode->i_size) {
3504 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3505 alloc_start);
3506 if (ret)
3507 goto out;
3508 } else if (offset + len > inode->i_size) {
3509 /*
3510 * If we are fallocating from the end of the file onward we
3511 * need to zero out the end of the block if i_size lands in the
3512 * middle of a block.
3513 */
3514 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3515 if (ret)
3516 goto out;
3517 }
3518
3519 /*
3520 * We have locked the inode at the VFS level (in exclusive mode) and we
3521 * have locked the i_mmap_lock lock (in exclusive mode). Now before
3522 * locking the file range, flush all dealloc in the range and wait for
3523 * all ordered extents in the range to complete. After this we can lock
3524 * the file range and, due to the previous locking we did, we know there
3525 * can't be more delalloc or ordered extents in the range.
3526 */
3527 ret = btrfs_wait_ordered_range(inode, alloc_start,
3528 alloc_end - alloc_start);
3529 if (ret)
3530 goto out;
3531
3532 if (mode & FALLOC_FL_ZERO_RANGE) {
3533 ret = btrfs_zero_range(inode, offset, len, mode);
3534 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3535 return ret;
3536 }
3537
3538 locked_end = alloc_end - 1;
3539 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3540 &cached_state);
3541
3542 btrfs_assert_inode_range_clean(BTRFS_I(inode), alloc_start, locked_end);
3543
3544 /* First, check if we exceed the qgroup limit */
3545 INIT_LIST_HEAD(&reserve_list);
3546 while (cur_offset < alloc_end) {
3547 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3548 alloc_end - cur_offset);
3549 if (IS_ERR(em)) {
3550 ret = PTR_ERR(em);
3551 break;
3552 }
3553 last_byte = min(extent_map_end(em), alloc_end);
3554 actual_end = min_t(u64, extent_map_end(em), offset + len);
3555 last_byte = ALIGN(last_byte, blocksize);
3556 if (em->block_start == EXTENT_MAP_HOLE ||
3557 (cur_offset >= inode->i_size &&
3558 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3559 const u64 range_len = last_byte - cur_offset;
3560
3561 ret = add_falloc_range(&reserve_list, cur_offset, range_len);
3562 if (ret < 0) {
3563 free_extent_map(em);
3564 break;
3565 }
3566 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3567 &data_reserved, cur_offset, range_len);
3568 if (ret < 0) {
3569 free_extent_map(em);
3570 break;
3571 }
3572 qgroup_reserved += range_len;
3573 data_space_needed += range_len;
3574 }
3575 free_extent_map(em);
3576 cur_offset = last_byte;
3577 }
3578
3579 if (!ret && data_space_needed > 0) {
3580 /*
3581 * We are safe to reserve space here as we can't have delalloc
3582 * in the range, see above.
3583 */
3584 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3585 data_space_needed);
3586 if (!ret)
3587 data_space_reserved = data_space_needed;
3588 }
3589
3590 /*
3591 * If ret is still 0, means we're OK to fallocate.
3592 * Or just cleanup the list and exit.
3593 */
3594 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3595 if (!ret) {
3596 ret = btrfs_prealloc_file_range(inode, mode,
3597 range->start,
3598 range->len, i_blocksize(inode),
3599 offset + len, &alloc_hint);
3600 /*
3601 * btrfs_prealloc_file_range() releases space even
3602 * if it returns an error.
3603 */
3604 data_space_reserved -= range->len;
3605 qgroup_reserved -= range->len;
3606 } else if (data_space_reserved > 0) {
3607 btrfs_free_reserved_data_space(BTRFS_I(inode),
3608 data_reserved, range->start,
3609 range->len);
3610 data_space_reserved -= range->len;
3611 qgroup_reserved -= range->len;
3612 } else if (qgroup_reserved > 0) {
3613 btrfs_qgroup_free_data(BTRFS_I(inode), data_reserved,
3614 range->start, range->len);
3615 qgroup_reserved -= range->len;
3616 }
3617 list_del(&range->list);
3618 kfree(range);
3619 }
3620 if (ret < 0)
3621 goto out_unlock;
3622
3623 /*
3624 * We didn't need to allocate any more space, but we still extended the
3625 * size of the file so we need to update i_size and the inode item.
3626 */
3627 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3628 out_unlock:
3629 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3630 &cached_state);
3631 out:
3632 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3633 extent_changeset_free(data_reserved);
3634 return ret;
3635 }
3636
3637 static loff_t find_desired_extent(struct btrfs_inode *inode, loff_t offset,
3638 int whence)
3639 {
3640 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3641 struct extent_map *em = NULL;
3642 struct extent_state *cached_state = NULL;
3643 loff_t i_size = inode->vfs_inode.i_size;
3644 u64 lockstart;
3645 u64 lockend;
3646 u64 start;
3647 u64 len;
3648 int ret = 0;
3649
3650 if (i_size == 0 || offset >= i_size)
3651 return -ENXIO;
3652
3653 /*
3654 * offset can be negative, in this case we start finding DATA/HOLE from
3655 * the very start of the file.
3656 */
3657 start = max_t(loff_t, 0, offset);
3658
3659 lockstart = round_down(start, fs_info->sectorsize);
3660 lockend = round_up(i_size, fs_info->sectorsize);
3661 if (lockend <= lockstart)
3662 lockend = lockstart + fs_info->sectorsize;
3663 lockend--;
3664 len = lockend - lockstart + 1;
3665
3666 lock_extent_bits(&inode->io_tree, lockstart, lockend, &cached_state);
3667
3668 while (start < i_size) {
3669 em = btrfs_get_extent_fiemap(inode, start, len);
3670 if (IS_ERR(em)) {
3671 ret = PTR_ERR(em);
3672 em = NULL;
3673 break;
3674 }
3675
3676 if (whence == SEEK_HOLE &&
3677 (em->block_start == EXTENT_MAP_HOLE ||
3678 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3679 break;
3680 else if (whence == SEEK_DATA &&
3681 (em->block_start != EXTENT_MAP_HOLE &&
3682 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3683 break;
3684
3685 start = em->start + em->len;
3686 free_extent_map(em);
3687 em = NULL;
3688 cond_resched();
3689 }
3690 free_extent_map(em);
3691 unlock_extent_cached(&inode->io_tree, lockstart, lockend,
3692 &cached_state);
3693 if (ret) {
3694 offset = ret;
3695 } else {
3696 if (whence == SEEK_DATA && start >= i_size)
3697 offset = -ENXIO;
3698 else
3699 offset = min_t(loff_t, start, i_size);
3700 }
3701
3702 return offset;
3703 }
3704
3705 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3706 {
3707 struct inode *inode = file->f_mapping->host;
3708
3709 switch (whence) {
3710 default:
3711 return generic_file_llseek(file, offset, whence);
3712 case SEEK_DATA:
3713 case SEEK_HOLE:
3714 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3715 offset = find_desired_extent(BTRFS_I(inode), offset, whence);
3716 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3717 break;
3718 }
3719
3720 if (offset < 0)
3721 return offset;
3722
3723 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3724 }
3725
3726 static int btrfs_file_open(struct inode *inode, struct file *filp)
3727 {
3728 int ret;
3729
3730 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3731
3732 ret = fsverity_file_open(inode, filp);
3733 if (ret)
3734 return ret;
3735 return generic_file_open(inode, filp);
3736 }
3737
3738 static int check_direct_read(struct btrfs_fs_info *fs_info,
3739 const struct iov_iter *iter, loff_t offset)
3740 {
3741 int ret;
3742 int i, seg;
3743
3744 ret = check_direct_IO(fs_info, iter, offset);
3745 if (ret < 0)
3746 return ret;
3747
3748 if (!iter_is_iovec(iter))
3749 return 0;
3750
3751 for (seg = 0; seg < iter->nr_segs; seg++)
3752 for (i = seg + 1; i < iter->nr_segs; i++)
3753 if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3754 return -EINVAL;
3755 return 0;
3756 }
3757
3758 static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3759 {
3760 struct inode *inode = file_inode(iocb->ki_filp);
3761 size_t prev_left = 0;
3762 ssize_t read = 0;
3763 ssize_t ret;
3764
3765 if (fsverity_active(inode))
3766 return 0;
3767
3768 if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3769 return 0;
3770
3771 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3772 again:
3773 /*
3774 * This is similar to what we do for direct IO writes, see the comment
3775 * at btrfs_direct_write(), but we also disable page faults in addition
3776 * to disabling them only at the iov_iter level. This is because when
3777 * reading from a hole or prealloc extent, iomap calls iov_iter_zero(),
3778 * which can still trigger page fault ins despite having set ->nofault
3779 * to true of our 'to' iov_iter.
3780 *
3781 * The difference to direct IO writes is that we deadlock when trying
3782 * to lock the extent range in the inode's tree during he page reads
3783 * triggered by the fault in (while for writes it is due to waiting for
3784 * our own ordered extent). This is because for direct IO reads,
3785 * btrfs_dio_iomap_begin() returns with the extent range locked, which
3786 * is only unlocked in the endio callback (end_bio_extent_readpage()).
3787 */
3788 pagefault_disable();
3789 to->nofault = true;
3790 ret = btrfs_dio_rw(iocb, to, read);
3791 to->nofault = false;
3792 pagefault_enable();
3793
3794 /* No increment (+=) because iomap returns a cumulative value. */
3795 if (ret > 0)
3796 read = ret;
3797
3798 if (iov_iter_count(to) > 0 && (ret == -EFAULT || ret > 0)) {
3799 const size_t left = iov_iter_count(to);
3800
3801 if (left == prev_left) {
3802 /*
3803 * We didn't make any progress since the last attempt,
3804 * fallback to a buffered read for the remainder of the
3805 * range. This is just to avoid any possibility of looping
3806 * for too long.
3807 */
3808 ret = read;
3809 } else {
3810 /*
3811 * We made some progress since the last retry or this is
3812 * the first time we are retrying. Fault in as many pages
3813 * as possible and retry.
3814 */
3815 fault_in_iov_iter_writeable(to, left);
3816 prev_left = left;
3817 goto again;
3818 }
3819 }
3820 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3821 return ret < 0 ? ret : read;
3822 }
3823
3824 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3825 {
3826 ssize_t ret = 0;
3827
3828 if (iocb->ki_flags & IOCB_DIRECT) {
3829 ret = btrfs_direct_read(iocb, to);
3830 if (ret < 0 || !iov_iter_count(to) ||
3831 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3832 return ret;
3833 }
3834
3835 return filemap_read(iocb, to, ret);
3836 }
3837
3838 const struct file_operations btrfs_file_operations = {
3839 .llseek = btrfs_file_llseek,
3840 .read_iter = btrfs_file_read_iter,
3841 .splice_read = generic_file_splice_read,
3842 .write_iter = btrfs_file_write_iter,
3843 .splice_write = iter_file_splice_write,
3844 .mmap = btrfs_file_mmap,
3845 .open = btrfs_file_open,
3846 .release = btrfs_release_file,
3847 .get_unmapped_area = thp_get_unmapped_area,
3848 .fsync = btrfs_sync_file,
3849 .fallocate = btrfs_fallocate,
3850 .unlocked_ioctl = btrfs_ioctl,
3851 #ifdef CONFIG_COMPAT
3852 .compat_ioctl = btrfs_compat_ioctl,
3853 #endif
3854 .remap_file_range = btrfs_remap_file_range,
3855 };
3856
3857 void __cold btrfs_auto_defrag_exit(void)
3858 {
3859 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3860 }
3861
3862 int __init btrfs_auto_defrag_init(void)
3863 {
3864 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3865 sizeof(struct inode_defrag), 0,
3866 SLAB_MEM_SPREAD,
3867 NULL);
3868 if (!btrfs_inode_defrag_cachep)
3869 return -ENOMEM;
3870
3871 return 0;
3872 }
3873
3874 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3875 {
3876 int ret;
3877
3878 /*
3879 * So with compression we will find and lock a dirty page and clear the
3880 * first one as dirty, setup an async extent, and immediately return
3881 * with the entire range locked but with nobody actually marked with
3882 * writeback. So we can't just filemap_write_and_wait_range() and
3883 * expect it to work since it will just kick off a thread to do the
3884 * actual work. So we need to call filemap_fdatawrite_range _again_
3885 * since it will wait on the page lock, which won't be unlocked until
3886 * after the pages have been marked as writeback and so we're good to go
3887 * from there. We have to do this otherwise we'll miss the ordered
3888 * extents and that results in badness. Please Josef, do not think you
3889 * know better and pull this out at some point in the future, it is
3890 * right and you are wrong.
3891 */
3892 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3893 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3894 &BTRFS_I(inode)->runtime_flags))
3895 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3896
3897 return ret;
3898 }
3899
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.