Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6#include <linux/fs.h>
7#include <linux/pagemap.h>
8#include <linux/time.h>
9#include <linux/init.h>
10#include <linux/string.h>
11#include <linux/backing-dev.h>
12#include <linux/falloc.h>
13#include <linux/writeback.h>
14#include <linux/compat.h>
15#include <linux/slab.h>
16#include <linux/btrfs.h>
17#include <linux/uio.h>
18#include <linux/iversion.h>
19#include <linux/fsverity.h>
20#include "ctree.h"
21#include "direct-io.h"
22#include "disk-io.h"
23#include "transaction.h"
24#include "btrfs_inode.h"
25#include "tree-log.h"
26#include "locking.h"
27#include "qgroup.h"
28#include "compression.h"
29#include "delalloc-space.h"
30#include "reflink.h"
31#include "subpage.h"
32#include "fs.h"
33#include "accessors.h"
34#include "extent-tree.h"
35#include "file-item.h"
36#include "ioctl.h"
37#include "file.h"
38#include "super.h"
39
40/*
41 * Helper to fault in page and copy. This should go away and be replaced with
42 * calls into generic code.
43 */
44static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
45 struct folio *folio, struct iov_iter *i)
46{
47 size_t copied = 0;
48 size_t total_copied = 0;
49 int offset = offset_in_page(pos);
50
51 while (write_bytes > 0) {
52 size_t count = min_t(size_t, PAGE_SIZE - offset, write_bytes);
53 /*
54 * Copy data from userspace to the current page
55 */
56 copied = copy_folio_from_iter_atomic(folio, offset, count, i);
57
58 /* Flush processor's dcache for this page */
59 flush_dcache_folio(folio);
60
61 /*
62 * if we get a partial write, we can end up with
63 * partially up to date page. These add
64 * a lot of complexity, so make sure they don't
65 * happen by forcing this copy to be retried.
66 *
67 * The rest of the btrfs_file_write code will fall
68 * back to page at a time copies after we return 0.
69 */
70 if (unlikely(copied < count)) {
71 if (!folio_test_uptodate(folio)) {
72 iov_iter_revert(i, copied);
73 copied = 0;
74 }
75 if (!copied)
76 break;
77 }
78
79 write_bytes -= copied;
80 total_copied += copied;
81 offset += copied;
82 }
83 return total_copied;
84}
85
86/*
87 * Unlock folio after btrfs_file_write() is done with it.
88 */
89static void btrfs_drop_folio(struct btrfs_fs_info *fs_info, struct folio *folio,
90 u64 pos, u64 copied)
91{
92 u64 block_start = round_down(pos, fs_info->sectorsize);
93 u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
94
95 ASSERT(block_len <= U32_MAX);
96 /*
97 * Folio checked is some magic around finding folios that have been
98 * modified without going through btrfs_dirty_folio(). Clear it here.
99 * There should be no need to mark the pages accessed as
100 * prepare_one_folio() should have marked them accessed in
101 * prepare_one_folio() via find_or_create_page()
102 */
103 btrfs_folio_clamp_clear_checked(fs_info, folio, block_start, block_len);
104 folio_unlock(folio);
105 folio_put(folio);
106}
107
108/*
109 * After btrfs_copy_from_user(), update the following things for delalloc:
110 * - Mark newly dirtied folio as DELALLOC in the io tree.
111 * Used to advise which range is to be written back.
112 * - Mark modified folio as Uptodate/Dirty and not needing COW fixup
113 * - Update inode size for past EOF write
114 */
115int btrfs_dirty_folio(struct btrfs_inode *inode, struct folio *folio, loff_t pos,
116 size_t write_bytes, struct extent_state **cached, bool noreserve)
117{
118 struct btrfs_fs_info *fs_info = inode->root->fs_info;
119 int ret = 0;
120 u64 num_bytes;
121 u64 start_pos;
122 u64 end_of_last_block;
123 u64 end_pos = pos + write_bytes;
124 loff_t isize = i_size_read(&inode->vfs_inode);
125 unsigned int extra_bits = 0;
126
127 if (write_bytes == 0)
128 return 0;
129
130 if (noreserve)
131 extra_bits |= EXTENT_NORESERVE;
132
133 start_pos = round_down(pos, fs_info->sectorsize);
134 num_bytes = round_up(write_bytes + pos - start_pos,
135 fs_info->sectorsize);
136 ASSERT(num_bytes <= U32_MAX);
137 ASSERT(folio_pos(folio) <= pos &&
138 folio_pos(folio) + folio_size(folio) >= pos + write_bytes);
139
140 end_of_last_block = start_pos + num_bytes - 1;
141
142 /*
143 * The pages may have already been dirty, clear out old accounting so
144 * we can set things up properly
145 */
146 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
147 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
148 cached);
149
150 ret = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
151 extra_bits, cached);
152 if (ret)
153 return ret;
154
155 btrfs_folio_clamp_set_uptodate(fs_info, folio, start_pos, num_bytes);
156 btrfs_folio_clamp_clear_checked(fs_info, folio, start_pos, num_bytes);
157 btrfs_folio_clamp_set_dirty(fs_info, folio, start_pos, num_bytes);
158
159 /*
160 * we've only changed i_size in ram, and we haven't updated
161 * the disk i_size. There is no need to log the inode
162 * at this time.
163 */
164 if (end_pos > isize)
165 i_size_write(&inode->vfs_inode, end_pos);
166 return 0;
167}
168
169/*
170 * this is very complex, but the basic idea is to drop all extents
171 * in the range start - end. hint_block is filled in with a block number
172 * that would be a good hint to the block allocator for this file.
173 *
174 * If an extent intersects the range but is not entirely inside the range
175 * it is either truncated or split. Anything entirely inside the range
176 * is deleted from the tree.
177 *
178 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
179 * to deal with that. We set the field 'bytes_found' of the arguments structure
180 * with the number of allocated bytes found in the target range, so that the
181 * caller can update the inode's number of bytes in an atomic way when
182 * replacing extents in a range to avoid races with stat(2).
183 */
184int btrfs_drop_extents(struct btrfs_trans_handle *trans,
185 struct btrfs_root *root, struct btrfs_inode *inode,
186 struct btrfs_drop_extents_args *args)
187{
188 struct btrfs_fs_info *fs_info = root->fs_info;
189 struct extent_buffer *leaf;
190 struct btrfs_file_extent_item *fi;
191 struct btrfs_key key;
192 struct btrfs_key new_key;
193 u64 ino = btrfs_ino(inode);
194 u64 search_start = args->start;
195 u64 disk_bytenr = 0;
196 u64 num_bytes = 0;
197 u64 extent_offset = 0;
198 u64 extent_end = 0;
199 u64 last_end = args->start;
200 int del_nr = 0;
201 int del_slot = 0;
202 int extent_type;
203 int recow;
204 int ret;
205 int modify_tree = -1;
206 int update_refs;
207 int found = 0;
208 struct btrfs_path *path = args->path;
209
210 args->bytes_found = 0;
211 args->extent_inserted = false;
212
213 /* Must always have a path if ->replace_extent is true */
214 ASSERT(!(args->replace_extent && !args->path));
215
216 if (!path) {
217 path = btrfs_alloc_path();
218 if (!path) {
219 ret = -ENOMEM;
220 goto out;
221 }
222 }
223
224 if (args->drop_cache)
225 btrfs_drop_extent_map_range(inode, args->start, args->end - 1, false);
226
227 if (data_race(args->start >= inode->disk_i_size) && !args->replace_extent)
228 modify_tree = 0;
229
230 update_refs = (btrfs_root_id(root) != BTRFS_TREE_LOG_OBJECTID);
231 while (1) {
232 recow = 0;
233 ret = btrfs_lookup_file_extent(trans, root, path, ino,
234 search_start, modify_tree);
235 if (ret < 0)
236 break;
237 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
238 leaf = path->nodes[0];
239 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
240 if (key.objectid == ino &&
241 key.type == BTRFS_EXTENT_DATA_KEY)
242 path->slots[0]--;
243 }
244 ret = 0;
245next_slot:
246 leaf = path->nodes[0];
247 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
248 BUG_ON(del_nr > 0);
249 ret = btrfs_next_leaf(root, path);
250 if (ret < 0)
251 break;
252 if (ret > 0) {
253 ret = 0;
254 break;
255 }
256 leaf = path->nodes[0];
257 recow = 1;
258 }
259
260 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
261
262 if (key.objectid > ino)
263 break;
264 if (WARN_ON_ONCE(key.objectid < ino) ||
265 key.type < BTRFS_EXTENT_DATA_KEY) {
266 ASSERT(del_nr == 0);
267 path->slots[0]++;
268 goto next_slot;
269 }
270 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
271 break;
272
273 fi = btrfs_item_ptr(leaf, path->slots[0],
274 struct btrfs_file_extent_item);
275 extent_type = btrfs_file_extent_type(leaf, fi);
276
277 if (extent_type == BTRFS_FILE_EXTENT_REG ||
278 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
279 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
280 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
281 extent_offset = btrfs_file_extent_offset(leaf, fi);
282 extent_end = key.offset +
283 btrfs_file_extent_num_bytes(leaf, fi);
284 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
285 extent_end = key.offset +
286 btrfs_file_extent_ram_bytes(leaf, fi);
287 } else {
288 /* can't happen */
289 BUG();
290 }
291
292 /*
293 * Don't skip extent items representing 0 byte lengths. They
294 * used to be created (bug) if while punching holes we hit
295 * -ENOSPC condition. So if we find one here, just ensure we
296 * delete it, otherwise we would insert a new file extent item
297 * with the same key (offset) as that 0 bytes length file
298 * extent item in the call to setup_items_for_insert() later
299 * in this function.
300 */
301 if (extent_end == key.offset && extent_end >= search_start) {
302 last_end = extent_end;
303 goto delete_extent_item;
304 }
305
306 if (extent_end <= search_start) {
307 path->slots[0]++;
308 goto next_slot;
309 }
310
311 found = 1;
312 search_start = max(key.offset, args->start);
313 if (recow || !modify_tree) {
314 modify_tree = -1;
315 btrfs_release_path(path);
316 continue;
317 }
318
319 /*
320 * | - range to drop - |
321 * | -------- extent -------- |
322 */
323 if (args->start > key.offset && args->end < extent_end) {
324 BUG_ON(del_nr > 0);
325 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
326 ret = -EOPNOTSUPP;
327 break;
328 }
329
330 memcpy(&new_key, &key, sizeof(new_key));
331 new_key.offset = args->start;
332 ret = btrfs_duplicate_item(trans, root, path,
333 &new_key);
334 if (ret == -EAGAIN) {
335 btrfs_release_path(path);
336 continue;
337 }
338 if (ret < 0)
339 break;
340
341 leaf = path->nodes[0];
342 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
343 struct btrfs_file_extent_item);
344 btrfs_set_file_extent_num_bytes(leaf, fi,
345 args->start - key.offset);
346
347 fi = btrfs_item_ptr(leaf, path->slots[0],
348 struct btrfs_file_extent_item);
349
350 extent_offset += args->start - key.offset;
351 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
352 btrfs_set_file_extent_num_bytes(leaf, fi,
353 extent_end - args->start);
354 btrfs_mark_buffer_dirty(trans, leaf);
355
356 if (update_refs && disk_bytenr > 0) {
357 struct btrfs_ref ref = {
358 .action = BTRFS_ADD_DELAYED_REF,
359 .bytenr = disk_bytenr,
360 .num_bytes = num_bytes,
361 .parent = 0,
362 .owning_root = btrfs_root_id(root),
363 .ref_root = btrfs_root_id(root),
364 };
365 btrfs_init_data_ref(&ref, new_key.objectid,
366 args->start - extent_offset,
367 0, false);
368 ret = btrfs_inc_extent_ref(trans, &ref);
369 if (ret) {
370 btrfs_abort_transaction(trans, ret);
371 break;
372 }
373 }
374 key.offset = args->start;
375 }
376 /*
377 * From here on out we will have actually dropped something, so
378 * last_end can be updated.
379 */
380 last_end = extent_end;
381
382 /*
383 * | ---- range to drop ----- |
384 * | -------- extent -------- |
385 */
386 if (args->start <= key.offset && args->end < extent_end) {
387 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
388 ret = -EOPNOTSUPP;
389 break;
390 }
391
392 memcpy(&new_key, &key, sizeof(new_key));
393 new_key.offset = args->end;
394 btrfs_set_item_key_safe(trans, path, &new_key);
395
396 extent_offset += args->end - key.offset;
397 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
398 btrfs_set_file_extent_num_bytes(leaf, fi,
399 extent_end - args->end);
400 btrfs_mark_buffer_dirty(trans, leaf);
401 if (update_refs && disk_bytenr > 0)
402 args->bytes_found += args->end - key.offset;
403 break;
404 }
405
406 search_start = extent_end;
407 /*
408 * | ---- range to drop ----- |
409 * | -------- extent -------- |
410 */
411 if (args->start > key.offset && args->end >= extent_end) {
412 BUG_ON(del_nr > 0);
413 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
414 ret = -EOPNOTSUPP;
415 break;
416 }
417
418 btrfs_set_file_extent_num_bytes(leaf, fi,
419 args->start - key.offset);
420 btrfs_mark_buffer_dirty(trans, leaf);
421 if (update_refs && disk_bytenr > 0)
422 args->bytes_found += extent_end - args->start;
423 if (args->end == extent_end)
424 break;
425
426 path->slots[0]++;
427 goto next_slot;
428 }
429
430 /*
431 * | ---- range to drop ----- |
432 * | ------ extent ------ |
433 */
434 if (args->start <= key.offset && args->end >= extent_end) {
435delete_extent_item:
436 if (del_nr == 0) {
437 del_slot = path->slots[0];
438 del_nr = 1;
439 } else {
440 BUG_ON(del_slot + del_nr != path->slots[0]);
441 del_nr++;
442 }
443
444 if (update_refs &&
445 extent_type == BTRFS_FILE_EXTENT_INLINE) {
446 args->bytes_found += extent_end - key.offset;
447 extent_end = ALIGN(extent_end,
448 fs_info->sectorsize);
449 } else if (update_refs && disk_bytenr > 0) {
450 struct btrfs_ref ref = {
451 .action = BTRFS_DROP_DELAYED_REF,
452 .bytenr = disk_bytenr,
453 .num_bytes = num_bytes,
454 .parent = 0,
455 .owning_root = btrfs_root_id(root),
456 .ref_root = btrfs_root_id(root),
457 };
458 btrfs_init_data_ref(&ref, key.objectid,
459 key.offset - extent_offset,
460 0, false);
461 ret = btrfs_free_extent(trans, &ref);
462 if (ret) {
463 btrfs_abort_transaction(trans, ret);
464 break;
465 }
466 args->bytes_found += extent_end - key.offset;
467 }
468
469 if (args->end == extent_end)
470 break;
471
472 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
473 path->slots[0]++;
474 goto next_slot;
475 }
476
477 ret = btrfs_del_items(trans, root, path, del_slot,
478 del_nr);
479 if (ret) {
480 btrfs_abort_transaction(trans, ret);
481 break;
482 }
483
484 del_nr = 0;
485 del_slot = 0;
486
487 btrfs_release_path(path);
488 continue;
489 }
490
491 BUG();
492 }
493
494 if (!ret && del_nr > 0) {
495 /*
496 * Set path->slots[0] to first slot, so that after the delete
497 * if items are move off from our leaf to its immediate left or
498 * right neighbor leafs, we end up with a correct and adjusted
499 * path->slots[0] for our insertion (if args->replace_extent).
500 */
501 path->slots[0] = del_slot;
502 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
503 if (ret)
504 btrfs_abort_transaction(trans, ret);
505 }
506
507 leaf = path->nodes[0];
508 /*
509 * If btrfs_del_items() was called, it might have deleted a leaf, in
510 * which case it unlocked our path, so check path->locks[0] matches a
511 * write lock.
512 */
513 if (!ret && args->replace_extent &&
514 path->locks[0] == BTRFS_WRITE_LOCK &&
515 btrfs_leaf_free_space(leaf) >=
516 sizeof(struct btrfs_item) + args->extent_item_size) {
517
518 key.objectid = ino;
519 key.type = BTRFS_EXTENT_DATA_KEY;
520 key.offset = args->start;
521 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
522 struct btrfs_key slot_key;
523
524 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
525 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
526 path->slots[0]++;
527 }
528 btrfs_setup_item_for_insert(trans, root, path, &key,
529 args->extent_item_size);
530 args->extent_inserted = true;
531 }
532
533 if (!args->path)
534 btrfs_free_path(path);
535 else if (!args->extent_inserted)
536 btrfs_release_path(path);
537out:
538 args->drop_end = found ? min(args->end, last_end) : args->end;
539
540 return ret;
541}
542
543static int extent_mergeable(struct extent_buffer *leaf, int slot,
544 u64 objectid, u64 bytenr, u64 orig_offset,
545 u64 *start, u64 *end)
546{
547 struct btrfs_file_extent_item *fi;
548 struct btrfs_key key;
549 u64 extent_end;
550
551 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
552 return 0;
553
554 btrfs_item_key_to_cpu(leaf, &key, slot);
555 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
556 return 0;
557
558 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
559 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
560 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
561 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
562 btrfs_file_extent_compression(leaf, fi) ||
563 btrfs_file_extent_encryption(leaf, fi) ||
564 btrfs_file_extent_other_encoding(leaf, fi))
565 return 0;
566
567 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
568 if ((*start && *start != key.offset) || (*end && *end != extent_end))
569 return 0;
570
571 *start = key.offset;
572 *end = extent_end;
573 return 1;
574}
575
576/*
577 * Mark extent in the range start - end as written.
578 *
579 * This changes extent type from 'pre-allocated' to 'regular'. If only
580 * part of extent is marked as written, the extent will be split into
581 * two or three.
582 */
583int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
584 struct btrfs_inode *inode, u64 start, u64 end)
585{
586 struct btrfs_root *root = inode->root;
587 struct extent_buffer *leaf;
588 struct btrfs_path *path;
589 struct btrfs_file_extent_item *fi;
590 struct btrfs_ref ref = { 0 };
591 struct btrfs_key key;
592 struct btrfs_key new_key;
593 u64 bytenr;
594 u64 num_bytes;
595 u64 extent_end;
596 u64 orig_offset;
597 u64 other_start;
598 u64 other_end;
599 u64 split;
600 int del_nr = 0;
601 int del_slot = 0;
602 int recow;
603 int ret = 0;
604 u64 ino = btrfs_ino(inode);
605
606 path = btrfs_alloc_path();
607 if (!path)
608 return -ENOMEM;
609again:
610 recow = 0;
611 split = start;
612 key.objectid = ino;
613 key.type = BTRFS_EXTENT_DATA_KEY;
614 key.offset = split;
615
616 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
617 if (ret < 0)
618 goto out;
619 if (ret > 0 && path->slots[0] > 0)
620 path->slots[0]--;
621
622 leaf = path->nodes[0];
623 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
624 if (key.objectid != ino ||
625 key.type != BTRFS_EXTENT_DATA_KEY) {
626 ret = -EINVAL;
627 btrfs_abort_transaction(trans, ret);
628 goto out;
629 }
630 fi = btrfs_item_ptr(leaf, path->slots[0],
631 struct btrfs_file_extent_item);
632 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
633 ret = -EINVAL;
634 btrfs_abort_transaction(trans, ret);
635 goto out;
636 }
637 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
638 if (key.offset > start || extent_end < end) {
639 ret = -EINVAL;
640 btrfs_abort_transaction(trans, ret);
641 goto out;
642 }
643
644 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
645 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
646 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
647 memcpy(&new_key, &key, sizeof(new_key));
648
649 if (start == key.offset && end < extent_end) {
650 other_start = 0;
651 other_end = start;
652 if (extent_mergeable(leaf, path->slots[0] - 1,
653 ino, bytenr, orig_offset,
654 &other_start, &other_end)) {
655 new_key.offset = end;
656 btrfs_set_item_key_safe(trans, path, &new_key);
657 fi = btrfs_item_ptr(leaf, path->slots[0],
658 struct btrfs_file_extent_item);
659 btrfs_set_file_extent_generation(leaf, fi,
660 trans->transid);
661 btrfs_set_file_extent_num_bytes(leaf, fi,
662 extent_end - end);
663 btrfs_set_file_extent_offset(leaf, fi,
664 end - orig_offset);
665 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
666 struct btrfs_file_extent_item);
667 btrfs_set_file_extent_generation(leaf, fi,
668 trans->transid);
669 btrfs_set_file_extent_num_bytes(leaf, fi,
670 end - other_start);
671 btrfs_mark_buffer_dirty(trans, leaf);
672 goto out;
673 }
674 }
675
676 if (start > key.offset && end == extent_end) {
677 other_start = end;
678 other_end = 0;
679 if (extent_mergeable(leaf, path->slots[0] + 1,
680 ino, bytenr, orig_offset,
681 &other_start, &other_end)) {
682 fi = btrfs_item_ptr(leaf, path->slots[0],
683 struct btrfs_file_extent_item);
684 btrfs_set_file_extent_num_bytes(leaf, fi,
685 start - key.offset);
686 btrfs_set_file_extent_generation(leaf, fi,
687 trans->transid);
688 path->slots[0]++;
689 new_key.offset = start;
690 btrfs_set_item_key_safe(trans, path, &new_key);
691
692 fi = btrfs_item_ptr(leaf, path->slots[0],
693 struct btrfs_file_extent_item);
694 btrfs_set_file_extent_generation(leaf, fi,
695 trans->transid);
696 btrfs_set_file_extent_num_bytes(leaf, fi,
697 other_end - start);
698 btrfs_set_file_extent_offset(leaf, fi,
699 start - orig_offset);
700 btrfs_mark_buffer_dirty(trans, leaf);
701 goto out;
702 }
703 }
704
705 while (start > key.offset || end < extent_end) {
706 if (key.offset == start)
707 split = end;
708
709 new_key.offset = split;
710 ret = btrfs_duplicate_item(trans, root, path, &new_key);
711 if (ret == -EAGAIN) {
712 btrfs_release_path(path);
713 goto again;
714 }
715 if (ret < 0) {
716 btrfs_abort_transaction(trans, ret);
717 goto out;
718 }
719
720 leaf = path->nodes[0];
721 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
722 struct btrfs_file_extent_item);
723 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
724 btrfs_set_file_extent_num_bytes(leaf, fi,
725 split - key.offset);
726
727 fi = btrfs_item_ptr(leaf, path->slots[0],
728 struct btrfs_file_extent_item);
729
730 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
731 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
732 btrfs_set_file_extent_num_bytes(leaf, fi,
733 extent_end - split);
734 btrfs_mark_buffer_dirty(trans, leaf);
735
736 ref.action = BTRFS_ADD_DELAYED_REF;
737 ref.bytenr = bytenr;
738 ref.num_bytes = num_bytes;
739 ref.parent = 0;
740 ref.owning_root = btrfs_root_id(root);
741 ref.ref_root = btrfs_root_id(root);
742 btrfs_init_data_ref(&ref, ino, orig_offset, 0, false);
743 ret = btrfs_inc_extent_ref(trans, &ref);
744 if (ret) {
745 btrfs_abort_transaction(trans, ret);
746 goto out;
747 }
748
749 if (split == start) {
750 key.offset = start;
751 } else {
752 if (start != key.offset) {
753 ret = -EINVAL;
754 btrfs_abort_transaction(trans, ret);
755 goto out;
756 }
757 path->slots[0]--;
758 extent_end = end;
759 }
760 recow = 1;
761 }
762
763 other_start = end;
764 other_end = 0;
765
766 ref.action = BTRFS_DROP_DELAYED_REF;
767 ref.bytenr = bytenr;
768 ref.num_bytes = num_bytes;
769 ref.parent = 0;
770 ref.owning_root = btrfs_root_id(root);
771 ref.ref_root = btrfs_root_id(root);
772 btrfs_init_data_ref(&ref, ino, orig_offset, 0, false);
773 if (extent_mergeable(leaf, path->slots[0] + 1,
774 ino, bytenr, orig_offset,
775 &other_start, &other_end)) {
776 if (recow) {
777 btrfs_release_path(path);
778 goto again;
779 }
780 extent_end = other_end;
781 del_slot = path->slots[0] + 1;
782 del_nr++;
783 ret = btrfs_free_extent(trans, &ref);
784 if (ret) {
785 btrfs_abort_transaction(trans, ret);
786 goto out;
787 }
788 }
789 other_start = 0;
790 other_end = start;
791 if (extent_mergeable(leaf, path->slots[0] - 1,
792 ino, bytenr, orig_offset,
793 &other_start, &other_end)) {
794 if (recow) {
795 btrfs_release_path(path);
796 goto again;
797 }
798 key.offset = other_start;
799 del_slot = path->slots[0];
800 del_nr++;
801 ret = btrfs_free_extent(trans, &ref);
802 if (ret) {
803 btrfs_abort_transaction(trans, ret);
804 goto out;
805 }
806 }
807 if (del_nr == 0) {
808 fi = btrfs_item_ptr(leaf, path->slots[0],
809 struct btrfs_file_extent_item);
810 btrfs_set_file_extent_type(leaf, fi,
811 BTRFS_FILE_EXTENT_REG);
812 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
813 btrfs_mark_buffer_dirty(trans, leaf);
814 } else {
815 fi = btrfs_item_ptr(leaf, del_slot - 1,
816 struct btrfs_file_extent_item);
817 btrfs_set_file_extent_type(leaf, fi,
818 BTRFS_FILE_EXTENT_REG);
819 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
820 btrfs_set_file_extent_num_bytes(leaf, fi,
821 extent_end - key.offset);
822 btrfs_mark_buffer_dirty(trans, leaf);
823
824 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
825 if (ret < 0) {
826 btrfs_abort_transaction(trans, ret);
827 goto out;
828 }
829 }
830out:
831 btrfs_free_path(path);
832 return ret;
833}
834
835/*
836 * On error return an unlocked folio and the error value
837 * On success return a locked folio and 0
838 */
839static int prepare_uptodate_folio(struct inode *inode, struct folio *folio, u64 pos,
840 u64 len, bool force_uptodate)
841{
842 u64 clamp_start = max_t(u64, pos, folio_pos(folio));
843 u64 clamp_end = min_t(u64, pos + len, folio_pos(folio) + folio_size(folio));
844 int ret = 0;
845
846 if (folio_test_uptodate(folio))
847 return 0;
848
849 if (!force_uptodate &&
850 IS_ALIGNED(clamp_start, PAGE_SIZE) &&
851 IS_ALIGNED(clamp_end, PAGE_SIZE))
852 return 0;
853
854 ret = btrfs_read_folio(NULL, folio);
855 if (ret)
856 return ret;
857 folio_lock(folio);
858 if (!folio_test_uptodate(folio)) {
859 folio_unlock(folio);
860 return -EIO;
861 }
862
863 /*
864 * Since btrfs_read_folio() will unlock the folio before it returns,
865 * there is a window where btrfs_release_folio() can be called to
866 * release the page. Here we check both inode mapping and page
867 * private to make sure the page was not released.
868 *
869 * The private flag check is essential for subpage as we need to store
870 * extra bitmap using folio private.
871 */
872 if (folio->mapping != inode->i_mapping || !folio_test_private(folio)) {
873 folio_unlock(folio);
874 return -EAGAIN;
875 }
876 return 0;
877}
878
879static gfp_t get_prepare_gfp_flags(struct inode *inode, bool nowait)
880{
881 gfp_t gfp;
882
883 gfp = btrfs_alloc_write_mask(inode->i_mapping);
884 if (nowait) {
885 gfp &= ~__GFP_DIRECT_RECLAIM;
886 gfp |= GFP_NOWAIT;
887 }
888
889 return gfp;
890}
891
892/*
893 * Get folio into the page cache and lock it.
894 */
895static noinline int prepare_one_folio(struct inode *inode, struct folio **folio_ret,
896 loff_t pos, size_t write_bytes,
897 bool force_uptodate, bool nowait)
898{
899 unsigned long index = pos >> PAGE_SHIFT;
900 gfp_t mask = get_prepare_gfp_flags(inode, nowait);
901 fgf_t fgp_flags = (nowait ? FGP_WRITEBEGIN | FGP_NOWAIT : FGP_WRITEBEGIN);
902 struct folio *folio;
903 int ret = 0;
904
905again:
906 folio = __filemap_get_folio(inode->i_mapping, index, fgp_flags, mask);
907 if (IS_ERR(folio)) {
908 if (nowait)
909 ret = -EAGAIN;
910 else
911 ret = PTR_ERR(folio);
912 return ret;
913 }
914 folio_wait_writeback(folio);
915 /* Only support page sized folio yet. */
916 ASSERT(folio_order(folio) == 0);
917 ret = set_folio_extent_mapped(folio);
918 if (ret < 0) {
919 folio_unlock(folio);
920 folio_put(folio);
921 return ret;
922 }
923 ret = prepare_uptodate_folio(inode, folio, pos, write_bytes, force_uptodate);
924 if (ret) {
925 /* The folio is already unlocked. */
926 folio_put(folio);
927 if (!nowait && ret == -EAGAIN) {
928 ret = 0;
929 goto again;
930 }
931 return ret;
932 }
933 *folio_ret = folio;
934 return 0;
935}
936
937/*
938 * Locks the extent and properly waits for data=ordered extents to finish
939 * before allowing the folios to be modified if need.
940 *
941 * Return:
942 * 1 - the extent is locked
943 * 0 - the extent is not locked, and everything is OK
944 * -EAGAIN - need to prepare the folios again
945 */
946static noinline int
947lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct folio *folio,
948 loff_t pos, size_t write_bytes,
949 u64 *lockstart, u64 *lockend, bool nowait,
950 struct extent_state **cached_state)
951{
952 struct btrfs_fs_info *fs_info = inode->root->fs_info;
953 u64 start_pos;
954 u64 last_pos;
955 int ret = 0;
956
957 start_pos = round_down(pos, fs_info->sectorsize);
958 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
959
960 if (start_pos < inode->vfs_inode.i_size) {
961 struct btrfs_ordered_extent *ordered;
962
963 if (nowait) {
964 if (!try_lock_extent(&inode->io_tree, start_pos, last_pos,
965 cached_state)) {
966 folio_unlock(folio);
967 folio_put(folio);
968 return -EAGAIN;
969 }
970 } else {
971 lock_extent(&inode->io_tree, start_pos, last_pos, cached_state);
972 }
973
974 ordered = btrfs_lookup_ordered_range(inode, start_pos,
975 last_pos - start_pos + 1);
976 if (ordered &&
977 ordered->file_offset + ordered->num_bytes > start_pos &&
978 ordered->file_offset <= last_pos) {
979 unlock_extent(&inode->io_tree, start_pos, last_pos,
980 cached_state);
981 folio_unlock(folio);
982 folio_put(folio);
983 btrfs_start_ordered_extent(ordered);
984 btrfs_put_ordered_extent(ordered);
985 return -EAGAIN;
986 }
987 if (ordered)
988 btrfs_put_ordered_extent(ordered);
989
990 *lockstart = start_pos;
991 *lockend = last_pos;
992 ret = 1;
993 }
994
995 /*
996 * We should be called after prepare_one_folio() which should have locked
997 * all pages in the range.
998 */
999 WARN_ON(!folio_test_locked(folio));
1000
1001 return ret;
1002}
1003
1004/*
1005 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1006 *
1007 * @pos: File offset.
1008 * @write_bytes: The length to write, will be updated to the nocow writeable
1009 * range.
1010 *
1011 * This function will flush ordered extents in the range to ensure proper
1012 * nocow checks.
1013 *
1014 * Return:
1015 * > 0 If we can nocow, and updates @write_bytes.
1016 * 0 If we can't do a nocow write.
1017 * -EAGAIN If we can't do a nocow write because snapshoting of the inode's
1018 * root is in progress.
1019 * < 0 If an error happened.
1020 *
1021 * NOTE: Callers need to call btrfs_check_nocow_unlock() if we return > 0.
1022 */
1023int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1024 size_t *write_bytes, bool nowait)
1025{
1026 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1027 struct btrfs_root *root = inode->root;
1028 struct extent_state *cached_state = NULL;
1029 u64 lockstart, lockend;
1030 u64 num_bytes;
1031 int ret;
1032
1033 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1034 return 0;
1035
1036 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1037 return -EAGAIN;
1038
1039 lockstart = round_down(pos, fs_info->sectorsize);
1040 lockend = round_up(pos + *write_bytes,
1041 fs_info->sectorsize) - 1;
1042 num_bytes = lockend - lockstart + 1;
1043
1044 if (nowait) {
1045 if (!btrfs_try_lock_ordered_range(inode, lockstart, lockend,
1046 &cached_state)) {
1047 btrfs_drew_write_unlock(&root->snapshot_lock);
1048 return -EAGAIN;
1049 }
1050 } else {
1051 btrfs_lock_and_flush_ordered_range(inode, lockstart, lockend,
1052 &cached_state);
1053 }
1054 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1055 NULL, nowait, false);
1056 if (ret <= 0)
1057 btrfs_drew_write_unlock(&root->snapshot_lock);
1058 else
1059 *write_bytes = min_t(size_t, *write_bytes ,
1060 num_bytes - pos + lockstart);
1061 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
1062
1063 return ret;
1064}
1065
1066void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1067{
1068 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1069}
1070
1071int btrfs_write_check(struct kiocb *iocb, size_t count)
1072{
1073 struct file *file = iocb->ki_filp;
1074 struct inode *inode = file_inode(file);
1075 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1076 loff_t pos = iocb->ki_pos;
1077 int ret;
1078 loff_t oldsize;
1079
1080 /*
1081 * Quickly bail out on NOWAIT writes if we don't have the nodatacow or
1082 * prealloc flags, as without those flags we always have to COW. We will
1083 * later check if we can really COW into the target range (using
1084 * can_nocow_extent() at btrfs_get_blocks_direct_write()).
1085 */
1086 if ((iocb->ki_flags & IOCB_NOWAIT) &&
1087 !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1088 return -EAGAIN;
1089
1090 ret = file_remove_privs(file);
1091 if (ret)
1092 return ret;
1093
1094 /*
1095 * We reserve space for updating the inode when we reserve space for the
1096 * extent we are going to write, so we will enospc out there. We don't
1097 * need to start yet another transaction to update the inode as we will
1098 * update the inode when we finish writing whatever data we write.
1099 */
1100 if (!IS_NOCMTIME(inode)) {
1101 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
1102 inode_inc_iversion(inode);
1103 }
1104
1105 oldsize = i_size_read(inode);
1106 if (pos > oldsize) {
1107 /* Expand hole size to cover write data, preventing empty gap */
1108 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1109
1110 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1111 if (ret)
1112 return ret;
1113 }
1114
1115 return 0;
1116}
1117
1118ssize_t btrfs_buffered_write(struct kiocb *iocb, struct iov_iter *i)
1119{
1120 struct file *file = iocb->ki_filp;
1121 loff_t pos;
1122 struct inode *inode = file_inode(file);
1123 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1124 struct extent_changeset *data_reserved = NULL;
1125 u64 release_bytes = 0;
1126 u64 lockstart;
1127 u64 lockend;
1128 size_t num_written = 0;
1129 ssize_t ret;
1130 loff_t old_isize;
1131 unsigned int ilock_flags = 0;
1132 const bool nowait = (iocb->ki_flags & IOCB_NOWAIT);
1133 unsigned int bdp_flags = (nowait ? BDP_ASYNC : 0);
1134 bool only_release_metadata = false;
1135
1136 if (nowait)
1137 ilock_flags |= BTRFS_ILOCK_TRY;
1138
1139 ret = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1140 if (ret < 0)
1141 return ret;
1142
1143 /*
1144 * We can only trust the isize with inode lock held, or it can race with
1145 * other buffered writes and cause incorrect call of
1146 * pagecache_isize_extended() to overwrite existing data.
1147 */
1148 old_isize = i_size_read(inode);
1149
1150 ret = generic_write_checks(iocb, i);
1151 if (ret <= 0)
1152 goto out;
1153
1154 ret = btrfs_write_check(iocb, ret);
1155 if (ret < 0)
1156 goto out;
1157
1158 pos = iocb->ki_pos;
1159 while (iov_iter_count(i) > 0) {
1160 struct extent_state *cached_state = NULL;
1161 size_t offset = offset_in_page(pos);
1162 size_t sector_offset;
1163 size_t write_bytes = min(iov_iter_count(i), PAGE_SIZE - offset);
1164 size_t reserve_bytes;
1165 size_t copied;
1166 size_t dirty_sectors;
1167 size_t num_sectors;
1168 struct folio *folio = NULL;
1169 int extents_locked;
1170 bool force_page_uptodate = false;
1171
1172 /*
1173 * Fault pages before locking them in prepare_one_folio()
1174 * to avoid recursive lock
1175 */
1176 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1177 ret = -EFAULT;
1178 break;
1179 }
1180
1181 only_release_metadata = false;
1182 sector_offset = pos & (fs_info->sectorsize - 1);
1183
1184 extent_changeset_release(data_reserved);
1185 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1186 &data_reserved, pos,
1187 write_bytes, nowait);
1188 if (ret < 0) {
1189 int can_nocow;
1190
1191 if (nowait && (ret == -ENOSPC || ret == -EAGAIN)) {
1192 ret = -EAGAIN;
1193 break;
1194 }
1195
1196 /*
1197 * If we don't have to COW at the offset, reserve
1198 * metadata only. write_bytes may get smaller than
1199 * requested here.
1200 */
1201 can_nocow = btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1202 &write_bytes, nowait);
1203 if (can_nocow < 0)
1204 ret = can_nocow;
1205 if (can_nocow > 0)
1206 ret = 0;
1207 if (ret)
1208 break;
1209 only_release_metadata = true;
1210 }
1211
1212 reserve_bytes = round_up(write_bytes + sector_offset,
1213 fs_info->sectorsize);
1214 WARN_ON(reserve_bytes == 0);
1215 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1216 reserve_bytes,
1217 reserve_bytes, nowait);
1218 if (ret) {
1219 if (!only_release_metadata)
1220 btrfs_free_reserved_data_space(BTRFS_I(inode),
1221 data_reserved, pos,
1222 write_bytes);
1223 else
1224 btrfs_check_nocow_unlock(BTRFS_I(inode));
1225
1226 if (nowait && ret == -ENOSPC)
1227 ret = -EAGAIN;
1228 break;
1229 }
1230
1231 release_bytes = reserve_bytes;
1232again:
1233 ret = balance_dirty_pages_ratelimited_flags(inode->i_mapping, bdp_flags);
1234 if (ret) {
1235 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1236 break;
1237 }
1238
1239 ret = prepare_one_folio(inode, &folio, pos, write_bytes,
1240 force_page_uptodate, false);
1241 if (ret) {
1242 btrfs_delalloc_release_extents(BTRFS_I(inode),
1243 reserve_bytes);
1244 break;
1245 }
1246
1247 extents_locked = lock_and_cleanup_extent_if_need(BTRFS_I(inode),
1248 folio, pos, write_bytes, &lockstart,
1249 &lockend, nowait, &cached_state);
1250 if (extents_locked < 0) {
1251 if (!nowait && extents_locked == -EAGAIN)
1252 goto again;
1253
1254 btrfs_delalloc_release_extents(BTRFS_I(inode),
1255 reserve_bytes);
1256 ret = extents_locked;
1257 break;
1258 }
1259
1260 copied = btrfs_copy_from_user(pos, write_bytes, folio, i);
1261
1262 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1263 dirty_sectors = round_up(copied + sector_offset,
1264 fs_info->sectorsize);
1265 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1266
1267 if (copied == 0) {
1268 force_page_uptodate = true;
1269 dirty_sectors = 0;
1270 } else {
1271 force_page_uptodate = false;
1272 }
1273
1274 if (num_sectors > dirty_sectors) {
1275 /* release everything except the sectors we dirtied */
1276 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1277 if (only_release_metadata) {
1278 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1279 release_bytes, true);
1280 } else {
1281 u64 release_start = round_up(pos + copied,
1282 fs_info->sectorsize);
1283 btrfs_delalloc_release_space(BTRFS_I(inode),
1284 data_reserved, release_start,
1285 release_bytes, true);
1286 }
1287 }
1288
1289 release_bytes = round_up(copied + sector_offset,
1290 fs_info->sectorsize);
1291
1292 ret = btrfs_dirty_folio(BTRFS_I(inode), folio, pos, copied,
1293 &cached_state, only_release_metadata);
1294
1295 /*
1296 * If we have not locked the extent range, because the range's
1297 * start offset is >= i_size, we might still have a non-NULL
1298 * cached extent state, acquired while marking the extent range
1299 * as delalloc through btrfs_dirty_page(). Therefore free any
1300 * possible cached extent state to avoid a memory leak.
1301 */
1302 if (extents_locked)
1303 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
1304 lockend, &cached_state);
1305 else
1306 free_extent_state(cached_state);
1307
1308 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1309 if (ret) {
1310 btrfs_drop_folio(fs_info, folio, pos, copied);
1311 break;
1312 }
1313
1314 release_bytes = 0;
1315 if (only_release_metadata)
1316 btrfs_check_nocow_unlock(BTRFS_I(inode));
1317
1318 btrfs_drop_folio(fs_info, folio, pos, copied);
1319
1320 cond_resched();
1321
1322 pos += copied;
1323 num_written += copied;
1324 }
1325
1326 if (release_bytes) {
1327 if (only_release_metadata) {
1328 btrfs_check_nocow_unlock(BTRFS_I(inode));
1329 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1330 release_bytes, true);
1331 } else {
1332 btrfs_delalloc_release_space(BTRFS_I(inode),
1333 data_reserved,
1334 round_down(pos, fs_info->sectorsize),
1335 release_bytes, true);
1336 }
1337 }
1338
1339 extent_changeset_free(data_reserved);
1340 if (num_written > 0) {
1341 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1342 iocb->ki_pos += num_written;
1343 }
1344out:
1345 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1346 return num_written ? num_written : ret;
1347}
1348
1349static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
1350 const struct btrfs_ioctl_encoded_io_args *encoded)
1351{
1352 struct file *file = iocb->ki_filp;
1353 struct inode *inode = file_inode(file);
1354 loff_t count;
1355 ssize_t ret;
1356
1357 btrfs_inode_lock(BTRFS_I(inode), 0);
1358 count = encoded->len;
1359 ret = generic_write_checks_count(iocb, &count);
1360 if (ret == 0 && count != encoded->len) {
1361 /*
1362 * The write got truncated by generic_write_checks_count(). We
1363 * can't do a partial encoded write.
1364 */
1365 ret = -EFBIG;
1366 }
1367 if (ret || encoded->len == 0)
1368 goto out;
1369
1370 ret = btrfs_write_check(iocb, encoded->len);
1371 if (ret < 0)
1372 goto out;
1373
1374 ret = btrfs_do_encoded_write(iocb, from, encoded);
1375out:
1376 btrfs_inode_unlock(BTRFS_I(inode), 0);
1377 return ret;
1378}
1379
1380ssize_t btrfs_do_write_iter(struct kiocb *iocb, struct iov_iter *from,
1381 const struct btrfs_ioctl_encoded_io_args *encoded)
1382{
1383 struct file *file = iocb->ki_filp;
1384 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1385 ssize_t num_written, num_sync;
1386
1387 /*
1388 * If the fs flips readonly due to some impossible error, although we
1389 * have opened a file as writable, we have to stop this write operation
1390 * to ensure consistency.
1391 */
1392 if (BTRFS_FS_ERROR(inode->root->fs_info))
1393 return -EROFS;
1394
1395 if (encoded && (iocb->ki_flags & IOCB_NOWAIT))
1396 return -EOPNOTSUPP;
1397
1398 if (encoded) {
1399 num_written = btrfs_encoded_write(iocb, from, encoded);
1400 num_sync = encoded->len;
1401 } else if (iocb->ki_flags & IOCB_DIRECT) {
1402 num_written = btrfs_direct_write(iocb, from);
1403 num_sync = num_written;
1404 } else {
1405 num_written = btrfs_buffered_write(iocb, from);
1406 num_sync = num_written;
1407 }
1408
1409 btrfs_set_inode_last_sub_trans(inode);
1410
1411 if (num_sync > 0) {
1412 num_sync = generic_write_sync(iocb, num_sync);
1413 if (num_sync < 0)
1414 num_written = num_sync;
1415 }
1416
1417 return num_written;
1418}
1419
1420static ssize_t btrfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1421{
1422 return btrfs_do_write_iter(iocb, from, NULL);
1423}
1424
1425int btrfs_release_file(struct inode *inode, struct file *filp)
1426{
1427 struct btrfs_file_private *private = filp->private_data;
1428
1429 if (private) {
1430 kfree(private->filldir_buf);
1431 free_extent_state(private->llseek_cached_state);
1432 kfree(private);
1433 filp->private_data = NULL;
1434 }
1435
1436 /*
1437 * Set by setattr when we are about to truncate a file from a non-zero
1438 * size to a zero size. This tries to flush down new bytes that may
1439 * have been written if the application were using truncate to replace
1440 * a file in place.
1441 */
1442 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
1443 &BTRFS_I(inode)->runtime_flags))
1444 filemap_flush(inode->i_mapping);
1445 return 0;
1446}
1447
1448static int start_ordered_ops(struct btrfs_inode *inode, loff_t start, loff_t end)
1449{
1450 int ret;
1451 struct blk_plug plug;
1452
1453 /*
1454 * This is only called in fsync, which would do synchronous writes, so
1455 * a plug can merge adjacent IOs as much as possible. Esp. in case of
1456 * multiple disks using raid profile, a large IO can be split to
1457 * several segments of stripe length (currently 64K).
1458 */
1459 blk_start_plug(&plug);
1460 ret = btrfs_fdatawrite_range(inode, start, end);
1461 blk_finish_plug(&plug);
1462
1463 return ret;
1464}
1465
1466static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
1467{
1468 struct btrfs_inode *inode = ctx->inode;
1469 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1470
1471 if (btrfs_inode_in_log(inode, btrfs_get_fs_generation(fs_info)) &&
1472 list_empty(&ctx->ordered_extents))
1473 return true;
1474
1475 /*
1476 * If we are doing a fast fsync we can not bail out if the inode's
1477 * last_trans is <= then the last committed transaction, because we only
1478 * update the last_trans of the inode during ordered extent completion,
1479 * and for a fast fsync we don't wait for that, we only wait for the
1480 * writeback to complete.
1481 */
1482 if (inode->last_trans <= btrfs_get_last_trans_committed(fs_info) &&
1483 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
1484 list_empty(&ctx->ordered_extents)))
1485 return true;
1486
1487 return false;
1488}
1489
1490/*
1491 * fsync call for both files and directories. This logs the inode into
1492 * the tree log instead of forcing full commits whenever possible.
1493 *
1494 * It needs to call filemap_fdatawait so that all ordered extent updates are
1495 * in the metadata btree are up to date for copying to the log.
1496 *
1497 * It drops the inode mutex before doing the tree log commit. This is an
1498 * important optimization for directories because holding the mutex prevents
1499 * new operations on the dir while we write to disk.
1500 */
1501int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1502{
1503 struct dentry *dentry = file_dentry(file);
1504 struct btrfs_inode *inode = BTRFS_I(d_inode(dentry));
1505 struct btrfs_root *root = inode->root;
1506 struct btrfs_fs_info *fs_info = root->fs_info;
1507 struct btrfs_trans_handle *trans;
1508 struct btrfs_log_ctx ctx;
1509 int ret = 0, err;
1510 u64 len;
1511 bool full_sync;
1512 bool skip_ilock = false;
1513
1514 if (current->journal_info == BTRFS_TRANS_DIO_WRITE_STUB) {
1515 skip_ilock = true;
1516 current->journal_info = NULL;
1517 btrfs_assert_inode_locked(inode);
1518 }
1519
1520 trace_btrfs_sync_file(file, datasync);
1521
1522 btrfs_init_log_ctx(&ctx, inode);
1523
1524 /*
1525 * Always set the range to a full range, otherwise we can get into
1526 * several problems, from missing file extent items to represent holes
1527 * when not using the NO_HOLES feature, to log tree corruption due to
1528 * races between hole detection during logging and completion of ordered
1529 * extents outside the range, to missing checksums due to ordered extents
1530 * for which we flushed only a subset of their pages.
1531 */
1532 start = 0;
1533 end = LLONG_MAX;
1534 len = (u64)LLONG_MAX + 1;
1535
1536 /*
1537 * We write the dirty pages in the range and wait until they complete
1538 * out of the ->i_mutex. If so, we can flush the dirty pages by
1539 * multi-task, and make the performance up. See
1540 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1541 */
1542 ret = start_ordered_ops(inode, start, end);
1543 if (ret)
1544 goto out;
1545
1546 if (skip_ilock)
1547 down_write(&inode->i_mmap_lock);
1548 else
1549 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
1550
1551 atomic_inc(&root->log_batch);
1552
1553 /*
1554 * Before we acquired the inode's lock and the mmap lock, someone may
1555 * have dirtied more pages in the target range. We need to make sure
1556 * that writeback for any such pages does not start while we are logging
1557 * the inode, because if it does, any of the following might happen when
1558 * we are not doing a full inode sync:
1559 *
1560 * 1) We log an extent after its writeback finishes but before its
1561 * checksums are added to the csum tree, leading to -EIO errors
1562 * when attempting to read the extent after a log replay.
1563 *
1564 * 2) We can end up logging an extent before its writeback finishes.
1565 * Therefore after the log replay we will have a file extent item
1566 * pointing to an unwritten extent (and no data checksums as well).
1567 *
1568 * So trigger writeback for any eventual new dirty pages and then we
1569 * wait for all ordered extents to complete below.
1570 */
1571 ret = start_ordered_ops(inode, start, end);
1572 if (ret) {
1573 if (skip_ilock)
1574 up_write(&inode->i_mmap_lock);
1575 else
1576 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
1577 goto out;
1578 }
1579
1580 /*
1581 * Always check for the full sync flag while holding the inode's lock,
1582 * to avoid races with other tasks. The flag must be either set all the
1583 * time during logging or always off all the time while logging.
1584 * We check the flag here after starting delalloc above, because when
1585 * running delalloc the full sync flag may be set if we need to drop
1586 * extra extent map ranges due to temporary memory allocation failures.
1587 */
1588 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
1589
1590 /*
1591 * We have to do this here to avoid the priority inversion of waiting on
1592 * IO of a lower priority task while holding a transaction open.
1593 *
1594 * For a full fsync we wait for the ordered extents to complete while
1595 * for a fast fsync we wait just for writeback to complete, and then
1596 * attach the ordered extents to the transaction so that a transaction
1597 * commit waits for their completion, to avoid data loss if we fsync,
1598 * the current transaction commits before the ordered extents complete
1599 * and a power failure happens right after that.
1600 *
1601 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
1602 * logical address recorded in the ordered extent may change. We need
1603 * to wait for the IO to stabilize the logical address.
1604 */
1605 if (full_sync || btrfs_is_zoned(fs_info)) {
1606 ret = btrfs_wait_ordered_range(inode, start, len);
1607 clear_bit(BTRFS_INODE_COW_WRITE_ERROR, &inode->runtime_flags);
1608 } else {
1609 /*
1610 * Get our ordered extents as soon as possible to avoid doing
1611 * checksum lookups in the csum tree, and use instead the
1612 * checksums attached to the ordered extents.
1613 */
1614 btrfs_get_ordered_extents_for_logging(inode, &ctx.ordered_extents);
1615 ret = filemap_fdatawait_range(inode->vfs_inode.i_mapping, start, end);
1616 if (ret)
1617 goto out_release_extents;
1618
1619 /*
1620 * Check and clear the BTRFS_INODE_COW_WRITE_ERROR now after
1621 * starting and waiting for writeback, because for buffered IO
1622 * it may have been set during the end IO callback
1623 * (end_bbio_data_write() -> btrfs_finish_ordered_extent()) in
1624 * case an error happened and we need to wait for ordered
1625 * extents to complete so that any extent maps that point to
1626 * unwritten locations are dropped and we don't log them.
1627 */
1628 if (test_and_clear_bit(BTRFS_INODE_COW_WRITE_ERROR, &inode->runtime_flags))
1629 ret = btrfs_wait_ordered_range(inode, start, len);
1630 }
1631
1632 if (ret)
1633 goto out_release_extents;
1634
1635 atomic_inc(&root->log_batch);
1636
1637 if (skip_inode_logging(&ctx)) {
1638 /*
1639 * We've had everything committed since the last time we were
1640 * modified so clear this flag in case it was set for whatever
1641 * reason, it's no longer relevant.
1642 */
1643 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
1644 /*
1645 * An ordered extent might have started before and completed
1646 * already with io errors, in which case the inode was not
1647 * updated and we end up here. So check the inode's mapping
1648 * for any errors that might have happened since we last
1649 * checked called fsync.
1650 */
1651 ret = filemap_check_wb_err(inode->vfs_inode.i_mapping, file->f_wb_err);
1652 goto out_release_extents;
1653 }
1654
1655 btrfs_init_log_ctx_scratch_eb(&ctx);
1656
1657 /*
1658 * We use start here because we will need to wait on the IO to complete
1659 * in btrfs_sync_log, which could require joining a transaction (for
1660 * example checking cross references in the nocow path). If we use join
1661 * here we could get into a situation where we're waiting on IO to
1662 * happen that is blocked on a transaction trying to commit. With start
1663 * we inc the extwriter counter, so we wait for all extwriters to exit
1664 * before we start blocking joiners. This comment is to keep somebody
1665 * from thinking they are super smart and changing this to
1666 * btrfs_join_transaction *cough*Josef*cough*.
1667 */
1668 trans = btrfs_start_transaction(root, 0);
1669 if (IS_ERR(trans)) {
1670 ret = PTR_ERR(trans);
1671 goto out_release_extents;
1672 }
1673 trans->in_fsync = true;
1674
1675 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
1676 /*
1677 * Scratch eb no longer needed, release before syncing log or commit
1678 * transaction, to avoid holding unnecessary memory during such long
1679 * operations.
1680 */
1681 if (ctx.scratch_eb) {
1682 free_extent_buffer(ctx.scratch_eb);
1683 ctx.scratch_eb = NULL;
1684 }
1685 btrfs_release_log_ctx_extents(&ctx);
1686 if (ret < 0) {
1687 /* Fallthrough and commit/free transaction. */
1688 ret = BTRFS_LOG_FORCE_COMMIT;
1689 }
1690
1691 /* we've logged all the items and now have a consistent
1692 * version of the file in the log. It is possible that
1693 * someone will come in and modify the file, but that's
1694 * fine because the log is consistent on disk, and we
1695 * have references to all of the file's extents
1696 *
1697 * It is possible that someone will come in and log the
1698 * file again, but that will end up using the synchronization
1699 * inside btrfs_sync_log to keep things safe.
1700 */
1701 if (skip_ilock)
1702 up_write(&inode->i_mmap_lock);
1703 else
1704 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
1705
1706 if (ret == BTRFS_NO_LOG_SYNC) {
1707 ret = btrfs_end_transaction(trans);
1708 goto out;
1709 }
1710
1711 /* We successfully logged the inode, attempt to sync the log. */
1712 if (!ret) {
1713 ret = btrfs_sync_log(trans, root, &ctx);
1714 if (!ret) {
1715 ret = btrfs_end_transaction(trans);
1716 goto out;
1717 }
1718 }
1719
1720 /*
1721 * At this point we need to commit the transaction because we had
1722 * btrfs_need_log_full_commit() or some other error.
1723 *
1724 * If we didn't do a full sync we have to stop the trans handle, wait on
1725 * the ordered extents, start it again and commit the transaction. If
1726 * we attempt to wait on the ordered extents here we could deadlock with
1727 * something like fallocate() that is holding the extent lock trying to
1728 * start a transaction while some other thread is trying to commit the
1729 * transaction while we (fsync) are currently holding the transaction
1730 * open.
1731 */
1732 if (!full_sync) {
1733 ret = btrfs_end_transaction(trans);
1734 if (ret)
1735 goto out;
1736 ret = btrfs_wait_ordered_range(inode, start, len);
1737 if (ret)
1738 goto out;
1739
1740 /*
1741 * This is safe to use here because we're only interested in
1742 * making sure the transaction that had the ordered extents is
1743 * committed. We aren't waiting on anything past this point,
1744 * we're purely getting the transaction and committing it.
1745 */
1746 trans = btrfs_attach_transaction_barrier(root);
1747 if (IS_ERR(trans)) {
1748 ret = PTR_ERR(trans);
1749
1750 /*
1751 * We committed the transaction and there's no currently
1752 * running transaction, this means everything we care
1753 * about made it to disk and we are done.
1754 */
1755 if (ret == -ENOENT)
1756 ret = 0;
1757 goto out;
1758 }
1759 }
1760
1761 ret = btrfs_commit_transaction(trans);
1762out:
1763 free_extent_buffer(ctx.scratch_eb);
1764 ASSERT(list_empty(&ctx.list));
1765 ASSERT(list_empty(&ctx.conflict_inodes));
1766 err = file_check_and_advance_wb_err(file);
1767 if (!ret)
1768 ret = err;
1769 return ret > 0 ? -EIO : ret;
1770
1771out_release_extents:
1772 btrfs_release_log_ctx_extents(&ctx);
1773 if (skip_ilock)
1774 up_write(&inode->i_mmap_lock);
1775 else
1776 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
1777 goto out;
1778}
1779
1780/*
1781 * btrfs_page_mkwrite() is not allowed to change the file size as it gets
1782 * called from a page fault handler when a page is first dirtied. Hence we must
1783 * be careful to check for EOF conditions here. We set the page up correctly
1784 * for a written page which means we get ENOSPC checking when writing into
1785 * holes and correct delalloc and unwritten extent mapping on filesystems that
1786 * support these features.
1787 *
1788 * We are not allowed to take the i_mutex here so we have to play games to
1789 * protect against truncate races as the page could now be beyond EOF. Because
1790 * truncate_setsize() writes the inode size before removing pages, once we have
1791 * the page lock we can determine safely if the page is beyond EOF. If it is not
1792 * beyond EOF, then the page is guaranteed safe against truncation until we
1793 * unlock the page.
1794 */
1795static vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
1796{
1797 struct page *page = vmf->page;
1798 struct folio *folio = page_folio(page);
1799 struct inode *inode = file_inode(vmf->vma->vm_file);
1800 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1801 struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1802 struct btrfs_ordered_extent *ordered;
1803 struct extent_state *cached_state = NULL;
1804 struct extent_changeset *data_reserved = NULL;
1805 unsigned long zero_start;
1806 loff_t size;
1807 vm_fault_t ret;
1808 int ret2;
1809 int reserved = 0;
1810 u64 reserved_space;
1811 u64 page_start;
1812 u64 page_end;
1813 u64 end;
1814
1815 ASSERT(folio_order(folio) == 0);
1816
1817 reserved_space = PAGE_SIZE;
1818
1819 sb_start_pagefault(inode->i_sb);
1820 page_start = folio_pos(folio);
1821 page_end = page_start + folio_size(folio) - 1;
1822 end = page_end;
1823
1824 /*
1825 * Reserving delalloc space after obtaining the page lock can lead to
1826 * deadlock. For example, if a dirty page is locked by this function
1827 * and the call to btrfs_delalloc_reserve_space() ends up triggering
1828 * dirty page write out, then the btrfs_writepages() function could
1829 * end up waiting indefinitely to get a lock on the page currently
1830 * being processed by btrfs_page_mkwrite() function.
1831 */
1832 ret2 = btrfs_delalloc_reserve_space(BTRFS_I(inode), &data_reserved,
1833 page_start, reserved_space);
1834 if (!ret2) {
1835 ret2 = file_update_time(vmf->vma->vm_file);
1836 reserved = 1;
1837 }
1838 if (ret2) {
1839 ret = vmf_error(ret2);
1840 if (reserved)
1841 goto out;
1842 goto out_noreserve;
1843 }
1844
1845 /* Make the VM retry the fault. */
1846 ret = VM_FAULT_NOPAGE;
1847again:
1848 down_read(&BTRFS_I(inode)->i_mmap_lock);
1849 folio_lock(folio);
1850 size = i_size_read(inode);
1851
1852 if ((folio->mapping != inode->i_mapping) ||
1853 (page_start >= size)) {
1854 /* Page got truncated out from underneath us. */
1855 goto out_unlock;
1856 }
1857 folio_wait_writeback(folio);
1858
1859 lock_extent(io_tree, page_start, page_end, &cached_state);
1860 ret2 = set_folio_extent_mapped(folio);
1861 if (ret2 < 0) {
1862 ret = vmf_error(ret2);
1863 unlock_extent(io_tree, page_start, page_end, &cached_state);
1864 goto out_unlock;
1865 }
1866
1867 /*
1868 * We can't set the delalloc bits if there are pending ordered
1869 * extents. Drop our locks and wait for them to finish.
1870 */
1871 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start, PAGE_SIZE);
1872 if (ordered) {
1873 unlock_extent(io_tree, page_start, page_end, &cached_state);
1874 folio_unlock(folio);
1875 up_read(&BTRFS_I(inode)->i_mmap_lock);
1876 btrfs_start_ordered_extent(ordered);
1877 btrfs_put_ordered_extent(ordered);
1878 goto again;
1879 }
1880
1881 if (folio->index == ((size - 1) >> PAGE_SHIFT)) {
1882 reserved_space = round_up(size - page_start, fs_info->sectorsize);
1883 if (reserved_space < PAGE_SIZE) {
1884 end = page_start + reserved_space - 1;
1885 btrfs_delalloc_release_space(BTRFS_I(inode),
1886 data_reserved, page_start,
1887 PAGE_SIZE - reserved_space, true);
1888 }
1889 }
1890
1891 /*
1892 * page_mkwrite gets called when the page is firstly dirtied after it's
1893 * faulted in, but write(2) could also dirty a page and set delalloc
1894 * bits, thus in this case for space account reason, we still need to
1895 * clear any delalloc bits within this page range since we have to
1896 * reserve data&meta space before lock_page() (see above comments).
1897 */
1898 clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
1899 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
1900 EXTENT_DEFRAG, &cached_state);
1901
1902 ret2 = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start, end, 0,
1903 &cached_state);
1904 if (ret2) {
1905 unlock_extent(io_tree, page_start, page_end, &cached_state);
1906 ret = VM_FAULT_SIGBUS;
1907 goto out_unlock;
1908 }
1909
1910 /* Page is wholly or partially inside EOF. */
1911 if (page_start + folio_size(folio) > size)
1912 zero_start = offset_in_folio(folio, size);
1913 else
1914 zero_start = PAGE_SIZE;
1915
1916 if (zero_start != PAGE_SIZE)
1917 folio_zero_range(folio, zero_start, folio_size(folio) - zero_start);
1918
1919 btrfs_folio_clear_checked(fs_info, folio, page_start, PAGE_SIZE);
1920 btrfs_folio_set_dirty(fs_info, folio, page_start, end + 1 - page_start);
1921 btrfs_folio_set_uptodate(fs_info, folio, page_start, end + 1 - page_start);
1922
1923 btrfs_set_inode_last_sub_trans(BTRFS_I(inode));
1924
1925 unlock_extent(io_tree, page_start, page_end, &cached_state);
1926 up_read(&BTRFS_I(inode)->i_mmap_lock);
1927
1928 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
1929 sb_end_pagefault(inode->i_sb);
1930 extent_changeset_free(data_reserved);
1931 return VM_FAULT_LOCKED;
1932
1933out_unlock:
1934 folio_unlock(folio);
1935 up_read(&BTRFS_I(inode)->i_mmap_lock);
1936out:
1937 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
1938 btrfs_delalloc_release_space(BTRFS_I(inode), data_reserved, page_start,
1939 reserved_space, (ret != 0));
1940out_noreserve:
1941 sb_end_pagefault(inode->i_sb);
1942 extent_changeset_free(data_reserved);
1943 return ret;
1944}
1945
1946static const struct vm_operations_struct btrfs_file_vm_ops = {
1947 .fault = filemap_fault,
1948 .map_pages = filemap_map_pages,
1949 .page_mkwrite = btrfs_page_mkwrite,
1950};
1951
1952static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
1953{
1954 struct address_space *mapping = filp->f_mapping;
1955
1956 if (!mapping->a_ops->read_folio)
1957 return -ENOEXEC;
1958
1959 file_accessed(filp);
1960 vma->vm_ops = &btrfs_file_vm_ops;
1961
1962 return 0;
1963}
1964
1965static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
1966 int slot, u64 start, u64 end)
1967{
1968 struct btrfs_file_extent_item *fi;
1969 struct btrfs_key key;
1970
1971 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1972 return 0;
1973
1974 btrfs_item_key_to_cpu(leaf, &key, slot);
1975 if (key.objectid != btrfs_ino(inode) ||
1976 key.type != BTRFS_EXTENT_DATA_KEY)
1977 return 0;
1978
1979 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1980
1981 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
1982 return 0;
1983
1984 if (btrfs_file_extent_disk_bytenr(leaf, fi))
1985 return 0;
1986
1987 if (key.offset == end)
1988 return 1;
1989 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
1990 return 1;
1991 return 0;
1992}
1993
1994static int fill_holes(struct btrfs_trans_handle *trans,
1995 struct btrfs_inode *inode,
1996 struct btrfs_path *path, u64 offset, u64 end)
1997{
1998 struct btrfs_fs_info *fs_info = trans->fs_info;
1999 struct btrfs_root *root = inode->root;
2000 struct extent_buffer *leaf;
2001 struct btrfs_file_extent_item *fi;
2002 struct extent_map *hole_em;
2003 struct btrfs_key key;
2004 int ret;
2005
2006 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2007 goto out;
2008
2009 key.objectid = btrfs_ino(inode);
2010 key.type = BTRFS_EXTENT_DATA_KEY;
2011 key.offset = offset;
2012
2013 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2014 if (ret <= 0) {
2015 /*
2016 * We should have dropped this offset, so if we find it then
2017 * something has gone horribly wrong.
2018 */
2019 if (ret == 0)
2020 ret = -EINVAL;
2021 return ret;
2022 }
2023
2024 leaf = path->nodes[0];
2025 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2026 u64 num_bytes;
2027
2028 path->slots[0]--;
2029 fi = btrfs_item_ptr(leaf, path->slots[0],
2030 struct btrfs_file_extent_item);
2031 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2032 end - offset;
2033 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2034 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2035 btrfs_set_file_extent_offset(leaf, fi, 0);
2036 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2037 btrfs_mark_buffer_dirty(trans, leaf);
2038 goto out;
2039 }
2040
2041 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2042 u64 num_bytes;
2043
2044 key.offset = offset;
2045 btrfs_set_item_key_safe(trans, path, &key);
2046 fi = btrfs_item_ptr(leaf, path->slots[0],
2047 struct btrfs_file_extent_item);
2048 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2049 offset;
2050 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2051 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2052 btrfs_set_file_extent_offset(leaf, fi, 0);
2053 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2054 btrfs_mark_buffer_dirty(trans, leaf);
2055 goto out;
2056 }
2057 btrfs_release_path(path);
2058
2059 ret = btrfs_insert_hole_extent(trans, root, btrfs_ino(inode), offset,
2060 end - offset);
2061 if (ret)
2062 return ret;
2063
2064out:
2065 btrfs_release_path(path);
2066
2067 hole_em = alloc_extent_map();
2068 if (!hole_em) {
2069 btrfs_drop_extent_map_range(inode, offset, end - 1, false);
2070 btrfs_set_inode_full_sync(inode);
2071 } else {
2072 hole_em->start = offset;
2073 hole_em->len = end - offset;
2074 hole_em->ram_bytes = hole_em->len;
2075
2076 hole_em->disk_bytenr = EXTENT_MAP_HOLE;
2077 hole_em->disk_num_bytes = 0;
2078 hole_em->generation = trans->transid;
2079
2080 ret = btrfs_replace_extent_map_range(inode, hole_em, true);
2081 free_extent_map(hole_em);
2082 if (ret)
2083 btrfs_set_inode_full_sync(inode);
2084 }
2085
2086 return 0;
2087}
2088
2089/*
2090 * Find a hole extent on given inode and change start/len to the end of hole
2091 * extent.(hole/vacuum extent whose em->start <= start &&
2092 * em->start + em->len > start)
2093 * When a hole extent is found, return 1 and modify start/len.
2094 */
2095static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2096{
2097 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2098 struct extent_map *em;
2099 int ret = 0;
2100
2101 em = btrfs_get_extent(inode, NULL,
2102 round_down(*start, fs_info->sectorsize),
2103 round_up(*len, fs_info->sectorsize));
2104 if (IS_ERR(em))
2105 return PTR_ERR(em);
2106
2107 /* Hole or vacuum extent(only exists in no-hole mode) */
2108 if (em->disk_bytenr == EXTENT_MAP_HOLE) {
2109 ret = 1;
2110 *len = em->start + em->len > *start + *len ?
2111 0 : *start + *len - em->start - em->len;
2112 *start = em->start + em->len;
2113 }
2114 free_extent_map(em);
2115 return ret;
2116}
2117
2118static void btrfs_punch_hole_lock_range(struct inode *inode,
2119 const u64 lockstart,
2120 const u64 lockend,
2121 struct extent_state **cached_state)
2122{
2123 /*
2124 * For subpage case, if the range is not at page boundary, we could
2125 * have pages at the leading/tailing part of the range.
2126 * This could lead to dead loop since filemap_range_has_page()
2127 * will always return true.
2128 * So here we need to do extra page alignment for
2129 * filemap_range_has_page().
2130 */
2131 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2132 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2133
2134 while (1) {
2135 truncate_pagecache_range(inode, lockstart, lockend);
2136
2137 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2138 cached_state);
2139 /*
2140 * We can't have ordered extents in the range, nor dirty/writeback
2141 * pages, because we have locked the inode's VFS lock in exclusive
2142 * mode, we have locked the inode's i_mmap_lock in exclusive mode,
2143 * we have flushed all delalloc in the range and we have waited
2144 * for any ordered extents in the range to complete.
2145 * We can race with anyone reading pages from this range, so after
2146 * locking the range check if we have pages in the range, and if
2147 * we do, unlock the range and retry.
2148 */
2149 if (!filemap_range_has_page(inode->i_mapping, page_lockstart,
2150 page_lockend))
2151 break;
2152
2153 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2154 cached_state);
2155 }
2156
2157 btrfs_assert_inode_range_clean(BTRFS_I(inode), lockstart, lockend);
2158}
2159
2160static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2161 struct btrfs_inode *inode,
2162 struct btrfs_path *path,
2163 struct btrfs_replace_extent_info *extent_info,
2164 const u64 replace_len,
2165 const u64 bytes_to_drop)
2166{
2167 struct btrfs_fs_info *fs_info = trans->fs_info;
2168 struct btrfs_root *root = inode->root;
2169 struct btrfs_file_extent_item *extent;
2170 struct extent_buffer *leaf;
2171 struct btrfs_key key;
2172 int slot;
2173 int ret;
2174
2175 if (replace_len == 0)
2176 return 0;
2177
2178 if (extent_info->disk_offset == 0 &&
2179 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2180 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2181 return 0;
2182 }
2183
2184 key.objectid = btrfs_ino(inode);
2185 key.type = BTRFS_EXTENT_DATA_KEY;
2186 key.offset = extent_info->file_offset;
2187 ret = btrfs_insert_empty_item(trans, root, path, &key,
2188 sizeof(struct btrfs_file_extent_item));
2189 if (ret)
2190 return ret;
2191 leaf = path->nodes[0];
2192 slot = path->slots[0];
2193 write_extent_buffer(leaf, extent_info->extent_buf,
2194 btrfs_item_ptr_offset(leaf, slot),
2195 sizeof(struct btrfs_file_extent_item));
2196 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2197 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2198 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2199 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2200 if (extent_info->is_new_extent)
2201 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2202 btrfs_mark_buffer_dirty(trans, leaf);
2203 btrfs_release_path(path);
2204
2205 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2206 replace_len);
2207 if (ret)
2208 return ret;
2209
2210 /* If it's a hole, nothing more needs to be done. */
2211 if (extent_info->disk_offset == 0) {
2212 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2213 return 0;
2214 }
2215
2216 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2217
2218 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2219 key.objectid = extent_info->disk_offset;
2220 key.type = BTRFS_EXTENT_ITEM_KEY;
2221 key.offset = extent_info->disk_len;
2222 ret = btrfs_alloc_reserved_file_extent(trans, root,
2223 btrfs_ino(inode),
2224 extent_info->file_offset,
2225 extent_info->qgroup_reserved,
2226 &key);
2227 } else {
2228 struct btrfs_ref ref = {
2229 .action = BTRFS_ADD_DELAYED_REF,
2230 .bytenr = extent_info->disk_offset,
2231 .num_bytes = extent_info->disk_len,
2232 .owning_root = btrfs_root_id(root),
2233 .ref_root = btrfs_root_id(root),
2234 };
2235 u64 ref_offset;
2236
2237 ref_offset = extent_info->file_offset - extent_info->data_offset;
2238 btrfs_init_data_ref(&ref, btrfs_ino(inode), ref_offset, 0, false);
2239 ret = btrfs_inc_extent_ref(trans, &ref);
2240 }
2241
2242 extent_info->insertions++;
2243
2244 return ret;
2245}
2246
2247/*
2248 * The respective range must have been previously locked, as well as the inode.
2249 * The end offset is inclusive (last byte of the range).
2250 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2251 * the file range with an extent.
2252 * When not punching a hole, we don't want to end up in a state where we dropped
2253 * extents without inserting a new one, so we must abort the transaction to avoid
2254 * a corruption.
2255 */
2256int btrfs_replace_file_extents(struct btrfs_inode *inode,
2257 struct btrfs_path *path, const u64 start,
2258 const u64 end,
2259 struct btrfs_replace_extent_info *extent_info,
2260 struct btrfs_trans_handle **trans_out)
2261{
2262 struct btrfs_drop_extents_args drop_args = { 0 };
2263 struct btrfs_root *root = inode->root;
2264 struct btrfs_fs_info *fs_info = root->fs_info;
2265 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2266 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2267 struct btrfs_trans_handle *trans = NULL;
2268 struct btrfs_block_rsv *rsv;
2269 unsigned int rsv_count;
2270 u64 cur_offset;
2271 u64 len = end - start;
2272 int ret = 0;
2273
2274 if (end <= start)
2275 return -EINVAL;
2276
2277 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2278 if (!rsv) {
2279 ret = -ENOMEM;
2280 goto out;
2281 }
2282 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2283 rsv->failfast = true;
2284
2285 /*
2286 * 1 - update the inode
2287 * 1 - removing the extents in the range
2288 * 1 - adding the hole extent if no_holes isn't set or if we are
2289 * replacing the range with a new extent
2290 */
2291 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2292 rsv_count = 3;
2293 else
2294 rsv_count = 2;
2295
2296 trans = btrfs_start_transaction(root, rsv_count);
2297 if (IS_ERR(trans)) {
2298 ret = PTR_ERR(trans);
2299 trans = NULL;
2300 goto out_free;
2301 }
2302
2303 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2304 min_size, false);
2305 if (WARN_ON(ret))
2306 goto out_trans;
2307 trans->block_rsv = rsv;
2308
2309 cur_offset = start;
2310 drop_args.path = path;
2311 drop_args.end = end + 1;
2312 drop_args.drop_cache = true;
2313 while (cur_offset < end) {
2314 drop_args.start = cur_offset;
2315 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2316 /* If we are punching a hole decrement the inode's byte count */
2317 if (!extent_info)
2318 btrfs_update_inode_bytes(inode, 0,
2319 drop_args.bytes_found);
2320 if (ret != -ENOSPC) {
2321 /*
2322 * The only time we don't want to abort is if we are
2323 * attempting to clone a partial inline extent, in which
2324 * case we'll get EOPNOTSUPP. However if we aren't
2325 * clone we need to abort no matter what, because if we
2326 * got EOPNOTSUPP via prealloc then we messed up and
2327 * need to abort.
2328 */
2329 if (ret &&
2330 (ret != -EOPNOTSUPP ||
2331 (extent_info && extent_info->is_new_extent)))
2332 btrfs_abort_transaction(trans, ret);
2333 break;
2334 }
2335
2336 trans->block_rsv = &fs_info->trans_block_rsv;
2337
2338 if (!extent_info && cur_offset < drop_args.drop_end &&
2339 cur_offset < ino_size) {
2340 ret = fill_holes(trans, inode, path, cur_offset,
2341 drop_args.drop_end);
2342 if (ret) {
2343 /*
2344 * If we failed then we didn't insert our hole
2345 * entries for the area we dropped, so now the
2346 * fs is corrupted, so we must abort the
2347 * transaction.
2348 */
2349 btrfs_abort_transaction(trans, ret);
2350 break;
2351 }
2352 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2353 /*
2354 * We are past the i_size here, but since we didn't
2355 * insert holes we need to clear the mapped area so we
2356 * know to not set disk_i_size in this area until a new
2357 * file extent is inserted here.
2358 */
2359 ret = btrfs_inode_clear_file_extent_range(inode,
2360 cur_offset,
2361 drop_args.drop_end - cur_offset);
2362 if (ret) {
2363 /*
2364 * We couldn't clear our area, so we could
2365 * presumably adjust up and corrupt the fs, so
2366 * we need to abort.
2367 */
2368 btrfs_abort_transaction(trans, ret);
2369 break;
2370 }
2371 }
2372
2373 if (extent_info &&
2374 drop_args.drop_end > extent_info->file_offset) {
2375 u64 replace_len = drop_args.drop_end -
2376 extent_info->file_offset;
2377
2378 ret = btrfs_insert_replace_extent(trans, inode, path,
2379 extent_info, replace_len,
2380 drop_args.bytes_found);
2381 if (ret) {
2382 btrfs_abort_transaction(trans, ret);
2383 break;
2384 }
2385 extent_info->data_len -= replace_len;
2386 extent_info->data_offset += replace_len;
2387 extent_info->file_offset += replace_len;
2388 }
2389
2390 /*
2391 * We are releasing our handle on the transaction, balance the
2392 * dirty pages of the btree inode and flush delayed items, and
2393 * then get a new transaction handle, which may now point to a
2394 * new transaction in case someone else may have committed the
2395 * transaction we used to replace/drop file extent items. So
2396 * bump the inode's iversion and update mtime and ctime except
2397 * if we are called from a dedupe context. This is because a
2398 * power failure/crash may happen after the transaction is
2399 * committed and before we finish replacing/dropping all the
2400 * file extent items we need.
2401 */
2402 inode_inc_iversion(&inode->vfs_inode);
2403
2404 if (!extent_info || extent_info->update_times)
2405 inode_set_mtime_to_ts(&inode->vfs_inode,
2406 inode_set_ctime_current(&inode->vfs_inode));
2407
2408 ret = btrfs_update_inode(trans, inode);
2409 if (ret)
2410 break;
2411
2412 btrfs_end_transaction(trans);
2413 btrfs_btree_balance_dirty(fs_info);
2414
2415 trans = btrfs_start_transaction(root, rsv_count);
2416 if (IS_ERR(trans)) {
2417 ret = PTR_ERR(trans);
2418 trans = NULL;
2419 break;
2420 }
2421
2422 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2423 rsv, min_size, false);
2424 if (WARN_ON(ret))
2425 break;
2426 trans->block_rsv = rsv;
2427
2428 cur_offset = drop_args.drop_end;
2429 len = end - cur_offset;
2430 if (!extent_info && len) {
2431 ret = find_first_non_hole(inode, &cur_offset, &len);
2432 if (unlikely(ret < 0))
2433 break;
2434 if (ret && !len) {
2435 ret = 0;
2436 break;
2437 }
2438 }
2439 }
2440
2441 /*
2442 * If we were cloning, force the next fsync to be a full one since we
2443 * we replaced (or just dropped in the case of cloning holes when
2444 * NO_HOLES is enabled) file extent items and did not setup new extent
2445 * maps for the replacement extents (or holes).
2446 */
2447 if (extent_info && !extent_info->is_new_extent)
2448 btrfs_set_inode_full_sync(inode);
2449
2450 if (ret)
2451 goto out_trans;
2452
2453 trans->block_rsv = &fs_info->trans_block_rsv;
2454 /*
2455 * If we are using the NO_HOLES feature we might have had already an
2456 * hole that overlaps a part of the region [lockstart, lockend] and
2457 * ends at (or beyond) lockend. Since we have no file extent items to
2458 * represent holes, drop_end can be less than lockend and so we must
2459 * make sure we have an extent map representing the existing hole (the
2460 * call to __btrfs_drop_extents() might have dropped the existing extent
2461 * map representing the existing hole), otherwise the fast fsync path
2462 * will not record the existence of the hole region
2463 * [existing_hole_start, lockend].
2464 */
2465 if (drop_args.drop_end <= end)
2466 drop_args.drop_end = end + 1;
2467 /*
2468 * Don't insert file hole extent item if it's for a range beyond eof
2469 * (because it's useless) or if it represents a 0 bytes range (when
2470 * cur_offset == drop_end).
2471 */
2472 if (!extent_info && cur_offset < ino_size &&
2473 cur_offset < drop_args.drop_end) {
2474 ret = fill_holes(trans, inode, path, cur_offset,
2475 drop_args.drop_end);
2476 if (ret) {
2477 /* Same comment as above. */
2478 btrfs_abort_transaction(trans, ret);
2479 goto out_trans;
2480 }
2481 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2482 /* See the comment in the loop above for the reasoning here. */
2483 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2484 drop_args.drop_end - cur_offset);
2485 if (ret) {
2486 btrfs_abort_transaction(trans, ret);
2487 goto out_trans;
2488 }
2489
2490 }
2491 if (extent_info) {
2492 ret = btrfs_insert_replace_extent(trans, inode, path,
2493 extent_info, extent_info->data_len,
2494 drop_args.bytes_found);
2495 if (ret) {
2496 btrfs_abort_transaction(trans, ret);
2497 goto out_trans;
2498 }
2499 }
2500
2501out_trans:
2502 if (!trans)
2503 goto out_free;
2504
2505 trans->block_rsv = &fs_info->trans_block_rsv;
2506 if (ret)
2507 btrfs_end_transaction(trans);
2508 else
2509 *trans_out = trans;
2510out_free:
2511 btrfs_free_block_rsv(fs_info, rsv);
2512out:
2513 return ret;
2514}
2515
2516static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2517{
2518 struct inode *inode = file_inode(file);
2519 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2520 struct btrfs_root *root = BTRFS_I(inode)->root;
2521 struct extent_state *cached_state = NULL;
2522 struct btrfs_path *path;
2523 struct btrfs_trans_handle *trans = NULL;
2524 u64 lockstart;
2525 u64 lockend;
2526 u64 tail_start;
2527 u64 tail_len;
2528 u64 orig_start = offset;
2529 int ret = 0;
2530 bool same_block;
2531 u64 ino_size;
2532 bool truncated_block = false;
2533 bool updated_inode = false;
2534
2535 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2536
2537 ret = btrfs_wait_ordered_range(BTRFS_I(inode), offset, len);
2538 if (ret)
2539 goto out_only_mutex;
2540
2541 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2542 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2543 if (ret < 0)
2544 goto out_only_mutex;
2545 if (ret && !len) {
2546 /* Already in a large hole */
2547 ret = 0;
2548 goto out_only_mutex;
2549 }
2550
2551 ret = file_modified(file);
2552 if (ret)
2553 goto out_only_mutex;
2554
2555 lockstart = round_up(offset, fs_info->sectorsize);
2556 lockend = round_down(offset + len, fs_info->sectorsize) - 1;
2557 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2558 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2559 /*
2560 * We needn't truncate any block which is beyond the end of the file
2561 * because we are sure there is no data there.
2562 */
2563 /*
2564 * Only do this if we are in the same block and we aren't doing the
2565 * entire block.
2566 */
2567 if (same_block && len < fs_info->sectorsize) {
2568 if (offset < ino_size) {
2569 truncated_block = true;
2570 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2571 0);
2572 } else {
2573 ret = 0;
2574 }
2575 goto out_only_mutex;
2576 }
2577
2578 /* zero back part of the first block */
2579 if (offset < ino_size) {
2580 truncated_block = true;
2581 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2582 if (ret) {
2583 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2584 return ret;
2585 }
2586 }
2587
2588 /* Check the aligned pages after the first unaligned page,
2589 * if offset != orig_start, which means the first unaligned page
2590 * including several following pages are already in holes,
2591 * the extra check can be skipped */
2592 if (offset == orig_start) {
2593 /* after truncate page, check hole again */
2594 len = offset + len - lockstart;
2595 offset = lockstart;
2596 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2597 if (ret < 0)
2598 goto out_only_mutex;
2599 if (ret && !len) {
2600 ret = 0;
2601 goto out_only_mutex;
2602 }
2603 lockstart = offset;
2604 }
2605
2606 /* Check the tail unaligned part is in a hole */
2607 tail_start = lockend + 1;
2608 tail_len = offset + len - tail_start;
2609 if (tail_len) {
2610 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2611 if (unlikely(ret < 0))
2612 goto out_only_mutex;
2613 if (!ret) {
2614 /* zero the front end of the last page */
2615 if (tail_start + tail_len < ino_size) {
2616 truncated_block = true;
2617 ret = btrfs_truncate_block(BTRFS_I(inode),
2618 tail_start + tail_len,
2619 0, 1);
2620 if (ret)
2621 goto out_only_mutex;
2622 }
2623 }
2624 }
2625
2626 if (lockend < lockstart) {
2627 ret = 0;
2628 goto out_only_mutex;
2629 }
2630
2631 btrfs_punch_hole_lock_range(inode, lockstart, lockend, &cached_state);
2632
2633 path = btrfs_alloc_path();
2634 if (!path) {
2635 ret = -ENOMEM;
2636 goto out;
2637 }
2638
2639 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
2640 lockend, NULL, &trans);
2641 btrfs_free_path(path);
2642 if (ret)
2643 goto out;
2644
2645 ASSERT(trans != NULL);
2646 inode_inc_iversion(inode);
2647 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
2648 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2649 updated_inode = true;
2650 btrfs_end_transaction(trans);
2651 btrfs_btree_balance_dirty(fs_info);
2652out:
2653 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2654 &cached_state);
2655out_only_mutex:
2656 if (!updated_inode && truncated_block && !ret) {
2657 /*
2658 * If we only end up zeroing part of a page, we still need to
2659 * update the inode item, so that all the time fields are
2660 * updated as well as the necessary btrfs inode in memory fields
2661 * for detecting, at fsync time, if the inode isn't yet in the
2662 * log tree or it's there but not up to date.
2663 */
2664 struct timespec64 now = inode_set_ctime_current(inode);
2665
2666 inode_inc_iversion(inode);
2667 inode_set_mtime_to_ts(inode, now);
2668 trans = btrfs_start_transaction(root, 1);
2669 if (IS_ERR(trans)) {
2670 ret = PTR_ERR(trans);
2671 } else {
2672 int ret2;
2673
2674 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2675 ret2 = btrfs_end_transaction(trans);
2676 if (!ret)
2677 ret = ret2;
2678 }
2679 }
2680 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2681 return ret;
2682}
2683
2684/* Helper structure to record which range is already reserved */
2685struct falloc_range {
2686 struct list_head list;
2687 u64 start;
2688 u64 len;
2689};
2690
2691/*
2692 * Helper function to add falloc range
2693 *
2694 * Caller should have locked the larger range of extent containing
2695 * [start, len)
2696 */
2697static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2698{
2699 struct falloc_range *range = NULL;
2700
2701 if (!list_empty(head)) {
2702 /*
2703 * As fallocate iterates by bytenr order, we only need to check
2704 * the last range.
2705 */
2706 range = list_last_entry(head, struct falloc_range, list);
2707 if (range->start + range->len == start) {
2708 range->len += len;
2709 return 0;
2710 }
2711 }
2712
2713 range = kmalloc(sizeof(*range), GFP_KERNEL);
2714 if (!range)
2715 return -ENOMEM;
2716 range->start = start;
2717 range->len = len;
2718 list_add_tail(&range->list, head);
2719 return 0;
2720}
2721
2722static int btrfs_fallocate_update_isize(struct inode *inode,
2723 const u64 end,
2724 const int mode)
2725{
2726 struct btrfs_trans_handle *trans;
2727 struct btrfs_root *root = BTRFS_I(inode)->root;
2728 int ret;
2729 int ret2;
2730
2731 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2732 return 0;
2733
2734 trans = btrfs_start_transaction(root, 1);
2735 if (IS_ERR(trans))
2736 return PTR_ERR(trans);
2737
2738 inode_set_ctime_current(inode);
2739 i_size_write(inode, end);
2740 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
2741 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2742 ret2 = btrfs_end_transaction(trans);
2743
2744 return ret ? ret : ret2;
2745}
2746
2747enum {
2748 RANGE_BOUNDARY_WRITTEN_EXTENT,
2749 RANGE_BOUNDARY_PREALLOC_EXTENT,
2750 RANGE_BOUNDARY_HOLE,
2751};
2752
2753static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
2754 u64 offset)
2755{
2756 const u64 sectorsize = inode->root->fs_info->sectorsize;
2757 struct extent_map *em;
2758 int ret;
2759
2760 offset = round_down(offset, sectorsize);
2761 em = btrfs_get_extent(inode, NULL, offset, sectorsize);
2762 if (IS_ERR(em))
2763 return PTR_ERR(em);
2764
2765 if (em->disk_bytenr == EXTENT_MAP_HOLE)
2766 ret = RANGE_BOUNDARY_HOLE;
2767 else if (em->flags & EXTENT_FLAG_PREALLOC)
2768 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2769 else
2770 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2771
2772 free_extent_map(em);
2773 return ret;
2774}
2775
2776static int btrfs_zero_range(struct inode *inode,
2777 loff_t offset,
2778 loff_t len,
2779 const int mode)
2780{
2781 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2782 struct extent_map *em;
2783 struct extent_changeset *data_reserved = NULL;
2784 int ret;
2785 u64 alloc_hint = 0;
2786 const u64 sectorsize = fs_info->sectorsize;
2787 u64 alloc_start = round_down(offset, sectorsize);
2788 u64 alloc_end = round_up(offset + len, sectorsize);
2789 u64 bytes_to_reserve = 0;
2790 bool space_reserved = false;
2791
2792 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start,
2793 alloc_end - alloc_start);
2794 if (IS_ERR(em)) {
2795 ret = PTR_ERR(em);
2796 goto out;
2797 }
2798
2799 /*
2800 * Avoid hole punching and extent allocation for some cases. More cases
2801 * could be considered, but these are unlikely common and we keep things
2802 * as simple as possible for now. Also, intentionally, if the target
2803 * range contains one or more prealloc extents together with regular
2804 * extents and holes, we drop all the existing extents and allocate a
2805 * new prealloc extent, so that we get a larger contiguous disk extent.
2806 */
2807 if (em->start <= alloc_start && (em->flags & EXTENT_FLAG_PREALLOC)) {
2808 const u64 em_end = em->start + em->len;
2809
2810 if (em_end >= offset + len) {
2811 /*
2812 * The whole range is already a prealloc extent,
2813 * do nothing except updating the inode's i_size if
2814 * needed.
2815 */
2816 free_extent_map(em);
2817 ret = btrfs_fallocate_update_isize(inode, offset + len,
2818 mode);
2819 goto out;
2820 }
2821 /*
2822 * Part of the range is already a prealloc extent, so operate
2823 * only on the remaining part of the range.
2824 */
2825 alloc_start = em_end;
2826 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2827 len = offset + len - alloc_start;
2828 offset = alloc_start;
2829 alloc_hint = extent_map_block_start(em) + em->len;
2830 }
2831 free_extent_map(em);
2832
2833 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2834 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2835 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start, sectorsize);
2836 if (IS_ERR(em)) {
2837 ret = PTR_ERR(em);
2838 goto out;
2839 }
2840
2841 if (em->flags & EXTENT_FLAG_PREALLOC) {
2842 free_extent_map(em);
2843 ret = btrfs_fallocate_update_isize(inode, offset + len,
2844 mode);
2845 goto out;
2846 }
2847 if (len < sectorsize && em->disk_bytenr != EXTENT_MAP_HOLE) {
2848 free_extent_map(em);
2849 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2850 0);
2851 if (!ret)
2852 ret = btrfs_fallocate_update_isize(inode,
2853 offset + len,
2854 mode);
2855 return ret;
2856 }
2857 free_extent_map(em);
2858 alloc_start = round_down(offset, sectorsize);
2859 alloc_end = alloc_start + sectorsize;
2860 goto reserve_space;
2861 }
2862
2863 alloc_start = round_up(offset, sectorsize);
2864 alloc_end = round_down(offset + len, sectorsize);
2865
2866 /*
2867 * For unaligned ranges, check the pages at the boundaries, they might
2868 * map to an extent, in which case we need to partially zero them, or
2869 * they might map to a hole, in which case we need our allocation range
2870 * to cover them.
2871 */
2872 if (!IS_ALIGNED(offset, sectorsize)) {
2873 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2874 offset);
2875 if (ret < 0)
2876 goto out;
2877 if (ret == RANGE_BOUNDARY_HOLE) {
2878 alloc_start = round_down(offset, sectorsize);
2879 ret = 0;
2880 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2881 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2882 if (ret)
2883 goto out;
2884 } else {
2885 ret = 0;
2886 }
2887 }
2888
2889 if (!IS_ALIGNED(offset + len, sectorsize)) {
2890 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2891 offset + len);
2892 if (ret < 0)
2893 goto out;
2894 if (ret == RANGE_BOUNDARY_HOLE) {
2895 alloc_end = round_up(offset + len, sectorsize);
2896 ret = 0;
2897 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2898 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
2899 0, 1);
2900 if (ret)
2901 goto out;
2902 } else {
2903 ret = 0;
2904 }
2905 }
2906
2907reserve_space:
2908 if (alloc_start < alloc_end) {
2909 struct extent_state *cached_state = NULL;
2910 const u64 lockstart = alloc_start;
2911 const u64 lockend = alloc_end - 1;
2912
2913 bytes_to_reserve = alloc_end - alloc_start;
2914 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2915 bytes_to_reserve);
2916 if (ret < 0)
2917 goto out;
2918 space_reserved = true;
2919 btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2920 &cached_state);
2921 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
2922 alloc_start, bytes_to_reserve);
2923 if (ret) {
2924 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
2925 lockend, &cached_state);
2926 goto out;
2927 }
2928 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
2929 alloc_end - alloc_start,
2930 fs_info->sectorsize,
2931 offset + len, &alloc_hint);
2932 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2933 &cached_state);
2934 /* btrfs_prealloc_file_range releases reserved space on error */
2935 if (ret) {
2936 space_reserved = false;
2937 goto out;
2938 }
2939 }
2940 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
2941 out:
2942 if (ret && space_reserved)
2943 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
2944 alloc_start, bytes_to_reserve);
2945 extent_changeset_free(data_reserved);
2946
2947 return ret;
2948}
2949
2950static long btrfs_fallocate(struct file *file, int mode,
2951 loff_t offset, loff_t len)
2952{
2953 struct inode *inode = file_inode(file);
2954 struct extent_state *cached_state = NULL;
2955 struct extent_changeset *data_reserved = NULL;
2956 struct falloc_range *range;
2957 struct falloc_range *tmp;
2958 LIST_HEAD(reserve_list);
2959 u64 cur_offset;
2960 u64 last_byte;
2961 u64 alloc_start;
2962 u64 alloc_end;
2963 u64 alloc_hint = 0;
2964 u64 locked_end;
2965 u64 actual_end = 0;
2966 u64 data_space_needed = 0;
2967 u64 data_space_reserved = 0;
2968 u64 qgroup_reserved = 0;
2969 struct extent_map *em;
2970 int blocksize = BTRFS_I(inode)->root->fs_info->sectorsize;
2971 int ret;
2972
2973 /* Do not allow fallocate in ZONED mode */
2974 if (btrfs_is_zoned(inode_to_fs_info(inode)))
2975 return -EOPNOTSUPP;
2976
2977 alloc_start = round_down(offset, blocksize);
2978 alloc_end = round_up(offset + len, blocksize);
2979 cur_offset = alloc_start;
2980
2981 /* Make sure we aren't being give some crap mode */
2982 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
2983 FALLOC_FL_ZERO_RANGE))
2984 return -EOPNOTSUPP;
2985
2986 if (mode & FALLOC_FL_PUNCH_HOLE)
2987 return btrfs_punch_hole(file, offset, len);
2988
2989 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2990
2991 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2992 ret = inode_newsize_ok(inode, offset + len);
2993 if (ret)
2994 goto out;
2995 }
2996
2997 ret = file_modified(file);
2998 if (ret)
2999 goto out;
3000
3001 /*
3002 * TODO: Move these two operations after we have checked
3003 * accurate reserved space, or fallocate can still fail but
3004 * with page truncated or size expanded.
3005 *
3006 * But that's a minor problem and won't do much harm BTW.
3007 */
3008 if (alloc_start > inode->i_size) {
3009 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3010 alloc_start);
3011 if (ret)
3012 goto out;
3013 } else if (offset + len > inode->i_size) {
3014 /*
3015 * If we are fallocating from the end of the file onward we
3016 * need to zero out the end of the block if i_size lands in the
3017 * middle of a block.
3018 */
3019 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3020 if (ret)
3021 goto out;
3022 }
3023
3024 /*
3025 * We have locked the inode at the VFS level (in exclusive mode) and we
3026 * have locked the i_mmap_lock lock (in exclusive mode). Now before
3027 * locking the file range, flush all dealloc in the range and wait for
3028 * all ordered extents in the range to complete. After this we can lock
3029 * the file range and, due to the previous locking we did, we know there
3030 * can't be more delalloc or ordered extents in the range.
3031 */
3032 ret = btrfs_wait_ordered_range(BTRFS_I(inode), alloc_start,
3033 alloc_end - alloc_start);
3034 if (ret)
3035 goto out;
3036
3037 if (mode & FALLOC_FL_ZERO_RANGE) {
3038 ret = btrfs_zero_range(inode, offset, len, mode);
3039 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3040 return ret;
3041 }
3042
3043 locked_end = alloc_end - 1;
3044 lock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3045 &cached_state);
3046
3047 btrfs_assert_inode_range_clean(BTRFS_I(inode), alloc_start, locked_end);
3048
3049 /* First, check if we exceed the qgroup limit */
3050 while (cur_offset < alloc_end) {
3051 em = btrfs_get_extent(BTRFS_I(inode), NULL, cur_offset,
3052 alloc_end - cur_offset);
3053 if (IS_ERR(em)) {
3054 ret = PTR_ERR(em);
3055 break;
3056 }
3057 last_byte = min(extent_map_end(em), alloc_end);
3058 actual_end = min_t(u64, extent_map_end(em), offset + len);
3059 last_byte = ALIGN(last_byte, blocksize);
3060 if (em->disk_bytenr == EXTENT_MAP_HOLE ||
3061 (cur_offset >= inode->i_size &&
3062 !(em->flags & EXTENT_FLAG_PREALLOC))) {
3063 const u64 range_len = last_byte - cur_offset;
3064
3065 ret = add_falloc_range(&reserve_list, cur_offset, range_len);
3066 if (ret < 0) {
3067 free_extent_map(em);
3068 break;
3069 }
3070 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3071 &data_reserved, cur_offset, range_len);
3072 if (ret < 0) {
3073 free_extent_map(em);
3074 break;
3075 }
3076 qgroup_reserved += range_len;
3077 data_space_needed += range_len;
3078 }
3079 free_extent_map(em);
3080 cur_offset = last_byte;
3081 }
3082
3083 if (!ret && data_space_needed > 0) {
3084 /*
3085 * We are safe to reserve space here as we can't have delalloc
3086 * in the range, see above.
3087 */
3088 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3089 data_space_needed);
3090 if (!ret)
3091 data_space_reserved = data_space_needed;
3092 }
3093
3094 /*
3095 * If ret is still 0, means we're OK to fallocate.
3096 * Or just cleanup the list and exit.
3097 */
3098 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3099 if (!ret) {
3100 ret = btrfs_prealloc_file_range(inode, mode,
3101 range->start,
3102 range->len, blocksize,
3103 offset + len, &alloc_hint);
3104 /*
3105 * btrfs_prealloc_file_range() releases space even
3106 * if it returns an error.
3107 */
3108 data_space_reserved -= range->len;
3109 qgroup_reserved -= range->len;
3110 } else if (data_space_reserved > 0) {
3111 btrfs_free_reserved_data_space(BTRFS_I(inode),
3112 data_reserved, range->start,
3113 range->len);
3114 data_space_reserved -= range->len;
3115 qgroup_reserved -= range->len;
3116 } else if (qgroup_reserved > 0) {
3117 btrfs_qgroup_free_data(BTRFS_I(inode), data_reserved,
3118 range->start, range->len, NULL);
3119 qgroup_reserved -= range->len;
3120 }
3121 list_del(&range->list);
3122 kfree(range);
3123 }
3124 if (ret < 0)
3125 goto out_unlock;
3126
3127 /*
3128 * We didn't need to allocate any more space, but we still extended the
3129 * size of the file so we need to update i_size and the inode item.
3130 */
3131 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3132out_unlock:
3133 unlock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3134 &cached_state);
3135out:
3136 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3137 extent_changeset_free(data_reserved);
3138 return ret;
3139}
3140
3141/*
3142 * Helper for btrfs_find_delalloc_in_range(). Find a subrange in a given range
3143 * that has unflushed and/or flushing delalloc. There might be other adjacent
3144 * subranges after the one it found, so btrfs_find_delalloc_in_range() keeps
3145 * looping while it gets adjacent subranges, and merging them together.
3146 */
3147static bool find_delalloc_subrange(struct btrfs_inode *inode, u64 start, u64 end,
3148 struct extent_state **cached_state,
3149 bool *search_io_tree,
3150 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3151{
3152 u64 len = end + 1 - start;
3153 u64 delalloc_len = 0;
3154 struct btrfs_ordered_extent *oe;
3155 u64 oe_start;
3156 u64 oe_end;
3157
3158 /*
3159 * Search the io tree first for EXTENT_DELALLOC. If we find any, it
3160 * means we have delalloc (dirty pages) for which writeback has not
3161 * started yet.
3162 */
3163 if (*search_io_tree) {
3164 spin_lock(&inode->lock);
3165 if (inode->delalloc_bytes > 0) {
3166 spin_unlock(&inode->lock);
3167 *delalloc_start_ret = start;
3168 delalloc_len = count_range_bits(&inode->io_tree,
3169 delalloc_start_ret, end,
3170 len, EXTENT_DELALLOC, 1,
3171 cached_state);
3172 } else {
3173 spin_unlock(&inode->lock);
3174 }
3175 }
3176
3177 if (delalloc_len > 0) {
3178 /*
3179 * If delalloc was found then *delalloc_start_ret has a sector size
3180 * aligned value (rounded down).
3181 */
3182 *delalloc_end_ret = *delalloc_start_ret + delalloc_len - 1;
3183
3184 if (*delalloc_start_ret == start) {
3185 /* Delalloc for the whole range, nothing more to do. */
3186 if (*delalloc_end_ret == end)
3187 return true;
3188 /* Else trim our search range for ordered extents. */
3189 start = *delalloc_end_ret + 1;
3190 len = end + 1 - start;
3191 }
3192 } else {
3193 /* No delalloc, future calls don't need to search again. */
3194 *search_io_tree = false;
3195 }
3196
3197 /*
3198 * Now also check if there's any ordered extent in the range.
3199 * We do this because:
3200 *
3201 * 1) When delalloc is flushed, the file range is locked, we clear the
3202 * EXTENT_DELALLOC bit from the io tree and create an extent map and
3203 * an ordered extent for the write. So we might just have been called
3204 * after delalloc is flushed and before the ordered extent completes
3205 * and inserts the new file extent item in the subvolume's btree;
3206 *
3207 * 2) We may have an ordered extent created by flushing delalloc for a
3208 * subrange that starts before the subrange we found marked with
3209 * EXTENT_DELALLOC in the io tree.
3210 *
3211 * We could also use the extent map tree to find such delalloc that is
3212 * being flushed, but using the ordered extents tree is more efficient
3213 * because it's usually much smaller as ordered extents are removed from
3214 * the tree once they complete. With the extent maps, we mau have them
3215 * in the extent map tree for a very long time, and they were either
3216 * created by previous writes or loaded by read operations.
3217 */
3218 oe = btrfs_lookup_first_ordered_range(inode, start, len);
3219 if (!oe)
3220 return (delalloc_len > 0);
3221
3222 /* The ordered extent may span beyond our search range. */
3223 oe_start = max(oe->file_offset, start);
3224 oe_end = min(oe->file_offset + oe->num_bytes - 1, end);
3225
3226 btrfs_put_ordered_extent(oe);
3227
3228 /* Don't have unflushed delalloc, return the ordered extent range. */
3229 if (delalloc_len == 0) {
3230 *delalloc_start_ret = oe_start;
3231 *delalloc_end_ret = oe_end;
3232 return true;
3233 }
3234
3235 /*
3236 * We have both unflushed delalloc (io_tree) and an ordered extent.
3237 * If the ranges are adjacent returned a combined range, otherwise
3238 * return the leftmost range.
3239 */
3240 if (oe_start < *delalloc_start_ret) {
3241 if (oe_end < *delalloc_start_ret)
3242 *delalloc_end_ret = oe_end;
3243 *delalloc_start_ret = oe_start;
3244 } else if (*delalloc_end_ret + 1 == oe_start) {
3245 *delalloc_end_ret = oe_end;
3246 }
3247
3248 return true;
3249}
3250
3251/*
3252 * Check if there's delalloc in a given range.
3253 *
3254 * @inode: The inode.
3255 * @start: The start offset of the range. It does not need to be
3256 * sector size aligned.
3257 * @end: The end offset (inclusive value) of the search range.
3258 * It does not need to be sector size aligned.
3259 * @cached_state: Extent state record used for speeding up delalloc
3260 * searches in the inode's io_tree. Can be NULL.
3261 * @delalloc_start_ret: Output argument, set to the start offset of the
3262 * subrange found with delalloc (may not be sector size
3263 * aligned).
3264 * @delalloc_end_ret: Output argument, set to he end offset (inclusive value)
3265 * of the subrange found with delalloc.
3266 *
3267 * Returns true if a subrange with delalloc is found within the given range, and
3268 * if so it sets @delalloc_start_ret and @delalloc_end_ret with the start and
3269 * end offsets of the subrange.
3270 */
3271bool btrfs_find_delalloc_in_range(struct btrfs_inode *inode, u64 start, u64 end,
3272 struct extent_state **cached_state,
3273 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3274{
3275 u64 cur_offset = round_down(start, inode->root->fs_info->sectorsize);
3276 u64 prev_delalloc_end = 0;
3277 bool search_io_tree = true;
3278 bool ret = false;
3279
3280 while (cur_offset <= end) {
3281 u64 delalloc_start;
3282 u64 delalloc_end;
3283 bool delalloc;
3284
3285 delalloc = find_delalloc_subrange(inode, cur_offset, end,
3286 cached_state, &search_io_tree,
3287 &delalloc_start,
3288 &delalloc_end);
3289 if (!delalloc)
3290 break;
3291
3292 if (prev_delalloc_end == 0) {
3293 /* First subrange found. */
3294 *delalloc_start_ret = max(delalloc_start, start);
3295 *delalloc_end_ret = delalloc_end;
3296 ret = true;
3297 } else if (delalloc_start == prev_delalloc_end + 1) {
3298 /* Subrange adjacent to the previous one, merge them. */
3299 *delalloc_end_ret = delalloc_end;
3300 } else {
3301 /* Subrange not adjacent to the previous one, exit. */
3302 break;
3303 }
3304
3305 prev_delalloc_end = delalloc_end;
3306 cur_offset = delalloc_end + 1;
3307 cond_resched();
3308 }
3309
3310 return ret;
3311}
3312
3313/*
3314 * Check if there's a hole or delalloc range in a range representing a hole (or
3315 * prealloc extent) found in the inode's subvolume btree.
3316 *
3317 * @inode: The inode.
3318 * @whence: Seek mode (SEEK_DATA or SEEK_HOLE).
3319 * @start: Start offset of the hole region. It does not need to be sector
3320 * size aligned.
3321 * @end: End offset (inclusive value) of the hole region. It does not
3322 * need to be sector size aligned.
3323 * @start_ret: Return parameter, used to set the start of the subrange in the
3324 * hole that matches the search criteria (seek mode), if such
3325 * subrange is found (return value of the function is true).
3326 * The value returned here may not be sector size aligned.
3327 *
3328 * Returns true if a subrange matching the given seek mode is found, and if one
3329 * is found, it updates @start_ret with the start of the subrange.
3330 */
3331static bool find_desired_extent_in_hole(struct btrfs_inode *inode, int whence,
3332 struct extent_state **cached_state,
3333 u64 start, u64 end, u64 *start_ret)
3334{
3335 u64 delalloc_start;
3336 u64 delalloc_end;
3337 bool delalloc;
3338
3339 delalloc = btrfs_find_delalloc_in_range(inode, start, end, cached_state,
3340 &delalloc_start, &delalloc_end);
3341 if (delalloc && whence == SEEK_DATA) {
3342 *start_ret = delalloc_start;
3343 return true;
3344 }
3345
3346 if (delalloc && whence == SEEK_HOLE) {
3347 /*
3348 * We found delalloc but it starts after out start offset. So we
3349 * have a hole between our start offset and the delalloc start.
3350 */
3351 if (start < delalloc_start) {
3352 *start_ret = start;
3353 return true;
3354 }
3355 /*
3356 * Delalloc range starts at our start offset.
3357 * If the delalloc range's length is smaller than our range,
3358 * then it means we have a hole that starts where the delalloc
3359 * subrange ends.
3360 */
3361 if (delalloc_end < end) {
3362 *start_ret = delalloc_end + 1;
3363 return true;
3364 }
3365
3366 /* There's delalloc for the whole range. */
3367 return false;
3368 }
3369
3370 if (!delalloc && whence == SEEK_HOLE) {
3371 *start_ret = start;
3372 return true;
3373 }
3374
3375 /*
3376 * No delalloc in the range and we are seeking for data. The caller has
3377 * to iterate to the next extent item in the subvolume btree.
3378 */
3379 return false;
3380}
3381
3382static loff_t find_desired_extent(struct file *file, loff_t offset, int whence)
3383{
3384 struct btrfs_inode *inode = BTRFS_I(file->f_mapping->host);
3385 struct btrfs_file_private *private;
3386 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3387 struct extent_state *cached_state = NULL;
3388 struct extent_state **delalloc_cached_state;
3389 const loff_t i_size = i_size_read(&inode->vfs_inode);
3390 const u64 ino = btrfs_ino(inode);
3391 struct btrfs_root *root = inode->root;
3392 struct btrfs_path *path;
3393 struct btrfs_key key;
3394 u64 last_extent_end;
3395 u64 lockstart;
3396 u64 lockend;
3397 u64 start;
3398 int ret;
3399 bool found = false;
3400
3401 if (i_size == 0 || offset >= i_size)
3402 return -ENXIO;
3403
3404 /*
3405 * Quick path. If the inode has no prealloc extents and its number of
3406 * bytes used matches its i_size, then it can not have holes.
3407 */
3408 if (whence == SEEK_HOLE &&
3409 !(inode->flags & BTRFS_INODE_PREALLOC) &&
3410 inode_get_bytes(&inode->vfs_inode) == i_size)
3411 return i_size;
3412
3413 spin_lock(&inode->lock);
3414 private = file->private_data;
3415 spin_unlock(&inode->lock);
3416
3417 if (private && private->owner_task != current) {
3418 /*
3419 * Not allocated by us, don't use it as its cached state is used
3420 * by the task that allocated it and we don't want neither to
3421 * mess with it nor get incorrect results because it reflects an
3422 * invalid state for the current task.
3423 */
3424 private = NULL;
3425 } else if (!private) {
3426 private = kzalloc(sizeof(*private), GFP_KERNEL);
3427 /*
3428 * No worries if memory allocation failed.
3429 * The private structure is used only for speeding up multiple
3430 * lseek SEEK_HOLE/DATA calls to a file when there's delalloc,
3431 * so everything will still be correct.
3432 */
3433 if (private) {
3434 bool free = false;
3435
3436 private->owner_task = current;
3437
3438 spin_lock(&inode->lock);
3439 if (file->private_data)
3440 free = true;
3441 else
3442 file->private_data = private;
3443 spin_unlock(&inode->lock);
3444
3445 if (free) {
3446 kfree(private);
3447 private = NULL;
3448 }
3449 }
3450 }
3451
3452 if (private)
3453 delalloc_cached_state = &private->llseek_cached_state;
3454 else
3455 delalloc_cached_state = NULL;
3456
3457 /*
3458 * offset can be negative, in this case we start finding DATA/HOLE from
3459 * the very start of the file.
3460 */
3461 start = max_t(loff_t, 0, offset);
3462
3463 lockstart = round_down(start, fs_info->sectorsize);
3464 lockend = round_up(i_size, fs_info->sectorsize);
3465 if (lockend <= lockstart)
3466 lockend = lockstart + fs_info->sectorsize;
3467 lockend--;
3468
3469 path = btrfs_alloc_path();
3470 if (!path)
3471 return -ENOMEM;
3472 path->reada = READA_FORWARD;
3473
3474 key.objectid = ino;
3475 key.type = BTRFS_EXTENT_DATA_KEY;
3476 key.offset = start;
3477
3478 last_extent_end = lockstart;
3479
3480 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3481
3482 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3483 if (ret < 0) {
3484 goto out;
3485 } else if (ret > 0 && path->slots[0] > 0) {
3486 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
3487 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
3488 path->slots[0]--;
3489 }
3490
3491 while (start < i_size) {
3492 struct extent_buffer *leaf = path->nodes[0];
3493 struct btrfs_file_extent_item *extent;
3494 u64 extent_end;
3495 u8 type;
3496
3497 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3498 ret = btrfs_next_leaf(root, path);
3499 if (ret < 0)
3500 goto out;
3501 else if (ret > 0)
3502 break;
3503
3504 leaf = path->nodes[0];
3505 }
3506
3507 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3508 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3509 break;
3510
3511 extent_end = btrfs_file_extent_end(path);
3512
3513 /*
3514 * In the first iteration we may have a slot that points to an
3515 * extent that ends before our start offset, so skip it.
3516 */
3517 if (extent_end <= start) {
3518 path->slots[0]++;
3519 continue;
3520 }
3521
3522 /* We have an implicit hole, NO_HOLES feature is likely set. */
3523 if (last_extent_end < key.offset) {
3524 u64 search_start = last_extent_end;
3525 u64 found_start;
3526
3527 /*
3528 * First iteration, @start matches @offset and it's
3529 * within the hole.
3530 */
3531 if (start == offset)
3532 search_start = offset;
3533
3534 found = find_desired_extent_in_hole(inode, whence,
3535 delalloc_cached_state,
3536 search_start,
3537 key.offset - 1,
3538 &found_start);
3539 if (found) {
3540 start = found_start;
3541 break;
3542 }
3543 /*
3544 * Didn't find data or a hole (due to delalloc) in the
3545 * implicit hole range, so need to analyze the extent.
3546 */
3547 }
3548
3549 extent = btrfs_item_ptr(leaf, path->slots[0],
3550 struct btrfs_file_extent_item);
3551 type = btrfs_file_extent_type(leaf, extent);
3552
3553 /*
3554 * Can't access the extent's disk_bytenr field if this is an
3555 * inline extent, since at that offset, it's where the extent
3556 * data starts.
3557 */
3558 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
3559 (type == BTRFS_FILE_EXTENT_REG &&
3560 btrfs_file_extent_disk_bytenr(leaf, extent) == 0)) {
3561 /*
3562 * Explicit hole or prealloc extent, search for delalloc.
3563 * A prealloc extent is treated like a hole.
3564 */
3565 u64 search_start = key.offset;
3566 u64 found_start;
3567
3568 /*
3569 * First iteration, @start matches @offset and it's
3570 * within the hole.
3571 */
3572 if (start == offset)
3573 search_start = offset;
3574
3575 found = find_desired_extent_in_hole(inode, whence,
3576 delalloc_cached_state,
3577 search_start,
3578 extent_end - 1,
3579 &found_start);
3580 if (found) {
3581 start = found_start;
3582 break;
3583 }
3584 /*
3585 * Didn't find data or a hole (due to delalloc) in the
3586 * implicit hole range, so need to analyze the next
3587 * extent item.
3588 */
3589 } else {
3590 /*
3591 * Found a regular or inline extent.
3592 * If we are seeking for data, adjust the start offset
3593 * and stop, we're done.
3594 */
3595 if (whence == SEEK_DATA) {
3596 start = max_t(u64, key.offset, offset);
3597 found = true;
3598 break;
3599 }
3600 /*
3601 * Else, we are seeking for a hole, check the next file
3602 * extent item.
3603 */
3604 }
3605
3606 start = extent_end;
3607 last_extent_end = extent_end;
3608 path->slots[0]++;
3609 if (fatal_signal_pending(current)) {
3610 ret = -EINTR;
3611 goto out;
3612 }
3613 cond_resched();
3614 }
3615
3616 /* We have an implicit hole from the last extent found up to i_size. */
3617 if (!found && start < i_size) {
3618 found = find_desired_extent_in_hole(inode, whence,
3619 delalloc_cached_state, start,
3620 i_size - 1, &start);
3621 if (!found)
3622 start = i_size;
3623 }
3624
3625out:
3626 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3627 btrfs_free_path(path);
3628
3629 if (ret < 0)
3630 return ret;
3631
3632 if (whence == SEEK_DATA && start >= i_size)
3633 return -ENXIO;
3634
3635 return min_t(loff_t, start, i_size);
3636}
3637
3638static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3639{
3640 struct inode *inode = file->f_mapping->host;
3641
3642 switch (whence) {
3643 default:
3644 return generic_file_llseek(file, offset, whence);
3645 case SEEK_DATA:
3646 case SEEK_HOLE:
3647 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3648 offset = find_desired_extent(file, offset, whence);
3649 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3650 break;
3651 }
3652
3653 if (offset < 0)
3654 return offset;
3655
3656 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3657}
3658
3659static int btrfs_file_open(struct inode *inode, struct file *filp)
3660{
3661 int ret;
3662
3663 filp->f_mode |= FMODE_NOWAIT | FMODE_CAN_ODIRECT;
3664
3665 ret = fsverity_file_open(inode, filp);
3666 if (ret)
3667 return ret;
3668 return generic_file_open(inode, filp);
3669}
3670
3671static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3672{
3673 ssize_t ret = 0;
3674
3675 if (iocb->ki_flags & IOCB_DIRECT) {
3676 ret = btrfs_direct_read(iocb, to);
3677 if (ret < 0 || !iov_iter_count(to) ||
3678 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3679 return ret;
3680 }
3681
3682 return filemap_read(iocb, to, ret);
3683}
3684
3685const struct file_operations btrfs_file_operations = {
3686 .llseek = btrfs_file_llseek,
3687 .read_iter = btrfs_file_read_iter,
3688 .splice_read = filemap_splice_read,
3689 .write_iter = btrfs_file_write_iter,
3690 .splice_write = iter_file_splice_write,
3691 .mmap = btrfs_file_mmap,
3692 .open = btrfs_file_open,
3693 .release = btrfs_release_file,
3694 .get_unmapped_area = thp_get_unmapped_area,
3695 .fsync = btrfs_sync_file,
3696 .fallocate = btrfs_fallocate,
3697 .unlocked_ioctl = btrfs_ioctl,
3698#ifdef CONFIG_COMPAT
3699 .compat_ioctl = btrfs_compat_ioctl,
3700#endif
3701 .remap_file_range = btrfs_remap_file_range,
3702 .uring_cmd = btrfs_uring_cmd,
3703 .fop_flags = FOP_BUFFER_RASYNC | FOP_BUFFER_WASYNC,
3704};
3705
3706int btrfs_fdatawrite_range(struct btrfs_inode *inode, loff_t start, loff_t end)
3707{
3708 struct address_space *mapping = inode->vfs_inode.i_mapping;
3709 int ret;
3710
3711 /*
3712 * So with compression we will find and lock a dirty page and clear the
3713 * first one as dirty, setup an async extent, and immediately return
3714 * with the entire range locked but with nobody actually marked with
3715 * writeback. So we can't just filemap_write_and_wait_range() and
3716 * expect it to work since it will just kick off a thread to do the
3717 * actual work. So we need to call filemap_fdatawrite_range _again_
3718 * since it will wait on the page lock, which won't be unlocked until
3719 * after the pages have been marked as writeback and so we're good to go
3720 * from there. We have to do this otherwise we'll miss the ordered
3721 * extents and that results in badness. Please Josef, do not think you
3722 * know better and pull this out at some point in the future, it is
3723 * right and you are wrong.
3724 */
3725 ret = filemap_fdatawrite_range(mapping, start, end);
3726 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT, &inode->runtime_flags))
3727 ret = filemap_fdatawrite_range(mapping, start, end);
3728
3729 return ret;
3730}
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 <linux/iomap.h>
21#include "ctree.h"
22#include "disk-io.h"
23#include "transaction.h"
24#include "btrfs_inode.h"
25#include "tree-log.h"
26#include "locking.h"
27#include "qgroup.h"
28#include "compression.h"
29#include "delalloc-space.h"
30#include "reflink.h"
31#include "subpage.h"
32#include "fs.h"
33#include "accessors.h"
34#include "extent-tree.h"
35#include "file-item.h"
36#include "ioctl.h"
37#include "file.h"
38#include "super.h"
39
40/* simple helper to fault in pages and copy. This should go away
41 * and be replaced with calls into generic code.
42 */
43static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
44 struct page **prepared_pages,
45 struct iov_iter *i)
46{
47 size_t copied = 0;
48 size_t total_copied = 0;
49 int pg = 0;
50 int offset = offset_in_page(pos);
51
52 while (write_bytes > 0) {
53 size_t count = min_t(size_t,
54 PAGE_SIZE - offset, write_bytes);
55 struct page *page = prepared_pages[pg];
56 /*
57 * Copy data from userspace to the current page
58 */
59 copied = copy_page_from_iter_atomic(page, offset, count, i);
60
61 /* Flush processor's dcache for this page */
62 flush_dcache_page(page);
63
64 /*
65 * if we get a partial write, we can end up with
66 * partially up to date pages. These add
67 * a lot of complexity, so make sure they don't
68 * happen by forcing this copy to be retried.
69 *
70 * The rest of the btrfs_file_write code will fall
71 * back to page at a time copies after we return 0.
72 */
73 if (unlikely(copied < count)) {
74 if (!PageUptodate(page)) {
75 iov_iter_revert(i, copied);
76 copied = 0;
77 }
78 if (!copied)
79 break;
80 }
81
82 write_bytes -= copied;
83 total_copied += copied;
84 offset += copied;
85 if (offset == PAGE_SIZE) {
86 pg++;
87 offset = 0;
88 }
89 }
90 return total_copied;
91}
92
93/*
94 * unlocks pages after btrfs_file_write is done with them
95 */
96static void btrfs_drop_pages(struct btrfs_fs_info *fs_info,
97 struct page **pages, size_t num_pages,
98 u64 pos, u64 copied)
99{
100 size_t i;
101 u64 block_start = round_down(pos, fs_info->sectorsize);
102 u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
103
104 ASSERT(block_len <= U32_MAX);
105 for (i = 0; i < num_pages; i++) {
106 /* page checked is some magic around finding pages that
107 * have been modified without going through btrfs_set_page_dirty
108 * clear it here. There should be no need to mark the pages
109 * accessed as prepare_pages should have marked them accessed
110 * in prepare_pages via find_or_create_page()
111 */
112 btrfs_folio_clamp_clear_checked(fs_info, page_folio(pages[i]),
113 block_start, block_len);
114 unlock_page(pages[i]);
115 put_page(pages[i]);
116 }
117}
118
119/*
120 * After btrfs_copy_from_user(), update the following things for delalloc:
121 * - Mark newly dirtied pages as DELALLOC in the io tree.
122 * Used to advise which range is to be written back.
123 * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
124 * - Update inode size for past EOF write
125 */
126int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
127 size_t num_pages, loff_t pos, size_t write_bytes,
128 struct extent_state **cached, bool noreserve)
129{
130 struct btrfs_fs_info *fs_info = inode->root->fs_info;
131 int err = 0;
132 int i;
133 u64 num_bytes;
134 u64 start_pos;
135 u64 end_of_last_block;
136 u64 end_pos = pos + write_bytes;
137 loff_t isize = i_size_read(&inode->vfs_inode);
138 unsigned int extra_bits = 0;
139
140 if (write_bytes == 0)
141 return 0;
142
143 if (noreserve)
144 extra_bits |= EXTENT_NORESERVE;
145
146 start_pos = round_down(pos, fs_info->sectorsize);
147 num_bytes = round_up(write_bytes + pos - start_pos,
148 fs_info->sectorsize);
149 ASSERT(num_bytes <= U32_MAX);
150
151 end_of_last_block = start_pos + num_bytes - 1;
152
153 /*
154 * The pages may have already been dirty, clear out old accounting so
155 * we can set things up properly
156 */
157 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
158 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
159 cached);
160
161 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
162 extra_bits, cached);
163 if (err)
164 return err;
165
166 for (i = 0; i < num_pages; i++) {
167 struct page *p = pages[i];
168
169 btrfs_folio_clamp_set_uptodate(fs_info, page_folio(p),
170 start_pos, num_bytes);
171 btrfs_folio_clamp_clear_checked(fs_info, page_folio(p),
172 start_pos, num_bytes);
173 btrfs_folio_clamp_set_dirty(fs_info, page_folio(p),
174 start_pos, num_bytes);
175 }
176
177 /*
178 * we've only changed i_size in ram, and we haven't updated
179 * the disk i_size. There is no need to log the inode
180 * at this time.
181 */
182 if (end_pos > isize)
183 i_size_write(&inode->vfs_inode, end_pos);
184 return 0;
185}
186
187/*
188 * this is very complex, but the basic idea is to drop all extents
189 * in the range start - end. hint_block is filled in with a block number
190 * that would be a good hint to the block allocator for this file.
191 *
192 * If an extent intersects the range but is not entirely inside the range
193 * it is either truncated or split. Anything entirely inside the range
194 * is deleted from the tree.
195 *
196 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
197 * to deal with that. We set the field 'bytes_found' of the arguments structure
198 * with the number of allocated bytes found in the target range, so that the
199 * caller can update the inode's number of bytes in an atomic way when
200 * replacing extents in a range to avoid races with stat(2).
201 */
202int btrfs_drop_extents(struct btrfs_trans_handle *trans,
203 struct btrfs_root *root, struct btrfs_inode *inode,
204 struct btrfs_drop_extents_args *args)
205{
206 struct btrfs_fs_info *fs_info = root->fs_info;
207 struct extent_buffer *leaf;
208 struct btrfs_file_extent_item *fi;
209 struct btrfs_ref ref = { 0 };
210 struct btrfs_key key;
211 struct btrfs_key new_key;
212 u64 ino = btrfs_ino(inode);
213 u64 search_start = args->start;
214 u64 disk_bytenr = 0;
215 u64 num_bytes = 0;
216 u64 extent_offset = 0;
217 u64 extent_end = 0;
218 u64 last_end = args->start;
219 int del_nr = 0;
220 int del_slot = 0;
221 int extent_type;
222 int recow;
223 int ret;
224 int modify_tree = -1;
225 int update_refs;
226 int found = 0;
227 struct btrfs_path *path = args->path;
228
229 args->bytes_found = 0;
230 args->extent_inserted = false;
231
232 /* Must always have a path if ->replace_extent is true */
233 ASSERT(!(args->replace_extent && !args->path));
234
235 if (!path) {
236 path = btrfs_alloc_path();
237 if (!path) {
238 ret = -ENOMEM;
239 goto out;
240 }
241 }
242
243 if (args->drop_cache)
244 btrfs_drop_extent_map_range(inode, args->start, args->end - 1, false);
245
246 if (args->start >= inode->disk_i_size && !args->replace_extent)
247 modify_tree = 0;
248
249 update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
250 while (1) {
251 recow = 0;
252 ret = btrfs_lookup_file_extent(trans, root, path, ino,
253 search_start, modify_tree);
254 if (ret < 0)
255 break;
256 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
257 leaf = path->nodes[0];
258 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
259 if (key.objectid == ino &&
260 key.type == BTRFS_EXTENT_DATA_KEY)
261 path->slots[0]--;
262 }
263 ret = 0;
264next_slot:
265 leaf = path->nodes[0];
266 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
267 BUG_ON(del_nr > 0);
268 ret = btrfs_next_leaf(root, path);
269 if (ret < 0)
270 break;
271 if (ret > 0) {
272 ret = 0;
273 break;
274 }
275 leaf = path->nodes[0];
276 recow = 1;
277 }
278
279 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
280
281 if (key.objectid > ino)
282 break;
283 if (WARN_ON_ONCE(key.objectid < ino) ||
284 key.type < BTRFS_EXTENT_DATA_KEY) {
285 ASSERT(del_nr == 0);
286 path->slots[0]++;
287 goto next_slot;
288 }
289 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
290 break;
291
292 fi = btrfs_item_ptr(leaf, path->slots[0],
293 struct btrfs_file_extent_item);
294 extent_type = btrfs_file_extent_type(leaf, fi);
295
296 if (extent_type == BTRFS_FILE_EXTENT_REG ||
297 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
298 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
299 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
300 extent_offset = btrfs_file_extent_offset(leaf, fi);
301 extent_end = key.offset +
302 btrfs_file_extent_num_bytes(leaf, fi);
303 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
304 extent_end = key.offset +
305 btrfs_file_extent_ram_bytes(leaf, fi);
306 } else {
307 /* can't happen */
308 BUG();
309 }
310
311 /*
312 * Don't skip extent items representing 0 byte lengths. They
313 * used to be created (bug) if while punching holes we hit
314 * -ENOSPC condition. So if we find one here, just ensure we
315 * delete it, otherwise we would insert a new file extent item
316 * with the same key (offset) as that 0 bytes length file
317 * extent item in the call to setup_items_for_insert() later
318 * in this function.
319 */
320 if (extent_end == key.offset && extent_end >= search_start) {
321 last_end = extent_end;
322 goto delete_extent_item;
323 }
324
325 if (extent_end <= search_start) {
326 path->slots[0]++;
327 goto next_slot;
328 }
329
330 found = 1;
331 search_start = max(key.offset, args->start);
332 if (recow || !modify_tree) {
333 modify_tree = -1;
334 btrfs_release_path(path);
335 continue;
336 }
337
338 /*
339 * | - range to drop - |
340 * | -------- extent -------- |
341 */
342 if (args->start > key.offset && args->end < extent_end) {
343 BUG_ON(del_nr > 0);
344 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
345 ret = -EOPNOTSUPP;
346 break;
347 }
348
349 memcpy(&new_key, &key, sizeof(new_key));
350 new_key.offset = args->start;
351 ret = btrfs_duplicate_item(trans, root, path,
352 &new_key);
353 if (ret == -EAGAIN) {
354 btrfs_release_path(path);
355 continue;
356 }
357 if (ret < 0)
358 break;
359
360 leaf = path->nodes[0];
361 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
362 struct btrfs_file_extent_item);
363 btrfs_set_file_extent_num_bytes(leaf, fi,
364 args->start - key.offset);
365
366 fi = btrfs_item_ptr(leaf, path->slots[0],
367 struct btrfs_file_extent_item);
368
369 extent_offset += args->start - key.offset;
370 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
371 btrfs_set_file_extent_num_bytes(leaf, fi,
372 extent_end - args->start);
373 btrfs_mark_buffer_dirty(trans, leaf);
374
375 if (update_refs && disk_bytenr > 0) {
376 btrfs_init_generic_ref(&ref,
377 BTRFS_ADD_DELAYED_REF,
378 disk_bytenr, num_bytes, 0,
379 root->root_key.objectid);
380 btrfs_init_data_ref(&ref,
381 root->root_key.objectid,
382 new_key.objectid,
383 args->start - extent_offset,
384 0, false);
385 ret = btrfs_inc_extent_ref(trans, &ref);
386 if (ret) {
387 btrfs_abort_transaction(trans, ret);
388 break;
389 }
390 }
391 key.offset = args->start;
392 }
393 /*
394 * From here on out we will have actually dropped something, so
395 * last_end can be updated.
396 */
397 last_end = extent_end;
398
399 /*
400 * | ---- range to drop ----- |
401 * | -------- extent -------- |
402 */
403 if (args->start <= key.offset && args->end < extent_end) {
404 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
405 ret = -EOPNOTSUPP;
406 break;
407 }
408
409 memcpy(&new_key, &key, sizeof(new_key));
410 new_key.offset = args->end;
411 btrfs_set_item_key_safe(trans, path, &new_key);
412
413 extent_offset += args->end - key.offset;
414 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
415 btrfs_set_file_extent_num_bytes(leaf, fi,
416 extent_end - args->end);
417 btrfs_mark_buffer_dirty(trans, leaf);
418 if (update_refs && disk_bytenr > 0)
419 args->bytes_found += args->end - key.offset;
420 break;
421 }
422
423 search_start = extent_end;
424 /*
425 * | ---- range to drop ----- |
426 * | -------- extent -------- |
427 */
428 if (args->start > key.offset && args->end >= extent_end) {
429 BUG_ON(del_nr > 0);
430 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
431 ret = -EOPNOTSUPP;
432 break;
433 }
434
435 btrfs_set_file_extent_num_bytes(leaf, fi,
436 args->start - key.offset);
437 btrfs_mark_buffer_dirty(trans, leaf);
438 if (update_refs && disk_bytenr > 0)
439 args->bytes_found += extent_end - args->start;
440 if (args->end == extent_end)
441 break;
442
443 path->slots[0]++;
444 goto next_slot;
445 }
446
447 /*
448 * | ---- range to drop ----- |
449 * | ------ extent ------ |
450 */
451 if (args->start <= key.offset && args->end >= extent_end) {
452delete_extent_item:
453 if (del_nr == 0) {
454 del_slot = path->slots[0];
455 del_nr = 1;
456 } else {
457 BUG_ON(del_slot + del_nr != path->slots[0]);
458 del_nr++;
459 }
460
461 if (update_refs &&
462 extent_type == BTRFS_FILE_EXTENT_INLINE) {
463 args->bytes_found += extent_end - key.offset;
464 extent_end = ALIGN(extent_end,
465 fs_info->sectorsize);
466 } else if (update_refs && disk_bytenr > 0) {
467 btrfs_init_generic_ref(&ref,
468 BTRFS_DROP_DELAYED_REF,
469 disk_bytenr, num_bytes, 0,
470 root->root_key.objectid);
471 btrfs_init_data_ref(&ref,
472 root->root_key.objectid,
473 key.objectid,
474 key.offset - extent_offset, 0,
475 false);
476 ret = btrfs_free_extent(trans, &ref);
477 if (ret) {
478 btrfs_abort_transaction(trans, ret);
479 break;
480 }
481 args->bytes_found += extent_end - key.offset;
482 }
483
484 if (args->end == extent_end)
485 break;
486
487 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
488 path->slots[0]++;
489 goto next_slot;
490 }
491
492 ret = btrfs_del_items(trans, root, path, del_slot,
493 del_nr);
494 if (ret) {
495 btrfs_abort_transaction(trans, ret);
496 break;
497 }
498
499 del_nr = 0;
500 del_slot = 0;
501
502 btrfs_release_path(path);
503 continue;
504 }
505
506 BUG();
507 }
508
509 if (!ret && del_nr > 0) {
510 /*
511 * Set path->slots[0] to first slot, so that after the delete
512 * if items are move off from our leaf to its immediate left or
513 * right neighbor leafs, we end up with a correct and adjusted
514 * path->slots[0] for our insertion (if args->replace_extent).
515 */
516 path->slots[0] = del_slot;
517 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
518 if (ret)
519 btrfs_abort_transaction(trans, ret);
520 }
521
522 leaf = path->nodes[0];
523 /*
524 * If btrfs_del_items() was called, it might have deleted a leaf, in
525 * which case it unlocked our path, so check path->locks[0] matches a
526 * write lock.
527 */
528 if (!ret && args->replace_extent &&
529 path->locks[0] == BTRFS_WRITE_LOCK &&
530 btrfs_leaf_free_space(leaf) >=
531 sizeof(struct btrfs_item) + args->extent_item_size) {
532
533 key.objectid = ino;
534 key.type = BTRFS_EXTENT_DATA_KEY;
535 key.offset = args->start;
536 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
537 struct btrfs_key slot_key;
538
539 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
540 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
541 path->slots[0]++;
542 }
543 btrfs_setup_item_for_insert(trans, root, path, &key,
544 args->extent_item_size);
545 args->extent_inserted = true;
546 }
547
548 if (!args->path)
549 btrfs_free_path(path);
550 else if (!args->extent_inserted)
551 btrfs_release_path(path);
552out:
553 args->drop_end = found ? min(args->end, last_end) : args->end;
554
555 return ret;
556}
557
558static int extent_mergeable(struct extent_buffer *leaf, int slot,
559 u64 objectid, u64 bytenr, u64 orig_offset,
560 u64 *start, u64 *end)
561{
562 struct btrfs_file_extent_item *fi;
563 struct btrfs_key key;
564 u64 extent_end;
565
566 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
567 return 0;
568
569 btrfs_item_key_to_cpu(leaf, &key, slot);
570 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
571 return 0;
572
573 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
574 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
575 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
576 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
577 btrfs_file_extent_compression(leaf, fi) ||
578 btrfs_file_extent_encryption(leaf, fi) ||
579 btrfs_file_extent_other_encoding(leaf, fi))
580 return 0;
581
582 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
583 if ((*start && *start != key.offset) || (*end && *end != extent_end))
584 return 0;
585
586 *start = key.offset;
587 *end = extent_end;
588 return 1;
589}
590
591/*
592 * Mark extent in the range start - end as written.
593 *
594 * This changes extent type from 'pre-allocated' to 'regular'. If only
595 * part of extent is marked as written, the extent will be split into
596 * two or three.
597 */
598int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
599 struct btrfs_inode *inode, u64 start, u64 end)
600{
601 struct btrfs_root *root = inode->root;
602 struct extent_buffer *leaf;
603 struct btrfs_path *path;
604 struct btrfs_file_extent_item *fi;
605 struct btrfs_ref ref = { 0 };
606 struct btrfs_key key;
607 struct btrfs_key new_key;
608 u64 bytenr;
609 u64 num_bytes;
610 u64 extent_end;
611 u64 orig_offset;
612 u64 other_start;
613 u64 other_end;
614 u64 split;
615 int del_nr = 0;
616 int del_slot = 0;
617 int recow;
618 int ret = 0;
619 u64 ino = btrfs_ino(inode);
620
621 path = btrfs_alloc_path();
622 if (!path)
623 return -ENOMEM;
624again:
625 recow = 0;
626 split = start;
627 key.objectid = ino;
628 key.type = BTRFS_EXTENT_DATA_KEY;
629 key.offset = split;
630
631 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
632 if (ret < 0)
633 goto out;
634 if (ret > 0 && path->slots[0] > 0)
635 path->slots[0]--;
636
637 leaf = path->nodes[0];
638 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
639 if (key.objectid != ino ||
640 key.type != BTRFS_EXTENT_DATA_KEY) {
641 ret = -EINVAL;
642 btrfs_abort_transaction(trans, ret);
643 goto out;
644 }
645 fi = btrfs_item_ptr(leaf, path->slots[0],
646 struct btrfs_file_extent_item);
647 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
648 ret = -EINVAL;
649 btrfs_abort_transaction(trans, ret);
650 goto out;
651 }
652 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
653 if (key.offset > start || extent_end < end) {
654 ret = -EINVAL;
655 btrfs_abort_transaction(trans, ret);
656 goto out;
657 }
658
659 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
660 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
661 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
662 memcpy(&new_key, &key, sizeof(new_key));
663
664 if (start == key.offset && end < extent_end) {
665 other_start = 0;
666 other_end = start;
667 if (extent_mergeable(leaf, path->slots[0] - 1,
668 ino, bytenr, orig_offset,
669 &other_start, &other_end)) {
670 new_key.offset = end;
671 btrfs_set_item_key_safe(trans, path, &new_key);
672 fi = btrfs_item_ptr(leaf, path->slots[0],
673 struct btrfs_file_extent_item);
674 btrfs_set_file_extent_generation(leaf, fi,
675 trans->transid);
676 btrfs_set_file_extent_num_bytes(leaf, fi,
677 extent_end - end);
678 btrfs_set_file_extent_offset(leaf, fi,
679 end - orig_offset);
680 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
681 struct btrfs_file_extent_item);
682 btrfs_set_file_extent_generation(leaf, fi,
683 trans->transid);
684 btrfs_set_file_extent_num_bytes(leaf, fi,
685 end - other_start);
686 btrfs_mark_buffer_dirty(trans, leaf);
687 goto out;
688 }
689 }
690
691 if (start > key.offset && end == extent_end) {
692 other_start = end;
693 other_end = 0;
694 if (extent_mergeable(leaf, path->slots[0] + 1,
695 ino, bytenr, orig_offset,
696 &other_start, &other_end)) {
697 fi = btrfs_item_ptr(leaf, path->slots[0],
698 struct btrfs_file_extent_item);
699 btrfs_set_file_extent_num_bytes(leaf, fi,
700 start - key.offset);
701 btrfs_set_file_extent_generation(leaf, fi,
702 trans->transid);
703 path->slots[0]++;
704 new_key.offset = start;
705 btrfs_set_item_key_safe(trans, path, &new_key);
706
707 fi = btrfs_item_ptr(leaf, path->slots[0],
708 struct btrfs_file_extent_item);
709 btrfs_set_file_extent_generation(leaf, fi,
710 trans->transid);
711 btrfs_set_file_extent_num_bytes(leaf, fi,
712 other_end - start);
713 btrfs_set_file_extent_offset(leaf, fi,
714 start - orig_offset);
715 btrfs_mark_buffer_dirty(trans, leaf);
716 goto out;
717 }
718 }
719
720 while (start > key.offset || end < extent_end) {
721 if (key.offset == start)
722 split = end;
723
724 new_key.offset = split;
725 ret = btrfs_duplicate_item(trans, root, path, &new_key);
726 if (ret == -EAGAIN) {
727 btrfs_release_path(path);
728 goto again;
729 }
730 if (ret < 0) {
731 btrfs_abort_transaction(trans, ret);
732 goto out;
733 }
734
735 leaf = path->nodes[0];
736 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
737 struct btrfs_file_extent_item);
738 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
739 btrfs_set_file_extent_num_bytes(leaf, fi,
740 split - key.offset);
741
742 fi = btrfs_item_ptr(leaf, path->slots[0],
743 struct btrfs_file_extent_item);
744
745 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
746 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
747 btrfs_set_file_extent_num_bytes(leaf, fi,
748 extent_end - split);
749 btrfs_mark_buffer_dirty(trans, leaf);
750
751 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
752 num_bytes, 0, root->root_key.objectid);
753 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
754 orig_offset, 0, false);
755 ret = btrfs_inc_extent_ref(trans, &ref);
756 if (ret) {
757 btrfs_abort_transaction(trans, ret);
758 goto out;
759 }
760
761 if (split == start) {
762 key.offset = start;
763 } else {
764 if (start != key.offset) {
765 ret = -EINVAL;
766 btrfs_abort_transaction(trans, ret);
767 goto out;
768 }
769 path->slots[0]--;
770 extent_end = end;
771 }
772 recow = 1;
773 }
774
775 other_start = end;
776 other_end = 0;
777 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
778 num_bytes, 0, root->root_key.objectid);
779 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset,
780 0, false);
781 if (extent_mergeable(leaf, path->slots[0] + 1,
782 ino, bytenr, orig_offset,
783 &other_start, &other_end)) {
784 if (recow) {
785 btrfs_release_path(path);
786 goto again;
787 }
788 extent_end = other_end;
789 del_slot = path->slots[0] + 1;
790 del_nr++;
791 ret = btrfs_free_extent(trans, &ref);
792 if (ret) {
793 btrfs_abort_transaction(trans, ret);
794 goto out;
795 }
796 }
797 other_start = 0;
798 other_end = start;
799 if (extent_mergeable(leaf, path->slots[0] - 1,
800 ino, bytenr, orig_offset,
801 &other_start, &other_end)) {
802 if (recow) {
803 btrfs_release_path(path);
804 goto again;
805 }
806 key.offset = other_start;
807 del_slot = path->slots[0];
808 del_nr++;
809 ret = btrfs_free_extent(trans, &ref);
810 if (ret) {
811 btrfs_abort_transaction(trans, ret);
812 goto out;
813 }
814 }
815 if (del_nr == 0) {
816 fi = btrfs_item_ptr(leaf, path->slots[0],
817 struct btrfs_file_extent_item);
818 btrfs_set_file_extent_type(leaf, fi,
819 BTRFS_FILE_EXTENT_REG);
820 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
821 btrfs_mark_buffer_dirty(trans, leaf);
822 } else {
823 fi = btrfs_item_ptr(leaf, del_slot - 1,
824 struct btrfs_file_extent_item);
825 btrfs_set_file_extent_type(leaf, fi,
826 BTRFS_FILE_EXTENT_REG);
827 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
828 btrfs_set_file_extent_num_bytes(leaf, fi,
829 extent_end - key.offset);
830 btrfs_mark_buffer_dirty(trans, leaf);
831
832 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
833 if (ret < 0) {
834 btrfs_abort_transaction(trans, ret);
835 goto out;
836 }
837 }
838out:
839 btrfs_free_path(path);
840 return ret;
841}
842
843/*
844 * on error we return an unlocked page and the error value
845 * on success we return a locked page and 0
846 */
847static int prepare_uptodate_page(struct inode *inode,
848 struct page *page, u64 pos,
849 bool force_uptodate)
850{
851 struct folio *folio = page_folio(page);
852 int ret = 0;
853
854 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
855 !PageUptodate(page)) {
856 ret = btrfs_read_folio(NULL, folio);
857 if (ret)
858 return ret;
859 lock_page(page);
860 if (!PageUptodate(page)) {
861 unlock_page(page);
862 return -EIO;
863 }
864
865 /*
866 * Since btrfs_read_folio() will unlock the folio before it
867 * returns, there is a window where btrfs_release_folio() can be
868 * called to release the page. Here we check both inode
869 * mapping and PagePrivate() to make sure the page was not
870 * released.
871 *
872 * The private flag check is essential for subpage as we need
873 * to store extra bitmap using folio private.
874 */
875 if (page->mapping != inode->i_mapping || !folio_test_private(folio)) {
876 unlock_page(page);
877 return -EAGAIN;
878 }
879 }
880 return 0;
881}
882
883static fgf_t get_prepare_fgp_flags(bool nowait)
884{
885 fgf_t fgp_flags = FGP_LOCK | FGP_ACCESSED | FGP_CREAT;
886
887 if (nowait)
888 fgp_flags |= FGP_NOWAIT;
889
890 return fgp_flags;
891}
892
893static gfp_t get_prepare_gfp_flags(struct inode *inode, bool nowait)
894{
895 gfp_t gfp;
896
897 gfp = btrfs_alloc_write_mask(inode->i_mapping);
898 if (nowait) {
899 gfp &= ~__GFP_DIRECT_RECLAIM;
900 gfp |= GFP_NOWAIT;
901 }
902
903 return gfp;
904}
905
906/*
907 * this just gets pages into the page cache and locks them down.
908 */
909static noinline int prepare_pages(struct inode *inode, struct page **pages,
910 size_t num_pages, loff_t pos,
911 size_t write_bytes, bool force_uptodate,
912 bool nowait)
913{
914 int i;
915 unsigned long index = pos >> PAGE_SHIFT;
916 gfp_t mask = get_prepare_gfp_flags(inode, nowait);
917 fgf_t fgp_flags = get_prepare_fgp_flags(nowait);
918 int err = 0;
919 int faili;
920
921 for (i = 0; i < num_pages; i++) {
922again:
923 pages[i] = pagecache_get_page(inode->i_mapping, index + i,
924 fgp_flags, mask | __GFP_WRITE);
925 if (!pages[i]) {
926 faili = i - 1;
927 if (nowait)
928 err = -EAGAIN;
929 else
930 err = -ENOMEM;
931 goto fail;
932 }
933
934 err = set_page_extent_mapped(pages[i]);
935 if (err < 0) {
936 faili = i;
937 goto fail;
938 }
939
940 if (i == 0)
941 err = prepare_uptodate_page(inode, pages[i], pos,
942 force_uptodate);
943 if (!err && i == num_pages - 1)
944 err = prepare_uptodate_page(inode, pages[i],
945 pos + write_bytes, false);
946 if (err) {
947 put_page(pages[i]);
948 if (!nowait && err == -EAGAIN) {
949 err = 0;
950 goto again;
951 }
952 faili = i - 1;
953 goto fail;
954 }
955 wait_on_page_writeback(pages[i]);
956 }
957
958 return 0;
959fail:
960 while (faili >= 0) {
961 unlock_page(pages[faili]);
962 put_page(pages[faili]);
963 faili--;
964 }
965 return err;
966
967}
968
969/*
970 * This function locks the extent and properly waits for data=ordered extents
971 * to finish before allowing the pages to be modified if need.
972 *
973 * The return value:
974 * 1 - the extent is locked
975 * 0 - the extent is not locked, and everything is OK
976 * -EAGAIN - need re-prepare the pages
977 * the other < 0 number - Something wrong happens
978 */
979static noinline int
980lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
981 size_t num_pages, loff_t pos,
982 size_t write_bytes,
983 u64 *lockstart, u64 *lockend, bool nowait,
984 struct extent_state **cached_state)
985{
986 struct btrfs_fs_info *fs_info = inode->root->fs_info;
987 u64 start_pos;
988 u64 last_pos;
989 int i;
990 int ret = 0;
991
992 start_pos = round_down(pos, fs_info->sectorsize);
993 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
994
995 if (start_pos < inode->vfs_inode.i_size) {
996 struct btrfs_ordered_extent *ordered;
997
998 if (nowait) {
999 if (!try_lock_extent(&inode->io_tree, start_pos, last_pos,
1000 cached_state)) {
1001 for (i = 0; i < num_pages; i++) {
1002 unlock_page(pages[i]);
1003 put_page(pages[i]);
1004 pages[i] = NULL;
1005 }
1006
1007 return -EAGAIN;
1008 }
1009 } else {
1010 lock_extent(&inode->io_tree, start_pos, last_pos, cached_state);
1011 }
1012
1013 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1014 last_pos - start_pos + 1);
1015 if (ordered &&
1016 ordered->file_offset + ordered->num_bytes > start_pos &&
1017 ordered->file_offset <= last_pos) {
1018 unlock_extent(&inode->io_tree, start_pos, last_pos,
1019 cached_state);
1020 for (i = 0; i < num_pages; i++) {
1021 unlock_page(pages[i]);
1022 put_page(pages[i]);
1023 }
1024 btrfs_start_ordered_extent(ordered);
1025 btrfs_put_ordered_extent(ordered);
1026 return -EAGAIN;
1027 }
1028 if (ordered)
1029 btrfs_put_ordered_extent(ordered);
1030
1031 *lockstart = start_pos;
1032 *lockend = last_pos;
1033 ret = 1;
1034 }
1035
1036 /*
1037 * We should be called after prepare_pages() which should have locked
1038 * all pages in the range.
1039 */
1040 for (i = 0; i < num_pages; i++)
1041 WARN_ON(!PageLocked(pages[i]));
1042
1043 return ret;
1044}
1045
1046/*
1047 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1048 *
1049 * @pos: File offset.
1050 * @write_bytes: The length to write, will be updated to the nocow writeable
1051 * range.
1052 *
1053 * This function will flush ordered extents in the range to ensure proper
1054 * nocow checks.
1055 *
1056 * Return:
1057 * > 0 If we can nocow, and updates @write_bytes.
1058 * 0 If we can't do a nocow write.
1059 * -EAGAIN If we can't do a nocow write because snapshoting of the inode's
1060 * root is in progress.
1061 * < 0 If an error happened.
1062 *
1063 * NOTE: Callers need to call btrfs_check_nocow_unlock() if we return > 0.
1064 */
1065int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1066 size_t *write_bytes, bool nowait)
1067{
1068 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1069 struct btrfs_root *root = inode->root;
1070 struct extent_state *cached_state = NULL;
1071 u64 lockstart, lockend;
1072 u64 num_bytes;
1073 int ret;
1074
1075 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1076 return 0;
1077
1078 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1079 return -EAGAIN;
1080
1081 lockstart = round_down(pos, fs_info->sectorsize);
1082 lockend = round_up(pos + *write_bytes,
1083 fs_info->sectorsize) - 1;
1084 num_bytes = lockend - lockstart + 1;
1085
1086 if (nowait) {
1087 if (!btrfs_try_lock_ordered_range(inode, lockstart, lockend,
1088 &cached_state)) {
1089 btrfs_drew_write_unlock(&root->snapshot_lock);
1090 return -EAGAIN;
1091 }
1092 } else {
1093 btrfs_lock_and_flush_ordered_range(inode, lockstart, lockend,
1094 &cached_state);
1095 }
1096 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1097 NULL, NULL, NULL, nowait, false);
1098 if (ret <= 0)
1099 btrfs_drew_write_unlock(&root->snapshot_lock);
1100 else
1101 *write_bytes = min_t(size_t, *write_bytes ,
1102 num_bytes - pos + lockstart);
1103 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
1104
1105 return ret;
1106}
1107
1108void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1109{
1110 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1111}
1112
1113static void update_time_for_write(struct inode *inode)
1114{
1115 struct timespec64 now, ts;
1116
1117 if (IS_NOCMTIME(inode))
1118 return;
1119
1120 now = current_time(inode);
1121 ts = inode_get_mtime(inode);
1122 if (!timespec64_equal(&ts, &now))
1123 inode_set_mtime_to_ts(inode, now);
1124
1125 ts = inode_get_ctime(inode);
1126 if (!timespec64_equal(&ts, &now))
1127 inode_set_ctime_to_ts(inode, now);
1128
1129 if (IS_I_VERSION(inode))
1130 inode_inc_iversion(inode);
1131}
1132
1133static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1134 size_t count)
1135{
1136 struct file *file = iocb->ki_filp;
1137 struct inode *inode = file_inode(file);
1138 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1139 loff_t pos = iocb->ki_pos;
1140 int ret;
1141 loff_t oldsize;
1142 loff_t start_pos;
1143
1144 /*
1145 * Quickly bail out on NOWAIT writes if we don't have the nodatacow or
1146 * prealloc flags, as without those flags we always have to COW. We will
1147 * later check if we can really COW into the target range (using
1148 * can_nocow_extent() at btrfs_get_blocks_direct_write()).
1149 */
1150 if ((iocb->ki_flags & IOCB_NOWAIT) &&
1151 !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1152 return -EAGAIN;
1153
1154 ret = file_remove_privs(file);
1155 if (ret)
1156 return ret;
1157
1158 /*
1159 * We reserve space for updating the inode when we reserve space for the
1160 * extent we are going to write, so we will enospc out there. We don't
1161 * need to start yet another transaction to update the inode as we will
1162 * update the inode when we finish writing whatever data we write.
1163 */
1164 update_time_for_write(inode);
1165
1166 start_pos = round_down(pos, fs_info->sectorsize);
1167 oldsize = i_size_read(inode);
1168 if (start_pos > oldsize) {
1169 /* Expand hole size to cover write data, preventing empty gap */
1170 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1171
1172 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1173 if (ret)
1174 return ret;
1175 }
1176
1177 return 0;
1178}
1179
1180static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1181 struct iov_iter *i)
1182{
1183 struct file *file = iocb->ki_filp;
1184 loff_t pos;
1185 struct inode *inode = file_inode(file);
1186 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1187 struct page **pages = NULL;
1188 struct extent_changeset *data_reserved = NULL;
1189 u64 release_bytes = 0;
1190 u64 lockstart;
1191 u64 lockend;
1192 size_t num_written = 0;
1193 int nrptrs;
1194 ssize_t ret;
1195 bool only_release_metadata = false;
1196 bool force_page_uptodate = false;
1197 loff_t old_isize = i_size_read(inode);
1198 unsigned int ilock_flags = 0;
1199 const bool nowait = (iocb->ki_flags & IOCB_NOWAIT);
1200 unsigned int bdp_flags = (nowait ? BDP_ASYNC : 0);
1201
1202 if (nowait)
1203 ilock_flags |= BTRFS_ILOCK_TRY;
1204
1205 ret = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1206 if (ret < 0)
1207 return ret;
1208
1209 ret = generic_write_checks(iocb, i);
1210 if (ret <= 0)
1211 goto out;
1212
1213 ret = btrfs_write_check(iocb, i, ret);
1214 if (ret < 0)
1215 goto out;
1216
1217 pos = iocb->ki_pos;
1218 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1219 PAGE_SIZE / (sizeof(struct page *)));
1220 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1221 nrptrs = max(nrptrs, 8);
1222 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1223 if (!pages) {
1224 ret = -ENOMEM;
1225 goto out;
1226 }
1227
1228 while (iov_iter_count(i) > 0) {
1229 struct extent_state *cached_state = NULL;
1230 size_t offset = offset_in_page(pos);
1231 size_t sector_offset;
1232 size_t write_bytes = min(iov_iter_count(i),
1233 nrptrs * (size_t)PAGE_SIZE -
1234 offset);
1235 size_t num_pages;
1236 size_t reserve_bytes;
1237 size_t dirty_pages;
1238 size_t copied;
1239 size_t dirty_sectors;
1240 size_t num_sectors;
1241 int extents_locked;
1242
1243 /*
1244 * Fault pages before locking them in prepare_pages
1245 * to avoid recursive lock
1246 */
1247 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1248 ret = -EFAULT;
1249 break;
1250 }
1251
1252 only_release_metadata = false;
1253 sector_offset = pos & (fs_info->sectorsize - 1);
1254
1255 extent_changeset_release(data_reserved);
1256 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1257 &data_reserved, pos,
1258 write_bytes, nowait);
1259 if (ret < 0) {
1260 int can_nocow;
1261
1262 if (nowait && (ret == -ENOSPC || ret == -EAGAIN)) {
1263 ret = -EAGAIN;
1264 break;
1265 }
1266
1267 /*
1268 * If we don't have to COW at the offset, reserve
1269 * metadata only. write_bytes may get smaller than
1270 * requested here.
1271 */
1272 can_nocow = btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1273 &write_bytes, nowait);
1274 if (can_nocow < 0)
1275 ret = can_nocow;
1276 if (can_nocow > 0)
1277 ret = 0;
1278 if (ret)
1279 break;
1280 only_release_metadata = true;
1281 }
1282
1283 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1284 WARN_ON(num_pages > nrptrs);
1285 reserve_bytes = round_up(write_bytes + sector_offset,
1286 fs_info->sectorsize);
1287 WARN_ON(reserve_bytes == 0);
1288 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1289 reserve_bytes,
1290 reserve_bytes, nowait);
1291 if (ret) {
1292 if (!only_release_metadata)
1293 btrfs_free_reserved_data_space(BTRFS_I(inode),
1294 data_reserved, pos,
1295 write_bytes);
1296 else
1297 btrfs_check_nocow_unlock(BTRFS_I(inode));
1298
1299 if (nowait && ret == -ENOSPC)
1300 ret = -EAGAIN;
1301 break;
1302 }
1303
1304 release_bytes = reserve_bytes;
1305again:
1306 ret = balance_dirty_pages_ratelimited_flags(inode->i_mapping, bdp_flags);
1307 if (ret) {
1308 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1309 break;
1310 }
1311
1312 /*
1313 * This is going to setup the pages array with the number of
1314 * pages we want, so we don't really need to worry about the
1315 * contents of pages from loop to loop
1316 */
1317 ret = prepare_pages(inode, pages, num_pages,
1318 pos, write_bytes, force_page_uptodate, false);
1319 if (ret) {
1320 btrfs_delalloc_release_extents(BTRFS_I(inode),
1321 reserve_bytes);
1322 break;
1323 }
1324
1325 extents_locked = lock_and_cleanup_extent_if_need(
1326 BTRFS_I(inode), pages,
1327 num_pages, pos, write_bytes, &lockstart,
1328 &lockend, nowait, &cached_state);
1329 if (extents_locked < 0) {
1330 if (!nowait && extents_locked == -EAGAIN)
1331 goto again;
1332
1333 btrfs_delalloc_release_extents(BTRFS_I(inode),
1334 reserve_bytes);
1335 ret = extents_locked;
1336 break;
1337 }
1338
1339 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1340
1341 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1342 dirty_sectors = round_up(copied + sector_offset,
1343 fs_info->sectorsize);
1344 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1345
1346 /*
1347 * if we have trouble faulting in the pages, fall
1348 * back to one page at a time
1349 */
1350 if (copied < write_bytes)
1351 nrptrs = 1;
1352
1353 if (copied == 0) {
1354 force_page_uptodate = true;
1355 dirty_sectors = 0;
1356 dirty_pages = 0;
1357 } else {
1358 force_page_uptodate = false;
1359 dirty_pages = DIV_ROUND_UP(copied + offset,
1360 PAGE_SIZE);
1361 }
1362
1363 if (num_sectors > dirty_sectors) {
1364 /* release everything except the sectors we dirtied */
1365 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1366 if (only_release_metadata) {
1367 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1368 release_bytes, true);
1369 } else {
1370 u64 __pos;
1371
1372 __pos = round_down(pos,
1373 fs_info->sectorsize) +
1374 (dirty_pages << PAGE_SHIFT);
1375 btrfs_delalloc_release_space(BTRFS_I(inode),
1376 data_reserved, __pos,
1377 release_bytes, true);
1378 }
1379 }
1380
1381 release_bytes = round_up(copied + sector_offset,
1382 fs_info->sectorsize);
1383
1384 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1385 dirty_pages, pos, copied,
1386 &cached_state, only_release_metadata);
1387
1388 /*
1389 * If we have not locked the extent range, because the range's
1390 * start offset is >= i_size, we might still have a non-NULL
1391 * cached extent state, acquired while marking the extent range
1392 * as delalloc through btrfs_dirty_pages(). Therefore free any
1393 * possible cached extent state to avoid a memory leak.
1394 */
1395 if (extents_locked)
1396 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
1397 lockend, &cached_state);
1398 else
1399 free_extent_state(cached_state);
1400
1401 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1402 if (ret) {
1403 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1404 break;
1405 }
1406
1407 release_bytes = 0;
1408 if (only_release_metadata)
1409 btrfs_check_nocow_unlock(BTRFS_I(inode));
1410
1411 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1412
1413 cond_resched();
1414
1415 pos += copied;
1416 num_written += copied;
1417 }
1418
1419 kfree(pages);
1420
1421 if (release_bytes) {
1422 if (only_release_metadata) {
1423 btrfs_check_nocow_unlock(BTRFS_I(inode));
1424 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1425 release_bytes, true);
1426 } else {
1427 btrfs_delalloc_release_space(BTRFS_I(inode),
1428 data_reserved,
1429 round_down(pos, fs_info->sectorsize),
1430 release_bytes, true);
1431 }
1432 }
1433
1434 extent_changeset_free(data_reserved);
1435 if (num_written > 0) {
1436 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1437 iocb->ki_pos += num_written;
1438 }
1439out:
1440 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1441 return num_written ? num_written : ret;
1442}
1443
1444static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1445 const struct iov_iter *iter, loff_t offset)
1446{
1447 const u32 blocksize_mask = fs_info->sectorsize - 1;
1448
1449 if (offset & blocksize_mask)
1450 return -EINVAL;
1451
1452 if (iov_iter_alignment(iter) & blocksize_mask)
1453 return -EINVAL;
1454
1455 return 0;
1456}
1457
1458static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1459{
1460 struct file *file = iocb->ki_filp;
1461 struct inode *inode = file_inode(file);
1462 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1463 loff_t pos;
1464 ssize_t written = 0;
1465 ssize_t written_buffered;
1466 size_t prev_left = 0;
1467 loff_t endbyte;
1468 ssize_t err;
1469 unsigned int ilock_flags = 0;
1470 struct iomap_dio *dio;
1471
1472 if (iocb->ki_flags & IOCB_NOWAIT)
1473 ilock_flags |= BTRFS_ILOCK_TRY;
1474
1475 /*
1476 * If the write DIO is within EOF, use a shared lock and also only if
1477 * security bits will likely not be dropped by file_remove_privs() called
1478 * from btrfs_write_check(). Either will need to be rechecked after the
1479 * lock was acquired.
1480 */
1481 if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode) && IS_NOSEC(inode))
1482 ilock_flags |= BTRFS_ILOCK_SHARED;
1483
1484relock:
1485 err = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1486 if (err < 0)
1487 return err;
1488
1489 /* Shared lock cannot be used with security bits set. */
1490 if ((ilock_flags & BTRFS_ILOCK_SHARED) && !IS_NOSEC(inode)) {
1491 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1492 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1493 goto relock;
1494 }
1495
1496 err = generic_write_checks(iocb, from);
1497 if (err <= 0) {
1498 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1499 return err;
1500 }
1501
1502 err = btrfs_write_check(iocb, from, err);
1503 if (err < 0) {
1504 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1505 goto out;
1506 }
1507
1508 pos = iocb->ki_pos;
1509 /*
1510 * Re-check since file size may have changed just before taking the
1511 * lock or pos may have changed because of O_APPEND in generic_write_check()
1512 */
1513 if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1514 pos + iov_iter_count(from) > i_size_read(inode)) {
1515 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1516 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1517 goto relock;
1518 }
1519
1520 if (check_direct_IO(fs_info, from, pos)) {
1521 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1522 goto buffered;
1523 }
1524
1525 /*
1526 * The iov_iter can be mapped to the same file range we are writing to.
1527 * If that's the case, then we will deadlock in the iomap code, because
1528 * it first calls our callback btrfs_dio_iomap_begin(), which will create
1529 * an ordered extent, and after that it will fault in the pages that the
1530 * iov_iter refers to. During the fault in we end up in the readahead
1531 * pages code (starting at btrfs_readahead()), which will lock the range,
1532 * find that ordered extent and then wait for it to complete (at
1533 * btrfs_lock_and_flush_ordered_range()), resulting in a deadlock since
1534 * obviously the ordered extent can never complete as we didn't submit
1535 * yet the respective bio(s). This always happens when the buffer is
1536 * memory mapped to the same file range, since the iomap DIO code always
1537 * invalidates pages in the target file range (after starting and waiting
1538 * for any writeback).
1539 *
1540 * So here we disable page faults in the iov_iter and then retry if we
1541 * got -EFAULT, faulting in the pages before the retry.
1542 */
1543 from->nofault = true;
1544 dio = btrfs_dio_write(iocb, from, written);
1545 from->nofault = false;
1546
1547 /*
1548 * iomap_dio_complete() will call btrfs_sync_file() if we have a dsync
1549 * iocb, and that needs to lock the inode. So unlock it before calling
1550 * iomap_dio_complete() to avoid a deadlock.
1551 */
1552 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1553
1554 if (IS_ERR_OR_NULL(dio))
1555 err = PTR_ERR_OR_ZERO(dio);
1556 else
1557 err = iomap_dio_complete(dio);
1558
1559 /* No increment (+=) because iomap returns a cumulative value. */
1560 if (err > 0)
1561 written = err;
1562
1563 if (iov_iter_count(from) > 0 && (err == -EFAULT || err > 0)) {
1564 const size_t left = iov_iter_count(from);
1565 /*
1566 * We have more data left to write. Try to fault in as many as
1567 * possible of the remainder pages and retry. We do this without
1568 * releasing and locking again the inode, to prevent races with
1569 * truncate.
1570 *
1571 * Also, in case the iov refers to pages in the file range of the
1572 * file we want to write to (due to a mmap), we could enter an
1573 * infinite loop if we retry after faulting the pages in, since
1574 * iomap will invalidate any pages in the range early on, before
1575 * it tries to fault in the pages of the iov. So we keep track of
1576 * how much was left of iov in the previous EFAULT and fallback
1577 * to buffered IO in case we haven't made any progress.
1578 */
1579 if (left == prev_left) {
1580 err = -ENOTBLK;
1581 } else {
1582 fault_in_iov_iter_readable(from, left);
1583 prev_left = left;
1584 goto relock;
1585 }
1586 }
1587
1588 /*
1589 * If 'err' is -ENOTBLK or we have not written all data, then it means
1590 * we must fallback to buffered IO.
1591 */
1592 if ((err < 0 && err != -ENOTBLK) || !iov_iter_count(from))
1593 goto out;
1594
1595buffered:
1596 /*
1597 * If we are in a NOWAIT context, then return -EAGAIN to signal the caller
1598 * it must retry the operation in a context where blocking is acceptable,
1599 * because even if we end up not blocking during the buffered IO attempt
1600 * below, we will block when flushing and waiting for the IO.
1601 */
1602 if (iocb->ki_flags & IOCB_NOWAIT) {
1603 err = -EAGAIN;
1604 goto out;
1605 }
1606
1607 pos = iocb->ki_pos;
1608 written_buffered = btrfs_buffered_write(iocb, from);
1609 if (written_buffered < 0) {
1610 err = written_buffered;
1611 goto out;
1612 }
1613 /*
1614 * Ensure all data is persisted. We want the next direct IO read to be
1615 * able to read what was just written.
1616 */
1617 endbyte = pos + written_buffered - 1;
1618 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1619 if (err)
1620 goto out;
1621 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1622 if (err)
1623 goto out;
1624 written += written_buffered;
1625 iocb->ki_pos = pos + written_buffered;
1626 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1627 endbyte >> PAGE_SHIFT);
1628out:
1629 return err < 0 ? err : written;
1630}
1631
1632static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
1633 const struct btrfs_ioctl_encoded_io_args *encoded)
1634{
1635 struct file *file = iocb->ki_filp;
1636 struct inode *inode = file_inode(file);
1637 loff_t count;
1638 ssize_t ret;
1639
1640 btrfs_inode_lock(BTRFS_I(inode), 0);
1641 count = encoded->len;
1642 ret = generic_write_checks_count(iocb, &count);
1643 if (ret == 0 && count != encoded->len) {
1644 /*
1645 * The write got truncated by generic_write_checks_count(). We
1646 * can't do a partial encoded write.
1647 */
1648 ret = -EFBIG;
1649 }
1650 if (ret || encoded->len == 0)
1651 goto out;
1652
1653 ret = btrfs_write_check(iocb, from, encoded->len);
1654 if (ret < 0)
1655 goto out;
1656
1657 ret = btrfs_do_encoded_write(iocb, from, encoded);
1658out:
1659 btrfs_inode_unlock(BTRFS_I(inode), 0);
1660 return ret;
1661}
1662
1663ssize_t btrfs_do_write_iter(struct kiocb *iocb, struct iov_iter *from,
1664 const struct btrfs_ioctl_encoded_io_args *encoded)
1665{
1666 struct file *file = iocb->ki_filp;
1667 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1668 ssize_t num_written, num_sync;
1669
1670 /*
1671 * If the fs flips readonly due to some impossible error, although we
1672 * have opened a file as writable, we have to stop this write operation
1673 * to ensure consistency.
1674 */
1675 if (BTRFS_FS_ERROR(inode->root->fs_info))
1676 return -EROFS;
1677
1678 if (encoded && (iocb->ki_flags & IOCB_NOWAIT))
1679 return -EOPNOTSUPP;
1680
1681 if (encoded) {
1682 num_written = btrfs_encoded_write(iocb, from, encoded);
1683 num_sync = encoded->len;
1684 } else if (iocb->ki_flags & IOCB_DIRECT) {
1685 num_written = btrfs_direct_write(iocb, from);
1686 num_sync = num_written;
1687 } else {
1688 num_written = btrfs_buffered_write(iocb, from);
1689 num_sync = num_written;
1690 }
1691
1692 btrfs_set_inode_last_sub_trans(inode);
1693
1694 if (num_sync > 0) {
1695 num_sync = generic_write_sync(iocb, num_sync);
1696 if (num_sync < 0)
1697 num_written = num_sync;
1698 }
1699
1700 return num_written;
1701}
1702
1703static ssize_t btrfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1704{
1705 return btrfs_do_write_iter(iocb, from, NULL);
1706}
1707
1708int btrfs_release_file(struct inode *inode, struct file *filp)
1709{
1710 struct btrfs_file_private *private = filp->private_data;
1711
1712 if (private) {
1713 kfree(private->filldir_buf);
1714 free_extent_state(private->llseek_cached_state);
1715 kfree(private);
1716 filp->private_data = NULL;
1717 }
1718
1719 /*
1720 * Set by setattr when we are about to truncate a file from a non-zero
1721 * size to a zero size. This tries to flush down new bytes that may
1722 * have been written if the application were using truncate to replace
1723 * a file in place.
1724 */
1725 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
1726 &BTRFS_I(inode)->runtime_flags))
1727 filemap_flush(inode->i_mapping);
1728 return 0;
1729}
1730
1731static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1732{
1733 int ret;
1734 struct blk_plug plug;
1735
1736 /*
1737 * This is only called in fsync, which would do synchronous writes, so
1738 * a plug can merge adjacent IOs as much as possible. Esp. in case of
1739 * multiple disks using raid profile, a large IO can be split to
1740 * several segments of stripe length (currently 64K).
1741 */
1742 blk_start_plug(&plug);
1743 ret = btrfs_fdatawrite_range(inode, start, end);
1744 blk_finish_plug(&plug);
1745
1746 return ret;
1747}
1748
1749static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
1750{
1751 struct btrfs_inode *inode = BTRFS_I(ctx->inode);
1752 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1753
1754 if (btrfs_inode_in_log(inode, btrfs_get_fs_generation(fs_info)) &&
1755 list_empty(&ctx->ordered_extents))
1756 return true;
1757
1758 /*
1759 * If we are doing a fast fsync we can not bail out if the inode's
1760 * last_trans is <= then the last committed transaction, because we only
1761 * update the last_trans of the inode during ordered extent completion,
1762 * and for a fast fsync we don't wait for that, we only wait for the
1763 * writeback to complete.
1764 */
1765 if (inode->last_trans <= btrfs_get_last_trans_committed(fs_info) &&
1766 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
1767 list_empty(&ctx->ordered_extents)))
1768 return true;
1769
1770 return false;
1771}
1772
1773/*
1774 * fsync call for both files and directories. This logs the inode into
1775 * the tree log instead of forcing full commits whenever possible.
1776 *
1777 * It needs to call filemap_fdatawait so that all ordered extent updates are
1778 * in the metadata btree are up to date for copying to the log.
1779 *
1780 * It drops the inode mutex before doing the tree log commit. This is an
1781 * important optimization for directories because holding the mutex prevents
1782 * new operations on the dir while we write to disk.
1783 */
1784int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1785{
1786 struct dentry *dentry = file_dentry(file);
1787 struct inode *inode = d_inode(dentry);
1788 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1789 struct btrfs_root *root = BTRFS_I(inode)->root;
1790 struct btrfs_trans_handle *trans;
1791 struct btrfs_log_ctx ctx;
1792 int ret = 0, err;
1793 u64 len;
1794 bool full_sync;
1795
1796 trace_btrfs_sync_file(file, datasync);
1797
1798 btrfs_init_log_ctx(&ctx, inode);
1799
1800 /*
1801 * Always set the range to a full range, otherwise we can get into
1802 * several problems, from missing file extent items to represent holes
1803 * when not using the NO_HOLES feature, to log tree corruption due to
1804 * races between hole detection during logging and completion of ordered
1805 * extents outside the range, to missing checksums due to ordered extents
1806 * for which we flushed only a subset of their pages.
1807 */
1808 start = 0;
1809 end = LLONG_MAX;
1810 len = (u64)LLONG_MAX + 1;
1811
1812 /*
1813 * We write the dirty pages in the range and wait until they complete
1814 * out of the ->i_mutex. If so, we can flush the dirty pages by
1815 * multi-task, and make the performance up. See
1816 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1817 */
1818 ret = start_ordered_ops(inode, start, end);
1819 if (ret)
1820 goto out;
1821
1822 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1823
1824 atomic_inc(&root->log_batch);
1825
1826 /*
1827 * Before we acquired the inode's lock and the mmap lock, someone may
1828 * have dirtied more pages in the target range. We need to make sure
1829 * that writeback for any such pages does not start while we are logging
1830 * the inode, because if it does, any of the following might happen when
1831 * we are not doing a full inode sync:
1832 *
1833 * 1) We log an extent after its writeback finishes but before its
1834 * checksums are added to the csum tree, leading to -EIO errors
1835 * when attempting to read the extent after a log replay.
1836 *
1837 * 2) We can end up logging an extent before its writeback finishes.
1838 * Therefore after the log replay we will have a file extent item
1839 * pointing to an unwritten extent (and no data checksums as well).
1840 *
1841 * So trigger writeback for any eventual new dirty pages and then we
1842 * wait for all ordered extents to complete below.
1843 */
1844 ret = start_ordered_ops(inode, start, end);
1845 if (ret) {
1846 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1847 goto out;
1848 }
1849
1850 /*
1851 * Always check for the full sync flag while holding the inode's lock,
1852 * to avoid races with other tasks. The flag must be either set all the
1853 * time during logging or always off all the time while logging.
1854 * We check the flag here after starting delalloc above, because when
1855 * running delalloc the full sync flag may be set if we need to drop
1856 * extra extent map ranges due to temporary memory allocation failures.
1857 */
1858 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1859 &BTRFS_I(inode)->runtime_flags);
1860
1861 /*
1862 * We have to do this here to avoid the priority inversion of waiting on
1863 * IO of a lower priority task while holding a transaction open.
1864 *
1865 * For a full fsync we wait for the ordered extents to complete while
1866 * for a fast fsync we wait just for writeback to complete, and then
1867 * attach the ordered extents to the transaction so that a transaction
1868 * commit waits for their completion, to avoid data loss if we fsync,
1869 * the current transaction commits before the ordered extents complete
1870 * and a power failure happens right after that.
1871 *
1872 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
1873 * logical address recorded in the ordered extent may change. We need
1874 * to wait for the IO to stabilize the logical address.
1875 */
1876 if (full_sync || btrfs_is_zoned(fs_info)) {
1877 ret = btrfs_wait_ordered_range(inode, start, len);
1878 } else {
1879 /*
1880 * Get our ordered extents as soon as possible to avoid doing
1881 * checksum lookups in the csum tree, and use instead the
1882 * checksums attached to the ordered extents.
1883 */
1884 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
1885 &ctx.ordered_extents);
1886 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
1887 }
1888
1889 if (ret)
1890 goto out_release_extents;
1891
1892 atomic_inc(&root->log_batch);
1893
1894 if (skip_inode_logging(&ctx)) {
1895 /*
1896 * We've had everything committed since the last time we were
1897 * modified so clear this flag in case it was set for whatever
1898 * reason, it's no longer relevant.
1899 */
1900 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1901 &BTRFS_I(inode)->runtime_flags);
1902 /*
1903 * An ordered extent might have started before and completed
1904 * already with io errors, in which case the inode was not
1905 * updated and we end up here. So check the inode's mapping
1906 * for any errors that might have happened since we last
1907 * checked called fsync.
1908 */
1909 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
1910 goto out_release_extents;
1911 }
1912
1913 btrfs_init_log_ctx_scratch_eb(&ctx);
1914
1915 /*
1916 * We use start here because we will need to wait on the IO to complete
1917 * in btrfs_sync_log, which could require joining a transaction (for
1918 * example checking cross references in the nocow path). If we use join
1919 * here we could get into a situation where we're waiting on IO to
1920 * happen that is blocked on a transaction trying to commit. With start
1921 * we inc the extwriter counter, so we wait for all extwriters to exit
1922 * before we start blocking joiners. This comment is to keep somebody
1923 * from thinking they are super smart and changing this to
1924 * btrfs_join_transaction *cough*Josef*cough*.
1925 */
1926 trans = btrfs_start_transaction(root, 0);
1927 if (IS_ERR(trans)) {
1928 ret = PTR_ERR(trans);
1929 goto out_release_extents;
1930 }
1931 trans->in_fsync = true;
1932
1933 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
1934 /*
1935 * Scratch eb no longer needed, release before syncing log or commit
1936 * transaction, to avoid holding unnecessary memory during such long
1937 * operations.
1938 */
1939 if (ctx.scratch_eb) {
1940 free_extent_buffer(ctx.scratch_eb);
1941 ctx.scratch_eb = NULL;
1942 }
1943 btrfs_release_log_ctx_extents(&ctx);
1944 if (ret < 0) {
1945 /* Fallthrough and commit/free transaction. */
1946 ret = BTRFS_LOG_FORCE_COMMIT;
1947 }
1948
1949 /* we've logged all the items and now have a consistent
1950 * version of the file in the log. It is possible that
1951 * someone will come in and modify the file, but that's
1952 * fine because the log is consistent on disk, and we
1953 * have references to all of the file's extents
1954 *
1955 * It is possible that someone will come in and log the
1956 * file again, but that will end up using the synchronization
1957 * inside btrfs_sync_log to keep things safe.
1958 */
1959 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1960
1961 if (ret == BTRFS_NO_LOG_SYNC) {
1962 ret = btrfs_end_transaction(trans);
1963 goto out;
1964 }
1965
1966 /* We successfully logged the inode, attempt to sync the log. */
1967 if (!ret) {
1968 ret = btrfs_sync_log(trans, root, &ctx);
1969 if (!ret) {
1970 ret = btrfs_end_transaction(trans);
1971 goto out;
1972 }
1973 }
1974
1975 /*
1976 * At this point we need to commit the transaction because we had
1977 * btrfs_need_log_full_commit() or some other error.
1978 *
1979 * If we didn't do a full sync we have to stop the trans handle, wait on
1980 * the ordered extents, start it again and commit the transaction. If
1981 * we attempt to wait on the ordered extents here we could deadlock with
1982 * something like fallocate() that is holding the extent lock trying to
1983 * start a transaction while some other thread is trying to commit the
1984 * transaction while we (fsync) are currently holding the transaction
1985 * open.
1986 */
1987 if (!full_sync) {
1988 ret = btrfs_end_transaction(trans);
1989 if (ret)
1990 goto out;
1991 ret = btrfs_wait_ordered_range(inode, start, len);
1992 if (ret)
1993 goto out;
1994
1995 /*
1996 * This is safe to use here because we're only interested in
1997 * making sure the transaction that had the ordered extents is
1998 * committed. We aren't waiting on anything past this point,
1999 * we're purely getting the transaction and committing it.
2000 */
2001 trans = btrfs_attach_transaction_barrier(root);
2002 if (IS_ERR(trans)) {
2003 ret = PTR_ERR(trans);
2004
2005 /*
2006 * We committed the transaction and there's no currently
2007 * running transaction, this means everything we care
2008 * about made it to disk and we are done.
2009 */
2010 if (ret == -ENOENT)
2011 ret = 0;
2012 goto out;
2013 }
2014 }
2015
2016 ret = btrfs_commit_transaction(trans);
2017out:
2018 free_extent_buffer(ctx.scratch_eb);
2019 ASSERT(list_empty(&ctx.list));
2020 ASSERT(list_empty(&ctx.conflict_inodes));
2021 err = file_check_and_advance_wb_err(file);
2022 if (!ret)
2023 ret = err;
2024 return ret > 0 ? -EIO : ret;
2025
2026out_release_extents:
2027 btrfs_release_log_ctx_extents(&ctx);
2028 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2029 goto out;
2030}
2031
2032static const struct vm_operations_struct btrfs_file_vm_ops = {
2033 .fault = filemap_fault,
2034 .map_pages = filemap_map_pages,
2035 .page_mkwrite = btrfs_page_mkwrite,
2036};
2037
2038static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2039{
2040 struct address_space *mapping = filp->f_mapping;
2041
2042 if (!mapping->a_ops->read_folio)
2043 return -ENOEXEC;
2044
2045 file_accessed(filp);
2046 vma->vm_ops = &btrfs_file_vm_ops;
2047
2048 return 0;
2049}
2050
2051static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2052 int slot, u64 start, u64 end)
2053{
2054 struct btrfs_file_extent_item *fi;
2055 struct btrfs_key key;
2056
2057 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2058 return 0;
2059
2060 btrfs_item_key_to_cpu(leaf, &key, slot);
2061 if (key.objectid != btrfs_ino(inode) ||
2062 key.type != BTRFS_EXTENT_DATA_KEY)
2063 return 0;
2064
2065 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2066
2067 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2068 return 0;
2069
2070 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2071 return 0;
2072
2073 if (key.offset == end)
2074 return 1;
2075 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2076 return 1;
2077 return 0;
2078}
2079
2080static int fill_holes(struct btrfs_trans_handle *trans,
2081 struct btrfs_inode *inode,
2082 struct btrfs_path *path, u64 offset, u64 end)
2083{
2084 struct btrfs_fs_info *fs_info = trans->fs_info;
2085 struct btrfs_root *root = inode->root;
2086 struct extent_buffer *leaf;
2087 struct btrfs_file_extent_item *fi;
2088 struct extent_map *hole_em;
2089 struct btrfs_key key;
2090 int ret;
2091
2092 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2093 goto out;
2094
2095 key.objectid = btrfs_ino(inode);
2096 key.type = BTRFS_EXTENT_DATA_KEY;
2097 key.offset = offset;
2098
2099 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2100 if (ret <= 0) {
2101 /*
2102 * We should have dropped this offset, so if we find it then
2103 * something has gone horribly wrong.
2104 */
2105 if (ret == 0)
2106 ret = -EINVAL;
2107 return ret;
2108 }
2109
2110 leaf = path->nodes[0];
2111 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2112 u64 num_bytes;
2113
2114 path->slots[0]--;
2115 fi = btrfs_item_ptr(leaf, path->slots[0],
2116 struct btrfs_file_extent_item);
2117 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2118 end - offset;
2119 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2120 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2121 btrfs_set_file_extent_offset(leaf, fi, 0);
2122 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2123 btrfs_mark_buffer_dirty(trans, leaf);
2124 goto out;
2125 }
2126
2127 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2128 u64 num_bytes;
2129
2130 key.offset = offset;
2131 btrfs_set_item_key_safe(trans, path, &key);
2132 fi = btrfs_item_ptr(leaf, path->slots[0],
2133 struct btrfs_file_extent_item);
2134 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2135 offset;
2136 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2137 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2138 btrfs_set_file_extent_offset(leaf, fi, 0);
2139 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2140 btrfs_mark_buffer_dirty(trans, leaf);
2141 goto out;
2142 }
2143 btrfs_release_path(path);
2144
2145 ret = btrfs_insert_hole_extent(trans, root, btrfs_ino(inode), offset,
2146 end - offset);
2147 if (ret)
2148 return ret;
2149
2150out:
2151 btrfs_release_path(path);
2152
2153 hole_em = alloc_extent_map();
2154 if (!hole_em) {
2155 btrfs_drop_extent_map_range(inode, offset, end - 1, false);
2156 btrfs_set_inode_full_sync(inode);
2157 } else {
2158 hole_em->start = offset;
2159 hole_em->len = end - offset;
2160 hole_em->ram_bytes = hole_em->len;
2161 hole_em->orig_start = offset;
2162
2163 hole_em->block_start = EXTENT_MAP_HOLE;
2164 hole_em->block_len = 0;
2165 hole_em->orig_block_len = 0;
2166 hole_em->generation = trans->transid;
2167
2168 ret = btrfs_replace_extent_map_range(inode, hole_em, true);
2169 free_extent_map(hole_em);
2170 if (ret)
2171 btrfs_set_inode_full_sync(inode);
2172 }
2173
2174 return 0;
2175}
2176
2177/*
2178 * Find a hole extent on given inode and change start/len to the end of hole
2179 * extent.(hole/vacuum extent whose em->start <= start &&
2180 * em->start + em->len > start)
2181 * When a hole extent is found, return 1 and modify start/len.
2182 */
2183static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2184{
2185 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2186 struct extent_map *em;
2187 int ret = 0;
2188
2189 em = btrfs_get_extent(inode, NULL,
2190 round_down(*start, fs_info->sectorsize),
2191 round_up(*len, fs_info->sectorsize));
2192 if (IS_ERR(em))
2193 return PTR_ERR(em);
2194
2195 /* Hole or vacuum extent(only exists in no-hole mode) */
2196 if (em->block_start == EXTENT_MAP_HOLE) {
2197 ret = 1;
2198 *len = em->start + em->len > *start + *len ?
2199 0 : *start + *len - em->start - em->len;
2200 *start = em->start + em->len;
2201 }
2202 free_extent_map(em);
2203 return ret;
2204}
2205
2206static void btrfs_punch_hole_lock_range(struct inode *inode,
2207 const u64 lockstart,
2208 const u64 lockend,
2209 struct extent_state **cached_state)
2210{
2211 /*
2212 * For subpage case, if the range is not at page boundary, we could
2213 * have pages at the leading/tailing part of the range.
2214 * This could lead to dead loop since filemap_range_has_page()
2215 * will always return true.
2216 * So here we need to do extra page alignment for
2217 * filemap_range_has_page().
2218 */
2219 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2220 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2221
2222 while (1) {
2223 truncate_pagecache_range(inode, lockstart, lockend);
2224
2225 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2226 cached_state);
2227 /*
2228 * We can't have ordered extents in the range, nor dirty/writeback
2229 * pages, because we have locked the inode's VFS lock in exclusive
2230 * mode, we have locked the inode's i_mmap_lock in exclusive mode,
2231 * we have flushed all delalloc in the range and we have waited
2232 * for any ordered extents in the range to complete.
2233 * We can race with anyone reading pages from this range, so after
2234 * locking the range check if we have pages in the range, and if
2235 * we do, unlock the range and retry.
2236 */
2237 if (!filemap_range_has_page(inode->i_mapping, page_lockstart,
2238 page_lockend))
2239 break;
2240
2241 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2242 cached_state);
2243 }
2244
2245 btrfs_assert_inode_range_clean(BTRFS_I(inode), lockstart, lockend);
2246}
2247
2248static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2249 struct btrfs_inode *inode,
2250 struct btrfs_path *path,
2251 struct btrfs_replace_extent_info *extent_info,
2252 const u64 replace_len,
2253 const u64 bytes_to_drop)
2254{
2255 struct btrfs_fs_info *fs_info = trans->fs_info;
2256 struct btrfs_root *root = inode->root;
2257 struct btrfs_file_extent_item *extent;
2258 struct extent_buffer *leaf;
2259 struct btrfs_key key;
2260 int slot;
2261 struct btrfs_ref ref = { 0 };
2262 int ret;
2263
2264 if (replace_len == 0)
2265 return 0;
2266
2267 if (extent_info->disk_offset == 0 &&
2268 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2269 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2270 return 0;
2271 }
2272
2273 key.objectid = btrfs_ino(inode);
2274 key.type = BTRFS_EXTENT_DATA_KEY;
2275 key.offset = extent_info->file_offset;
2276 ret = btrfs_insert_empty_item(trans, root, path, &key,
2277 sizeof(struct btrfs_file_extent_item));
2278 if (ret)
2279 return ret;
2280 leaf = path->nodes[0];
2281 slot = path->slots[0];
2282 write_extent_buffer(leaf, extent_info->extent_buf,
2283 btrfs_item_ptr_offset(leaf, slot),
2284 sizeof(struct btrfs_file_extent_item));
2285 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2286 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2287 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2288 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2289 if (extent_info->is_new_extent)
2290 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2291 btrfs_mark_buffer_dirty(trans, leaf);
2292 btrfs_release_path(path);
2293
2294 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2295 replace_len);
2296 if (ret)
2297 return ret;
2298
2299 /* If it's a hole, nothing more needs to be done. */
2300 if (extent_info->disk_offset == 0) {
2301 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2302 return 0;
2303 }
2304
2305 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2306
2307 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2308 key.objectid = extent_info->disk_offset;
2309 key.type = BTRFS_EXTENT_ITEM_KEY;
2310 key.offset = extent_info->disk_len;
2311 ret = btrfs_alloc_reserved_file_extent(trans, root,
2312 btrfs_ino(inode),
2313 extent_info->file_offset,
2314 extent_info->qgroup_reserved,
2315 &key);
2316 } else {
2317 u64 ref_offset;
2318
2319 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2320 extent_info->disk_offset,
2321 extent_info->disk_len, 0,
2322 root->root_key.objectid);
2323 ref_offset = extent_info->file_offset - extent_info->data_offset;
2324 btrfs_init_data_ref(&ref, root->root_key.objectid,
2325 btrfs_ino(inode), ref_offset, 0, false);
2326 ret = btrfs_inc_extent_ref(trans, &ref);
2327 }
2328
2329 extent_info->insertions++;
2330
2331 return ret;
2332}
2333
2334/*
2335 * The respective range must have been previously locked, as well as the inode.
2336 * The end offset is inclusive (last byte of the range).
2337 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2338 * the file range with an extent.
2339 * When not punching a hole, we don't want to end up in a state where we dropped
2340 * extents without inserting a new one, so we must abort the transaction to avoid
2341 * a corruption.
2342 */
2343int btrfs_replace_file_extents(struct btrfs_inode *inode,
2344 struct btrfs_path *path, const u64 start,
2345 const u64 end,
2346 struct btrfs_replace_extent_info *extent_info,
2347 struct btrfs_trans_handle **trans_out)
2348{
2349 struct btrfs_drop_extents_args drop_args = { 0 };
2350 struct btrfs_root *root = inode->root;
2351 struct btrfs_fs_info *fs_info = root->fs_info;
2352 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2353 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2354 struct btrfs_trans_handle *trans = NULL;
2355 struct btrfs_block_rsv *rsv;
2356 unsigned int rsv_count;
2357 u64 cur_offset;
2358 u64 len = end - start;
2359 int ret = 0;
2360
2361 if (end <= start)
2362 return -EINVAL;
2363
2364 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2365 if (!rsv) {
2366 ret = -ENOMEM;
2367 goto out;
2368 }
2369 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2370 rsv->failfast = true;
2371
2372 /*
2373 * 1 - update the inode
2374 * 1 - removing the extents in the range
2375 * 1 - adding the hole extent if no_holes isn't set or if we are
2376 * replacing the range with a new extent
2377 */
2378 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2379 rsv_count = 3;
2380 else
2381 rsv_count = 2;
2382
2383 trans = btrfs_start_transaction(root, rsv_count);
2384 if (IS_ERR(trans)) {
2385 ret = PTR_ERR(trans);
2386 trans = NULL;
2387 goto out_free;
2388 }
2389
2390 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2391 min_size, false);
2392 if (WARN_ON(ret))
2393 goto out_trans;
2394 trans->block_rsv = rsv;
2395
2396 cur_offset = start;
2397 drop_args.path = path;
2398 drop_args.end = end + 1;
2399 drop_args.drop_cache = true;
2400 while (cur_offset < end) {
2401 drop_args.start = cur_offset;
2402 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2403 /* If we are punching a hole decrement the inode's byte count */
2404 if (!extent_info)
2405 btrfs_update_inode_bytes(inode, 0,
2406 drop_args.bytes_found);
2407 if (ret != -ENOSPC) {
2408 /*
2409 * The only time we don't want to abort is if we are
2410 * attempting to clone a partial inline extent, in which
2411 * case we'll get EOPNOTSUPP. However if we aren't
2412 * clone we need to abort no matter what, because if we
2413 * got EOPNOTSUPP via prealloc then we messed up and
2414 * need to abort.
2415 */
2416 if (ret &&
2417 (ret != -EOPNOTSUPP ||
2418 (extent_info && extent_info->is_new_extent)))
2419 btrfs_abort_transaction(trans, ret);
2420 break;
2421 }
2422
2423 trans->block_rsv = &fs_info->trans_block_rsv;
2424
2425 if (!extent_info && cur_offset < drop_args.drop_end &&
2426 cur_offset < ino_size) {
2427 ret = fill_holes(trans, inode, path, cur_offset,
2428 drop_args.drop_end);
2429 if (ret) {
2430 /*
2431 * If we failed then we didn't insert our hole
2432 * entries for the area we dropped, so now the
2433 * fs is corrupted, so we must abort the
2434 * transaction.
2435 */
2436 btrfs_abort_transaction(trans, ret);
2437 break;
2438 }
2439 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2440 /*
2441 * We are past the i_size here, but since we didn't
2442 * insert holes we need to clear the mapped area so we
2443 * know to not set disk_i_size in this area until a new
2444 * file extent is inserted here.
2445 */
2446 ret = btrfs_inode_clear_file_extent_range(inode,
2447 cur_offset,
2448 drop_args.drop_end - cur_offset);
2449 if (ret) {
2450 /*
2451 * We couldn't clear our area, so we could
2452 * presumably adjust up and corrupt the fs, so
2453 * we need to abort.
2454 */
2455 btrfs_abort_transaction(trans, ret);
2456 break;
2457 }
2458 }
2459
2460 if (extent_info &&
2461 drop_args.drop_end > extent_info->file_offset) {
2462 u64 replace_len = drop_args.drop_end -
2463 extent_info->file_offset;
2464
2465 ret = btrfs_insert_replace_extent(trans, inode, path,
2466 extent_info, replace_len,
2467 drop_args.bytes_found);
2468 if (ret) {
2469 btrfs_abort_transaction(trans, ret);
2470 break;
2471 }
2472 extent_info->data_len -= replace_len;
2473 extent_info->data_offset += replace_len;
2474 extent_info->file_offset += replace_len;
2475 }
2476
2477 /*
2478 * We are releasing our handle on the transaction, balance the
2479 * dirty pages of the btree inode and flush delayed items, and
2480 * then get a new transaction handle, which may now point to a
2481 * new transaction in case someone else may have committed the
2482 * transaction we used to replace/drop file extent items. So
2483 * bump the inode's iversion and update mtime and ctime except
2484 * if we are called from a dedupe context. This is because a
2485 * power failure/crash may happen after the transaction is
2486 * committed and before we finish replacing/dropping all the
2487 * file extent items we need.
2488 */
2489 inode_inc_iversion(&inode->vfs_inode);
2490
2491 if (!extent_info || extent_info->update_times)
2492 inode_set_mtime_to_ts(&inode->vfs_inode,
2493 inode_set_ctime_current(&inode->vfs_inode));
2494
2495 ret = btrfs_update_inode(trans, inode);
2496 if (ret)
2497 break;
2498
2499 btrfs_end_transaction(trans);
2500 btrfs_btree_balance_dirty(fs_info);
2501
2502 trans = btrfs_start_transaction(root, rsv_count);
2503 if (IS_ERR(trans)) {
2504 ret = PTR_ERR(trans);
2505 trans = NULL;
2506 break;
2507 }
2508
2509 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2510 rsv, min_size, false);
2511 if (WARN_ON(ret))
2512 break;
2513 trans->block_rsv = rsv;
2514
2515 cur_offset = drop_args.drop_end;
2516 len = end - cur_offset;
2517 if (!extent_info && len) {
2518 ret = find_first_non_hole(inode, &cur_offset, &len);
2519 if (unlikely(ret < 0))
2520 break;
2521 if (ret && !len) {
2522 ret = 0;
2523 break;
2524 }
2525 }
2526 }
2527
2528 /*
2529 * If we were cloning, force the next fsync to be a full one since we
2530 * we replaced (or just dropped in the case of cloning holes when
2531 * NO_HOLES is enabled) file extent items and did not setup new extent
2532 * maps for the replacement extents (or holes).
2533 */
2534 if (extent_info && !extent_info->is_new_extent)
2535 btrfs_set_inode_full_sync(inode);
2536
2537 if (ret)
2538 goto out_trans;
2539
2540 trans->block_rsv = &fs_info->trans_block_rsv;
2541 /*
2542 * If we are using the NO_HOLES feature we might have had already an
2543 * hole that overlaps a part of the region [lockstart, lockend] and
2544 * ends at (or beyond) lockend. Since we have no file extent items to
2545 * represent holes, drop_end can be less than lockend and so we must
2546 * make sure we have an extent map representing the existing hole (the
2547 * call to __btrfs_drop_extents() might have dropped the existing extent
2548 * map representing the existing hole), otherwise the fast fsync path
2549 * will not record the existence of the hole region
2550 * [existing_hole_start, lockend].
2551 */
2552 if (drop_args.drop_end <= end)
2553 drop_args.drop_end = end + 1;
2554 /*
2555 * Don't insert file hole extent item if it's for a range beyond eof
2556 * (because it's useless) or if it represents a 0 bytes range (when
2557 * cur_offset == drop_end).
2558 */
2559 if (!extent_info && cur_offset < ino_size &&
2560 cur_offset < drop_args.drop_end) {
2561 ret = fill_holes(trans, inode, path, cur_offset,
2562 drop_args.drop_end);
2563 if (ret) {
2564 /* Same comment as above. */
2565 btrfs_abort_transaction(trans, ret);
2566 goto out_trans;
2567 }
2568 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2569 /* See the comment in the loop above for the reasoning here. */
2570 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2571 drop_args.drop_end - cur_offset);
2572 if (ret) {
2573 btrfs_abort_transaction(trans, ret);
2574 goto out_trans;
2575 }
2576
2577 }
2578 if (extent_info) {
2579 ret = btrfs_insert_replace_extent(trans, inode, path,
2580 extent_info, extent_info->data_len,
2581 drop_args.bytes_found);
2582 if (ret) {
2583 btrfs_abort_transaction(trans, ret);
2584 goto out_trans;
2585 }
2586 }
2587
2588out_trans:
2589 if (!trans)
2590 goto out_free;
2591
2592 trans->block_rsv = &fs_info->trans_block_rsv;
2593 if (ret)
2594 btrfs_end_transaction(trans);
2595 else
2596 *trans_out = trans;
2597out_free:
2598 btrfs_free_block_rsv(fs_info, rsv);
2599out:
2600 return ret;
2601}
2602
2603static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2604{
2605 struct inode *inode = file_inode(file);
2606 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2607 struct btrfs_root *root = BTRFS_I(inode)->root;
2608 struct extent_state *cached_state = NULL;
2609 struct btrfs_path *path;
2610 struct btrfs_trans_handle *trans = NULL;
2611 u64 lockstart;
2612 u64 lockend;
2613 u64 tail_start;
2614 u64 tail_len;
2615 u64 orig_start = offset;
2616 int ret = 0;
2617 bool same_block;
2618 u64 ino_size;
2619 bool truncated_block = false;
2620 bool updated_inode = false;
2621
2622 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2623
2624 ret = btrfs_wait_ordered_range(inode, offset, len);
2625 if (ret)
2626 goto out_only_mutex;
2627
2628 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2629 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2630 if (ret < 0)
2631 goto out_only_mutex;
2632 if (ret && !len) {
2633 /* Already in a large hole */
2634 ret = 0;
2635 goto out_only_mutex;
2636 }
2637
2638 ret = file_modified(file);
2639 if (ret)
2640 goto out_only_mutex;
2641
2642 lockstart = round_up(offset, fs_info->sectorsize);
2643 lockend = round_down(offset + len, fs_info->sectorsize) - 1;
2644 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2645 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2646 /*
2647 * We needn't truncate any block which is beyond the end of the file
2648 * because we are sure there is no data there.
2649 */
2650 /*
2651 * Only do this if we are in the same block and we aren't doing the
2652 * entire block.
2653 */
2654 if (same_block && len < fs_info->sectorsize) {
2655 if (offset < ino_size) {
2656 truncated_block = true;
2657 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2658 0);
2659 } else {
2660 ret = 0;
2661 }
2662 goto out_only_mutex;
2663 }
2664
2665 /* zero back part of the first block */
2666 if (offset < ino_size) {
2667 truncated_block = true;
2668 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2669 if (ret) {
2670 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2671 return ret;
2672 }
2673 }
2674
2675 /* Check the aligned pages after the first unaligned page,
2676 * if offset != orig_start, which means the first unaligned page
2677 * including several following pages are already in holes,
2678 * the extra check can be skipped */
2679 if (offset == orig_start) {
2680 /* after truncate page, check hole again */
2681 len = offset + len - lockstart;
2682 offset = lockstart;
2683 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2684 if (ret < 0)
2685 goto out_only_mutex;
2686 if (ret && !len) {
2687 ret = 0;
2688 goto out_only_mutex;
2689 }
2690 lockstart = offset;
2691 }
2692
2693 /* Check the tail unaligned part is in a hole */
2694 tail_start = lockend + 1;
2695 tail_len = offset + len - tail_start;
2696 if (tail_len) {
2697 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2698 if (unlikely(ret < 0))
2699 goto out_only_mutex;
2700 if (!ret) {
2701 /* zero the front end of the last page */
2702 if (tail_start + tail_len < ino_size) {
2703 truncated_block = true;
2704 ret = btrfs_truncate_block(BTRFS_I(inode),
2705 tail_start + tail_len,
2706 0, 1);
2707 if (ret)
2708 goto out_only_mutex;
2709 }
2710 }
2711 }
2712
2713 if (lockend < lockstart) {
2714 ret = 0;
2715 goto out_only_mutex;
2716 }
2717
2718 btrfs_punch_hole_lock_range(inode, lockstart, lockend, &cached_state);
2719
2720 path = btrfs_alloc_path();
2721 if (!path) {
2722 ret = -ENOMEM;
2723 goto out;
2724 }
2725
2726 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
2727 lockend, NULL, &trans);
2728 btrfs_free_path(path);
2729 if (ret)
2730 goto out;
2731
2732 ASSERT(trans != NULL);
2733 inode_inc_iversion(inode);
2734 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
2735 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2736 updated_inode = true;
2737 btrfs_end_transaction(trans);
2738 btrfs_btree_balance_dirty(fs_info);
2739out:
2740 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2741 &cached_state);
2742out_only_mutex:
2743 if (!updated_inode && truncated_block && !ret) {
2744 /*
2745 * If we only end up zeroing part of a page, we still need to
2746 * update the inode item, so that all the time fields are
2747 * updated as well as the necessary btrfs inode in memory fields
2748 * for detecting, at fsync time, if the inode isn't yet in the
2749 * log tree or it's there but not up to date.
2750 */
2751 struct timespec64 now = inode_set_ctime_current(inode);
2752
2753 inode_inc_iversion(inode);
2754 inode_set_mtime_to_ts(inode, now);
2755 trans = btrfs_start_transaction(root, 1);
2756 if (IS_ERR(trans)) {
2757 ret = PTR_ERR(trans);
2758 } else {
2759 int ret2;
2760
2761 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2762 ret2 = btrfs_end_transaction(trans);
2763 if (!ret)
2764 ret = ret2;
2765 }
2766 }
2767 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2768 return ret;
2769}
2770
2771/* Helper structure to record which range is already reserved */
2772struct falloc_range {
2773 struct list_head list;
2774 u64 start;
2775 u64 len;
2776};
2777
2778/*
2779 * Helper function to add falloc range
2780 *
2781 * Caller should have locked the larger range of extent containing
2782 * [start, len)
2783 */
2784static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2785{
2786 struct falloc_range *range = NULL;
2787
2788 if (!list_empty(head)) {
2789 /*
2790 * As fallocate iterates by bytenr order, we only need to check
2791 * the last range.
2792 */
2793 range = list_last_entry(head, struct falloc_range, list);
2794 if (range->start + range->len == start) {
2795 range->len += len;
2796 return 0;
2797 }
2798 }
2799
2800 range = kmalloc(sizeof(*range), GFP_KERNEL);
2801 if (!range)
2802 return -ENOMEM;
2803 range->start = start;
2804 range->len = len;
2805 list_add_tail(&range->list, head);
2806 return 0;
2807}
2808
2809static int btrfs_fallocate_update_isize(struct inode *inode,
2810 const u64 end,
2811 const int mode)
2812{
2813 struct btrfs_trans_handle *trans;
2814 struct btrfs_root *root = BTRFS_I(inode)->root;
2815 int ret;
2816 int ret2;
2817
2818 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2819 return 0;
2820
2821 trans = btrfs_start_transaction(root, 1);
2822 if (IS_ERR(trans))
2823 return PTR_ERR(trans);
2824
2825 inode_set_ctime_current(inode);
2826 i_size_write(inode, end);
2827 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
2828 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2829 ret2 = btrfs_end_transaction(trans);
2830
2831 return ret ? ret : ret2;
2832}
2833
2834enum {
2835 RANGE_BOUNDARY_WRITTEN_EXTENT,
2836 RANGE_BOUNDARY_PREALLOC_EXTENT,
2837 RANGE_BOUNDARY_HOLE,
2838};
2839
2840static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
2841 u64 offset)
2842{
2843 const u64 sectorsize = inode->root->fs_info->sectorsize;
2844 struct extent_map *em;
2845 int ret;
2846
2847 offset = round_down(offset, sectorsize);
2848 em = btrfs_get_extent(inode, NULL, offset, sectorsize);
2849 if (IS_ERR(em))
2850 return PTR_ERR(em);
2851
2852 if (em->block_start == EXTENT_MAP_HOLE)
2853 ret = RANGE_BOUNDARY_HOLE;
2854 else if (em->flags & EXTENT_FLAG_PREALLOC)
2855 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2856 else
2857 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2858
2859 free_extent_map(em);
2860 return ret;
2861}
2862
2863static int btrfs_zero_range(struct inode *inode,
2864 loff_t offset,
2865 loff_t len,
2866 const int mode)
2867{
2868 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2869 struct extent_map *em;
2870 struct extent_changeset *data_reserved = NULL;
2871 int ret;
2872 u64 alloc_hint = 0;
2873 const u64 sectorsize = fs_info->sectorsize;
2874 u64 alloc_start = round_down(offset, sectorsize);
2875 u64 alloc_end = round_up(offset + len, sectorsize);
2876 u64 bytes_to_reserve = 0;
2877 bool space_reserved = false;
2878
2879 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start,
2880 alloc_end - alloc_start);
2881 if (IS_ERR(em)) {
2882 ret = PTR_ERR(em);
2883 goto out;
2884 }
2885
2886 /*
2887 * Avoid hole punching and extent allocation for some cases. More cases
2888 * could be considered, but these are unlikely common and we keep things
2889 * as simple as possible for now. Also, intentionally, if the target
2890 * range contains one or more prealloc extents together with regular
2891 * extents and holes, we drop all the existing extents and allocate a
2892 * new prealloc extent, so that we get a larger contiguous disk extent.
2893 */
2894 if (em->start <= alloc_start && (em->flags & EXTENT_FLAG_PREALLOC)) {
2895 const u64 em_end = em->start + em->len;
2896
2897 if (em_end >= offset + len) {
2898 /*
2899 * The whole range is already a prealloc extent,
2900 * do nothing except updating the inode's i_size if
2901 * needed.
2902 */
2903 free_extent_map(em);
2904 ret = btrfs_fallocate_update_isize(inode, offset + len,
2905 mode);
2906 goto out;
2907 }
2908 /*
2909 * Part of the range is already a prealloc extent, so operate
2910 * only on the remaining part of the range.
2911 */
2912 alloc_start = em_end;
2913 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2914 len = offset + len - alloc_start;
2915 offset = alloc_start;
2916 alloc_hint = em->block_start + em->len;
2917 }
2918 free_extent_map(em);
2919
2920 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2921 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2922 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start, sectorsize);
2923 if (IS_ERR(em)) {
2924 ret = PTR_ERR(em);
2925 goto out;
2926 }
2927
2928 if (em->flags & EXTENT_FLAG_PREALLOC) {
2929 free_extent_map(em);
2930 ret = btrfs_fallocate_update_isize(inode, offset + len,
2931 mode);
2932 goto out;
2933 }
2934 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2935 free_extent_map(em);
2936 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2937 0);
2938 if (!ret)
2939 ret = btrfs_fallocate_update_isize(inode,
2940 offset + len,
2941 mode);
2942 return ret;
2943 }
2944 free_extent_map(em);
2945 alloc_start = round_down(offset, sectorsize);
2946 alloc_end = alloc_start + sectorsize;
2947 goto reserve_space;
2948 }
2949
2950 alloc_start = round_up(offset, sectorsize);
2951 alloc_end = round_down(offset + len, sectorsize);
2952
2953 /*
2954 * For unaligned ranges, check the pages at the boundaries, they might
2955 * map to an extent, in which case we need to partially zero them, or
2956 * they might map to a hole, in which case we need our allocation range
2957 * to cover them.
2958 */
2959 if (!IS_ALIGNED(offset, sectorsize)) {
2960 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2961 offset);
2962 if (ret < 0)
2963 goto out;
2964 if (ret == RANGE_BOUNDARY_HOLE) {
2965 alloc_start = round_down(offset, sectorsize);
2966 ret = 0;
2967 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2968 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2969 if (ret)
2970 goto out;
2971 } else {
2972 ret = 0;
2973 }
2974 }
2975
2976 if (!IS_ALIGNED(offset + len, sectorsize)) {
2977 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2978 offset + len);
2979 if (ret < 0)
2980 goto out;
2981 if (ret == RANGE_BOUNDARY_HOLE) {
2982 alloc_end = round_up(offset + len, sectorsize);
2983 ret = 0;
2984 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2985 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
2986 0, 1);
2987 if (ret)
2988 goto out;
2989 } else {
2990 ret = 0;
2991 }
2992 }
2993
2994reserve_space:
2995 if (alloc_start < alloc_end) {
2996 struct extent_state *cached_state = NULL;
2997 const u64 lockstart = alloc_start;
2998 const u64 lockend = alloc_end - 1;
2999
3000 bytes_to_reserve = alloc_end - alloc_start;
3001 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3002 bytes_to_reserve);
3003 if (ret < 0)
3004 goto out;
3005 space_reserved = true;
3006 btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3007 &cached_state);
3008 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3009 alloc_start, bytes_to_reserve);
3010 if (ret) {
3011 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
3012 lockend, &cached_state);
3013 goto out;
3014 }
3015 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3016 alloc_end - alloc_start,
3017 fs_info->sectorsize,
3018 offset + len, &alloc_hint);
3019 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3020 &cached_state);
3021 /* btrfs_prealloc_file_range releases reserved space on error */
3022 if (ret) {
3023 space_reserved = false;
3024 goto out;
3025 }
3026 }
3027 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3028 out:
3029 if (ret && space_reserved)
3030 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3031 alloc_start, bytes_to_reserve);
3032 extent_changeset_free(data_reserved);
3033
3034 return ret;
3035}
3036
3037static long btrfs_fallocate(struct file *file, int mode,
3038 loff_t offset, loff_t len)
3039{
3040 struct inode *inode = file_inode(file);
3041 struct extent_state *cached_state = NULL;
3042 struct extent_changeset *data_reserved = NULL;
3043 struct falloc_range *range;
3044 struct falloc_range *tmp;
3045 LIST_HEAD(reserve_list);
3046 u64 cur_offset;
3047 u64 last_byte;
3048 u64 alloc_start;
3049 u64 alloc_end;
3050 u64 alloc_hint = 0;
3051 u64 locked_end;
3052 u64 actual_end = 0;
3053 u64 data_space_needed = 0;
3054 u64 data_space_reserved = 0;
3055 u64 qgroup_reserved = 0;
3056 struct extent_map *em;
3057 int blocksize = BTRFS_I(inode)->root->fs_info->sectorsize;
3058 int ret;
3059
3060 /* Do not allow fallocate in ZONED mode */
3061 if (btrfs_is_zoned(inode_to_fs_info(inode)))
3062 return -EOPNOTSUPP;
3063
3064 alloc_start = round_down(offset, blocksize);
3065 alloc_end = round_up(offset + len, blocksize);
3066 cur_offset = alloc_start;
3067
3068 /* Make sure we aren't being give some crap mode */
3069 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3070 FALLOC_FL_ZERO_RANGE))
3071 return -EOPNOTSUPP;
3072
3073 if (mode & FALLOC_FL_PUNCH_HOLE)
3074 return btrfs_punch_hole(file, offset, len);
3075
3076 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3077
3078 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3079 ret = inode_newsize_ok(inode, offset + len);
3080 if (ret)
3081 goto out;
3082 }
3083
3084 ret = file_modified(file);
3085 if (ret)
3086 goto out;
3087
3088 /*
3089 * TODO: Move these two operations after we have checked
3090 * accurate reserved space, or fallocate can still fail but
3091 * with page truncated or size expanded.
3092 *
3093 * But that's a minor problem and won't do much harm BTW.
3094 */
3095 if (alloc_start > inode->i_size) {
3096 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3097 alloc_start);
3098 if (ret)
3099 goto out;
3100 } else if (offset + len > inode->i_size) {
3101 /*
3102 * If we are fallocating from the end of the file onward we
3103 * need to zero out the end of the block if i_size lands in the
3104 * middle of a block.
3105 */
3106 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3107 if (ret)
3108 goto out;
3109 }
3110
3111 /*
3112 * We have locked the inode at the VFS level (in exclusive mode) and we
3113 * have locked the i_mmap_lock lock (in exclusive mode). Now before
3114 * locking the file range, flush all dealloc in the range and wait for
3115 * all ordered extents in the range to complete. After this we can lock
3116 * the file range and, due to the previous locking we did, we know there
3117 * can't be more delalloc or ordered extents in the range.
3118 */
3119 ret = btrfs_wait_ordered_range(inode, alloc_start,
3120 alloc_end - alloc_start);
3121 if (ret)
3122 goto out;
3123
3124 if (mode & FALLOC_FL_ZERO_RANGE) {
3125 ret = btrfs_zero_range(inode, offset, len, mode);
3126 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3127 return ret;
3128 }
3129
3130 locked_end = alloc_end - 1;
3131 lock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3132 &cached_state);
3133
3134 btrfs_assert_inode_range_clean(BTRFS_I(inode), alloc_start, locked_end);
3135
3136 /* First, check if we exceed the qgroup limit */
3137 while (cur_offset < alloc_end) {
3138 em = btrfs_get_extent(BTRFS_I(inode), NULL, cur_offset,
3139 alloc_end - cur_offset);
3140 if (IS_ERR(em)) {
3141 ret = PTR_ERR(em);
3142 break;
3143 }
3144 last_byte = min(extent_map_end(em), alloc_end);
3145 actual_end = min_t(u64, extent_map_end(em), offset + len);
3146 last_byte = ALIGN(last_byte, blocksize);
3147 if (em->block_start == EXTENT_MAP_HOLE ||
3148 (cur_offset >= inode->i_size &&
3149 !(em->flags & EXTENT_FLAG_PREALLOC))) {
3150 const u64 range_len = last_byte - cur_offset;
3151
3152 ret = add_falloc_range(&reserve_list, cur_offset, range_len);
3153 if (ret < 0) {
3154 free_extent_map(em);
3155 break;
3156 }
3157 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3158 &data_reserved, cur_offset, range_len);
3159 if (ret < 0) {
3160 free_extent_map(em);
3161 break;
3162 }
3163 qgroup_reserved += range_len;
3164 data_space_needed += range_len;
3165 }
3166 free_extent_map(em);
3167 cur_offset = last_byte;
3168 }
3169
3170 if (!ret && data_space_needed > 0) {
3171 /*
3172 * We are safe to reserve space here as we can't have delalloc
3173 * in the range, see above.
3174 */
3175 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3176 data_space_needed);
3177 if (!ret)
3178 data_space_reserved = data_space_needed;
3179 }
3180
3181 /*
3182 * If ret is still 0, means we're OK to fallocate.
3183 * Or just cleanup the list and exit.
3184 */
3185 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3186 if (!ret) {
3187 ret = btrfs_prealloc_file_range(inode, mode,
3188 range->start,
3189 range->len, blocksize,
3190 offset + len, &alloc_hint);
3191 /*
3192 * btrfs_prealloc_file_range() releases space even
3193 * if it returns an error.
3194 */
3195 data_space_reserved -= range->len;
3196 qgroup_reserved -= range->len;
3197 } else if (data_space_reserved > 0) {
3198 btrfs_free_reserved_data_space(BTRFS_I(inode),
3199 data_reserved, range->start,
3200 range->len);
3201 data_space_reserved -= range->len;
3202 qgroup_reserved -= range->len;
3203 } else if (qgroup_reserved > 0) {
3204 btrfs_qgroup_free_data(BTRFS_I(inode), data_reserved,
3205 range->start, range->len, NULL);
3206 qgroup_reserved -= range->len;
3207 }
3208 list_del(&range->list);
3209 kfree(range);
3210 }
3211 if (ret < 0)
3212 goto out_unlock;
3213
3214 /*
3215 * We didn't need to allocate any more space, but we still extended the
3216 * size of the file so we need to update i_size and the inode item.
3217 */
3218 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3219out_unlock:
3220 unlock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3221 &cached_state);
3222out:
3223 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3224 extent_changeset_free(data_reserved);
3225 return ret;
3226}
3227
3228/*
3229 * Helper for btrfs_find_delalloc_in_range(). Find a subrange in a given range
3230 * that has unflushed and/or flushing delalloc. There might be other adjacent
3231 * subranges after the one it found, so btrfs_find_delalloc_in_range() keeps
3232 * looping while it gets adjacent subranges, and merging them together.
3233 */
3234static bool find_delalloc_subrange(struct btrfs_inode *inode, u64 start, u64 end,
3235 struct extent_state **cached_state,
3236 bool *search_io_tree,
3237 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3238{
3239 u64 len = end + 1 - start;
3240 u64 delalloc_len = 0;
3241 struct btrfs_ordered_extent *oe;
3242 u64 oe_start;
3243 u64 oe_end;
3244
3245 /*
3246 * Search the io tree first for EXTENT_DELALLOC. If we find any, it
3247 * means we have delalloc (dirty pages) for which writeback has not
3248 * started yet.
3249 */
3250 if (*search_io_tree) {
3251 spin_lock(&inode->lock);
3252 if (inode->delalloc_bytes > 0) {
3253 spin_unlock(&inode->lock);
3254 *delalloc_start_ret = start;
3255 delalloc_len = count_range_bits(&inode->io_tree,
3256 delalloc_start_ret, end,
3257 len, EXTENT_DELALLOC, 1,
3258 cached_state);
3259 } else {
3260 spin_unlock(&inode->lock);
3261 }
3262 }
3263
3264 if (delalloc_len > 0) {
3265 /*
3266 * If delalloc was found then *delalloc_start_ret has a sector size
3267 * aligned value (rounded down).
3268 */
3269 *delalloc_end_ret = *delalloc_start_ret + delalloc_len - 1;
3270
3271 if (*delalloc_start_ret == start) {
3272 /* Delalloc for the whole range, nothing more to do. */
3273 if (*delalloc_end_ret == end)
3274 return true;
3275 /* Else trim our search range for ordered extents. */
3276 start = *delalloc_end_ret + 1;
3277 len = end + 1 - start;
3278 }
3279 } else {
3280 /* No delalloc, future calls don't need to search again. */
3281 *search_io_tree = false;
3282 }
3283
3284 /*
3285 * Now also check if there's any ordered extent in the range.
3286 * We do this because:
3287 *
3288 * 1) When delalloc is flushed, the file range is locked, we clear the
3289 * EXTENT_DELALLOC bit from the io tree and create an extent map and
3290 * an ordered extent for the write. So we might just have been called
3291 * after delalloc is flushed and before the ordered extent completes
3292 * and inserts the new file extent item in the subvolume's btree;
3293 *
3294 * 2) We may have an ordered extent created by flushing delalloc for a
3295 * subrange that starts before the subrange we found marked with
3296 * EXTENT_DELALLOC in the io tree.
3297 *
3298 * We could also use the extent map tree to find such delalloc that is
3299 * being flushed, but using the ordered extents tree is more efficient
3300 * because it's usually much smaller as ordered extents are removed from
3301 * the tree once they complete. With the extent maps, we mau have them
3302 * in the extent map tree for a very long time, and they were either
3303 * created by previous writes or loaded by read operations.
3304 */
3305 oe = btrfs_lookup_first_ordered_range(inode, start, len);
3306 if (!oe)
3307 return (delalloc_len > 0);
3308
3309 /* The ordered extent may span beyond our search range. */
3310 oe_start = max(oe->file_offset, start);
3311 oe_end = min(oe->file_offset + oe->num_bytes - 1, end);
3312
3313 btrfs_put_ordered_extent(oe);
3314
3315 /* Don't have unflushed delalloc, return the ordered extent range. */
3316 if (delalloc_len == 0) {
3317 *delalloc_start_ret = oe_start;
3318 *delalloc_end_ret = oe_end;
3319 return true;
3320 }
3321
3322 /*
3323 * We have both unflushed delalloc (io_tree) and an ordered extent.
3324 * If the ranges are adjacent returned a combined range, otherwise
3325 * return the leftmost range.
3326 */
3327 if (oe_start < *delalloc_start_ret) {
3328 if (oe_end < *delalloc_start_ret)
3329 *delalloc_end_ret = oe_end;
3330 *delalloc_start_ret = oe_start;
3331 } else if (*delalloc_end_ret + 1 == oe_start) {
3332 *delalloc_end_ret = oe_end;
3333 }
3334
3335 return true;
3336}
3337
3338/*
3339 * Check if there's delalloc in a given range.
3340 *
3341 * @inode: The inode.
3342 * @start: The start offset of the range. It does not need to be
3343 * sector size aligned.
3344 * @end: The end offset (inclusive value) of the search range.
3345 * It does not need to be sector size aligned.
3346 * @cached_state: Extent state record used for speeding up delalloc
3347 * searches in the inode's io_tree. Can be NULL.
3348 * @delalloc_start_ret: Output argument, set to the start offset of the
3349 * subrange found with delalloc (may not be sector size
3350 * aligned).
3351 * @delalloc_end_ret: Output argument, set to he end offset (inclusive value)
3352 * of the subrange found with delalloc.
3353 *
3354 * Returns true if a subrange with delalloc is found within the given range, and
3355 * if so it sets @delalloc_start_ret and @delalloc_end_ret with the start and
3356 * end offsets of the subrange.
3357 */
3358bool btrfs_find_delalloc_in_range(struct btrfs_inode *inode, u64 start, u64 end,
3359 struct extent_state **cached_state,
3360 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3361{
3362 u64 cur_offset = round_down(start, inode->root->fs_info->sectorsize);
3363 u64 prev_delalloc_end = 0;
3364 bool search_io_tree = true;
3365 bool ret = false;
3366
3367 while (cur_offset <= end) {
3368 u64 delalloc_start;
3369 u64 delalloc_end;
3370 bool delalloc;
3371
3372 delalloc = find_delalloc_subrange(inode, cur_offset, end,
3373 cached_state, &search_io_tree,
3374 &delalloc_start,
3375 &delalloc_end);
3376 if (!delalloc)
3377 break;
3378
3379 if (prev_delalloc_end == 0) {
3380 /* First subrange found. */
3381 *delalloc_start_ret = max(delalloc_start, start);
3382 *delalloc_end_ret = delalloc_end;
3383 ret = true;
3384 } else if (delalloc_start == prev_delalloc_end + 1) {
3385 /* Subrange adjacent to the previous one, merge them. */
3386 *delalloc_end_ret = delalloc_end;
3387 } else {
3388 /* Subrange not adjacent to the previous one, exit. */
3389 break;
3390 }
3391
3392 prev_delalloc_end = delalloc_end;
3393 cur_offset = delalloc_end + 1;
3394 cond_resched();
3395 }
3396
3397 return ret;
3398}
3399
3400/*
3401 * Check if there's a hole or delalloc range in a range representing a hole (or
3402 * prealloc extent) found in the inode's subvolume btree.
3403 *
3404 * @inode: The inode.
3405 * @whence: Seek mode (SEEK_DATA or SEEK_HOLE).
3406 * @start: Start offset of the hole region. It does not need to be sector
3407 * size aligned.
3408 * @end: End offset (inclusive value) of the hole region. It does not
3409 * need to be sector size aligned.
3410 * @start_ret: Return parameter, used to set the start of the subrange in the
3411 * hole that matches the search criteria (seek mode), if such
3412 * subrange is found (return value of the function is true).
3413 * The value returned here may not be sector size aligned.
3414 *
3415 * Returns true if a subrange matching the given seek mode is found, and if one
3416 * is found, it updates @start_ret with the start of the subrange.
3417 */
3418static bool find_desired_extent_in_hole(struct btrfs_inode *inode, int whence,
3419 struct extent_state **cached_state,
3420 u64 start, u64 end, u64 *start_ret)
3421{
3422 u64 delalloc_start;
3423 u64 delalloc_end;
3424 bool delalloc;
3425
3426 delalloc = btrfs_find_delalloc_in_range(inode, start, end, cached_state,
3427 &delalloc_start, &delalloc_end);
3428 if (delalloc && whence == SEEK_DATA) {
3429 *start_ret = delalloc_start;
3430 return true;
3431 }
3432
3433 if (delalloc && whence == SEEK_HOLE) {
3434 /*
3435 * We found delalloc but it starts after out start offset. So we
3436 * have a hole between our start offset and the delalloc start.
3437 */
3438 if (start < delalloc_start) {
3439 *start_ret = start;
3440 return true;
3441 }
3442 /*
3443 * Delalloc range starts at our start offset.
3444 * If the delalloc range's length is smaller than our range,
3445 * then it means we have a hole that starts where the delalloc
3446 * subrange ends.
3447 */
3448 if (delalloc_end < end) {
3449 *start_ret = delalloc_end + 1;
3450 return true;
3451 }
3452
3453 /* There's delalloc for the whole range. */
3454 return false;
3455 }
3456
3457 if (!delalloc && whence == SEEK_HOLE) {
3458 *start_ret = start;
3459 return true;
3460 }
3461
3462 /*
3463 * No delalloc in the range and we are seeking for data. The caller has
3464 * to iterate to the next extent item in the subvolume btree.
3465 */
3466 return false;
3467}
3468
3469static loff_t find_desired_extent(struct file *file, loff_t offset, int whence)
3470{
3471 struct btrfs_inode *inode = BTRFS_I(file->f_mapping->host);
3472 struct btrfs_file_private *private = file->private_data;
3473 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3474 struct extent_state *cached_state = NULL;
3475 struct extent_state **delalloc_cached_state;
3476 const loff_t i_size = i_size_read(&inode->vfs_inode);
3477 const u64 ino = btrfs_ino(inode);
3478 struct btrfs_root *root = inode->root;
3479 struct btrfs_path *path;
3480 struct btrfs_key key;
3481 u64 last_extent_end;
3482 u64 lockstart;
3483 u64 lockend;
3484 u64 start;
3485 int ret;
3486 bool found = false;
3487
3488 if (i_size == 0 || offset >= i_size)
3489 return -ENXIO;
3490
3491 /*
3492 * Quick path. If the inode has no prealloc extents and its number of
3493 * bytes used matches its i_size, then it can not have holes.
3494 */
3495 if (whence == SEEK_HOLE &&
3496 !(inode->flags & BTRFS_INODE_PREALLOC) &&
3497 inode_get_bytes(&inode->vfs_inode) == i_size)
3498 return i_size;
3499
3500 if (!private) {
3501 private = kzalloc(sizeof(*private), GFP_KERNEL);
3502 /*
3503 * No worries if memory allocation failed.
3504 * The private structure is used only for speeding up multiple
3505 * lseek SEEK_HOLE/DATA calls to a file when there's delalloc,
3506 * so everything will still be correct.
3507 */
3508 file->private_data = private;
3509 }
3510
3511 if (private)
3512 delalloc_cached_state = &private->llseek_cached_state;
3513 else
3514 delalloc_cached_state = NULL;
3515
3516 /*
3517 * offset can be negative, in this case we start finding DATA/HOLE from
3518 * the very start of the file.
3519 */
3520 start = max_t(loff_t, 0, offset);
3521
3522 lockstart = round_down(start, fs_info->sectorsize);
3523 lockend = round_up(i_size, fs_info->sectorsize);
3524 if (lockend <= lockstart)
3525 lockend = lockstart + fs_info->sectorsize;
3526 lockend--;
3527
3528 path = btrfs_alloc_path();
3529 if (!path)
3530 return -ENOMEM;
3531 path->reada = READA_FORWARD;
3532
3533 key.objectid = ino;
3534 key.type = BTRFS_EXTENT_DATA_KEY;
3535 key.offset = start;
3536
3537 last_extent_end = lockstart;
3538
3539 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3540
3541 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3542 if (ret < 0) {
3543 goto out;
3544 } else if (ret > 0 && path->slots[0] > 0) {
3545 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
3546 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
3547 path->slots[0]--;
3548 }
3549
3550 while (start < i_size) {
3551 struct extent_buffer *leaf = path->nodes[0];
3552 struct btrfs_file_extent_item *extent;
3553 u64 extent_end;
3554 u8 type;
3555
3556 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3557 ret = btrfs_next_leaf(root, path);
3558 if (ret < 0)
3559 goto out;
3560 else if (ret > 0)
3561 break;
3562
3563 leaf = path->nodes[0];
3564 }
3565
3566 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3567 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3568 break;
3569
3570 extent_end = btrfs_file_extent_end(path);
3571
3572 /*
3573 * In the first iteration we may have a slot that points to an
3574 * extent that ends before our start offset, so skip it.
3575 */
3576 if (extent_end <= start) {
3577 path->slots[0]++;
3578 continue;
3579 }
3580
3581 /* We have an implicit hole, NO_HOLES feature is likely set. */
3582 if (last_extent_end < key.offset) {
3583 u64 search_start = last_extent_end;
3584 u64 found_start;
3585
3586 /*
3587 * First iteration, @start matches @offset and it's
3588 * within the hole.
3589 */
3590 if (start == offset)
3591 search_start = offset;
3592
3593 found = find_desired_extent_in_hole(inode, whence,
3594 delalloc_cached_state,
3595 search_start,
3596 key.offset - 1,
3597 &found_start);
3598 if (found) {
3599 start = found_start;
3600 break;
3601 }
3602 /*
3603 * Didn't find data or a hole (due to delalloc) in the
3604 * implicit hole range, so need to analyze the extent.
3605 */
3606 }
3607
3608 extent = btrfs_item_ptr(leaf, path->slots[0],
3609 struct btrfs_file_extent_item);
3610 type = btrfs_file_extent_type(leaf, extent);
3611
3612 /*
3613 * Can't access the extent's disk_bytenr field if this is an
3614 * inline extent, since at that offset, it's where the extent
3615 * data starts.
3616 */
3617 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
3618 (type == BTRFS_FILE_EXTENT_REG &&
3619 btrfs_file_extent_disk_bytenr(leaf, extent) == 0)) {
3620 /*
3621 * Explicit hole or prealloc extent, search for delalloc.
3622 * A prealloc extent is treated like a hole.
3623 */
3624 u64 search_start = key.offset;
3625 u64 found_start;
3626
3627 /*
3628 * First iteration, @start matches @offset and it's
3629 * within the hole.
3630 */
3631 if (start == offset)
3632 search_start = offset;
3633
3634 found = find_desired_extent_in_hole(inode, whence,
3635 delalloc_cached_state,
3636 search_start,
3637 extent_end - 1,
3638 &found_start);
3639 if (found) {
3640 start = found_start;
3641 break;
3642 }
3643 /*
3644 * Didn't find data or a hole (due to delalloc) in the
3645 * implicit hole range, so need to analyze the next
3646 * extent item.
3647 */
3648 } else {
3649 /*
3650 * Found a regular or inline extent.
3651 * If we are seeking for data, adjust the start offset
3652 * and stop, we're done.
3653 */
3654 if (whence == SEEK_DATA) {
3655 start = max_t(u64, key.offset, offset);
3656 found = true;
3657 break;
3658 }
3659 /*
3660 * Else, we are seeking for a hole, check the next file
3661 * extent item.
3662 */
3663 }
3664
3665 start = extent_end;
3666 last_extent_end = extent_end;
3667 path->slots[0]++;
3668 if (fatal_signal_pending(current)) {
3669 ret = -EINTR;
3670 goto out;
3671 }
3672 cond_resched();
3673 }
3674
3675 /* We have an implicit hole from the last extent found up to i_size. */
3676 if (!found && start < i_size) {
3677 found = find_desired_extent_in_hole(inode, whence,
3678 delalloc_cached_state, start,
3679 i_size - 1, &start);
3680 if (!found)
3681 start = i_size;
3682 }
3683
3684out:
3685 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3686 btrfs_free_path(path);
3687
3688 if (ret < 0)
3689 return ret;
3690
3691 if (whence == SEEK_DATA && start >= i_size)
3692 return -ENXIO;
3693
3694 return min_t(loff_t, start, i_size);
3695}
3696
3697static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3698{
3699 struct inode *inode = file->f_mapping->host;
3700
3701 switch (whence) {
3702 default:
3703 return generic_file_llseek(file, offset, whence);
3704 case SEEK_DATA:
3705 case SEEK_HOLE:
3706 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3707 offset = find_desired_extent(file, offset, whence);
3708 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3709 break;
3710 }
3711
3712 if (offset < 0)
3713 return offset;
3714
3715 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3716}
3717
3718static int btrfs_file_open(struct inode *inode, struct file *filp)
3719{
3720 int ret;
3721
3722 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC |
3723 FMODE_CAN_ODIRECT;
3724
3725 ret = fsverity_file_open(inode, filp);
3726 if (ret)
3727 return ret;
3728 return generic_file_open(inode, filp);
3729}
3730
3731static int check_direct_read(struct btrfs_fs_info *fs_info,
3732 const struct iov_iter *iter, loff_t offset)
3733{
3734 int ret;
3735 int i, seg;
3736
3737 ret = check_direct_IO(fs_info, iter, offset);
3738 if (ret < 0)
3739 return ret;
3740
3741 if (!iter_is_iovec(iter))
3742 return 0;
3743
3744 for (seg = 0; seg < iter->nr_segs; seg++) {
3745 for (i = seg + 1; i < iter->nr_segs; i++) {
3746 const struct iovec *iov1 = iter_iov(iter) + seg;
3747 const struct iovec *iov2 = iter_iov(iter) + i;
3748
3749 if (iov1->iov_base == iov2->iov_base)
3750 return -EINVAL;
3751 }
3752 }
3753 return 0;
3754}
3755
3756static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3757{
3758 struct inode *inode = file_inode(iocb->ki_filp);
3759 size_t prev_left = 0;
3760 ssize_t read = 0;
3761 ssize_t ret;
3762
3763 if (fsverity_active(inode))
3764 return 0;
3765
3766 if (check_direct_read(inode_to_fs_info(inode), to, iocb->ki_pos))
3767 return 0;
3768
3769 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3770again:
3771 /*
3772 * This is similar to what we do for direct IO writes, see the comment
3773 * at btrfs_direct_write(), but we also disable page faults in addition
3774 * to disabling them only at the iov_iter level. This is because when
3775 * reading from a hole or prealloc extent, iomap calls iov_iter_zero(),
3776 * which can still trigger page fault ins despite having set ->nofault
3777 * to true of our 'to' iov_iter.
3778 *
3779 * The difference to direct IO writes is that we deadlock when trying
3780 * to lock the extent range in the inode's tree during he page reads
3781 * triggered by the fault in (while for writes it is due to waiting for
3782 * our own ordered extent). This is because for direct IO reads,
3783 * btrfs_dio_iomap_begin() returns with the extent range locked, which
3784 * is only unlocked in the endio callback (end_bio_extent_readpage()).
3785 */
3786 pagefault_disable();
3787 to->nofault = true;
3788 ret = btrfs_dio_read(iocb, to, read);
3789 to->nofault = false;
3790 pagefault_enable();
3791
3792 /* No increment (+=) because iomap returns a cumulative value. */
3793 if (ret > 0)
3794 read = ret;
3795
3796 if (iov_iter_count(to) > 0 && (ret == -EFAULT || ret > 0)) {
3797 const size_t left = iov_iter_count(to);
3798
3799 if (left == prev_left) {
3800 /*
3801 * We didn't make any progress since the last attempt,
3802 * fallback to a buffered read for the remainder of the
3803 * range. This is just to avoid any possibility of looping
3804 * for too long.
3805 */
3806 ret = read;
3807 } else {
3808 /*
3809 * We made some progress since the last retry or this is
3810 * the first time we are retrying. Fault in as many pages
3811 * as possible and retry.
3812 */
3813 fault_in_iov_iter_writeable(to, left);
3814 prev_left = left;
3815 goto again;
3816 }
3817 }
3818 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3819 return ret < 0 ? ret : read;
3820}
3821
3822static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3823{
3824 ssize_t ret = 0;
3825
3826 if (iocb->ki_flags & IOCB_DIRECT) {
3827 ret = btrfs_direct_read(iocb, to);
3828 if (ret < 0 || !iov_iter_count(to) ||
3829 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3830 return ret;
3831 }
3832
3833 return filemap_read(iocb, to, ret);
3834}
3835
3836const struct file_operations btrfs_file_operations = {
3837 .llseek = btrfs_file_llseek,
3838 .read_iter = btrfs_file_read_iter,
3839 .splice_read = filemap_splice_read,
3840 .write_iter = btrfs_file_write_iter,
3841 .splice_write = iter_file_splice_write,
3842 .mmap = btrfs_file_mmap,
3843 .open = btrfs_file_open,
3844 .release = btrfs_release_file,
3845 .get_unmapped_area = thp_get_unmapped_area,
3846 .fsync = btrfs_sync_file,
3847 .fallocate = btrfs_fallocate,
3848 .unlocked_ioctl = btrfs_ioctl,
3849#ifdef CONFIG_COMPAT
3850 .compat_ioctl = btrfs_compat_ioctl,
3851#endif
3852 .remap_file_range = btrfs_remap_file_range,
3853};
3854
3855int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3856{
3857 int ret;
3858
3859 /*
3860 * So with compression we will find and lock a dirty page and clear the
3861 * first one as dirty, setup an async extent, and immediately return
3862 * with the entire range locked but with nobody actually marked with
3863 * writeback. So we can't just filemap_write_and_wait_range() and
3864 * expect it to work since it will just kick off a thread to do the
3865 * actual work. So we need to call filemap_fdatawrite_range _again_
3866 * since it will wait on the page lock, which won't be unlocked until
3867 * after the pages have been marked as writeback and so we're good to go
3868 * from there. We have to do this otherwise we'll miss the ordered
3869 * extents and that results in badness. Please Josef, do not think you
3870 * know better and pull this out at some point in the future, it is
3871 * right and you are wrong.
3872 */
3873 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3874 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3875 &BTRFS_I(inode)->runtime_flags))
3876 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3877
3878 return ret;
3879}