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