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