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