Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6#include <linux/fs.h>
7#include <linux/pagemap.h>
8#include <linux/time.h>
9#include <linux/init.h>
10#include <linux/string.h>
11#include <linux/backing-dev.h>
12#include <linux/falloc.h>
13#include <linux/writeback.h>
14#include <linux/compat.h>
15#include <linux/slab.h>
16#include <linux/btrfs.h>
17#include <linux/uio.h>
18#include <linux/iversion.h>
19#include <linux/fsverity.h>
20#include "ctree.h"
21#include "direct-io.h"
22#include "disk-io.h"
23#include "transaction.h"
24#include "btrfs_inode.h"
25#include "tree-log.h"
26#include "locking.h"
27#include "qgroup.h"
28#include "compression.h"
29#include "delalloc-space.h"
30#include "reflink.h"
31#include "subpage.h"
32#include "fs.h"
33#include "accessors.h"
34#include "extent-tree.h"
35#include "file-item.h"
36#include "ioctl.h"
37#include "file.h"
38#include "super.h"
39
40/*
41 * Helper to fault in page and copy. This should go away and be replaced with
42 * calls into generic code.
43 */
44static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
45 struct folio *folio, struct iov_iter *i)
46{
47 size_t copied = 0;
48 size_t total_copied = 0;
49 int offset = offset_in_page(pos);
50
51 while (write_bytes > 0) {
52 size_t count = min_t(size_t, PAGE_SIZE - offset, write_bytes);
53 /*
54 * Copy data from userspace to the current page
55 */
56 copied = copy_folio_from_iter_atomic(folio, offset, count, i);
57
58 /* Flush processor's dcache for this page */
59 flush_dcache_folio(folio);
60
61 /*
62 * if we get a partial write, we can end up with
63 * partially up to date page. These add
64 * a lot of complexity, so make sure they don't
65 * happen by forcing this copy to be retried.
66 *
67 * The rest of the btrfs_file_write code will fall
68 * back to page at a time copies after we return 0.
69 */
70 if (unlikely(copied < count)) {
71 if (!folio_test_uptodate(folio)) {
72 iov_iter_revert(i, copied);
73 copied = 0;
74 }
75 if (!copied)
76 break;
77 }
78
79 write_bytes -= copied;
80 total_copied += copied;
81 offset += copied;
82 }
83 return total_copied;
84}
85
86/*
87 * Unlock folio after btrfs_file_write() is done with it.
88 */
89static void btrfs_drop_folio(struct btrfs_fs_info *fs_info, struct folio *folio,
90 u64 pos, u64 copied)
91{
92 u64 block_start = round_down(pos, fs_info->sectorsize);
93 u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
94
95 ASSERT(block_len <= U32_MAX);
96 /*
97 * Folio checked is some magic around finding folios that have been
98 * modified without going through btrfs_dirty_folio(). Clear it here.
99 * There should be no need to mark the pages accessed as
100 * prepare_one_folio() should have marked them accessed in
101 * prepare_one_folio() via find_or_create_page()
102 */
103 btrfs_folio_clamp_clear_checked(fs_info, folio, block_start, block_len);
104 folio_unlock(folio);
105 folio_put(folio);
106}
107
108/*
109 * After btrfs_copy_from_user(), update the following things for delalloc:
110 * - Mark newly dirtied folio as DELALLOC in the io tree.
111 * Used to advise which range is to be written back.
112 * - Mark modified folio as Uptodate/Dirty and not needing COW fixup
113 * - Update inode size for past EOF write
114 */
115int btrfs_dirty_folio(struct btrfs_inode *inode, struct folio *folio, loff_t pos,
116 size_t write_bytes, struct extent_state **cached, bool noreserve)
117{
118 struct btrfs_fs_info *fs_info = inode->root->fs_info;
119 int ret = 0;
120 u64 num_bytes;
121 u64 start_pos;
122 u64 end_of_last_block;
123 u64 end_pos = pos + write_bytes;
124 loff_t isize = i_size_read(&inode->vfs_inode);
125 unsigned int extra_bits = 0;
126
127 if (write_bytes == 0)
128 return 0;
129
130 if (noreserve)
131 extra_bits |= EXTENT_NORESERVE;
132
133 start_pos = round_down(pos, fs_info->sectorsize);
134 num_bytes = round_up(write_bytes + pos - start_pos,
135 fs_info->sectorsize);
136 ASSERT(num_bytes <= U32_MAX);
137 ASSERT(folio_pos(folio) <= pos &&
138 folio_pos(folio) + folio_size(folio) >= pos + write_bytes);
139
140 end_of_last_block = start_pos + num_bytes - 1;
141
142 /*
143 * The pages may have already been dirty, clear out old accounting so
144 * we can set things up properly
145 */
146 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
147 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
148 cached);
149
150 ret = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
151 extra_bits, cached);
152 if (ret)
153 return ret;
154
155 btrfs_folio_clamp_set_uptodate(fs_info, folio, start_pos, num_bytes);
156 btrfs_folio_clamp_clear_checked(fs_info, folio, start_pos, num_bytes);
157 btrfs_folio_clamp_set_dirty(fs_info, folio, start_pos, num_bytes);
158
159 /*
160 * we've only changed i_size in ram, and we haven't updated
161 * the disk i_size. There is no need to log the inode
162 * at this time.
163 */
164 if (end_pos > isize)
165 i_size_write(&inode->vfs_inode, end_pos);
166 return 0;
167}
168
169/*
170 * this is very complex, but the basic idea is to drop all extents
171 * in the range start - end. hint_block is filled in with a block number
172 * that would be a good hint to the block allocator for this file.
173 *
174 * If an extent intersects the range but is not entirely inside the range
175 * it is either truncated or split. Anything entirely inside the range
176 * is deleted from the tree.
177 *
178 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
179 * to deal with that. We set the field 'bytes_found' of the arguments structure
180 * with the number of allocated bytes found in the target range, so that the
181 * caller can update the inode's number of bytes in an atomic way when
182 * replacing extents in a range to avoid races with stat(2).
183 */
184int btrfs_drop_extents(struct btrfs_trans_handle *trans,
185 struct btrfs_root *root, struct btrfs_inode *inode,
186 struct btrfs_drop_extents_args *args)
187{
188 struct btrfs_fs_info *fs_info = root->fs_info;
189 struct extent_buffer *leaf;
190 struct btrfs_file_extent_item *fi;
191 struct btrfs_key key;
192 struct btrfs_key new_key;
193 u64 ino = btrfs_ino(inode);
194 u64 search_start = args->start;
195 u64 disk_bytenr = 0;
196 u64 num_bytes = 0;
197 u64 extent_offset = 0;
198 u64 extent_end = 0;
199 u64 last_end = args->start;
200 int del_nr = 0;
201 int del_slot = 0;
202 int extent_type;
203 int recow;
204 int ret;
205 int modify_tree = -1;
206 int update_refs;
207 int found = 0;
208 struct btrfs_path *path = args->path;
209
210 args->bytes_found = 0;
211 args->extent_inserted = false;
212
213 /* Must always have a path if ->replace_extent is true */
214 ASSERT(!(args->replace_extent && !args->path));
215
216 if (!path) {
217 path = btrfs_alloc_path();
218 if (!path) {
219 ret = -ENOMEM;
220 goto out;
221 }
222 }
223
224 if (args->drop_cache)
225 btrfs_drop_extent_map_range(inode, args->start, args->end - 1, false);
226
227 if (data_race(args->start >= inode->disk_i_size) && !args->replace_extent)
228 modify_tree = 0;
229
230 update_refs = (btrfs_root_id(root) != BTRFS_TREE_LOG_OBJECTID);
231 while (1) {
232 recow = 0;
233 ret = btrfs_lookup_file_extent(trans, root, path, ino,
234 search_start, modify_tree);
235 if (ret < 0)
236 break;
237 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
238 leaf = path->nodes[0];
239 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
240 if (key.objectid == ino &&
241 key.type == BTRFS_EXTENT_DATA_KEY)
242 path->slots[0]--;
243 }
244 ret = 0;
245next_slot:
246 leaf = path->nodes[0];
247 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
248 BUG_ON(del_nr > 0);
249 ret = btrfs_next_leaf(root, path);
250 if (ret < 0)
251 break;
252 if (ret > 0) {
253 ret = 0;
254 break;
255 }
256 leaf = path->nodes[0];
257 recow = 1;
258 }
259
260 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
261
262 if (key.objectid > ino)
263 break;
264 if (WARN_ON_ONCE(key.objectid < ino) ||
265 key.type < BTRFS_EXTENT_DATA_KEY) {
266 ASSERT(del_nr == 0);
267 path->slots[0]++;
268 goto next_slot;
269 }
270 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
271 break;
272
273 fi = btrfs_item_ptr(leaf, path->slots[0],
274 struct btrfs_file_extent_item);
275 extent_type = btrfs_file_extent_type(leaf, fi);
276
277 if (extent_type == BTRFS_FILE_EXTENT_REG ||
278 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
279 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
280 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
281 extent_offset = btrfs_file_extent_offset(leaf, fi);
282 extent_end = key.offset +
283 btrfs_file_extent_num_bytes(leaf, fi);
284 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
285 extent_end = key.offset +
286 btrfs_file_extent_ram_bytes(leaf, fi);
287 } else {
288 /* can't happen */
289 BUG();
290 }
291
292 /*
293 * Don't skip extent items representing 0 byte lengths. They
294 * used to be created (bug) if while punching holes we hit
295 * -ENOSPC condition. So if we find one here, just ensure we
296 * delete it, otherwise we would insert a new file extent item
297 * with the same key (offset) as that 0 bytes length file
298 * extent item in the call to setup_items_for_insert() later
299 * in this function.
300 */
301 if (extent_end == key.offset && extent_end >= search_start) {
302 last_end = extent_end;
303 goto delete_extent_item;
304 }
305
306 if (extent_end <= search_start) {
307 path->slots[0]++;
308 goto next_slot;
309 }
310
311 found = 1;
312 search_start = max(key.offset, args->start);
313 if (recow || !modify_tree) {
314 modify_tree = -1;
315 btrfs_release_path(path);
316 continue;
317 }
318
319 /*
320 * | - range to drop - |
321 * | -------- extent -------- |
322 */
323 if (args->start > key.offset && args->end < extent_end) {
324 BUG_ON(del_nr > 0);
325 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
326 ret = -EOPNOTSUPP;
327 break;
328 }
329
330 memcpy(&new_key, &key, sizeof(new_key));
331 new_key.offset = args->start;
332 ret = btrfs_duplicate_item(trans, root, path,
333 &new_key);
334 if (ret == -EAGAIN) {
335 btrfs_release_path(path);
336 continue;
337 }
338 if (ret < 0)
339 break;
340
341 leaf = path->nodes[0];
342 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
343 struct btrfs_file_extent_item);
344 btrfs_set_file_extent_num_bytes(leaf, fi,
345 args->start - key.offset);
346
347 fi = btrfs_item_ptr(leaf, path->slots[0],
348 struct btrfs_file_extent_item);
349
350 extent_offset += args->start - key.offset;
351 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
352 btrfs_set_file_extent_num_bytes(leaf, fi,
353 extent_end - args->start);
354 btrfs_mark_buffer_dirty(trans, leaf);
355
356 if (update_refs && disk_bytenr > 0) {
357 struct btrfs_ref ref = {
358 .action = BTRFS_ADD_DELAYED_REF,
359 .bytenr = disk_bytenr,
360 .num_bytes = num_bytes,
361 .parent = 0,
362 .owning_root = btrfs_root_id(root),
363 .ref_root = btrfs_root_id(root),
364 };
365 btrfs_init_data_ref(&ref, new_key.objectid,
366 args->start - extent_offset,
367 0, false);
368 ret = btrfs_inc_extent_ref(trans, &ref);
369 if (ret) {
370 btrfs_abort_transaction(trans, ret);
371 break;
372 }
373 }
374 key.offset = args->start;
375 }
376 /*
377 * From here on out we will have actually dropped something, so
378 * last_end can be updated.
379 */
380 last_end = extent_end;
381
382 /*
383 * | ---- range to drop ----- |
384 * | -------- extent -------- |
385 */
386 if (args->start <= key.offset && args->end < extent_end) {
387 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
388 ret = -EOPNOTSUPP;
389 break;
390 }
391
392 memcpy(&new_key, &key, sizeof(new_key));
393 new_key.offset = args->end;
394 btrfs_set_item_key_safe(trans, path, &new_key);
395
396 extent_offset += args->end - key.offset;
397 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
398 btrfs_set_file_extent_num_bytes(leaf, fi,
399 extent_end - args->end);
400 btrfs_mark_buffer_dirty(trans, leaf);
401 if (update_refs && disk_bytenr > 0)
402 args->bytes_found += args->end - key.offset;
403 break;
404 }
405
406 search_start = extent_end;
407 /*
408 * | ---- range to drop ----- |
409 * | -------- extent -------- |
410 */
411 if (args->start > key.offset && args->end >= extent_end) {
412 BUG_ON(del_nr > 0);
413 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
414 ret = -EOPNOTSUPP;
415 break;
416 }
417
418 btrfs_set_file_extent_num_bytes(leaf, fi,
419 args->start - key.offset);
420 btrfs_mark_buffer_dirty(trans, leaf);
421 if (update_refs && disk_bytenr > 0)
422 args->bytes_found += extent_end - args->start;
423 if (args->end == extent_end)
424 break;
425
426 path->slots[0]++;
427 goto next_slot;
428 }
429
430 /*
431 * | ---- range to drop ----- |
432 * | ------ extent ------ |
433 */
434 if (args->start <= key.offset && args->end >= extent_end) {
435delete_extent_item:
436 if (del_nr == 0) {
437 del_slot = path->slots[0];
438 del_nr = 1;
439 } else {
440 BUG_ON(del_slot + del_nr != path->slots[0]);
441 del_nr++;
442 }
443
444 if (update_refs &&
445 extent_type == BTRFS_FILE_EXTENT_INLINE) {
446 args->bytes_found += extent_end - key.offset;
447 extent_end = ALIGN(extent_end,
448 fs_info->sectorsize);
449 } else if (update_refs && disk_bytenr > 0) {
450 struct btrfs_ref ref = {
451 .action = BTRFS_DROP_DELAYED_REF,
452 .bytenr = disk_bytenr,
453 .num_bytes = num_bytes,
454 .parent = 0,
455 .owning_root = btrfs_root_id(root),
456 .ref_root = btrfs_root_id(root),
457 };
458 btrfs_init_data_ref(&ref, key.objectid,
459 key.offset - extent_offset,
460 0, false);
461 ret = btrfs_free_extent(trans, &ref);
462 if (ret) {
463 btrfs_abort_transaction(trans, ret);
464 break;
465 }
466 args->bytes_found += extent_end - key.offset;
467 }
468
469 if (args->end == extent_end)
470 break;
471
472 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
473 path->slots[0]++;
474 goto next_slot;
475 }
476
477 ret = btrfs_del_items(trans, root, path, del_slot,
478 del_nr);
479 if (ret) {
480 btrfs_abort_transaction(trans, ret);
481 break;
482 }
483
484 del_nr = 0;
485 del_slot = 0;
486
487 btrfs_release_path(path);
488 continue;
489 }
490
491 BUG();
492 }
493
494 if (!ret && del_nr > 0) {
495 /*
496 * Set path->slots[0] to first slot, so that after the delete
497 * if items are move off from our leaf to its immediate left or
498 * right neighbor leafs, we end up with a correct and adjusted
499 * path->slots[0] for our insertion (if args->replace_extent).
500 */
501 path->slots[0] = del_slot;
502 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
503 if (ret)
504 btrfs_abort_transaction(trans, ret);
505 }
506
507 leaf = path->nodes[0];
508 /*
509 * If btrfs_del_items() was called, it might have deleted a leaf, in
510 * which case it unlocked our path, so check path->locks[0] matches a
511 * write lock.
512 */
513 if (!ret && args->replace_extent &&
514 path->locks[0] == BTRFS_WRITE_LOCK &&
515 btrfs_leaf_free_space(leaf) >=
516 sizeof(struct btrfs_item) + args->extent_item_size) {
517
518 key.objectid = ino;
519 key.type = BTRFS_EXTENT_DATA_KEY;
520 key.offset = args->start;
521 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
522 struct btrfs_key slot_key;
523
524 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
525 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
526 path->slots[0]++;
527 }
528 btrfs_setup_item_for_insert(trans, root, path, &key,
529 args->extent_item_size);
530 args->extent_inserted = true;
531 }
532
533 if (!args->path)
534 btrfs_free_path(path);
535 else if (!args->extent_inserted)
536 btrfs_release_path(path);
537out:
538 args->drop_end = found ? min(args->end, last_end) : args->end;
539
540 return ret;
541}
542
543static int extent_mergeable(struct extent_buffer *leaf, int slot,
544 u64 objectid, u64 bytenr, u64 orig_offset,
545 u64 *start, u64 *end)
546{
547 struct btrfs_file_extent_item *fi;
548 struct btrfs_key key;
549 u64 extent_end;
550
551 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
552 return 0;
553
554 btrfs_item_key_to_cpu(leaf, &key, slot);
555 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
556 return 0;
557
558 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
559 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
560 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
561 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
562 btrfs_file_extent_compression(leaf, fi) ||
563 btrfs_file_extent_encryption(leaf, fi) ||
564 btrfs_file_extent_other_encoding(leaf, fi))
565 return 0;
566
567 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
568 if ((*start && *start != key.offset) || (*end && *end != extent_end))
569 return 0;
570
571 *start = key.offset;
572 *end = extent_end;
573 return 1;
574}
575
576/*
577 * Mark extent in the range start - end as written.
578 *
579 * This changes extent type from 'pre-allocated' to 'regular'. If only
580 * part of extent is marked as written, the extent will be split into
581 * two or three.
582 */
583int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
584 struct btrfs_inode *inode, u64 start, u64 end)
585{
586 struct btrfs_root *root = inode->root;
587 struct extent_buffer *leaf;
588 struct btrfs_path *path;
589 struct btrfs_file_extent_item *fi;
590 struct btrfs_ref ref = { 0 };
591 struct btrfs_key key;
592 struct btrfs_key new_key;
593 u64 bytenr;
594 u64 num_bytes;
595 u64 extent_end;
596 u64 orig_offset;
597 u64 other_start;
598 u64 other_end;
599 u64 split;
600 int del_nr = 0;
601 int del_slot = 0;
602 int recow;
603 int ret = 0;
604 u64 ino = btrfs_ino(inode);
605
606 path = btrfs_alloc_path();
607 if (!path)
608 return -ENOMEM;
609again:
610 recow = 0;
611 split = start;
612 key.objectid = ino;
613 key.type = BTRFS_EXTENT_DATA_KEY;
614 key.offset = split;
615
616 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
617 if (ret < 0)
618 goto out;
619 if (ret > 0 && path->slots[0] > 0)
620 path->slots[0]--;
621
622 leaf = path->nodes[0];
623 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
624 if (key.objectid != ino ||
625 key.type != BTRFS_EXTENT_DATA_KEY) {
626 ret = -EINVAL;
627 btrfs_abort_transaction(trans, ret);
628 goto out;
629 }
630 fi = btrfs_item_ptr(leaf, path->slots[0],
631 struct btrfs_file_extent_item);
632 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
633 ret = -EINVAL;
634 btrfs_abort_transaction(trans, ret);
635 goto out;
636 }
637 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
638 if (key.offset > start || extent_end < end) {
639 ret = -EINVAL;
640 btrfs_abort_transaction(trans, ret);
641 goto out;
642 }
643
644 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
645 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
646 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
647 memcpy(&new_key, &key, sizeof(new_key));
648
649 if (start == key.offset && end < extent_end) {
650 other_start = 0;
651 other_end = start;
652 if (extent_mergeable(leaf, path->slots[0] - 1,
653 ino, bytenr, orig_offset,
654 &other_start, &other_end)) {
655 new_key.offset = end;
656 btrfs_set_item_key_safe(trans, path, &new_key);
657 fi = btrfs_item_ptr(leaf, path->slots[0],
658 struct btrfs_file_extent_item);
659 btrfs_set_file_extent_generation(leaf, fi,
660 trans->transid);
661 btrfs_set_file_extent_num_bytes(leaf, fi,
662 extent_end - end);
663 btrfs_set_file_extent_offset(leaf, fi,
664 end - orig_offset);
665 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
666 struct btrfs_file_extent_item);
667 btrfs_set_file_extent_generation(leaf, fi,
668 trans->transid);
669 btrfs_set_file_extent_num_bytes(leaf, fi,
670 end - other_start);
671 btrfs_mark_buffer_dirty(trans, leaf);
672 goto out;
673 }
674 }
675
676 if (start > key.offset && end == extent_end) {
677 other_start = end;
678 other_end = 0;
679 if (extent_mergeable(leaf, path->slots[0] + 1,
680 ino, bytenr, orig_offset,
681 &other_start, &other_end)) {
682 fi = btrfs_item_ptr(leaf, path->slots[0],
683 struct btrfs_file_extent_item);
684 btrfs_set_file_extent_num_bytes(leaf, fi,
685 start - key.offset);
686 btrfs_set_file_extent_generation(leaf, fi,
687 trans->transid);
688 path->slots[0]++;
689 new_key.offset = start;
690 btrfs_set_item_key_safe(trans, path, &new_key);
691
692 fi = btrfs_item_ptr(leaf, path->slots[0],
693 struct btrfs_file_extent_item);
694 btrfs_set_file_extent_generation(leaf, fi,
695 trans->transid);
696 btrfs_set_file_extent_num_bytes(leaf, fi,
697 other_end - start);
698 btrfs_set_file_extent_offset(leaf, fi,
699 start - orig_offset);
700 btrfs_mark_buffer_dirty(trans, leaf);
701 goto out;
702 }
703 }
704
705 while (start > key.offset || end < extent_end) {
706 if (key.offset == start)
707 split = end;
708
709 new_key.offset = split;
710 ret = btrfs_duplicate_item(trans, root, path, &new_key);
711 if (ret == -EAGAIN) {
712 btrfs_release_path(path);
713 goto again;
714 }
715 if (ret < 0) {
716 btrfs_abort_transaction(trans, ret);
717 goto out;
718 }
719
720 leaf = path->nodes[0];
721 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
722 struct btrfs_file_extent_item);
723 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
724 btrfs_set_file_extent_num_bytes(leaf, fi,
725 split - key.offset);
726
727 fi = btrfs_item_ptr(leaf, path->slots[0],
728 struct btrfs_file_extent_item);
729
730 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
731 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
732 btrfs_set_file_extent_num_bytes(leaf, fi,
733 extent_end - split);
734 btrfs_mark_buffer_dirty(trans, leaf);
735
736 ref.action = BTRFS_ADD_DELAYED_REF;
737 ref.bytenr = bytenr;
738 ref.num_bytes = num_bytes;
739 ref.parent = 0;
740 ref.owning_root = btrfs_root_id(root);
741 ref.ref_root = btrfs_root_id(root);
742 btrfs_init_data_ref(&ref, ino, orig_offset, 0, false);
743 ret = btrfs_inc_extent_ref(trans, &ref);
744 if (ret) {
745 btrfs_abort_transaction(trans, ret);
746 goto out;
747 }
748
749 if (split == start) {
750 key.offset = start;
751 } else {
752 if (start != key.offset) {
753 ret = -EINVAL;
754 btrfs_abort_transaction(trans, ret);
755 goto out;
756 }
757 path->slots[0]--;
758 extent_end = end;
759 }
760 recow = 1;
761 }
762
763 other_start = end;
764 other_end = 0;
765
766 ref.action = BTRFS_DROP_DELAYED_REF;
767 ref.bytenr = bytenr;
768 ref.num_bytes = num_bytes;
769 ref.parent = 0;
770 ref.owning_root = btrfs_root_id(root);
771 ref.ref_root = btrfs_root_id(root);
772 btrfs_init_data_ref(&ref, ino, orig_offset, 0, false);
773 if (extent_mergeable(leaf, path->slots[0] + 1,
774 ino, bytenr, orig_offset,
775 &other_start, &other_end)) {
776 if (recow) {
777 btrfs_release_path(path);
778 goto again;
779 }
780 extent_end = other_end;
781 del_slot = path->slots[0] + 1;
782 del_nr++;
783 ret = btrfs_free_extent(trans, &ref);
784 if (ret) {
785 btrfs_abort_transaction(trans, ret);
786 goto out;
787 }
788 }
789 other_start = 0;
790 other_end = start;
791 if (extent_mergeable(leaf, path->slots[0] - 1,
792 ino, bytenr, orig_offset,
793 &other_start, &other_end)) {
794 if (recow) {
795 btrfs_release_path(path);
796 goto again;
797 }
798 key.offset = other_start;
799 del_slot = path->slots[0];
800 del_nr++;
801 ret = btrfs_free_extent(trans, &ref);
802 if (ret) {
803 btrfs_abort_transaction(trans, ret);
804 goto out;
805 }
806 }
807 if (del_nr == 0) {
808 fi = btrfs_item_ptr(leaf, path->slots[0],
809 struct btrfs_file_extent_item);
810 btrfs_set_file_extent_type(leaf, fi,
811 BTRFS_FILE_EXTENT_REG);
812 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
813 btrfs_mark_buffer_dirty(trans, leaf);
814 } else {
815 fi = btrfs_item_ptr(leaf, del_slot - 1,
816 struct btrfs_file_extent_item);
817 btrfs_set_file_extent_type(leaf, fi,
818 BTRFS_FILE_EXTENT_REG);
819 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
820 btrfs_set_file_extent_num_bytes(leaf, fi,
821 extent_end - key.offset);
822 btrfs_mark_buffer_dirty(trans, leaf);
823
824 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
825 if (ret < 0) {
826 btrfs_abort_transaction(trans, ret);
827 goto out;
828 }
829 }
830out:
831 btrfs_free_path(path);
832 return ret;
833}
834
835/*
836 * On error return an unlocked folio and the error value
837 * On success return a locked folio and 0
838 */
839static int prepare_uptodate_folio(struct inode *inode, struct folio *folio, u64 pos,
840 u64 len, bool force_uptodate)
841{
842 u64 clamp_start = max_t(u64, pos, folio_pos(folio));
843 u64 clamp_end = min_t(u64, pos + len, folio_pos(folio) + folio_size(folio));
844 int ret = 0;
845
846 if (folio_test_uptodate(folio))
847 return 0;
848
849 if (!force_uptodate &&
850 IS_ALIGNED(clamp_start, PAGE_SIZE) &&
851 IS_ALIGNED(clamp_end, PAGE_SIZE))
852 return 0;
853
854 ret = btrfs_read_folio(NULL, folio);
855 if (ret)
856 return ret;
857 folio_lock(folio);
858 if (!folio_test_uptodate(folio)) {
859 folio_unlock(folio);
860 return -EIO;
861 }
862
863 /*
864 * Since btrfs_read_folio() will unlock the folio before it returns,
865 * there is a window where btrfs_release_folio() can be called to
866 * release the page. Here we check both inode mapping and page
867 * private to make sure the page was not released.
868 *
869 * The private flag check is essential for subpage as we need to store
870 * extra bitmap using folio private.
871 */
872 if (folio->mapping != inode->i_mapping || !folio_test_private(folio)) {
873 folio_unlock(folio);
874 return -EAGAIN;
875 }
876 return 0;
877}
878
879static gfp_t get_prepare_gfp_flags(struct inode *inode, bool nowait)
880{
881 gfp_t gfp;
882
883 gfp = btrfs_alloc_write_mask(inode->i_mapping);
884 if (nowait) {
885 gfp &= ~__GFP_DIRECT_RECLAIM;
886 gfp |= GFP_NOWAIT;
887 }
888
889 return gfp;
890}
891
892/*
893 * Get folio into the page cache and lock it.
894 */
895static noinline int prepare_one_folio(struct inode *inode, struct folio **folio_ret,
896 loff_t pos, size_t write_bytes,
897 bool force_uptodate, bool nowait)
898{
899 unsigned long index = pos >> PAGE_SHIFT;
900 gfp_t mask = get_prepare_gfp_flags(inode, nowait);
901 fgf_t fgp_flags = (nowait ? FGP_WRITEBEGIN | FGP_NOWAIT : FGP_WRITEBEGIN);
902 struct folio *folio;
903 int ret = 0;
904
905again:
906 folio = __filemap_get_folio(inode->i_mapping, index, fgp_flags, mask);
907 if (IS_ERR(folio)) {
908 if (nowait)
909 ret = -EAGAIN;
910 else
911 ret = PTR_ERR(folio);
912 return ret;
913 }
914 folio_wait_writeback(folio);
915 /* Only support page sized folio yet. */
916 ASSERT(folio_order(folio) == 0);
917 ret = set_folio_extent_mapped(folio);
918 if (ret < 0) {
919 folio_unlock(folio);
920 folio_put(folio);
921 return ret;
922 }
923 ret = prepare_uptodate_folio(inode, folio, pos, write_bytes, force_uptodate);
924 if (ret) {
925 /* The folio is already unlocked. */
926 folio_put(folio);
927 if (!nowait && ret == -EAGAIN) {
928 ret = 0;
929 goto again;
930 }
931 return ret;
932 }
933 *folio_ret = folio;
934 return 0;
935}
936
937/*
938 * Locks the extent and properly waits for data=ordered extents to finish
939 * before allowing the folios to be modified if need.
940 *
941 * Return:
942 * 1 - the extent is locked
943 * 0 - the extent is not locked, and everything is OK
944 * -EAGAIN - need to prepare the folios again
945 */
946static noinline int
947lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct folio *folio,
948 loff_t pos, size_t write_bytes,
949 u64 *lockstart, u64 *lockend, bool nowait,
950 struct extent_state **cached_state)
951{
952 struct btrfs_fs_info *fs_info = inode->root->fs_info;
953 u64 start_pos;
954 u64 last_pos;
955 int ret = 0;
956
957 start_pos = round_down(pos, fs_info->sectorsize);
958 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
959
960 if (start_pos < inode->vfs_inode.i_size) {
961 struct btrfs_ordered_extent *ordered;
962
963 if (nowait) {
964 if (!try_lock_extent(&inode->io_tree, start_pos, last_pos,
965 cached_state)) {
966 folio_unlock(folio);
967 folio_put(folio);
968 return -EAGAIN;
969 }
970 } else {
971 lock_extent(&inode->io_tree, start_pos, last_pos, cached_state);
972 }
973
974 ordered = btrfs_lookup_ordered_range(inode, start_pos,
975 last_pos - start_pos + 1);
976 if (ordered &&
977 ordered->file_offset + ordered->num_bytes > start_pos &&
978 ordered->file_offset <= last_pos) {
979 unlock_extent(&inode->io_tree, start_pos, last_pos,
980 cached_state);
981 folio_unlock(folio);
982 folio_put(folio);
983 btrfs_start_ordered_extent(ordered);
984 btrfs_put_ordered_extent(ordered);
985 return -EAGAIN;
986 }
987 if (ordered)
988 btrfs_put_ordered_extent(ordered);
989
990 *lockstart = start_pos;
991 *lockend = last_pos;
992 ret = 1;
993 }
994
995 /*
996 * We should be called after prepare_one_folio() which should have locked
997 * all pages in the range.
998 */
999 WARN_ON(!folio_test_locked(folio));
1000
1001 return ret;
1002}
1003
1004/*
1005 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1006 *
1007 * @pos: File offset.
1008 * @write_bytes: The length to write, will be updated to the nocow writeable
1009 * range.
1010 *
1011 * This function will flush ordered extents in the range to ensure proper
1012 * nocow checks.
1013 *
1014 * Return:
1015 * > 0 If we can nocow, and updates @write_bytes.
1016 * 0 If we can't do a nocow write.
1017 * -EAGAIN If we can't do a nocow write because snapshoting of the inode's
1018 * root is in progress.
1019 * < 0 If an error happened.
1020 *
1021 * NOTE: Callers need to call btrfs_check_nocow_unlock() if we return > 0.
1022 */
1023int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1024 size_t *write_bytes, bool nowait)
1025{
1026 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1027 struct btrfs_root *root = inode->root;
1028 struct extent_state *cached_state = NULL;
1029 u64 lockstart, lockend;
1030 u64 num_bytes;
1031 int ret;
1032
1033 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1034 return 0;
1035
1036 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1037 return -EAGAIN;
1038
1039 lockstart = round_down(pos, fs_info->sectorsize);
1040 lockend = round_up(pos + *write_bytes,
1041 fs_info->sectorsize) - 1;
1042 num_bytes = lockend - lockstart + 1;
1043
1044 if (nowait) {
1045 if (!btrfs_try_lock_ordered_range(inode, lockstart, lockend,
1046 &cached_state)) {
1047 btrfs_drew_write_unlock(&root->snapshot_lock);
1048 return -EAGAIN;
1049 }
1050 } else {
1051 btrfs_lock_and_flush_ordered_range(inode, lockstart, lockend,
1052 &cached_state);
1053 }
1054 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1055 NULL, nowait, false);
1056 if (ret <= 0)
1057 btrfs_drew_write_unlock(&root->snapshot_lock);
1058 else
1059 *write_bytes = min_t(size_t, *write_bytes ,
1060 num_bytes - pos + lockstart);
1061 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
1062
1063 return ret;
1064}
1065
1066void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1067{
1068 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1069}
1070
1071int btrfs_write_check(struct kiocb *iocb, size_t count)
1072{
1073 struct file *file = iocb->ki_filp;
1074 struct inode *inode = file_inode(file);
1075 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1076 loff_t pos = iocb->ki_pos;
1077 int ret;
1078 loff_t oldsize;
1079
1080 /*
1081 * Quickly bail out on NOWAIT writes if we don't have the nodatacow or
1082 * prealloc flags, as without those flags we always have to COW. We will
1083 * later check if we can really COW into the target range (using
1084 * can_nocow_extent() at btrfs_get_blocks_direct_write()).
1085 */
1086 if ((iocb->ki_flags & IOCB_NOWAIT) &&
1087 !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1088 return -EAGAIN;
1089
1090 ret = file_remove_privs(file);
1091 if (ret)
1092 return ret;
1093
1094 /*
1095 * We reserve space for updating the inode when we reserve space for the
1096 * extent we are going to write, so we will enospc out there. We don't
1097 * need to start yet another transaction to update the inode as we will
1098 * update the inode when we finish writing whatever data we write.
1099 */
1100 if (!IS_NOCMTIME(inode)) {
1101 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
1102 inode_inc_iversion(inode);
1103 }
1104
1105 oldsize = i_size_read(inode);
1106 if (pos > oldsize) {
1107 /* Expand hole size to cover write data, preventing empty gap */
1108 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1109
1110 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1111 if (ret)
1112 return ret;
1113 }
1114
1115 return 0;
1116}
1117
1118ssize_t btrfs_buffered_write(struct kiocb *iocb, struct iov_iter *i)
1119{
1120 struct file *file = iocb->ki_filp;
1121 loff_t pos;
1122 struct inode *inode = file_inode(file);
1123 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1124 struct extent_changeset *data_reserved = NULL;
1125 u64 release_bytes = 0;
1126 u64 lockstart;
1127 u64 lockend;
1128 size_t num_written = 0;
1129 ssize_t ret;
1130 loff_t old_isize;
1131 unsigned int ilock_flags = 0;
1132 const bool nowait = (iocb->ki_flags & IOCB_NOWAIT);
1133 unsigned int bdp_flags = (nowait ? BDP_ASYNC : 0);
1134 bool only_release_metadata = false;
1135
1136 if (nowait)
1137 ilock_flags |= BTRFS_ILOCK_TRY;
1138
1139 ret = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1140 if (ret < 0)
1141 return ret;
1142
1143 /*
1144 * We can only trust the isize with inode lock held, or it can race with
1145 * other buffered writes and cause incorrect call of
1146 * pagecache_isize_extended() to overwrite existing data.
1147 */
1148 old_isize = i_size_read(inode);
1149
1150 ret = generic_write_checks(iocb, i);
1151 if (ret <= 0)
1152 goto out;
1153
1154 ret = btrfs_write_check(iocb, ret);
1155 if (ret < 0)
1156 goto out;
1157
1158 pos = iocb->ki_pos;
1159 while (iov_iter_count(i) > 0) {
1160 struct extent_state *cached_state = NULL;
1161 size_t offset = offset_in_page(pos);
1162 size_t sector_offset;
1163 size_t write_bytes = min(iov_iter_count(i), PAGE_SIZE - offset);
1164 size_t reserve_bytes;
1165 size_t copied;
1166 size_t dirty_sectors;
1167 size_t num_sectors;
1168 struct folio *folio = NULL;
1169 int extents_locked;
1170 bool force_page_uptodate = false;
1171
1172 /*
1173 * Fault pages before locking them in prepare_one_folio()
1174 * to avoid recursive lock
1175 */
1176 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1177 ret = -EFAULT;
1178 break;
1179 }
1180
1181 only_release_metadata = false;
1182 sector_offset = pos & (fs_info->sectorsize - 1);
1183
1184 extent_changeset_release(data_reserved);
1185 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1186 &data_reserved, pos,
1187 write_bytes, nowait);
1188 if (ret < 0) {
1189 int can_nocow;
1190
1191 if (nowait && (ret == -ENOSPC || ret == -EAGAIN)) {
1192 ret = -EAGAIN;
1193 break;
1194 }
1195
1196 /*
1197 * If we don't have to COW at the offset, reserve
1198 * metadata only. write_bytes may get smaller than
1199 * requested here.
1200 */
1201 can_nocow = btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1202 &write_bytes, nowait);
1203 if (can_nocow < 0)
1204 ret = can_nocow;
1205 if (can_nocow > 0)
1206 ret = 0;
1207 if (ret)
1208 break;
1209 only_release_metadata = true;
1210 }
1211
1212 reserve_bytes = round_up(write_bytes + sector_offset,
1213 fs_info->sectorsize);
1214 WARN_ON(reserve_bytes == 0);
1215 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1216 reserve_bytes,
1217 reserve_bytes, nowait);
1218 if (ret) {
1219 if (!only_release_metadata)
1220 btrfs_free_reserved_data_space(BTRFS_I(inode),
1221 data_reserved, pos,
1222 write_bytes);
1223 else
1224 btrfs_check_nocow_unlock(BTRFS_I(inode));
1225
1226 if (nowait && ret == -ENOSPC)
1227 ret = -EAGAIN;
1228 break;
1229 }
1230
1231 release_bytes = reserve_bytes;
1232again:
1233 ret = balance_dirty_pages_ratelimited_flags(inode->i_mapping, bdp_flags);
1234 if (ret) {
1235 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1236 break;
1237 }
1238
1239 ret = prepare_one_folio(inode, &folio, pos, write_bytes,
1240 force_page_uptodate, false);
1241 if (ret) {
1242 btrfs_delalloc_release_extents(BTRFS_I(inode),
1243 reserve_bytes);
1244 break;
1245 }
1246
1247 extents_locked = lock_and_cleanup_extent_if_need(BTRFS_I(inode),
1248 folio, pos, write_bytes, &lockstart,
1249 &lockend, nowait, &cached_state);
1250 if (extents_locked < 0) {
1251 if (!nowait && extents_locked == -EAGAIN)
1252 goto again;
1253
1254 btrfs_delalloc_release_extents(BTRFS_I(inode),
1255 reserve_bytes);
1256 ret = extents_locked;
1257 break;
1258 }
1259
1260 copied = btrfs_copy_from_user(pos, write_bytes, folio, i);
1261
1262 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1263 dirty_sectors = round_up(copied + sector_offset,
1264 fs_info->sectorsize);
1265 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1266
1267 if (copied == 0) {
1268 force_page_uptodate = true;
1269 dirty_sectors = 0;
1270 } else {
1271 force_page_uptodate = false;
1272 }
1273
1274 if (num_sectors > dirty_sectors) {
1275 /* release everything except the sectors we dirtied */
1276 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1277 if (only_release_metadata) {
1278 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1279 release_bytes, true);
1280 } else {
1281 u64 release_start = round_up(pos + copied,
1282 fs_info->sectorsize);
1283 btrfs_delalloc_release_space(BTRFS_I(inode),
1284 data_reserved, release_start,
1285 release_bytes, true);
1286 }
1287 }
1288
1289 release_bytes = round_up(copied + sector_offset,
1290 fs_info->sectorsize);
1291
1292 ret = btrfs_dirty_folio(BTRFS_I(inode), folio, pos, copied,
1293 &cached_state, only_release_metadata);
1294
1295 /*
1296 * If we have not locked the extent range, because the range's
1297 * start offset is >= i_size, we might still have a non-NULL
1298 * cached extent state, acquired while marking the extent range
1299 * as delalloc through btrfs_dirty_page(). Therefore free any
1300 * possible cached extent state to avoid a memory leak.
1301 */
1302 if (extents_locked)
1303 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
1304 lockend, &cached_state);
1305 else
1306 free_extent_state(cached_state);
1307
1308 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1309 if (ret) {
1310 btrfs_drop_folio(fs_info, folio, pos, copied);
1311 break;
1312 }
1313
1314 release_bytes = 0;
1315 if (only_release_metadata)
1316 btrfs_check_nocow_unlock(BTRFS_I(inode));
1317
1318 btrfs_drop_folio(fs_info, folio, pos, copied);
1319
1320 cond_resched();
1321
1322 pos += copied;
1323 num_written += copied;
1324 }
1325
1326 if (release_bytes) {
1327 if (only_release_metadata) {
1328 btrfs_check_nocow_unlock(BTRFS_I(inode));
1329 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1330 release_bytes, true);
1331 } else {
1332 btrfs_delalloc_release_space(BTRFS_I(inode),
1333 data_reserved,
1334 round_down(pos, fs_info->sectorsize),
1335 release_bytes, true);
1336 }
1337 }
1338
1339 extent_changeset_free(data_reserved);
1340 if (num_written > 0) {
1341 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1342 iocb->ki_pos += num_written;
1343 }
1344out:
1345 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1346 return num_written ? num_written : ret;
1347}
1348
1349static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
1350 const struct btrfs_ioctl_encoded_io_args *encoded)
1351{
1352 struct file *file = iocb->ki_filp;
1353 struct inode *inode = file_inode(file);
1354 loff_t count;
1355 ssize_t ret;
1356
1357 btrfs_inode_lock(BTRFS_I(inode), 0);
1358 count = encoded->len;
1359 ret = generic_write_checks_count(iocb, &count);
1360 if (ret == 0 && count != encoded->len) {
1361 /*
1362 * The write got truncated by generic_write_checks_count(). We
1363 * can't do a partial encoded write.
1364 */
1365 ret = -EFBIG;
1366 }
1367 if (ret || encoded->len == 0)
1368 goto out;
1369
1370 ret = btrfs_write_check(iocb, encoded->len);
1371 if (ret < 0)
1372 goto out;
1373
1374 ret = btrfs_do_encoded_write(iocb, from, encoded);
1375out:
1376 btrfs_inode_unlock(BTRFS_I(inode), 0);
1377 return ret;
1378}
1379
1380ssize_t btrfs_do_write_iter(struct kiocb *iocb, struct iov_iter *from,
1381 const struct btrfs_ioctl_encoded_io_args *encoded)
1382{
1383 struct file *file = iocb->ki_filp;
1384 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1385 ssize_t num_written, num_sync;
1386
1387 /*
1388 * If the fs flips readonly due to some impossible error, although we
1389 * have opened a file as writable, we have to stop this write operation
1390 * to ensure consistency.
1391 */
1392 if (BTRFS_FS_ERROR(inode->root->fs_info))
1393 return -EROFS;
1394
1395 if (encoded && (iocb->ki_flags & IOCB_NOWAIT))
1396 return -EOPNOTSUPP;
1397
1398 if (encoded) {
1399 num_written = btrfs_encoded_write(iocb, from, encoded);
1400 num_sync = encoded->len;
1401 } else if (iocb->ki_flags & IOCB_DIRECT) {
1402 num_written = btrfs_direct_write(iocb, from);
1403 num_sync = num_written;
1404 } else {
1405 num_written = btrfs_buffered_write(iocb, from);
1406 num_sync = num_written;
1407 }
1408
1409 btrfs_set_inode_last_sub_trans(inode);
1410
1411 if (num_sync > 0) {
1412 num_sync = generic_write_sync(iocb, num_sync);
1413 if (num_sync < 0)
1414 num_written = num_sync;
1415 }
1416
1417 return num_written;
1418}
1419
1420static ssize_t btrfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1421{
1422 return btrfs_do_write_iter(iocb, from, NULL);
1423}
1424
1425int btrfs_release_file(struct inode *inode, struct file *filp)
1426{
1427 struct btrfs_file_private *private = filp->private_data;
1428
1429 if (private) {
1430 kfree(private->filldir_buf);
1431 free_extent_state(private->llseek_cached_state);
1432 kfree(private);
1433 filp->private_data = NULL;
1434 }
1435
1436 /*
1437 * Set by setattr when we are about to truncate a file from a non-zero
1438 * size to a zero size. This tries to flush down new bytes that may
1439 * have been written if the application were using truncate to replace
1440 * a file in place.
1441 */
1442 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
1443 &BTRFS_I(inode)->runtime_flags))
1444 filemap_flush(inode->i_mapping);
1445 return 0;
1446}
1447
1448static int start_ordered_ops(struct btrfs_inode *inode, loff_t start, loff_t end)
1449{
1450 int ret;
1451 struct blk_plug plug;
1452
1453 /*
1454 * This is only called in fsync, which would do synchronous writes, so
1455 * a plug can merge adjacent IOs as much as possible. Esp. in case of
1456 * multiple disks using raid profile, a large IO can be split to
1457 * several segments of stripe length (currently 64K).
1458 */
1459 blk_start_plug(&plug);
1460 ret = btrfs_fdatawrite_range(inode, start, end);
1461 blk_finish_plug(&plug);
1462
1463 return ret;
1464}
1465
1466static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
1467{
1468 struct btrfs_inode *inode = ctx->inode;
1469 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1470
1471 if (btrfs_inode_in_log(inode, btrfs_get_fs_generation(fs_info)) &&
1472 list_empty(&ctx->ordered_extents))
1473 return true;
1474
1475 /*
1476 * If we are doing a fast fsync we can not bail out if the inode's
1477 * last_trans is <= then the last committed transaction, because we only
1478 * update the last_trans of the inode during ordered extent completion,
1479 * and for a fast fsync we don't wait for that, we only wait for the
1480 * writeback to complete.
1481 */
1482 if (inode->last_trans <= btrfs_get_last_trans_committed(fs_info) &&
1483 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
1484 list_empty(&ctx->ordered_extents)))
1485 return true;
1486
1487 return false;
1488}
1489
1490/*
1491 * fsync call for both files and directories. This logs the inode into
1492 * the tree log instead of forcing full commits whenever possible.
1493 *
1494 * It needs to call filemap_fdatawait so that all ordered extent updates are
1495 * in the metadata btree are up to date for copying to the log.
1496 *
1497 * It drops the inode mutex before doing the tree log commit. This is an
1498 * important optimization for directories because holding the mutex prevents
1499 * new operations on the dir while we write to disk.
1500 */
1501int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1502{
1503 struct dentry *dentry = file_dentry(file);
1504 struct btrfs_inode *inode = BTRFS_I(d_inode(dentry));
1505 struct btrfs_root *root = inode->root;
1506 struct btrfs_fs_info *fs_info = root->fs_info;
1507 struct btrfs_trans_handle *trans;
1508 struct btrfs_log_ctx ctx;
1509 int ret = 0, err;
1510 u64 len;
1511 bool full_sync;
1512 bool skip_ilock = false;
1513
1514 if (current->journal_info == BTRFS_TRANS_DIO_WRITE_STUB) {
1515 skip_ilock = true;
1516 current->journal_info = NULL;
1517 btrfs_assert_inode_locked(inode);
1518 }
1519
1520 trace_btrfs_sync_file(file, datasync);
1521
1522 btrfs_init_log_ctx(&ctx, inode);
1523
1524 /*
1525 * Always set the range to a full range, otherwise we can get into
1526 * several problems, from missing file extent items to represent holes
1527 * when not using the NO_HOLES feature, to log tree corruption due to
1528 * races between hole detection during logging and completion of ordered
1529 * extents outside the range, to missing checksums due to ordered extents
1530 * for which we flushed only a subset of their pages.
1531 */
1532 start = 0;
1533 end = LLONG_MAX;
1534 len = (u64)LLONG_MAX + 1;
1535
1536 /*
1537 * We write the dirty pages in the range and wait until they complete
1538 * out of the ->i_mutex. If so, we can flush the dirty pages by
1539 * multi-task, and make the performance up. See
1540 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1541 */
1542 ret = start_ordered_ops(inode, start, end);
1543 if (ret)
1544 goto out;
1545
1546 if (skip_ilock)
1547 down_write(&inode->i_mmap_lock);
1548 else
1549 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
1550
1551 atomic_inc(&root->log_batch);
1552
1553 /*
1554 * Before we acquired the inode's lock and the mmap lock, someone may
1555 * have dirtied more pages in the target range. We need to make sure
1556 * that writeback for any such pages does not start while we are logging
1557 * the inode, because if it does, any of the following might happen when
1558 * we are not doing a full inode sync:
1559 *
1560 * 1) We log an extent after its writeback finishes but before its
1561 * checksums are added to the csum tree, leading to -EIO errors
1562 * when attempting to read the extent after a log replay.
1563 *
1564 * 2) We can end up logging an extent before its writeback finishes.
1565 * Therefore after the log replay we will have a file extent item
1566 * pointing to an unwritten extent (and no data checksums as well).
1567 *
1568 * So trigger writeback for any eventual new dirty pages and then we
1569 * wait for all ordered extents to complete below.
1570 */
1571 ret = start_ordered_ops(inode, start, end);
1572 if (ret) {
1573 if (skip_ilock)
1574 up_write(&inode->i_mmap_lock);
1575 else
1576 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
1577 goto out;
1578 }
1579
1580 /*
1581 * Always check for the full sync flag while holding the inode's lock,
1582 * to avoid races with other tasks. The flag must be either set all the
1583 * time during logging or always off all the time while logging.
1584 * We check the flag here after starting delalloc above, because when
1585 * running delalloc the full sync flag may be set if we need to drop
1586 * extra extent map ranges due to temporary memory allocation failures.
1587 */
1588 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
1589
1590 /*
1591 * We have to do this here to avoid the priority inversion of waiting on
1592 * IO of a lower priority task while holding a transaction open.
1593 *
1594 * For a full fsync we wait for the ordered extents to complete while
1595 * for a fast fsync we wait just for writeback to complete, and then
1596 * attach the ordered extents to the transaction so that a transaction
1597 * commit waits for their completion, to avoid data loss if we fsync,
1598 * the current transaction commits before the ordered extents complete
1599 * and a power failure happens right after that.
1600 *
1601 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
1602 * logical address recorded in the ordered extent may change. We need
1603 * to wait for the IO to stabilize the logical address.
1604 */
1605 if (full_sync || btrfs_is_zoned(fs_info)) {
1606 ret = btrfs_wait_ordered_range(inode, start, len);
1607 clear_bit(BTRFS_INODE_COW_WRITE_ERROR, &inode->runtime_flags);
1608 } else {
1609 /*
1610 * Get our ordered extents as soon as possible to avoid doing
1611 * checksum lookups in the csum tree, and use instead the
1612 * checksums attached to the ordered extents.
1613 */
1614 btrfs_get_ordered_extents_for_logging(inode, &ctx.ordered_extents);
1615 ret = filemap_fdatawait_range(inode->vfs_inode.i_mapping, start, end);
1616 if (ret)
1617 goto out_release_extents;
1618
1619 /*
1620 * Check and clear the BTRFS_INODE_COW_WRITE_ERROR now after
1621 * starting and waiting for writeback, because for buffered IO
1622 * it may have been set during the end IO callback
1623 * (end_bbio_data_write() -> btrfs_finish_ordered_extent()) in
1624 * case an error happened and we need to wait for ordered
1625 * extents to complete so that any extent maps that point to
1626 * unwritten locations are dropped and we don't log them.
1627 */
1628 if (test_and_clear_bit(BTRFS_INODE_COW_WRITE_ERROR, &inode->runtime_flags))
1629 ret = btrfs_wait_ordered_range(inode, start, len);
1630 }
1631
1632 if (ret)
1633 goto out_release_extents;
1634
1635 atomic_inc(&root->log_batch);
1636
1637 if (skip_inode_logging(&ctx)) {
1638 /*
1639 * We've had everything committed since the last time we were
1640 * modified so clear this flag in case it was set for whatever
1641 * reason, it's no longer relevant.
1642 */
1643 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
1644 /*
1645 * An ordered extent might have started before and completed
1646 * already with io errors, in which case the inode was not
1647 * updated and we end up here. So check the inode's mapping
1648 * for any errors that might have happened since we last
1649 * checked called fsync.
1650 */
1651 ret = filemap_check_wb_err(inode->vfs_inode.i_mapping, file->f_wb_err);
1652 goto out_release_extents;
1653 }
1654
1655 btrfs_init_log_ctx_scratch_eb(&ctx);
1656
1657 /*
1658 * We use start here because we will need to wait on the IO to complete
1659 * in btrfs_sync_log, which could require joining a transaction (for
1660 * example checking cross references in the nocow path). If we use join
1661 * here we could get into a situation where we're waiting on IO to
1662 * happen that is blocked on a transaction trying to commit. With start
1663 * we inc the extwriter counter, so we wait for all extwriters to exit
1664 * before we start blocking joiners. This comment is to keep somebody
1665 * from thinking they are super smart and changing this to
1666 * btrfs_join_transaction *cough*Josef*cough*.
1667 */
1668 trans = btrfs_start_transaction(root, 0);
1669 if (IS_ERR(trans)) {
1670 ret = PTR_ERR(trans);
1671 goto out_release_extents;
1672 }
1673 trans->in_fsync = true;
1674
1675 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
1676 /*
1677 * Scratch eb no longer needed, release before syncing log or commit
1678 * transaction, to avoid holding unnecessary memory during such long
1679 * operations.
1680 */
1681 if (ctx.scratch_eb) {
1682 free_extent_buffer(ctx.scratch_eb);
1683 ctx.scratch_eb = NULL;
1684 }
1685 btrfs_release_log_ctx_extents(&ctx);
1686 if (ret < 0) {
1687 /* Fallthrough and commit/free transaction. */
1688 ret = BTRFS_LOG_FORCE_COMMIT;
1689 }
1690
1691 /* we've logged all the items and now have a consistent
1692 * version of the file in the log. It is possible that
1693 * someone will come in and modify the file, but that's
1694 * fine because the log is consistent on disk, and we
1695 * have references to all of the file's extents
1696 *
1697 * It is possible that someone will come in and log the
1698 * file again, but that will end up using the synchronization
1699 * inside btrfs_sync_log to keep things safe.
1700 */
1701 if (skip_ilock)
1702 up_write(&inode->i_mmap_lock);
1703 else
1704 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
1705
1706 if (ret == BTRFS_NO_LOG_SYNC) {
1707 ret = btrfs_end_transaction(trans);
1708 goto out;
1709 }
1710
1711 /* We successfully logged the inode, attempt to sync the log. */
1712 if (!ret) {
1713 ret = btrfs_sync_log(trans, root, &ctx);
1714 if (!ret) {
1715 ret = btrfs_end_transaction(trans);
1716 goto out;
1717 }
1718 }
1719
1720 /*
1721 * At this point we need to commit the transaction because we had
1722 * btrfs_need_log_full_commit() or some other error.
1723 *
1724 * If we didn't do a full sync we have to stop the trans handle, wait on
1725 * the ordered extents, start it again and commit the transaction. If
1726 * we attempt to wait on the ordered extents here we could deadlock with
1727 * something like fallocate() that is holding the extent lock trying to
1728 * start a transaction while some other thread is trying to commit the
1729 * transaction while we (fsync) are currently holding the transaction
1730 * open.
1731 */
1732 if (!full_sync) {
1733 ret = btrfs_end_transaction(trans);
1734 if (ret)
1735 goto out;
1736 ret = btrfs_wait_ordered_range(inode, start, len);
1737 if (ret)
1738 goto out;
1739
1740 /*
1741 * This is safe to use here because we're only interested in
1742 * making sure the transaction that had the ordered extents is
1743 * committed. We aren't waiting on anything past this point,
1744 * we're purely getting the transaction and committing it.
1745 */
1746 trans = btrfs_attach_transaction_barrier(root);
1747 if (IS_ERR(trans)) {
1748 ret = PTR_ERR(trans);
1749
1750 /*
1751 * We committed the transaction and there's no currently
1752 * running transaction, this means everything we care
1753 * about made it to disk and we are done.
1754 */
1755 if (ret == -ENOENT)
1756 ret = 0;
1757 goto out;
1758 }
1759 }
1760
1761 ret = btrfs_commit_transaction(trans);
1762out:
1763 free_extent_buffer(ctx.scratch_eb);
1764 ASSERT(list_empty(&ctx.list));
1765 ASSERT(list_empty(&ctx.conflict_inodes));
1766 err = file_check_and_advance_wb_err(file);
1767 if (!ret)
1768 ret = err;
1769 return ret > 0 ? -EIO : ret;
1770
1771out_release_extents:
1772 btrfs_release_log_ctx_extents(&ctx);
1773 if (skip_ilock)
1774 up_write(&inode->i_mmap_lock);
1775 else
1776 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
1777 goto out;
1778}
1779
1780/*
1781 * btrfs_page_mkwrite() is not allowed to change the file size as it gets
1782 * called from a page fault handler when a page is first dirtied. Hence we must
1783 * be careful to check for EOF conditions here. We set the page up correctly
1784 * for a written page which means we get ENOSPC checking when writing into
1785 * holes and correct delalloc and unwritten extent mapping on filesystems that
1786 * support these features.
1787 *
1788 * We are not allowed to take the i_mutex here so we have to play games to
1789 * protect against truncate races as the page could now be beyond EOF. Because
1790 * truncate_setsize() writes the inode size before removing pages, once we have
1791 * the page lock we can determine safely if the page is beyond EOF. If it is not
1792 * beyond EOF, then the page is guaranteed safe against truncation until we
1793 * unlock the page.
1794 */
1795static vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
1796{
1797 struct page *page = vmf->page;
1798 struct folio *folio = page_folio(page);
1799 struct inode *inode = file_inode(vmf->vma->vm_file);
1800 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1801 struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1802 struct btrfs_ordered_extent *ordered;
1803 struct extent_state *cached_state = NULL;
1804 struct extent_changeset *data_reserved = NULL;
1805 unsigned long zero_start;
1806 loff_t size;
1807 vm_fault_t ret;
1808 int ret2;
1809 int reserved = 0;
1810 u64 reserved_space;
1811 u64 page_start;
1812 u64 page_end;
1813 u64 end;
1814
1815 ASSERT(folio_order(folio) == 0);
1816
1817 reserved_space = PAGE_SIZE;
1818
1819 sb_start_pagefault(inode->i_sb);
1820 page_start = folio_pos(folio);
1821 page_end = page_start + folio_size(folio) - 1;
1822 end = page_end;
1823
1824 /*
1825 * Reserving delalloc space after obtaining the page lock can lead to
1826 * deadlock. For example, if a dirty page is locked by this function
1827 * and the call to btrfs_delalloc_reserve_space() ends up triggering
1828 * dirty page write out, then the btrfs_writepages() function could
1829 * end up waiting indefinitely to get a lock on the page currently
1830 * being processed by btrfs_page_mkwrite() function.
1831 */
1832 ret2 = btrfs_delalloc_reserve_space(BTRFS_I(inode), &data_reserved,
1833 page_start, reserved_space);
1834 if (!ret2) {
1835 ret2 = file_update_time(vmf->vma->vm_file);
1836 reserved = 1;
1837 }
1838 if (ret2) {
1839 ret = vmf_error(ret2);
1840 if (reserved)
1841 goto out;
1842 goto out_noreserve;
1843 }
1844
1845 /* Make the VM retry the fault. */
1846 ret = VM_FAULT_NOPAGE;
1847again:
1848 down_read(&BTRFS_I(inode)->i_mmap_lock);
1849 folio_lock(folio);
1850 size = i_size_read(inode);
1851
1852 if ((folio->mapping != inode->i_mapping) ||
1853 (page_start >= size)) {
1854 /* Page got truncated out from underneath us. */
1855 goto out_unlock;
1856 }
1857 folio_wait_writeback(folio);
1858
1859 lock_extent(io_tree, page_start, page_end, &cached_state);
1860 ret2 = set_folio_extent_mapped(folio);
1861 if (ret2 < 0) {
1862 ret = vmf_error(ret2);
1863 unlock_extent(io_tree, page_start, page_end, &cached_state);
1864 goto out_unlock;
1865 }
1866
1867 /*
1868 * We can't set the delalloc bits if there are pending ordered
1869 * extents. Drop our locks and wait for them to finish.
1870 */
1871 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start, PAGE_SIZE);
1872 if (ordered) {
1873 unlock_extent(io_tree, page_start, page_end, &cached_state);
1874 folio_unlock(folio);
1875 up_read(&BTRFS_I(inode)->i_mmap_lock);
1876 btrfs_start_ordered_extent(ordered);
1877 btrfs_put_ordered_extent(ordered);
1878 goto again;
1879 }
1880
1881 if (folio->index == ((size - 1) >> PAGE_SHIFT)) {
1882 reserved_space = round_up(size - page_start, fs_info->sectorsize);
1883 if (reserved_space < PAGE_SIZE) {
1884 end = page_start + reserved_space - 1;
1885 btrfs_delalloc_release_space(BTRFS_I(inode),
1886 data_reserved, page_start,
1887 PAGE_SIZE - reserved_space, true);
1888 }
1889 }
1890
1891 /*
1892 * page_mkwrite gets called when the page is firstly dirtied after it's
1893 * faulted in, but write(2) could also dirty a page and set delalloc
1894 * bits, thus in this case for space account reason, we still need to
1895 * clear any delalloc bits within this page range since we have to
1896 * reserve data&meta space before lock_page() (see above comments).
1897 */
1898 clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
1899 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
1900 EXTENT_DEFRAG, &cached_state);
1901
1902 ret2 = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start, end, 0,
1903 &cached_state);
1904 if (ret2) {
1905 unlock_extent(io_tree, page_start, page_end, &cached_state);
1906 ret = VM_FAULT_SIGBUS;
1907 goto out_unlock;
1908 }
1909
1910 /* Page is wholly or partially inside EOF. */
1911 if (page_start + folio_size(folio) > size)
1912 zero_start = offset_in_folio(folio, size);
1913 else
1914 zero_start = PAGE_SIZE;
1915
1916 if (zero_start != PAGE_SIZE)
1917 folio_zero_range(folio, zero_start, folio_size(folio) - zero_start);
1918
1919 btrfs_folio_clear_checked(fs_info, folio, page_start, PAGE_SIZE);
1920 btrfs_folio_set_dirty(fs_info, folio, page_start, end + 1 - page_start);
1921 btrfs_folio_set_uptodate(fs_info, folio, page_start, end + 1 - page_start);
1922
1923 btrfs_set_inode_last_sub_trans(BTRFS_I(inode));
1924
1925 unlock_extent(io_tree, page_start, page_end, &cached_state);
1926 up_read(&BTRFS_I(inode)->i_mmap_lock);
1927
1928 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
1929 sb_end_pagefault(inode->i_sb);
1930 extent_changeset_free(data_reserved);
1931 return VM_FAULT_LOCKED;
1932
1933out_unlock:
1934 folio_unlock(folio);
1935 up_read(&BTRFS_I(inode)->i_mmap_lock);
1936out:
1937 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
1938 btrfs_delalloc_release_space(BTRFS_I(inode), data_reserved, page_start,
1939 reserved_space, (ret != 0));
1940out_noreserve:
1941 sb_end_pagefault(inode->i_sb);
1942 extent_changeset_free(data_reserved);
1943 return ret;
1944}
1945
1946static const struct vm_operations_struct btrfs_file_vm_ops = {
1947 .fault = filemap_fault,
1948 .map_pages = filemap_map_pages,
1949 .page_mkwrite = btrfs_page_mkwrite,
1950};
1951
1952static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
1953{
1954 struct address_space *mapping = filp->f_mapping;
1955
1956 if (!mapping->a_ops->read_folio)
1957 return -ENOEXEC;
1958
1959 file_accessed(filp);
1960 vma->vm_ops = &btrfs_file_vm_ops;
1961
1962 return 0;
1963}
1964
1965static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
1966 int slot, u64 start, u64 end)
1967{
1968 struct btrfs_file_extent_item *fi;
1969 struct btrfs_key key;
1970
1971 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1972 return 0;
1973
1974 btrfs_item_key_to_cpu(leaf, &key, slot);
1975 if (key.objectid != btrfs_ino(inode) ||
1976 key.type != BTRFS_EXTENT_DATA_KEY)
1977 return 0;
1978
1979 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1980
1981 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
1982 return 0;
1983
1984 if (btrfs_file_extent_disk_bytenr(leaf, fi))
1985 return 0;
1986
1987 if (key.offset == end)
1988 return 1;
1989 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
1990 return 1;
1991 return 0;
1992}
1993
1994static int fill_holes(struct btrfs_trans_handle *trans,
1995 struct btrfs_inode *inode,
1996 struct btrfs_path *path, u64 offset, u64 end)
1997{
1998 struct btrfs_fs_info *fs_info = trans->fs_info;
1999 struct btrfs_root *root = inode->root;
2000 struct extent_buffer *leaf;
2001 struct btrfs_file_extent_item *fi;
2002 struct extent_map *hole_em;
2003 struct btrfs_key key;
2004 int ret;
2005
2006 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2007 goto out;
2008
2009 key.objectid = btrfs_ino(inode);
2010 key.type = BTRFS_EXTENT_DATA_KEY;
2011 key.offset = offset;
2012
2013 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2014 if (ret <= 0) {
2015 /*
2016 * We should have dropped this offset, so if we find it then
2017 * something has gone horribly wrong.
2018 */
2019 if (ret == 0)
2020 ret = -EINVAL;
2021 return ret;
2022 }
2023
2024 leaf = path->nodes[0];
2025 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2026 u64 num_bytes;
2027
2028 path->slots[0]--;
2029 fi = btrfs_item_ptr(leaf, path->slots[0],
2030 struct btrfs_file_extent_item);
2031 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2032 end - offset;
2033 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2034 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2035 btrfs_set_file_extent_offset(leaf, fi, 0);
2036 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2037 btrfs_mark_buffer_dirty(trans, leaf);
2038 goto out;
2039 }
2040
2041 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2042 u64 num_bytes;
2043
2044 key.offset = offset;
2045 btrfs_set_item_key_safe(trans, path, &key);
2046 fi = btrfs_item_ptr(leaf, path->slots[0],
2047 struct btrfs_file_extent_item);
2048 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2049 offset;
2050 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2051 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2052 btrfs_set_file_extent_offset(leaf, fi, 0);
2053 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2054 btrfs_mark_buffer_dirty(trans, leaf);
2055 goto out;
2056 }
2057 btrfs_release_path(path);
2058
2059 ret = btrfs_insert_hole_extent(trans, root, btrfs_ino(inode), offset,
2060 end - offset);
2061 if (ret)
2062 return ret;
2063
2064out:
2065 btrfs_release_path(path);
2066
2067 hole_em = alloc_extent_map();
2068 if (!hole_em) {
2069 btrfs_drop_extent_map_range(inode, offset, end - 1, false);
2070 btrfs_set_inode_full_sync(inode);
2071 } else {
2072 hole_em->start = offset;
2073 hole_em->len = end - offset;
2074 hole_em->ram_bytes = hole_em->len;
2075
2076 hole_em->disk_bytenr = EXTENT_MAP_HOLE;
2077 hole_em->disk_num_bytes = 0;
2078 hole_em->generation = trans->transid;
2079
2080 ret = btrfs_replace_extent_map_range(inode, hole_em, true);
2081 free_extent_map(hole_em);
2082 if (ret)
2083 btrfs_set_inode_full_sync(inode);
2084 }
2085
2086 return 0;
2087}
2088
2089/*
2090 * Find a hole extent on given inode and change start/len to the end of hole
2091 * extent.(hole/vacuum extent whose em->start <= start &&
2092 * em->start + em->len > start)
2093 * When a hole extent is found, return 1 and modify start/len.
2094 */
2095static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2096{
2097 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2098 struct extent_map *em;
2099 int ret = 0;
2100
2101 em = btrfs_get_extent(inode, NULL,
2102 round_down(*start, fs_info->sectorsize),
2103 round_up(*len, fs_info->sectorsize));
2104 if (IS_ERR(em))
2105 return PTR_ERR(em);
2106
2107 /* Hole or vacuum extent(only exists in no-hole mode) */
2108 if (em->disk_bytenr == EXTENT_MAP_HOLE) {
2109 ret = 1;
2110 *len = em->start + em->len > *start + *len ?
2111 0 : *start + *len - em->start - em->len;
2112 *start = em->start + em->len;
2113 }
2114 free_extent_map(em);
2115 return ret;
2116}
2117
2118static void btrfs_punch_hole_lock_range(struct inode *inode,
2119 const u64 lockstart,
2120 const u64 lockend,
2121 struct extent_state **cached_state)
2122{
2123 /*
2124 * For subpage case, if the range is not at page boundary, we could
2125 * have pages at the leading/tailing part of the range.
2126 * This could lead to dead loop since filemap_range_has_page()
2127 * will always return true.
2128 * So here we need to do extra page alignment for
2129 * filemap_range_has_page().
2130 */
2131 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2132 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2133
2134 while (1) {
2135 truncate_pagecache_range(inode, lockstart, lockend);
2136
2137 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2138 cached_state);
2139 /*
2140 * We can't have ordered extents in the range, nor dirty/writeback
2141 * pages, because we have locked the inode's VFS lock in exclusive
2142 * mode, we have locked the inode's i_mmap_lock in exclusive mode,
2143 * we have flushed all delalloc in the range and we have waited
2144 * for any ordered extents in the range to complete.
2145 * We can race with anyone reading pages from this range, so after
2146 * locking the range check if we have pages in the range, and if
2147 * we do, unlock the range and retry.
2148 */
2149 if (!filemap_range_has_page(inode->i_mapping, page_lockstart,
2150 page_lockend))
2151 break;
2152
2153 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2154 cached_state);
2155 }
2156
2157 btrfs_assert_inode_range_clean(BTRFS_I(inode), lockstart, lockend);
2158}
2159
2160static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2161 struct btrfs_inode *inode,
2162 struct btrfs_path *path,
2163 struct btrfs_replace_extent_info *extent_info,
2164 const u64 replace_len,
2165 const u64 bytes_to_drop)
2166{
2167 struct btrfs_fs_info *fs_info = trans->fs_info;
2168 struct btrfs_root *root = inode->root;
2169 struct btrfs_file_extent_item *extent;
2170 struct extent_buffer *leaf;
2171 struct btrfs_key key;
2172 int slot;
2173 int ret;
2174
2175 if (replace_len == 0)
2176 return 0;
2177
2178 if (extent_info->disk_offset == 0 &&
2179 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2180 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2181 return 0;
2182 }
2183
2184 key.objectid = btrfs_ino(inode);
2185 key.type = BTRFS_EXTENT_DATA_KEY;
2186 key.offset = extent_info->file_offset;
2187 ret = btrfs_insert_empty_item(trans, root, path, &key,
2188 sizeof(struct btrfs_file_extent_item));
2189 if (ret)
2190 return ret;
2191 leaf = path->nodes[0];
2192 slot = path->slots[0];
2193 write_extent_buffer(leaf, extent_info->extent_buf,
2194 btrfs_item_ptr_offset(leaf, slot),
2195 sizeof(struct btrfs_file_extent_item));
2196 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2197 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2198 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2199 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2200 if (extent_info->is_new_extent)
2201 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2202 btrfs_mark_buffer_dirty(trans, leaf);
2203 btrfs_release_path(path);
2204
2205 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2206 replace_len);
2207 if (ret)
2208 return ret;
2209
2210 /* If it's a hole, nothing more needs to be done. */
2211 if (extent_info->disk_offset == 0) {
2212 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2213 return 0;
2214 }
2215
2216 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2217
2218 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2219 key.objectid = extent_info->disk_offset;
2220 key.type = BTRFS_EXTENT_ITEM_KEY;
2221 key.offset = extent_info->disk_len;
2222 ret = btrfs_alloc_reserved_file_extent(trans, root,
2223 btrfs_ino(inode),
2224 extent_info->file_offset,
2225 extent_info->qgroup_reserved,
2226 &key);
2227 } else {
2228 struct btrfs_ref ref = {
2229 .action = BTRFS_ADD_DELAYED_REF,
2230 .bytenr = extent_info->disk_offset,
2231 .num_bytes = extent_info->disk_len,
2232 .owning_root = btrfs_root_id(root),
2233 .ref_root = btrfs_root_id(root),
2234 };
2235 u64 ref_offset;
2236
2237 ref_offset = extent_info->file_offset - extent_info->data_offset;
2238 btrfs_init_data_ref(&ref, btrfs_ino(inode), ref_offset, 0, false);
2239 ret = btrfs_inc_extent_ref(trans, &ref);
2240 }
2241
2242 extent_info->insertions++;
2243
2244 return ret;
2245}
2246
2247/*
2248 * The respective range must have been previously locked, as well as the inode.
2249 * The end offset is inclusive (last byte of the range).
2250 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2251 * the file range with an extent.
2252 * When not punching a hole, we don't want to end up in a state where we dropped
2253 * extents without inserting a new one, so we must abort the transaction to avoid
2254 * a corruption.
2255 */
2256int btrfs_replace_file_extents(struct btrfs_inode *inode,
2257 struct btrfs_path *path, const u64 start,
2258 const u64 end,
2259 struct btrfs_replace_extent_info *extent_info,
2260 struct btrfs_trans_handle **trans_out)
2261{
2262 struct btrfs_drop_extents_args drop_args = { 0 };
2263 struct btrfs_root *root = inode->root;
2264 struct btrfs_fs_info *fs_info = root->fs_info;
2265 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2266 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2267 struct btrfs_trans_handle *trans = NULL;
2268 struct btrfs_block_rsv *rsv;
2269 unsigned int rsv_count;
2270 u64 cur_offset;
2271 u64 len = end - start;
2272 int ret = 0;
2273
2274 if (end <= start)
2275 return -EINVAL;
2276
2277 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2278 if (!rsv) {
2279 ret = -ENOMEM;
2280 goto out;
2281 }
2282 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2283 rsv->failfast = true;
2284
2285 /*
2286 * 1 - update the inode
2287 * 1 - removing the extents in the range
2288 * 1 - adding the hole extent if no_holes isn't set or if we are
2289 * replacing the range with a new extent
2290 */
2291 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2292 rsv_count = 3;
2293 else
2294 rsv_count = 2;
2295
2296 trans = btrfs_start_transaction(root, rsv_count);
2297 if (IS_ERR(trans)) {
2298 ret = PTR_ERR(trans);
2299 trans = NULL;
2300 goto out_free;
2301 }
2302
2303 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2304 min_size, false);
2305 if (WARN_ON(ret))
2306 goto out_trans;
2307 trans->block_rsv = rsv;
2308
2309 cur_offset = start;
2310 drop_args.path = path;
2311 drop_args.end = end + 1;
2312 drop_args.drop_cache = true;
2313 while (cur_offset < end) {
2314 drop_args.start = cur_offset;
2315 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2316 /* If we are punching a hole decrement the inode's byte count */
2317 if (!extent_info)
2318 btrfs_update_inode_bytes(inode, 0,
2319 drop_args.bytes_found);
2320 if (ret != -ENOSPC) {
2321 /*
2322 * The only time we don't want to abort is if we are
2323 * attempting to clone a partial inline extent, in which
2324 * case we'll get EOPNOTSUPP. However if we aren't
2325 * clone we need to abort no matter what, because if we
2326 * got EOPNOTSUPP via prealloc then we messed up and
2327 * need to abort.
2328 */
2329 if (ret &&
2330 (ret != -EOPNOTSUPP ||
2331 (extent_info && extent_info->is_new_extent)))
2332 btrfs_abort_transaction(trans, ret);
2333 break;
2334 }
2335
2336 trans->block_rsv = &fs_info->trans_block_rsv;
2337
2338 if (!extent_info && cur_offset < drop_args.drop_end &&
2339 cur_offset < ino_size) {
2340 ret = fill_holes(trans, inode, path, cur_offset,
2341 drop_args.drop_end);
2342 if (ret) {
2343 /*
2344 * If we failed then we didn't insert our hole
2345 * entries for the area we dropped, so now the
2346 * fs is corrupted, so we must abort the
2347 * transaction.
2348 */
2349 btrfs_abort_transaction(trans, ret);
2350 break;
2351 }
2352 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2353 /*
2354 * We are past the i_size here, but since we didn't
2355 * insert holes we need to clear the mapped area so we
2356 * know to not set disk_i_size in this area until a new
2357 * file extent is inserted here.
2358 */
2359 ret = btrfs_inode_clear_file_extent_range(inode,
2360 cur_offset,
2361 drop_args.drop_end - cur_offset);
2362 if (ret) {
2363 /*
2364 * We couldn't clear our area, so we could
2365 * presumably adjust up and corrupt the fs, so
2366 * we need to abort.
2367 */
2368 btrfs_abort_transaction(trans, ret);
2369 break;
2370 }
2371 }
2372
2373 if (extent_info &&
2374 drop_args.drop_end > extent_info->file_offset) {
2375 u64 replace_len = drop_args.drop_end -
2376 extent_info->file_offset;
2377
2378 ret = btrfs_insert_replace_extent(trans, inode, path,
2379 extent_info, replace_len,
2380 drop_args.bytes_found);
2381 if (ret) {
2382 btrfs_abort_transaction(trans, ret);
2383 break;
2384 }
2385 extent_info->data_len -= replace_len;
2386 extent_info->data_offset += replace_len;
2387 extent_info->file_offset += replace_len;
2388 }
2389
2390 /*
2391 * We are releasing our handle on the transaction, balance the
2392 * dirty pages of the btree inode and flush delayed items, and
2393 * then get a new transaction handle, which may now point to a
2394 * new transaction in case someone else may have committed the
2395 * transaction we used to replace/drop file extent items. So
2396 * bump the inode's iversion and update mtime and ctime except
2397 * if we are called from a dedupe context. This is because a
2398 * power failure/crash may happen after the transaction is
2399 * committed and before we finish replacing/dropping all the
2400 * file extent items we need.
2401 */
2402 inode_inc_iversion(&inode->vfs_inode);
2403
2404 if (!extent_info || extent_info->update_times)
2405 inode_set_mtime_to_ts(&inode->vfs_inode,
2406 inode_set_ctime_current(&inode->vfs_inode));
2407
2408 ret = btrfs_update_inode(trans, inode);
2409 if (ret)
2410 break;
2411
2412 btrfs_end_transaction(trans);
2413 btrfs_btree_balance_dirty(fs_info);
2414
2415 trans = btrfs_start_transaction(root, rsv_count);
2416 if (IS_ERR(trans)) {
2417 ret = PTR_ERR(trans);
2418 trans = NULL;
2419 break;
2420 }
2421
2422 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2423 rsv, min_size, false);
2424 if (WARN_ON(ret))
2425 break;
2426 trans->block_rsv = rsv;
2427
2428 cur_offset = drop_args.drop_end;
2429 len = end - cur_offset;
2430 if (!extent_info && len) {
2431 ret = find_first_non_hole(inode, &cur_offset, &len);
2432 if (unlikely(ret < 0))
2433 break;
2434 if (ret && !len) {
2435 ret = 0;
2436 break;
2437 }
2438 }
2439 }
2440
2441 /*
2442 * If we were cloning, force the next fsync to be a full one since we
2443 * we replaced (or just dropped in the case of cloning holes when
2444 * NO_HOLES is enabled) file extent items and did not setup new extent
2445 * maps for the replacement extents (or holes).
2446 */
2447 if (extent_info && !extent_info->is_new_extent)
2448 btrfs_set_inode_full_sync(inode);
2449
2450 if (ret)
2451 goto out_trans;
2452
2453 trans->block_rsv = &fs_info->trans_block_rsv;
2454 /*
2455 * If we are using the NO_HOLES feature we might have had already an
2456 * hole that overlaps a part of the region [lockstart, lockend] and
2457 * ends at (or beyond) lockend. Since we have no file extent items to
2458 * represent holes, drop_end can be less than lockend and so we must
2459 * make sure we have an extent map representing the existing hole (the
2460 * call to __btrfs_drop_extents() might have dropped the existing extent
2461 * map representing the existing hole), otherwise the fast fsync path
2462 * will not record the existence of the hole region
2463 * [existing_hole_start, lockend].
2464 */
2465 if (drop_args.drop_end <= end)
2466 drop_args.drop_end = end + 1;
2467 /*
2468 * Don't insert file hole extent item if it's for a range beyond eof
2469 * (because it's useless) or if it represents a 0 bytes range (when
2470 * cur_offset == drop_end).
2471 */
2472 if (!extent_info && cur_offset < ino_size &&
2473 cur_offset < drop_args.drop_end) {
2474 ret = fill_holes(trans, inode, path, cur_offset,
2475 drop_args.drop_end);
2476 if (ret) {
2477 /* Same comment as above. */
2478 btrfs_abort_transaction(trans, ret);
2479 goto out_trans;
2480 }
2481 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2482 /* See the comment in the loop above for the reasoning here. */
2483 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2484 drop_args.drop_end - cur_offset);
2485 if (ret) {
2486 btrfs_abort_transaction(trans, ret);
2487 goto out_trans;
2488 }
2489
2490 }
2491 if (extent_info) {
2492 ret = btrfs_insert_replace_extent(trans, inode, path,
2493 extent_info, extent_info->data_len,
2494 drop_args.bytes_found);
2495 if (ret) {
2496 btrfs_abort_transaction(trans, ret);
2497 goto out_trans;
2498 }
2499 }
2500
2501out_trans:
2502 if (!trans)
2503 goto out_free;
2504
2505 trans->block_rsv = &fs_info->trans_block_rsv;
2506 if (ret)
2507 btrfs_end_transaction(trans);
2508 else
2509 *trans_out = trans;
2510out_free:
2511 btrfs_free_block_rsv(fs_info, rsv);
2512out:
2513 return ret;
2514}
2515
2516static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2517{
2518 struct inode *inode = file_inode(file);
2519 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2520 struct btrfs_root *root = BTRFS_I(inode)->root;
2521 struct extent_state *cached_state = NULL;
2522 struct btrfs_path *path;
2523 struct btrfs_trans_handle *trans = NULL;
2524 u64 lockstart;
2525 u64 lockend;
2526 u64 tail_start;
2527 u64 tail_len;
2528 u64 orig_start = offset;
2529 int ret = 0;
2530 bool same_block;
2531 u64 ino_size;
2532 bool truncated_block = false;
2533 bool updated_inode = false;
2534
2535 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2536
2537 ret = btrfs_wait_ordered_range(BTRFS_I(inode), offset, len);
2538 if (ret)
2539 goto out_only_mutex;
2540
2541 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2542 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2543 if (ret < 0)
2544 goto out_only_mutex;
2545 if (ret && !len) {
2546 /* Already in a large hole */
2547 ret = 0;
2548 goto out_only_mutex;
2549 }
2550
2551 ret = file_modified(file);
2552 if (ret)
2553 goto out_only_mutex;
2554
2555 lockstart = round_up(offset, fs_info->sectorsize);
2556 lockend = round_down(offset + len, fs_info->sectorsize) - 1;
2557 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2558 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2559 /*
2560 * We needn't truncate any block which is beyond the end of the file
2561 * because we are sure there is no data there.
2562 */
2563 /*
2564 * Only do this if we are in the same block and we aren't doing the
2565 * entire block.
2566 */
2567 if (same_block && len < fs_info->sectorsize) {
2568 if (offset < ino_size) {
2569 truncated_block = true;
2570 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2571 0);
2572 } else {
2573 ret = 0;
2574 }
2575 goto out_only_mutex;
2576 }
2577
2578 /* zero back part of the first block */
2579 if (offset < ino_size) {
2580 truncated_block = true;
2581 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2582 if (ret) {
2583 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2584 return ret;
2585 }
2586 }
2587
2588 /* Check the aligned pages after the first unaligned page,
2589 * if offset != orig_start, which means the first unaligned page
2590 * including several following pages are already in holes,
2591 * the extra check can be skipped */
2592 if (offset == orig_start) {
2593 /* after truncate page, check hole again */
2594 len = offset + len - lockstart;
2595 offset = lockstart;
2596 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2597 if (ret < 0)
2598 goto out_only_mutex;
2599 if (ret && !len) {
2600 ret = 0;
2601 goto out_only_mutex;
2602 }
2603 lockstart = offset;
2604 }
2605
2606 /* Check the tail unaligned part is in a hole */
2607 tail_start = lockend + 1;
2608 tail_len = offset + len - tail_start;
2609 if (tail_len) {
2610 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2611 if (unlikely(ret < 0))
2612 goto out_only_mutex;
2613 if (!ret) {
2614 /* zero the front end of the last page */
2615 if (tail_start + tail_len < ino_size) {
2616 truncated_block = true;
2617 ret = btrfs_truncate_block(BTRFS_I(inode),
2618 tail_start + tail_len,
2619 0, 1);
2620 if (ret)
2621 goto out_only_mutex;
2622 }
2623 }
2624 }
2625
2626 if (lockend < lockstart) {
2627 ret = 0;
2628 goto out_only_mutex;
2629 }
2630
2631 btrfs_punch_hole_lock_range(inode, lockstart, lockend, &cached_state);
2632
2633 path = btrfs_alloc_path();
2634 if (!path) {
2635 ret = -ENOMEM;
2636 goto out;
2637 }
2638
2639 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
2640 lockend, NULL, &trans);
2641 btrfs_free_path(path);
2642 if (ret)
2643 goto out;
2644
2645 ASSERT(trans != NULL);
2646 inode_inc_iversion(inode);
2647 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
2648 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2649 updated_inode = true;
2650 btrfs_end_transaction(trans);
2651 btrfs_btree_balance_dirty(fs_info);
2652out:
2653 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2654 &cached_state);
2655out_only_mutex:
2656 if (!updated_inode && truncated_block && !ret) {
2657 /*
2658 * If we only end up zeroing part of a page, we still need to
2659 * update the inode item, so that all the time fields are
2660 * updated as well as the necessary btrfs inode in memory fields
2661 * for detecting, at fsync time, if the inode isn't yet in the
2662 * log tree or it's there but not up to date.
2663 */
2664 struct timespec64 now = inode_set_ctime_current(inode);
2665
2666 inode_inc_iversion(inode);
2667 inode_set_mtime_to_ts(inode, now);
2668 trans = btrfs_start_transaction(root, 1);
2669 if (IS_ERR(trans)) {
2670 ret = PTR_ERR(trans);
2671 } else {
2672 int ret2;
2673
2674 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2675 ret2 = btrfs_end_transaction(trans);
2676 if (!ret)
2677 ret = ret2;
2678 }
2679 }
2680 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2681 return ret;
2682}
2683
2684/* Helper structure to record which range is already reserved */
2685struct falloc_range {
2686 struct list_head list;
2687 u64 start;
2688 u64 len;
2689};
2690
2691/*
2692 * Helper function to add falloc range
2693 *
2694 * Caller should have locked the larger range of extent containing
2695 * [start, len)
2696 */
2697static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2698{
2699 struct falloc_range *range = NULL;
2700
2701 if (!list_empty(head)) {
2702 /*
2703 * As fallocate iterates by bytenr order, we only need to check
2704 * the last range.
2705 */
2706 range = list_last_entry(head, struct falloc_range, list);
2707 if (range->start + range->len == start) {
2708 range->len += len;
2709 return 0;
2710 }
2711 }
2712
2713 range = kmalloc(sizeof(*range), GFP_KERNEL);
2714 if (!range)
2715 return -ENOMEM;
2716 range->start = start;
2717 range->len = len;
2718 list_add_tail(&range->list, head);
2719 return 0;
2720}
2721
2722static int btrfs_fallocate_update_isize(struct inode *inode,
2723 const u64 end,
2724 const int mode)
2725{
2726 struct btrfs_trans_handle *trans;
2727 struct btrfs_root *root = BTRFS_I(inode)->root;
2728 int ret;
2729 int ret2;
2730
2731 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2732 return 0;
2733
2734 trans = btrfs_start_transaction(root, 1);
2735 if (IS_ERR(trans))
2736 return PTR_ERR(trans);
2737
2738 inode_set_ctime_current(inode);
2739 i_size_write(inode, end);
2740 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
2741 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2742 ret2 = btrfs_end_transaction(trans);
2743
2744 return ret ? ret : ret2;
2745}
2746
2747enum {
2748 RANGE_BOUNDARY_WRITTEN_EXTENT,
2749 RANGE_BOUNDARY_PREALLOC_EXTENT,
2750 RANGE_BOUNDARY_HOLE,
2751};
2752
2753static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
2754 u64 offset)
2755{
2756 const u64 sectorsize = inode->root->fs_info->sectorsize;
2757 struct extent_map *em;
2758 int ret;
2759
2760 offset = round_down(offset, sectorsize);
2761 em = btrfs_get_extent(inode, NULL, offset, sectorsize);
2762 if (IS_ERR(em))
2763 return PTR_ERR(em);
2764
2765 if (em->disk_bytenr == EXTENT_MAP_HOLE)
2766 ret = RANGE_BOUNDARY_HOLE;
2767 else if (em->flags & EXTENT_FLAG_PREALLOC)
2768 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2769 else
2770 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2771
2772 free_extent_map(em);
2773 return ret;
2774}
2775
2776static int btrfs_zero_range(struct inode *inode,
2777 loff_t offset,
2778 loff_t len,
2779 const int mode)
2780{
2781 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2782 struct extent_map *em;
2783 struct extent_changeset *data_reserved = NULL;
2784 int ret;
2785 u64 alloc_hint = 0;
2786 const u64 sectorsize = fs_info->sectorsize;
2787 u64 alloc_start = round_down(offset, sectorsize);
2788 u64 alloc_end = round_up(offset + len, sectorsize);
2789 u64 bytes_to_reserve = 0;
2790 bool space_reserved = false;
2791
2792 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start,
2793 alloc_end - alloc_start);
2794 if (IS_ERR(em)) {
2795 ret = PTR_ERR(em);
2796 goto out;
2797 }
2798
2799 /*
2800 * Avoid hole punching and extent allocation for some cases. More cases
2801 * could be considered, but these are unlikely common and we keep things
2802 * as simple as possible for now. Also, intentionally, if the target
2803 * range contains one or more prealloc extents together with regular
2804 * extents and holes, we drop all the existing extents and allocate a
2805 * new prealloc extent, so that we get a larger contiguous disk extent.
2806 */
2807 if (em->start <= alloc_start && (em->flags & EXTENT_FLAG_PREALLOC)) {
2808 const u64 em_end = em->start + em->len;
2809
2810 if (em_end >= offset + len) {
2811 /*
2812 * The whole range is already a prealloc extent,
2813 * do nothing except updating the inode's i_size if
2814 * needed.
2815 */
2816 free_extent_map(em);
2817 ret = btrfs_fallocate_update_isize(inode, offset + len,
2818 mode);
2819 goto out;
2820 }
2821 /*
2822 * Part of the range is already a prealloc extent, so operate
2823 * only on the remaining part of the range.
2824 */
2825 alloc_start = em_end;
2826 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2827 len = offset + len - alloc_start;
2828 offset = alloc_start;
2829 alloc_hint = extent_map_block_start(em) + em->len;
2830 }
2831 free_extent_map(em);
2832
2833 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2834 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2835 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start, sectorsize);
2836 if (IS_ERR(em)) {
2837 ret = PTR_ERR(em);
2838 goto out;
2839 }
2840
2841 if (em->flags & EXTENT_FLAG_PREALLOC) {
2842 free_extent_map(em);
2843 ret = btrfs_fallocate_update_isize(inode, offset + len,
2844 mode);
2845 goto out;
2846 }
2847 if (len < sectorsize && em->disk_bytenr != EXTENT_MAP_HOLE) {
2848 free_extent_map(em);
2849 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2850 0);
2851 if (!ret)
2852 ret = btrfs_fallocate_update_isize(inode,
2853 offset + len,
2854 mode);
2855 return ret;
2856 }
2857 free_extent_map(em);
2858 alloc_start = round_down(offset, sectorsize);
2859 alloc_end = alloc_start + sectorsize;
2860 goto reserve_space;
2861 }
2862
2863 alloc_start = round_up(offset, sectorsize);
2864 alloc_end = round_down(offset + len, sectorsize);
2865
2866 /*
2867 * For unaligned ranges, check the pages at the boundaries, they might
2868 * map to an extent, in which case we need to partially zero them, or
2869 * they might map to a hole, in which case we need our allocation range
2870 * to cover them.
2871 */
2872 if (!IS_ALIGNED(offset, sectorsize)) {
2873 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2874 offset);
2875 if (ret < 0)
2876 goto out;
2877 if (ret == RANGE_BOUNDARY_HOLE) {
2878 alloc_start = round_down(offset, sectorsize);
2879 ret = 0;
2880 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2881 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2882 if (ret)
2883 goto out;
2884 } else {
2885 ret = 0;
2886 }
2887 }
2888
2889 if (!IS_ALIGNED(offset + len, sectorsize)) {
2890 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2891 offset + len);
2892 if (ret < 0)
2893 goto out;
2894 if (ret == RANGE_BOUNDARY_HOLE) {
2895 alloc_end = round_up(offset + len, sectorsize);
2896 ret = 0;
2897 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2898 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
2899 0, 1);
2900 if (ret)
2901 goto out;
2902 } else {
2903 ret = 0;
2904 }
2905 }
2906
2907reserve_space:
2908 if (alloc_start < alloc_end) {
2909 struct extent_state *cached_state = NULL;
2910 const u64 lockstart = alloc_start;
2911 const u64 lockend = alloc_end - 1;
2912
2913 bytes_to_reserve = alloc_end - alloc_start;
2914 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2915 bytes_to_reserve);
2916 if (ret < 0)
2917 goto out;
2918 space_reserved = true;
2919 btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2920 &cached_state);
2921 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
2922 alloc_start, bytes_to_reserve);
2923 if (ret) {
2924 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
2925 lockend, &cached_state);
2926 goto out;
2927 }
2928 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
2929 alloc_end - alloc_start,
2930 fs_info->sectorsize,
2931 offset + len, &alloc_hint);
2932 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2933 &cached_state);
2934 /* btrfs_prealloc_file_range releases reserved space on error */
2935 if (ret) {
2936 space_reserved = false;
2937 goto out;
2938 }
2939 }
2940 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
2941 out:
2942 if (ret && space_reserved)
2943 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
2944 alloc_start, bytes_to_reserve);
2945 extent_changeset_free(data_reserved);
2946
2947 return ret;
2948}
2949
2950static long btrfs_fallocate(struct file *file, int mode,
2951 loff_t offset, loff_t len)
2952{
2953 struct inode *inode = file_inode(file);
2954 struct extent_state *cached_state = NULL;
2955 struct extent_changeset *data_reserved = NULL;
2956 struct falloc_range *range;
2957 struct falloc_range *tmp;
2958 LIST_HEAD(reserve_list);
2959 u64 cur_offset;
2960 u64 last_byte;
2961 u64 alloc_start;
2962 u64 alloc_end;
2963 u64 alloc_hint = 0;
2964 u64 locked_end;
2965 u64 actual_end = 0;
2966 u64 data_space_needed = 0;
2967 u64 data_space_reserved = 0;
2968 u64 qgroup_reserved = 0;
2969 struct extent_map *em;
2970 int blocksize = BTRFS_I(inode)->root->fs_info->sectorsize;
2971 int ret;
2972
2973 /* Do not allow fallocate in ZONED mode */
2974 if (btrfs_is_zoned(inode_to_fs_info(inode)))
2975 return -EOPNOTSUPP;
2976
2977 alloc_start = round_down(offset, blocksize);
2978 alloc_end = round_up(offset + len, blocksize);
2979 cur_offset = alloc_start;
2980
2981 /* Make sure we aren't being give some crap mode */
2982 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
2983 FALLOC_FL_ZERO_RANGE))
2984 return -EOPNOTSUPP;
2985
2986 if (mode & FALLOC_FL_PUNCH_HOLE)
2987 return btrfs_punch_hole(file, offset, len);
2988
2989 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2990
2991 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2992 ret = inode_newsize_ok(inode, offset + len);
2993 if (ret)
2994 goto out;
2995 }
2996
2997 ret = file_modified(file);
2998 if (ret)
2999 goto out;
3000
3001 /*
3002 * TODO: Move these two operations after we have checked
3003 * accurate reserved space, or fallocate can still fail but
3004 * with page truncated or size expanded.
3005 *
3006 * But that's a minor problem and won't do much harm BTW.
3007 */
3008 if (alloc_start > inode->i_size) {
3009 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3010 alloc_start);
3011 if (ret)
3012 goto out;
3013 } else if (offset + len > inode->i_size) {
3014 /*
3015 * If we are fallocating from the end of the file onward we
3016 * need to zero out the end of the block if i_size lands in the
3017 * middle of a block.
3018 */
3019 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3020 if (ret)
3021 goto out;
3022 }
3023
3024 /*
3025 * We have locked the inode at the VFS level (in exclusive mode) and we
3026 * have locked the i_mmap_lock lock (in exclusive mode). Now before
3027 * locking the file range, flush all dealloc in the range and wait for
3028 * all ordered extents in the range to complete. After this we can lock
3029 * the file range and, due to the previous locking we did, we know there
3030 * can't be more delalloc or ordered extents in the range.
3031 */
3032 ret = btrfs_wait_ordered_range(BTRFS_I(inode), alloc_start,
3033 alloc_end - alloc_start);
3034 if (ret)
3035 goto out;
3036
3037 if (mode & FALLOC_FL_ZERO_RANGE) {
3038 ret = btrfs_zero_range(inode, offset, len, mode);
3039 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3040 return ret;
3041 }
3042
3043 locked_end = alloc_end - 1;
3044 lock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3045 &cached_state);
3046
3047 btrfs_assert_inode_range_clean(BTRFS_I(inode), alloc_start, locked_end);
3048
3049 /* First, check if we exceed the qgroup limit */
3050 while (cur_offset < alloc_end) {
3051 em = btrfs_get_extent(BTRFS_I(inode), NULL, cur_offset,
3052 alloc_end - cur_offset);
3053 if (IS_ERR(em)) {
3054 ret = PTR_ERR(em);
3055 break;
3056 }
3057 last_byte = min(extent_map_end(em), alloc_end);
3058 actual_end = min_t(u64, extent_map_end(em), offset + len);
3059 last_byte = ALIGN(last_byte, blocksize);
3060 if (em->disk_bytenr == EXTENT_MAP_HOLE ||
3061 (cur_offset >= inode->i_size &&
3062 !(em->flags & EXTENT_FLAG_PREALLOC))) {
3063 const u64 range_len = last_byte - cur_offset;
3064
3065 ret = add_falloc_range(&reserve_list, cur_offset, range_len);
3066 if (ret < 0) {
3067 free_extent_map(em);
3068 break;
3069 }
3070 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3071 &data_reserved, cur_offset, range_len);
3072 if (ret < 0) {
3073 free_extent_map(em);
3074 break;
3075 }
3076 qgroup_reserved += range_len;
3077 data_space_needed += range_len;
3078 }
3079 free_extent_map(em);
3080 cur_offset = last_byte;
3081 }
3082
3083 if (!ret && data_space_needed > 0) {
3084 /*
3085 * We are safe to reserve space here as we can't have delalloc
3086 * in the range, see above.
3087 */
3088 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3089 data_space_needed);
3090 if (!ret)
3091 data_space_reserved = data_space_needed;
3092 }
3093
3094 /*
3095 * If ret is still 0, means we're OK to fallocate.
3096 * Or just cleanup the list and exit.
3097 */
3098 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3099 if (!ret) {
3100 ret = btrfs_prealloc_file_range(inode, mode,
3101 range->start,
3102 range->len, blocksize,
3103 offset + len, &alloc_hint);
3104 /*
3105 * btrfs_prealloc_file_range() releases space even
3106 * if it returns an error.
3107 */
3108 data_space_reserved -= range->len;
3109 qgroup_reserved -= range->len;
3110 } else if (data_space_reserved > 0) {
3111 btrfs_free_reserved_data_space(BTRFS_I(inode),
3112 data_reserved, range->start,
3113 range->len);
3114 data_space_reserved -= range->len;
3115 qgroup_reserved -= range->len;
3116 } else if (qgroup_reserved > 0) {
3117 btrfs_qgroup_free_data(BTRFS_I(inode), data_reserved,
3118 range->start, range->len, NULL);
3119 qgroup_reserved -= range->len;
3120 }
3121 list_del(&range->list);
3122 kfree(range);
3123 }
3124 if (ret < 0)
3125 goto out_unlock;
3126
3127 /*
3128 * We didn't need to allocate any more space, but we still extended the
3129 * size of the file so we need to update i_size and the inode item.
3130 */
3131 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3132out_unlock:
3133 unlock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3134 &cached_state);
3135out:
3136 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3137 extent_changeset_free(data_reserved);
3138 return ret;
3139}
3140
3141/*
3142 * Helper for btrfs_find_delalloc_in_range(). Find a subrange in a given range
3143 * that has unflushed and/or flushing delalloc. There might be other adjacent
3144 * subranges after the one it found, so btrfs_find_delalloc_in_range() keeps
3145 * looping while it gets adjacent subranges, and merging them together.
3146 */
3147static bool find_delalloc_subrange(struct btrfs_inode *inode, u64 start, u64 end,
3148 struct extent_state **cached_state,
3149 bool *search_io_tree,
3150 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3151{
3152 u64 len = end + 1 - start;
3153 u64 delalloc_len = 0;
3154 struct btrfs_ordered_extent *oe;
3155 u64 oe_start;
3156 u64 oe_end;
3157
3158 /*
3159 * Search the io tree first for EXTENT_DELALLOC. If we find any, it
3160 * means we have delalloc (dirty pages) for which writeback has not
3161 * started yet.
3162 */
3163 if (*search_io_tree) {
3164 spin_lock(&inode->lock);
3165 if (inode->delalloc_bytes > 0) {
3166 spin_unlock(&inode->lock);
3167 *delalloc_start_ret = start;
3168 delalloc_len = count_range_bits(&inode->io_tree,
3169 delalloc_start_ret, end,
3170 len, EXTENT_DELALLOC, 1,
3171 cached_state);
3172 } else {
3173 spin_unlock(&inode->lock);
3174 }
3175 }
3176
3177 if (delalloc_len > 0) {
3178 /*
3179 * If delalloc was found then *delalloc_start_ret has a sector size
3180 * aligned value (rounded down).
3181 */
3182 *delalloc_end_ret = *delalloc_start_ret + delalloc_len - 1;
3183
3184 if (*delalloc_start_ret == start) {
3185 /* Delalloc for the whole range, nothing more to do. */
3186 if (*delalloc_end_ret == end)
3187 return true;
3188 /* Else trim our search range for ordered extents. */
3189 start = *delalloc_end_ret + 1;
3190 len = end + 1 - start;
3191 }
3192 } else {
3193 /* No delalloc, future calls don't need to search again. */
3194 *search_io_tree = false;
3195 }
3196
3197 /*
3198 * Now also check if there's any ordered extent in the range.
3199 * We do this because:
3200 *
3201 * 1) When delalloc is flushed, the file range is locked, we clear the
3202 * EXTENT_DELALLOC bit from the io tree and create an extent map and
3203 * an ordered extent for the write. So we might just have been called
3204 * after delalloc is flushed and before the ordered extent completes
3205 * and inserts the new file extent item in the subvolume's btree;
3206 *
3207 * 2) We may have an ordered extent created by flushing delalloc for a
3208 * subrange that starts before the subrange we found marked with
3209 * EXTENT_DELALLOC in the io tree.
3210 *
3211 * We could also use the extent map tree to find such delalloc that is
3212 * being flushed, but using the ordered extents tree is more efficient
3213 * because it's usually much smaller as ordered extents are removed from
3214 * the tree once they complete. With the extent maps, we mau have them
3215 * in the extent map tree for a very long time, and they were either
3216 * created by previous writes or loaded by read operations.
3217 */
3218 oe = btrfs_lookup_first_ordered_range(inode, start, len);
3219 if (!oe)
3220 return (delalloc_len > 0);
3221
3222 /* The ordered extent may span beyond our search range. */
3223 oe_start = max(oe->file_offset, start);
3224 oe_end = min(oe->file_offset + oe->num_bytes - 1, end);
3225
3226 btrfs_put_ordered_extent(oe);
3227
3228 /* Don't have unflushed delalloc, return the ordered extent range. */
3229 if (delalloc_len == 0) {
3230 *delalloc_start_ret = oe_start;
3231 *delalloc_end_ret = oe_end;
3232 return true;
3233 }
3234
3235 /*
3236 * We have both unflushed delalloc (io_tree) and an ordered extent.
3237 * If the ranges are adjacent returned a combined range, otherwise
3238 * return the leftmost range.
3239 */
3240 if (oe_start < *delalloc_start_ret) {
3241 if (oe_end < *delalloc_start_ret)
3242 *delalloc_end_ret = oe_end;
3243 *delalloc_start_ret = oe_start;
3244 } else if (*delalloc_end_ret + 1 == oe_start) {
3245 *delalloc_end_ret = oe_end;
3246 }
3247
3248 return true;
3249}
3250
3251/*
3252 * Check if there's delalloc in a given range.
3253 *
3254 * @inode: The inode.
3255 * @start: The start offset of the range. It does not need to be
3256 * sector size aligned.
3257 * @end: The end offset (inclusive value) of the search range.
3258 * It does not need to be sector size aligned.
3259 * @cached_state: Extent state record used for speeding up delalloc
3260 * searches in the inode's io_tree. Can be NULL.
3261 * @delalloc_start_ret: Output argument, set to the start offset of the
3262 * subrange found with delalloc (may not be sector size
3263 * aligned).
3264 * @delalloc_end_ret: Output argument, set to he end offset (inclusive value)
3265 * of the subrange found with delalloc.
3266 *
3267 * Returns true if a subrange with delalloc is found within the given range, and
3268 * if so it sets @delalloc_start_ret and @delalloc_end_ret with the start and
3269 * end offsets of the subrange.
3270 */
3271bool btrfs_find_delalloc_in_range(struct btrfs_inode *inode, u64 start, u64 end,
3272 struct extent_state **cached_state,
3273 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3274{
3275 u64 cur_offset = round_down(start, inode->root->fs_info->sectorsize);
3276 u64 prev_delalloc_end = 0;
3277 bool search_io_tree = true;
3278 bool ret = false;
3279
3280 while (cur_offset <= end) {
3281 u64 delalloc_start;
3282 u64 delalloc_end;
3283 bool delalloc;
3284
3285 delalloc = find_delalloc_subrange(inode, cur_offset, end,
3286 cached_state, &search_io_tree,
3287 &delalloc_start,
3288 &delalloc_end);
3289 if (!delalloc)
3290 break;
3291
3292 if (prev_delalloc_end == 0) {
3293 /* First subrange found. */
3294 *delalloc_start_ret = max(delalloc_start, start);
3295 *delalloc_end_ret = delalloc_end;
3296 ret = true;
3297 } else if (delalloc_start == prev_delalloc_end + 1) {
3298 /* Subrange adjacent to the previous one, merge them. */
3299 *delalloc_end_ret = delalloc_end;
3300 } else {
3301 /* Subrange not adjacent to the previous one, exit. */
3302 break;
3303 }
3304
3305 prev_delalloc_end = delalloc_end;
3306 cur_offset = delalloc_end + 1;
3307 cond_resched();
3308 }
3309
3310 return ret;
3311}
3312
3313/*
3314 * Check if there's a hole or delalloc range in a range representing a hole (or
3315 * prealloc extent) found in the inode's subvolume btree.
3316 *
3317 * @inode: The inode.
3318 * @whence: Seek mode (SEEK_DATA or SEEK_HOLE).
3319 * @start: Start offset of the hole region. It does not need to be sector
3320 * size aligned.
3321 * @end: End offset (inclusive value) of the hole region. It does not
3322 * need to be sector size aligned.
3323 * @start_ret: Return parameter, used to set the start of the subrange in the
3324 * hole that matches the search criteria (seek mode), if such
3325 * subrange is found (return value of the function is true).
3326 * The value returned here may not be sector size aligned.
3327 *
3328 * Returns true if a subrange matching the given seek mode is found, and if one
3329 * is found, it updates @start_ret with the start of the subrange.
3330 */
3331static bool find_desired_extent_in_hole(struct btrfs_inode *inode, int whence,
3332 struct extent_state **cached_state,
3333 u64 start, u64 end, u64 *start_ret)
3334{
3335 u64 delalloc_start;
3336 u64 delalloc_end;
3337 bool delalloc;
3338
3339 delalloc = btrfs_find_delalloc_in_range(inode, start, end, cached_state,
3340 &delalloc_start, &delalloc_end);
3341 if (delalloc && whence == SEEK_DATA) {
3342 *start_ret = delalloc_start;
3343 return true;
3344 }
3345
3346 if (delalloc && whence == SEEK_HOLE) {
3347 /*
3348 * We found delalloc but it starts after out start offset. So we
3349 * have a hole between our start offset and the delalloc start.
3350 */
3351 if (start < delalloc_start) {
3352 *start_ret = start;
3353 return true;
3354 }
3355 /*
3356 * Delalloc range starts at our start offset.
3357 * If the delalloc range's length is smaller than our range,
3358 * then it means we have a hole that starts where the delalloc
3359 * subrange ends.
3360 */
3361 if (delalloc_end < end) {
3362 *start_ret = delalloc_end + 1;
3363 return true;
3364 }
3365
3366 /* There's delalloc for the whole range. */
3367 return false;
3368 }
3369
3370 if (!delalloc && whence == SEEK_HOLE) {
3371 *start_ret = start;
3372 return true;
3373 }
3374
3375 /*
3376 * No delalloc in the range and we are seeking for data. The caller has
3377 * to iterate to the next extent item in the subvolume btree.
3378 */
3379 return false;
3380}
3381
3382static loff_t find_desired_extent(struct file *file, loff_t offset, int whence)
3383{
3384 struct btrfs_inode *inode = BTRFS_I(file->f_mapping->host);
3385 struct btrfs_file_private *private;
3386 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3387 struct extent_state *cached_state = NULL;
3388 struct extent_state **delalloc_cached_state;
3389 const loff_t i_size = i_size_read(&inode->vfs_inode);
3390 const u64 ino = btrfs_ino(inode);
3391 struct btrfs_root *root = inode->root;
3392 struct btrfs_path *path;
3393 struct btrfs_key key;
3394 u64 last_extent_end;
3395 u64 lockstart;
3396 u64 lockend;
3397 u64 start;
3398 int ret;
3399 bool found = false;
3400
3401 if (i_size == 0 || offset >= i_size)
3402 return -ENXIO;
3403
3404 /*
3405 * Quick path. If the inode has no prealloc extents and its number of
3406 * bytes used matches its i_size, then it can not have holes.
3407 */
3408 if (whence == SEEK_HOLE &&
3409 !(inode->flags & BTRFS_INODE_PREALLOC) &&
3410 inode_get_bytes(&inode->vfs_inode) == i_size)
3411 return i_size;
3412
3413 spin_lock(&inode->lock);
3414 private = file->private_data;
3415 spin_unlock(&inode->lock);
3416
3417 if (private && private->owner_task != current) {
3418 /*
3419 * Not allocated by us, don't use it as its cached state is used
3420 * by the task that allocated it and we don't want neither to
3421 * mess with it nor get incorrect results because it reflects an
3422 * invalid state for the current task.
3423 */
3424 private = NULL;
3425 } else if (!private) {
3426 private = kzalloc(sizeof(*private), GFP_KERNEL);
3427 /*
3428 * No worries if memory allocation failed.
3429 * The private structure is used only for speeding up multiple
3430 * lseek SEEK_HOLE/DATA calls to a file when there's delalloc,
3431 * so everything will still be correct.
3432 */
3433 if (private) {
3434 bool free = false;
3435
3436 private->owner_task = current;
3437
3438 spin_lock(&inode->lock);
3439 if (file->private_data)
3440 free = true;
3441 else
3442 file->private_data = private;
3443 spin_unlock(&inode->lock);
3444
3445 if (free) {
3446 kfree(private);
3447 private = NULL;
3448 }
3449 }
3450 }
3451
3452 if (private)
3453 delalloc_cached_state = &private->llseek_cached_state;
3454 else
3455 delalloc_cached_state = NULL;
3456
3457 /*
3458 * offset can be negative, in this case we start finding DATA/HOLE from
3459 * the very start of the file.
3460 */
3461 start = max_t(loff_t, 0, offset);
3462
3463 lockstart = round_down(start, fs_info->sectorsize);
3464 lockend = round_up(i_size, fs_info->sectorsize);
3465 if (lockend <= lockstart)
3466 lockend = lockstart + fs_info->sectorsize;
3467 lockend--;
3468
3469 path = btrfs_alloc_path();
3470 if (!path)
3471 return -ENOMEM;
3472 path->reada = READA_FORWARD;
3473
3474 key.objectid = ino;
3475 key.type = BTRFS_EXTENT_DATA_KEY;
3476 key.offset = start;
3477
3478 last_extent_end = lockstart;
3479
3480 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3481
3482 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3483 if (ret < 0) {
3484 goto out;
3485 } else if (ret > 0 && path->slots[0] > 0) {
3486 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
3487 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
3488 path->slots[0]--;
3489 }
3490
3491 while (start < i_size) {
3492 struct extent_buffer *leaf = path->nodes[0];
3493 struct btrfs_file_extent_item *extent;
3494 u64 extent_end;
3495 u8 type;
3496
3497 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3498 ret = btrfs_next_leaf(root, path);
3499 if (ret < 0)
3500 goto out;
3501 else if (ret > 0)
3502 break;
3503
3504 leaf = path->nodes[0];
3505 }
3506
3507 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3508 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3509 break;
3510
3511 extent_end = btrfs_file_extent_end(path);
3512
3513 /*
3514 * In the first iteration we may have a slot that points to an
3515 * extent that ends before our start offset, so skip it.
3516 */
3517 if (extent_end <= start) {
3518 path->slots[0]++;
3519 continue;
3520 }
3521
3522 /* We have an implicit hole, NO_HOLES feature is likely set. */
3523 if (last_extent_end < key.offset) {
3524 u64 search_start = last_extent_end;
3525 u64 found_start;
3526
3527 /*
3528 * First iteration, @start matches @offset and it's
3529 * within the hole.
3530 */
3531 if (start == offset)
3532 search_start = offset;
3533
3534 found = find_desired_extent_in_hole(inode, whence,
3535 delalloc_cached_state,
3536 search_start,
3537 key.offset - 1,
3538 &found_start);
3539 if (found) {
3540 start = found_start;
3541 break;
3542 }
3543 /*
3544 * Didn't find data or a hole (due to delalloc) in the
3545 * implicit hole range, so need to analyze the extent.
3546 */
3547 }
3548
3549 extent = btrfs_item_ptr(leaf, path->slots[0],
3550 struct btrfs_file_extent_item);
3551 type = btrfs_file_extent_type(leaf, extent);
3552
3553 /*
3554 * Can't access the extent's disk_bytenr field if this is an
3555 * inline extent, since at that offset, it's where the extent
3556 * data starts.
3557 */
3558 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
3559 (type == BTRFS_FILE_EXTENT_REG &&
3560 btrfs_file_extent_disk_bytenr(leaf, extent) == 0)) {
3561 /*
3562 * Explicit hole or prealloc extent, search for delalloc.
3563 * A prealloc extent is treated like a hole.
3564 */
3565 u64 search_start = key.offset;
3566 u64 found_start;
3567
3568 /*
3569 * First iteration, @start matches @offset and it's
3570 * within the hole.
3571 */
3572 if (start == offset)
3573 search_start = offset;
3574
3575 found = find_desired_extent_in_hole(inode, whence,
3576 delalloc_cached_state,
3577 search_start,
3578 extent_end - 1,
3579 &found_start);
3580 if (found) {
3581 start = found_start;
3582 break;
3583 }
3584 /*
3585 * Didn't find data or a hole (due to delalloc) in the
3586 * implicit hole range, so need to analyze the next
3587 * extent item.
3588 */
3589 } else {
3590 /*
3591 * Found a regular or inline extent.
3592 * If we are seeking for data, adjust the start offset
3593 * and stop, we're done.
3594 */
3595 if (whence == SEEK_DATA) {
3596 start = max_t(u64, key.offset, offset);
3597 found = true;
3598 break;
3599 }
3600 /*
3601 * Else, we are seeking for a hole, check the next file
3602 * extent item.
3603 */
3604 }
3605
3606 start = extent_end;
3607 last_extent_end = extent_end;
3608 path->slots[0]++;
3609 if (fatal_signal_pending(current)) {
3610 ret = -EINTR;
3611 goto out;
3612 }
3613 cond_resched();
3614 }
3615
3616 /* We have an implicit hole from the last extent found up to i_size. */
3617 if (!found && start < i_size) {
3618 found = find_desired_extent_in_hole(inode, whence,
3619 delalloc_cached_state, start,
3620 i_size - 1, &start);
3621 if (!found)
3622 start = i_size;
3623 }
3624
3625out:
3626 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3627 btrfs_free_path(path);
3628
3629 if (ret < 0)
3630 return ret;
3631
3632 if (whence == SEEK_DATA && start >= i_size)
3633 return -ENXIO;
3634
3635 return min_t(loff_t, start, i_size);
3636}
3637
3638static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3639{
3640 struct inode *inode = file->f_mapping->host;
3641
3642 switch (whence) {
3643 default:
3644 return generic_file_llseek(file, offset, whence);
3645 case SEEK_DATA:
3646 case SEEK_HOLE:
3647 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3648 offset = find_desired_extent(file, offset, whence);
3649 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3650 break;
3651 }
3652
3653 if (offset < 0)
3654 return offset;
3655
3656 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3657}
3658
3659static int btrfs_file_open(struct inode *inode, struct file *filp)
3660{
3661 int ret;
3662
3663 filp->f_mode |= FMODE_NOWAIT | FMODE_CAN_ODIRECT;
3664
3665 ret = fsverity_file_open(inode, filp);
3666 if (ret)
3667 return ret;
3668 return generic_file_open(inode, filp);
3669}
3670
3671static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3672{
3673 ssize_t ret = 0;
3674
3675 if (iocb->ki_flags & IOCB_DIRECT) {
3676 ret = btrfs_direct_read(iocb, to);
3677 if (ret < 0 || !iov_iter_count(to) ||
3678 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3679 return ret;
3680 }
3681
3682 return filemap_read(iocb, to, ret);
3683}
3684
3685const struct file_operations btrfs_file_operations = {
3686 .llseek = btrfs_file_llseek,
3687 .read_iter = btrfs_file_read_iter,
3688 .splice_read = filemap_splice_read,
3689 .write_iter = btrfs_file_write_iter,
3690 .splice_write = iter_file_splice_write,
3691 .mmap = btrfs_file_mmap,
3692 .open = btrfs_file_open,
3693 .release = btrfs_release_file,
3694 .get_unmapped_area = thp_get_unmapped_area,
3695 .fsync = btrfs_sync_file,
3696 .fallocate = btrfs_fallocate,
3697 .unlocked_ioctl = btrfs_ioctl,
3698#ifdef CONFIG_COMPAT
3699 .compat_ioctl = btrfs_compat_ioctl,
3700#endif
3701 .remap_file_range = btrfs_remap_file_range,
3702 .uring_cmd = btrfs_uring_cmd,
3703 .fop_flags = FOP_BUFFER_RASYNC | FOP_BUFFER_WASYNC,
3704};
3705
3706int btrfs_fdatawrite_range(struct btrfs_inode *inode, loff_t start, loff_t end)
3707{
3708 struct address_space *mapping = inode->vfs_inode.i_mapping;
3709 int ret;
3710
3711 /*
3712 * So with compression we will find and lock a dirty page and clear the
3713 * first one as dirty, setup an async extent, and immediately return
3714 * with the entire range locked but with nobody actually marked with
3715 * writeback. So we can't just filemap_write_and_wait_range() and
3716 * expect it to work since it will just kick off a thread to do the
3717 * actual work. So we need to call filemap_fdatawrite_range _again_
3718 * since it will wait on the page lock, which won't be unlocked until
3719 * after the pages have been marked as writeback and so we're good to go
3720 * from there. We have to do this otherwise we'll miss the ordered
3721 * extents and that results in badness. Please Josef, do not think you
3722 * know better and pull this out at some point in the future, it is
3723 * right and you are wrong.
3724 */
3725 ret = filemap_fdatawrite_range(mapping, start, end);
3726 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT, &inode->runtime_flags))
3727 ret = filemap_fdatawrite_range(mapping, start, end);
3728
3729 return ret;
3730}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6#include <linux/fs.h>
7#include <linux/pagemap.h>
8#include <linux/time.h>
9#include <linux/init.h>
10#include <linux/string.h>
11#include <linux/backing-dev.h>
12#include <linux/falloc.h>
13#include <linux/writeback.h>
14#include <linux/compat.h>
15#include <linux/slab.h>
16#include <linux/btrfs.h>
17#include <linux/uio.h>
18#include <linux/iversion.h>
19#include "ctree.h"
20#include "disk-io.h"
21#include "transaction.h"
22#include "btrfs_inode.h"
23#include "print-tree.h"
24#include "tree-log.h"
25#include "locking.h"
26#include "volumes.h"
27#include "qgroup.h"
28#include "compression.h"
29#include "delalloc-space.h"
30#include "reflink.h"
31
32static struct kmem_cache *btrfs_inode_defrag_cachep;
33/*
34 * when auto defrag is enabled we
35 * queue up these defrag structs to remember which
36 * inodes need defragging passes
37 */
38struct inode_defrag {
39 struct rb_node rb_node;
40 /* objectid */
41 u64 ino;
42 /*
43 * transid where the defrag was added, we search for
44 * extents newer than this
45 */
46 u64 transid;
47
48 /* root objectid */
49 u64 root;
50
51 /* last offset we were able to defrag */
52 u64 last_offset;
53
54 /* if we've wrapped around back to zero once already */
55 int cycled;
56};
57
58static int __compare_inode_defrag(struct inode_defrag *defrag1,
59 struct inode_defrag *defrag2)
60{
61 if (defrag1->root > defrag2->root)
62 return 1;
63 else if (defrag1->root < defrag2->root)
64 return -1;
65 else if (defrag1->ino > defrag2->ino)
66 return 1;
67 else if (defrag1->ino < defrag2->ino)
68 return -1;
69 else
70 return 0;
71}
72
73/* pop a record for an inode into the defrag tree. The lock
74 * must be held already
75 *
76 * If you're inserting a record for an older transid than an
77 * existing record, the transid already in the tree is lowered
78 *
79 * If an existing record is found the defrag item you
80 * pass in is freed
81 */
82static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
83 struct inode_defrag *defrag)
84{
85 struct btrfs_fs_info *fs_info = inode->root->fs_info;
86 struct inode_defrag *entry;
87 struct rb_node **p;
88 struct rb_node *parent = NULL;
89 int ret;
90
91 p = &fs_info->defrag_inodes.rb_node;
92 while (*p) {
93 parent = *p;
94 entry = rb_entry(parent, struct inode_defrag, rb_node);
95
96 ret = __compare_inode_defrag(defrag, entry);
97 if (ret < 0)
98 p = &parent->rb_left;
99 else if (ret > 0)
100 p = &parent->rb_right;
101 else {
102 /* if we're reinserting an entry for
103 * an old defrag run, make sure to
104 * lower the transid of our existing record
105 */
106 if (defrag->transid < entry->transid)
107 entry->transid = defrag->transid;
108 if (defrag->last_offset > entry->last_offset)
109 entry->last_offset = defrag->last_offset;
110 return -EEXIST;
111 }
112 }
113 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
114 rb_link_node(&defrag->rb_node, parent, p);
115 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
116 return 0;
117}
118
119static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
120{
121 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
122 return 0;
123
124 if (btrfs_fs_closing(fs_info))
125 return 0;
126
127 return 1;
128}
129
130/*
131 * insert a defrag record for this inode if auto defrag is
132 * enabled
133 */
134int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
135 struct btrfs_inode *inode)
136{
137 struct btrfs_root *root = inode->root;
138 struct btrfs_fs_info *fs_info = root->fs_info;
139 struct inode_defrag *defrag;
140 u64 transid;
141 int ret;
142
143 if (!__need_auto_defrag(fs_info))
144 return 0;
145
146 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
147 return 0;
148
149 if (trans)
150 transid = trans->transid;
151 else
152 transid = inode->root->last_trans;
153
154 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
155 if (!defrag)
156 return -ENOMEM;
157
158 defrag->ino = btrfs_ino(inode);
159 defrag->transid = transid;
160 defrag->root = root->root_key.objectid;
161
162 spin_lock(&fs_info->defrag_inodes_lock);
163 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
164 /*
165 * If we set IN_DEFRAG flag and evict the inode from memory,
166 * and then re-read this inode, this new inode doesn't have
167 * IN_DEFRAG flag. At the case, we may find the existed defrag.
168 */
169 ret = __btrfs_add_inode_defrag(inode, defrag);
170 if (ret)
171 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
172 } else {
173 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
174 }
175 spin_unlock(&fs_info->defrag_inodes_lock);
176 return 0;
177}
178
179/*
180 * Requeue the defrag object. If there is a defrag object that points to
181 * the same inode in the tree, we will merge them together (by
182 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
183 */
184static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
185 struct inode_defrag *defrag)
186{
187 struct btrfs_fs_info *fs_info = inode->root->fs_info;
188 int ret;
189
190 if (!__need_auto_defrag(fs_info))
191 goto out;
192
193 /*
194 * Here we don't check the IN_DEFRAG flag, because we need merge
195 * them together.
196 */
197 spin_lock(&fs_info->defrag_inodes_lock);
198 ret = __btrfs_add_inode_defrag(inode, defrag);
199 spin_unlock(&fs_info->defrag_inodes_lock);
200 if (ret)
201 goto out;
202 return;
203out:
204 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
205}
206
207/*
208 * pick the defragable inode that we want, if it doesn't exist, we will get
209 * the next one.
210 */
211static struct inode_defrag *
212btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
213{
214 struct inode_defrag *entry = NULL;
215 struct inode_defrag tmp;
216 struct rb_node *p;
217 struct rb_node *parent = NULL;
218 int ret;
219
220 tmp.ino = ino;
221 tmp.root = root;
222
223 spin_lock(&fs_info->defrag_inodes_lock);
224 p = fs_info->defrag_inodes.rb_node;
225 while (p) {
226 parent = p;
227 entry = rb_entry(parent, struct inode_defrag, rb_node);
228
229 ret = __compare_inode_defrag(&tmp, entry);
230 if (ret < 0)
231 p = parent->rb_left;
232 else if (ret > 0)
233 p = parent->rb_right;
234 else
235 goto out;
236 }
237
238 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
239 parent = rb_next(parent);
240 if (parent)
241 entry = rb_entry(parent, struct inode_defrag, rb_node);
242 else
243 entry = NULL;
244 }
245out:
246 if (entry)
247 rb_erase(parent, &fs_info->defrag_inodes);
248 spin_unlock(&fs_info->defrag_inodes_lock);
249 return entry;
250}
251
252void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
253{
254 struct inode_defrag *defrag;
255 struct rb_node *node;
256
257 spin_lock(&fs_info->defrag_inodes_lock);
258 node = rb_first(&fs_info->defrag_inodes);
259 while (node) {
260 rb_erase(node, &fs_info->defrag_inodes);
261 defrag = rb_entry(node, struct inode_defrag, rb_node);
262 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
263
264 cond_resched_lock(&fs_info->defrag_inodes_lock);
265
266 node = rb_first(&fs_info->defrag_inodes);
267 }
268 spin_unlock(&fs_info->defrag_inodes_lock);
269}
270
271#define BTRFS_DEFRAG_BATCH 1024
272
273static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
274 struct inode_defrag *defrag)
275{
276 struct btrfs_root *inode_root;
277 struct inode *inode;
278 struct btrfs_ioctl_defrag_range_args range;
279 int num_defrag;
280 int ret;
281
282 /* get the inode */
283 inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
284 if (IS_ERR(inode_root)) {
285 ret = PTR_ERR(inode_root);
286 goto cleanup;
287 }
288
289 inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
290 btrfs_put_root(inode_root);
291 if (IS_ERR(inode)) {
292 ret = PTR_ERR(inode);
293 goto cleanup;
294 }
295
296 /* do a chunk of defrag */
297 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
298 memset(&range, 0, sizeof(range));
299 range.len = (u64)-1;
300 range.start = defrag->last_offset;
301
302 sb_start_write(fs_info->sb);
303 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
304 BTRFS_DEFRAG_BATCH);
305 sb_end_write(fs_info->sb);
306 /*
307 * if we filled the whole defrag batch, there
308 * must be more work to do. Queue this defrag
309 * again
310 */
311 if (num_defrag == BTRFS_DEFRAG_BATCH) {
312 defrag->last_offset = range.start;
313 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
314 } else if (defrag->last_offset && !defrag->cycled) {
315 /*
316 * we didn't fill our defrag batch, but
317 * we didn't start at zero. Make sure we loop
318 * around to the start of the file.
319 */
320 defrag->last_offset = 0;
321 defrag->cycled = 1;
322 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
323 } else {
324 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
325 }
326
327 iput(inode);
328 return 0;
329cleanup:
330 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
331 return ret;
332}
333
334/*
335 * run through the list of inodes in the FS that need
336 * defragging
337 */
338int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
339{
340 struct inode_defrag *defrag;
341 u64 first_ino = 0;
342 u64 root_objectid = 0;
343
344 atomic_inc(&fs_info->defrag_running);
345 while (1) {
346 /* Pause the auto defragger. */
347 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
348 &fs_info->fs_state))
349 break;
350
351 if (!__need_auto_defrag(fs_info))
352 break;
353
354 /* find an inode to defrag */
355 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
356 first_ino);
357 if (!defrag) {
358 if (root_objectid || first_ino) {
359 root_objectid = 0;
360 first_ino = 0;
361 continue;
362 } else {
363 break;
364 }
365 }
366
367 first_ino = defrag->ino + 1;
368 root_objectid = defrag->root;
369
370 __btrfs_run_defrag_inode(fs_info, defrag);
371 }
372 atomic_dec(&fs_info->defrag_running);
373
374 /*
375 * during unmount, we use the transaction_wait queue to
376 * wait for the defragger to stop
377 */
378 wake_up(&fs_info->transaction_wait);
379 return 0;
380}
381
382/* simple helper to fault in pages and copy. This should go away
383 * and be replaced with calls into generic code.
384 */
385static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
386 struct page **prepared_pages,
387 struct iov_iter *i)
388{
389 size_t copied = 0;
390 size_t total_copied = 0;
391 int pg = 0;
392 int offset = offset_in_page(pos);
393
394 while (write_bytes > 0) {
395 size_t count = min_t(size_t,
396 PAGE_SIZE - offset, write_bytes);
397 struct page *page = prepared_pages[pg];
398 /*
399 * Copy data from userspace to the current page
400 */
401 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
402
403 /* Flush processor's dcache for this page */
404 flush_dcache_page(page);
405
406 /*
407 * if we get a partial write, we can end up with
408 * partially up to date pages. These add
409 * a lot of complexity, so make sure they don't
410 * happen by forcing this copy to be retried.
411 *
412 * The rest of the btrfs_file_write code will fall
413 * back to page at a time copies after we return 0.
414 */
415 if (!PageUptodate(page) && copied < count)
416 copied = 0;
417
418 iov_iter_advance(i, copied);
419 write_bytes -= copied;
420 total_copied += copied;
421
422 /* Return to btrfs_file_write_iter to fault page */
423 if (unlikely(copied == 0))
424 break;
425
426 if (copied < PAGE_SIZE - offset) {
427 offset += copied;
428 } else {
429 pg++;
430 offset = 0;
431 }
432 }
433 return total_copied;
434}
435
436/*
437 * unlocks pages after btrfs_file_write is done with them
438 */
439static void btrfs_drop_pages(struct page **pages, size_t num_pages)
440{
441 size_t i;
442 for (i = 0; i < num_pages; i++) {
443 /* page checked is some magic around finding pages that
444 * have been modified without going through btrfs_set_page_dirty
445 * clear it here. There should be no need to mark the pages
446 * accessed as prepare_pages should have marked them accessed
447 * in prepare_pages via find_or_create_page()
448 */
449 ClearPageChecked(pages[i]);
450 unlock_page(pages[i]);
451 put_page(pages[i]);
452 }
453}
454
455static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
456 const u64 start,
457 const u64 len,
458 struct extent_state **cached_state)
459{
460 u64 search_start = start;
461 const u64 end = start + len - 1;
462
463 while (search_start < end) {
464 const u64 search_len = end - search_start + 1;
465 struct extent_map *em;
466 u64 em_len;
467 int ret = 0;
468
469 em = btrfs_get_extent(inode, NULL, 0, search_start, search_len);
470 if (IS_ERR(em))
471 return PTR_ERR(em);
472
473 if (em->block_start != EXTENT_MAP_HOLE)
474 goto next;
475
476 em_len = em->len;
477 if (em->start < search_start)
478 em_len -= search_start - em->start;
479 if (em_len > search_len)
480 em_len = search_len;
481
482 ret = set_extent_bit(&inode->io_tree, search_start,
483 search_start + em_len - 1,
484 EXTENT_DELALLOC_NEW,
485 NULL, cached_state, GFP_NOFS);
486next:
487 search_start = extent_map_end(em);
488 free_extent_map(em);
489 if (ret)
490 return ret;
491 }
492 return 0;
493}
494
495/*
496 * after copy_from_user, pages need to be dirtied and we need to make
497 * sure holes are created between the current EOF and the start of
498 * any next extents (if required).
499 *
500 * this also makes the decision about creating an inline extent vs
501 * doing real data extents, marking pages dirty and delalloc as required.
502 */
503int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
504 size_t num_pages, loff_t pos, size_t write_bytes,
505 struct extent_state **cached)
506{
507 struct btrfs_fs_info *fs_info = inode->root->fs_info;
508 int err = 0;
509 int i;
510 u64 num_bytes;
511 u64 start_pos;
512 u64 end_of_last_block;
513 u64 end_pos = pos + write_bytes;
514 loff_t isize = i_size_read(&inode->vfs_inode);
515 unsigned int extra_bits = 0;
516
517 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
518 num_bytes = round_up(write_bytes + pos - start_pos,
519 fs_info->sectorsize);
520
521 end_of_last_block = start_pos + num_bytes - 1;
522
523 /*
524 * The pages may have already been dirty, clear out old accounting so
525 * we can set things up properly
526 */
527 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
528 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
529 0, 0, cached);
530
531 if (!btrfs_is_free_space_inode(inode)) {
532 if (start_pos >= isize &&
533 !(inode->flags & BTRFS_INODE_PREALLOC)) {
534 /*
535 * There can't be any extents following eof in this case
536 * so just set the delalloc new bit for the range
537 * directly.
538 */
539 extra_bits |= EXTENT_DELALLOC_NEW;
540 } else {
541 err = btrfs_find_new_delalloc_bytes(inode, start_pos,
542 num_bytes, cached);
543 if (err)
544 return err;
545 }
546 }
547
548 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
549 extra_bits, cached);
550 if (err)
551 return err;
552
553 for (i = 0; i < num_pages; i++) {
554 struct page *p = pages[i];
555 SetPageUptodate(p);
556 ClearPageChecked(p);
557 set_page_dirty(p);
558 }
559
560 /*
561 * we've only changed i_size in ram, and we haven't updated
562 * the disk i_size. There is no need to log the inode
563 * at this time.
564 */
565 if (end_pos > isize)
566 i_size_write(&inode->vfs_inode, end_pos);
567 return 0;
568}
569
570/*
571 * this drops all the extents in the cache that intersect the range
572 * [start, end]. Existing extents are split as required.
573 */
574void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
575 int skip_pinned)
576{
577 struct extent_map *em;
578 struct extent_map *split = NULL;
579 struct extent_map *split2 = NULL;
580 struct extent_map_tree *em_tree = &inode->extent_tree;
581 u64 len = end - start + 1;
582 u64 gen;
583 int ret;
584 int testend = 1;
585 unsigned long flags;
586 int compressed = 0;
587 bool modified;
588
589 WARN_ON(end < start);
590 if (end == (u64)-1) {
591 len = (u64)-1;
592 testend = 0;
593 }
594 while (1) {
595 int no_splits = 0;
596
597 modified = false;
598 if (!split)
599 split = alloc_extent_map();
600 if (!split2)
601 split2 = alloc_extent_map();
602 if (!split || !split2)
603 no_splits = 1;
604
605 write_lock(&em_tree->lock);
606 em = lookup_extent_mapping(em_tree, start, len);
607 if (!em) {
608 write_unlock(&em_tree->lock);
609 break;
610 }
611 flags = em->flags;
612 gen = em->generation;
613 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
614 if (testend && em->start + em->len >= start + len) {
615 free_extent_map(em);
616 write_unlock(&em_tree->lock);
617 break;
618 }
619 start = em->start + em->len;
620 if (testend)
621 len = start + len - (em->start + em->len);
622 free_extent_map(em);
623 write_unlock(&em_tree->lock);
624 continue;
625 }
626 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
627 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
628 clear_bit(EXTENT_FLAG_LOGGING, &flags);
629 modified = !list_empty(&em->list);
630 if (no_splits)
631 goto next;
632
633 if (em->start < start) {
634 split->start = em->start;
635 split->len = start - em->start;
636
637 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
638 split->orig_start = em->orig_start;
639 split->block_start = em->block_start;
640
641 if (compressed)
642 split->block_len = em->block_len;
643 else
644 split->block_len = split->len;
645 split->orig_block_len = max(split->block_len,
646 em->orig_block_len);
647 split->ram_bytes = em->ram_bytes;
648 } else {
649 split->orig_start = split->start;
650 split->block_len = 0;
651 split->block_start = em->block_start;
652 split->orig_block_len = 0;
653 split->ram_bytes = split->len;
654 }
655
656 split->generation = gen;
657 split->flags = flags;
658 split->compress_type = em->compress_type;
659 replace_extent_mapping(em_tree, em, split, modified);
660 free_extent_map(split);
661 split = split2;
662 split2 = NULL;
663 }
664 if (testend && em->start + em->len > start + len) {
665 u64 diff = start + len - em->start;
666
667 split->start = start + len;
668 split->len = em->start + em->len - (start + len);
669 split->flags = flags;
670 split->compress_type = em->compress_type;
671 split->generation = gen;
672
673 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
674 split->orig_block_len = max(em->block_len,
675 em->orig_block_len);
676
677 split->ram_bytes = em->ram_bytes;
678 if (compressed) {
679 split->block_len = em->block_len;
680 split->block_start = em->block_start;
681 split->orig_start = em->orig_start;
682 } else {
683 split->block_len = split->len;
684 split->block_start = em->block_start
685 + diff;
686 split->orig_start = em->orig_start;
687 }
688 } else {
689 split->ram_bytes = split->len;
690 split->orig_start = split->start;
691 split->block_len = 0;
692 split->block_start = em->block_start;
693 split->orig_block_len = 0;
694 }
695
696 if (extent_map_in_tree(em)) {
697 replace_extent_mapping(em_tree, em, split,
698 modified);
699 } else {
700 ret = add_extent_mapping(em_tree, split,
701 modified);
702 ASSERT(ret == 0); /* Logic error */
703 }
704 free_extent_map(split);
705 split = NULL;
706 }
707next:
708 if (extent_map_in_tree(em))
709 remove_extent_mapping(em_tree, em);
710 write_unlock(&em_tree->lock);
711
712 /* once for us */
713 free_extent_map(em);
714 /* once for the tree*/
715 free_extent_map(em);
716 }
717 if (split)
718 free_extent_map(split);
719 if (split2)
720 free_extent_map(split2);
721}
722
723/*
724 * this is very complex, but the basic idea is to drop all extents
725 * in the range start - end. hint_block is filled in with a block number
726 * that would be a good hint to the block allocator for this file.
727 *
728 * If an extent intersects the range but is not entirely inside the range
729 * it is either truncated or split. Anything entirely inside the range
730 * is deleted from the tree.
731 */
732int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
733 struct btrfs_root *root, struct btrfs_inode *inode,
734 struct btrfs_path *path, u64 start, u64 end,
735 u64 *drop_end, int drop_cache,
736 int replace_extent,
737 u32 extent_item_size,
738 int *key_inserted)
739{
740 struct btrfs_fs_info *fs_info = root->fs_info;
741 struct extent_buffer *leaf;
742 struct btrfs_file_extent_item *fi;
743 struct btrfs_ref ref = { 0 };
744 struct btrfs_key key;
745 struct btrfs_key new_key;
746 struct inode *vfs_inode = &inode->vfs_inode;
747 u64 ino = btrfs_ino(inode);
748 u64 search_start = start;
749 u64 disk_bytenr = 0;
750 u64 num_bytes = 0;
751 u64 extent_offset = 0;
752 u64 extent_end = 0;
753 u64 last_end = start;
754 int del_nr = 0;
755 int del_slot = 0;
756 int extent_type;
757 int recow;
758 int ret;
759 int modify_tree = -1;
760 int update_refs;
761 int found = 0;
762 int leafs_visited = 0;
763
764 if (drop_cache)
765 btrfs_drop_extent_cache(inode, start, end - 1, 0);
766
767 if (start >= inode->disk_i_size && !replace_extent)
768 modify_tree = 0;
769
770 update_refs = (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
771 root == fs_info->tree_root);
772 while (1) {
773 recow = 0;
774 ret = btrfs_lookup_file_extent(trans, root, path, ino,
775 search_start, modify_tree);
776 if (ret < 0)
777 break;
778 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
779 leaf = path->nodes[0];
780 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
781 if (key.objectid == ino &&
782 key.type == BTRFS_EXTENT_DATA_KEY)
783 path->slots[0]--;
784 }
785 ret = 0;
786 leafs_visited++;
787next_slot:
788 leaf = path->nodes[0];
789 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
790 BUG_ON(del_nr > 0);
791 ret = btrfs_next_leaf(root, path);
792 if (ret < 0)
793 break;
794 if (ret > 0) {
795 ret = 0;
796 break;
797 }
798 leafs_visited++;
799 leaf = path->nodes[0];
800 recow = 1;
801 }
802
803 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
804
805 if (key.objectid > ino)
806 break;
807 if (WARN_ON_ONCE(key.objectid < ino) ||
808 key.type < BTRFS_EXTENT_DATA_KEY) {
809 ASSERT(del_nr == 0);
810 path->slots[0]++;
811 goto next_slot;
812 }
813 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
814 break;
815
816 fi = btrfs_item_ptr(leaf, path->slots[0],
817 struct btrfs_file_extent_item);
818 extent_type = btrfs_file_extent_type(leaf, fi);
819
820 if (extent_type == BTRFS_FILE_EXTENT_REG ||
821 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
822 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
823 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
824 extent_offset = btrfs_file_extent_offset(leaf, fi);
825 extent_end = key.offset +
826 btrfs_file_extent_num_bytes(leaf, fi);
827 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
828 extent_end = key.offset +
829 btrfs_file_extent_ram_bytes(leaf, fi);
830 } else {
831 /* can't happen */
832 BUG();
833 }
834
835 /*
836 * Don't skip extent items representing 0 byte lengths. They
837 * used to be created (bug) if while punching holes we hit
838 * -ENOSPC condition. So if we find one here, just ensure we
839 * delete it, otherwise we would insert a new file extent item
840 * with the same key (offset) as that 0 bytes length file
841 * extent item in the call to setup_items_for_insert() later
842 * in this function.
843 */
844 if (extent_end == key.offset && extent_end >= search_start) {
845 last_end = extent_end;
846 goto delete_extent_item;
847 }
848
849 if (extent_end <= search_start) {
850 path->slots[0]++;
851 goto next_slot;
852 }
853
854 found = 1;
855 search_start = max(key.offset, start);
856 if (recow || !modify_tree) {
857 modify_tree = -1;
858 btrfs_release_path(path);
859 continue;
860 }
861
862 /*
863 * | - range to drop - |
864 * | -------- extent -------- |
865 */
866 if (start > key.offset && end < extent_end) {
867 BUG_ON(del_nr > 0);
868 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
869 ret = -EOPNOTSUPP;
870 break;
871 }
872
873 memcpy(&new_key, &key, sizeof(new_key));
874 new_key.offset = start;
875 ret = btrfs_duplicate_item(trans, root, path,
876 &new_key);
877 if (ret == -EAGAIN) {
878 btrfs_release_path(path);
879 continue;
880 }
881 if (ret < 0)
882 break;
883
884 leaf = path->nodes[0];
885 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
886 struct btrfs_file_extent_item);
887 btrfs_set_file_extent_num_bytes(leaf, fi,
888 start - key.offset);
889
890 fi = btrfs_item_ptr(leaf, path->slots[0],
891 struct btrfs_file_extent_item);
892
893 extent_offset += start - key.offset;
894 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
895 btrfs_set_file_extent_num_bytes(leaf, fi,
896 extent_end - start);
897 btrfs_mark_buffer_dirty(leaf);
898
899 if (update_refs && disk_bytenr > 0) {
900 btrfs_init_generic_ref(&ref,
901 BTRFS_ADD_DELAYED_REF,
902 disk_bytenr, num_bytes, 0);
903 btrfs_init_data_ref(&ref,
904 root->root_key.objectid,
905 new_key.objectid,
906 start - extent_offset);
907 ret = btrfs_inc_extent_ref(trans, &ref);
908 BUG_ON(ret); /* -ENOMEM */
909 }
910 key.offset = start;
911 }
912 /*
913 * From here on out we will have actually dropped something, so
914 * last_end can be updated.
915 */
916 last_end = extent_end;
917
918 /*
919 * | ---- range to drop ----- |
920 * | -------- extent -------- |
921 */
922 if (start <= key.offset && end < extent_end) {
923 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
924 ret = -EOPNOTSUPP;
925 break;
926 }
927
928 memcpy(&new_key, &key, sizeof(new_key));
929 new_key.offset = end;
930 btrfs_set_item_key_safe(fs_info, path, &new_key);
931
932 extent_offset += end - key.offset;
933 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
934 btrfs_set_file_extent_num_bytes(leaf, fi,
935 extent_end - end);
936 btrfs_mark_buffer_dirty(leaf);
937 if (update_refs && disk_bytenr > 0)
938 inode_sub_bytes(vfs_inode, end - key.offset);
939 break;
940 }
941
942 search_start = extent_end;
943 /*
944 * | ---- range to drop ----- |
945 * | -------- extent -------- |
946 */
947 if (start > key.offset && end >= extent_end) {
948 BUG_ON(del_nr > 0);
949 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
950 ret = -EOPNOTSUPP;
951 break;
952 }
953
954 btrfs_set_file_extent_num_bytes(leaf, fi,
955 start - key.offset);
956 btrfs_mark_buffer_dirty(leaf);
957 if (update_refs && disk_bytenr > 0)
958 inode_sub_bytes(vfs_inode, extent_end - start);
959 if (end == extent_end)
960 break;
961
962 path->slots[0]++;
963 goto next_slot;
964 }
965
966 /*
967 * | ---- range to drop ----- |
968 * | ------ extent ------ |
969 */
970 if (start <= key.offset && end >= extent_end) {
971delete_extent_item:
972 if (del_nr == 0) {
973 del_slot = path->slots[0];
974 del_nr = 1;
975 } else {
976 BUG_ON(del_slot + del_nr != path->slots[0]);
977 del_nr++;
978 }
979
980 if (update_refs &&
981 extent_type == BTRFS_FILE_EXTENT_INLINE) {
982 inode_sub_bytes(vfs_inode,
983 extent_end - key.offset);
984 extent_end = ALIGN(extent_end,
985 fs_info->sectorsize);
986 } else if (update_refs && disk_bytenr > 0) {
987 btrfs_init_generic_ref(&ref,
988 BTRFS_DROP_DELAYED_REF,
989 disk_bytenr, num_bytes, 0);
990 btrfs_init_data_ref(&ref,
991 root->root_key.objectid,
992 key.objectid,
993 key.offset - extent_offset);
994 ret = btrfs_free_extent(trans, &ref);
995 BUG_ON(ret); /* -ENOMEM */
996 inode_sub_bytes(vfs_inode,
997 extent_end - key.offset);
998 }
999
1000 if (end == extent_end)
1001 break;
1002
1003 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1004 path->slots[0]++;
1005 goto next_slot;
1006 }
1007
1008 ret = btrfs_del_items(trans, root, path, del_slot,
1009 del_nr);
1010 if (ret) {
1011 btrfs_abort_transaction(trans, ret);
1012 break;
1013 }
1014
1015 del_nr = 0;
1016 del_slot = 0;
1017
1018 btrfs_release_path(path);
1019 continue;
1020 }
1021
1022 BUG();
1023 }
1024
1025 if (!ret && del_nr > 0) {
1026 /*
1027 * Set path->slots[0] to first slot, so that after the delete
1028 * if items are move off from our leaf to its immediate left or
1029 * right neighbor leafs, we end up with a correct and adjusted
1030 * path->slots[0] for our insertion (if replace_extent != 0).
1031 */
1032 path->slots[0] = del_slot;
1033 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1034 if (ret)
1035 btrfs_abort_transaction(trans, ret);
1036 }
1037
1038 leaf = path->nodes[0];
1039 /*
1040 * If btrfs_del_items() was called, it might have deleted a leaf, in
1041 * which case it unlocked our path, so check path->locks[0] matches a
1042 * write lock.
1043 */
1044 if (!ret && replace_extent && leafs_visited == 1 &&
1045 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1046 path->locks[0] == BTRFS_WRITE_LOCK) &&
1047 btrfs_leaf_free_space(leaf) >=
1048 sizeof(struct btrfs_item) + extent_item_size) {
1049
1050 key.objectid = ino;
1051 key.type = BTRFS_EXTENT_DATA_KEY;
1052 key.offset = start;
1053 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1054 struct btrfs_key slot_key;
1055
1056 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1057 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1058 path->slots[0]++;
1059 }
1060 setup_items_for_insert(root, path, &key,
1061 &extent_item_size,
1062 extent_item_size,
1063 sizeof(struct btrfs_item) +
1064 extent_item_size, 1);
1065 *key_inserted = 1;
1066 }
1067
1068 if (!replace_extent || !(*key_inserted))
1069 btrfs_release_path(path);
1070 if (drop_end)
1071 *drop_end = found ? min(end, last_end) : end;
1072 return ret;
1073}
1074
1075int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1076 struct btrfs_root *root, struct inode *inode, u64 start,
1077 u64 end, int drop_cache)
1078{
1079 struct btrfs_path *path;
1080 int ret;
1081
1082 path = btrfs_alloc_path();
1083 if (!path)
1084 return -ENOMEM;
1085 ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path, start,
1086 end, NULL, drop_cache, 0, 0, NULL);
1087 btrfs_free_path(path);
1088 return ret;
1089}
1090
1091static int extent_mergeable(struct extent_buffer *leaf, int slot,
1092 u64 objectid, u64 bytenr, u64 orig_offset,
1093 u64 *start, u64 *end)
1094{
1095 struct btrfs_file_extent_item *fi;
1096 struct btrfs_key key;
1097 u64 extent_end;
1098
1099 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1100 return 0;
1101
1102 btrfs_item_key_to_cpu(leaf, &key, slot);
1103 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1104 return 0;
1105
1106 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1107 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1108 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1109 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1110 btrfs_file_extent_compression(leaf, fi) ||
1111 btrfs_file_extent_encryption(leaf, fi) ||
1112 btrfs_file_extent_other_encoding(leaf, fi))
1113 return 0;
1114
1115 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1116 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1117 return 0;
1118
1119 *start = key.offset;
1120 *end = extent_end;
1121 return 1;
1122}
1123
1124/*
1125 * Mark extent in the range start - end as written.
1126 *
1127 * This changes extent type from 'pre-allocated' to 'regular'. If only
1128 * part of extent is marked as written, the extent will be split into
1129 * two or three.
1130 */
1131int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1132 struct btrfs_inode *inode, u64 start, u64 end)
1133{
1134 struct btrfs_fs_info *fs_info = trans->fs_info;
1135 struct btrfs_root *root = inode->root;
1136 struct extent_buffer *leaf;
1137 struct btrfs_path *path;
1138 struct btrfs_file_extent_item *fi;
1139 struct btrfs_ref ref = { 0 };
1140 struct btrfs_key key;
1141 struct btrfs_key new_key;
1142 u64 bytenr;
1143 u64 num_bytes;
1144 u64 extent_end;
1145 u64 orig_offset;
1146 u64 other_start;
1147 u64 other_end;
1148 u64 split;
1149 int del_nr = 0;
1150 int del_slot = 0;
1151 int recow;
1152 int ret;
1153 u64 ino = btrfs_ino(inode);
1154
1155 path = btrfs_alloc_path();
1156 if (!path)
1157 return -ENOMEM;
1158again:
1159 recow = 0;
1160 split = start;
1161 key.objectid = ino;
1162 key.type = BTRFS_EXTENT_DATA_KEY;
1163 key.offset = split;
1164
1165 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1166 if (ret < 0)
1167 goto out;
1168 if (ret > 0 && path->slots[0] > 0)
1169 path->slots[0]--;
1170
1171 leaf = path->nodes[0];
1172 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1173 if (key.objectid != ino ||
1174 key.type != BTRFS_EXTENT_DATA_KEY) {
1175 ret = -EINVAL;
1176 btrfs_abort_transaction(trans, ret);
1177 goto out;
1178 }
1179 fi = btrfs_item_ptr(leaf, path->slots[0],
1180 struct btrfs_file_extent_item);
1181 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1182 ret = -EINVAL;
1183 btrfs_abort_transaction(trans, ret);
1184 goto out;
1185 }
1186 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1187 if (key.offset > start || extent_end < end) {
1188 ret = -EINVAL;
1189 btrfs_abort_transaction(trans, ret);
1190 goto out;
1191 }
1192
1193 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1194 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1195 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1196 memcpy(&new_key, &key, sizeof(new_key));
1197
1198 if (start == key.offset && end < extent_end) {
1199 other_start = 0;
1200 other_end = start;
1201 if (extent_mergeable(leaf, path->slots[0] - 1,
1202 ino, bytenr, orig_offset,
1203 &other_start, &other_end)) {
1204 new_key.offset = end;
1205 btrfs_set_item_key_safe(fs_info, path, &new_key);
1206 fi = btrfs_item_ptr(leaf, path->slots[0],
1207 struct btrfs_file_extent_item);
1208 btrfs_set_file_extent_generation(leaf, fi,
1209 trans->transid);
1210 btrfs_set_file_extent_num_bytes(leaf, fi,
1211 extent_end - end);
1212 btrfs_set_file_extent_offset(leaf, fi,
1213 end - orig_offset);
1214 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1215 struct btrfs_file_extent_item);
1216 btrfs_set_file_extent_generation(leaf, fi,
1217 trans->transid);
1218 btrfs_set_file_extent_num_bytes(leaf, fi,
1219 end - other_start);
1220 btrfs_mark_buffer_dirty(leaf);
1221 goto out;
1222 }
1223 }
1224
1225 if (start > key.offset && end == extent_end) {
1226 other_start = end;
1227 other_end = 0;
1228 if (extent_mergeable(leaf, path->slots[0] + 1,
1229 ino, bytenr, orig_offset,
1230 &other_start, &other_end)) {
1231 fi = btrfs_item_ptr(leaf, path->slots[0],
1232 struct btrfs_file_extent_item);
1233 btrfs_set_file_extent_num_bytes(leaf, fi,
1234 start - key.offset);
1235 btrfs_set_file_extent_generation(leaf, fi,
1236 trans->transid);
1237 path->slots[0]++;
1238 new_key.offset = start;
1239 btrfs_set_item_key_safe(fs_info, path, &new_key);
1240
1241 fi = btrfs_item_ptr(leaf, path->slots[0],
1242 struct btrfs_file_extent_item);
1243 btrfs_set_file_extent_generation(leaf, fi,
1244 trans->transid);
1245 btrfs_set_file_extent_num_bytes(leaf, fi,
1246 other_end - start);
1247 btrfs_set_file_extent_offset(leaf, fi,
1248 start - orig_offset);
1249 btrfs_mark_buffer_dirty(leaf);
1250 goto out;
1251 }
1252 }
1253
1254 while (start > key.offset || end < extent_end) {
1255 if (key.offset == start)
1256 split = end;
1257
1258 new_key.offset = split;
1259 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1260 if (ret == -EAGAIN) {
1261 btrfs_release_path(path);
1262 goto again;
1263 }
1264 if (ret < 0) {
1265 btrfs_abort_transaction(trans, ret);
1266 goto out;
1267 }
1268
1269 leaf = path->nodes[0];
1270 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1271 struct btrfs_file_extent_item);
1272 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1273 btrfs_set_file_extent_num_bytes(leaf, fi,
1274 split - key.offset);
1275
1276 fi = btrfs_item_ptr(leaf, path->slots[0],
1277 struct btrfs_file_extent_item);
1278
1279 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1280 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1281 btrfs_set_file_extent_num_bytes(leaf, fi,
1282 extent_end - split);
1283 btrfs_mark_buffer_dirty(leaf);
1284
1285 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1286 num_bytes, 0);
1287 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1288 orig_offset);
1289 ret = btrfs_inc_extent_ref(trans, &ref);
1290 if (ret) {
1291 btrfs_abort_transaction(trans, ret);
1292 goto out;
1293 }
1294
1295 if (split == start) {
1296 key.offset = start;
1297 } else {
1298 if (start != key.offset) {
1299 ret = -EINVAL;
1300 btrfs_abort_transaction(trans, ret);
1301 goto out;
1302 }
1303 path->slots[0]--;
1304 extent_end = end;
1305 }
1306 recow = 1;
1307 }
1308
1309 other_start = end;
1310 other_end = 0;
1311 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1312 num_bytes, 0);
1313 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1314 if (extent_mergeable(leaf, path->slots[0] + 1,
1315 ino, bytenr, orig_offset,
1316 &other_start, &other_end)) {
1317 if (recow) {
1318 btrfs_release_path(path);
1319 goto again;
1320 }
1321 extent_end = other_end;
1322 del_slot = path->slots[0] + 1;
1323 del_nr++;
1324 ret = btrfs_free_extent(trans, &ref);
1325 if (ret) {
1326 btrfs_abort_transaction(trans, ret);
1327 goto out;
1328 }
1329 }
1330 other_start = 0;
1331 other_end = start;
1332 if (extent_mergeable(leaf, path->slots[0] - 1,
1333 ino, bytenr, orig_offset,
1334 &other_start, &other_end)) {
1335 if (recow) {
1336 btrfs_release_path(path);
1337 goto again;
1338 }
1339 key.offset = other_start;
1340 del_slot = path->slots[0];
1341 del_nr++;
1342 ret = btrfs_free_extent(trans, &ref);
1343 if (ret) {
1344 btrfs_abort_transaction(trans, ret);
1345 goto out;
1346 }
1347 }
1348 if (del_nr == 0) {
1349 fi = btrfs_item_ptr(leaf, path->slots[0],
1350 struct btrfs_file_extent_item);
1351 btrfs_set_file_extent_type(leaf, fi,
1352 BTRFS_FILE_EXTENT_REG);
1353 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1354 btrfs_mark_buffer_dirty(leaf);
1355 } else {
1356 fi = btrfs_item_ptr(leaf, del_slot - 1,
1357 struct btrfs_file_extent_item);
1358 btrfs_set_file_extent_type(leaf, fi,
1359 BTRFS_FILE_EXTENT_REG);
1360 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1361 btrfs_set_file_extent_num_bytes(leaf, fi,
1362 extent_end - key.offset);
1363 btrfs_mark_buffer_dirty(leaf);
1364
1365 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1366 if (ret < 0) {
1367 btrfs_abort_transaction(trans, ret);
1368 goto out;
1369 }
1370 }
1371out:
1372 btrfs_free_path(path);
1373 return 0;
1374}
1375
1376/*
1377 * on error we return an unlocked page and the error value
1378 * on success we return a locked page and 0
1379 */
1380static int prepare_uptodate_page(struct inode *inode,
1381 struct page *page, u64 pos,
1382 bool force_uptodate)
1383{
1384 int ret = 0;
1385
1386 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1387 !PageUptodate(page)) {
1388 ret = btrfs_readpage(NULL, page);
1389 if (ret)
1390 return ret;
1391 lock_page(page);
1392 if (!PageUptodate(page)) {
1393 unlock_page(page);
1394 return -EIO;
1395 }
1396 if (page->mapping != inode->i_mapping) {
1397 unlock_page(page);
1398 return -EAGAIN;
1399 }
1400 }
1401 return 0;
1402}
1403
1404/*
1405 * this just gets pages into the page cache and locks them down.
1406 */
1407static noinline int prepare_pages(struct inode *inode, struct page **pages,
1408 size_t num_pages, loff_t pos,
1409 size_t write_bytes, bool force_uptodate)
1410{
1411 int i;
1412 unsigned long index = pos >> PAGE_SHIFT;
1413 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1414 int err = 0;
1415 int faili;
1416
1417 for (i = 0; i < num_pages; i++) {
1418again:
1419 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1420 mask | __GFP_WRITE);
1421 if (!pages[i]) {
1422 faili = i - 1;
1423 err = -ENOMEM;
1424 goto fail;
1425 }
1426
1427 if (i == 0)
1428 err = prepare_uptodate_page(inode, pages[i], pos,
1429 force_uptodate);
1430 if (!err && i == num_pages - 1)
1431 err = prepare_uptodate_page(inode, pages[i],
1432 pos + write_bytes, false);
1433 if (err) {
1434 put_page(pages[i]);
1435 if (err == -EAGAIN) {
1436 err = 0;
1437 goto again;
1438 }
1439 faili = i - 1;
1440 goto fail;
1441 }
1442 wait_on_page_writeback(pages[i]);
1443 }
1444
1445 return 0;
1446fail:
1447 while (faili >= 0) {
1448 unlock_page(pages[faili]);
1449 put_page(pages[faili]);
1450 faili--;
1451 }
1452 return err;
1453
1454}
1455
1456/*
1457 * This function locks the extent and properly waits for data=ordered extents
1458 * to finish before allowing the pages to be modified if need.
1459 *
1460 * The return value:
1461 * 1 - the extent is locked
1462 * 0 - the extent is not locked, and everything is OK
1463 * -EAGAIN - need re-prepare the pages
1464 * the other < 0 number - Something wrong happens
1465 */
1466static noinline int
1467lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1468 size_t num_pages, loff_t pos,
1469 size_t write_bytes,
1470 u64 *lockstart, u64 *lockend,
1471 struct extent_state **cached_state)
1472{
1473 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1474 u64 start_pos;
1475 u64 last_pos;
1476 int i;
1477 int ret = 0;
1478
1479 start_pos = round_down(pos, fs_info->sectorsize);
1480 last_pos = start_pos
1481 + round_up(pos + write_bytes - start_pos,
1482 fs_info->sectorsize) - 1;
1483
1484 if (start_pos < inode->vfs_inode.i_size) {
1485 struct btrfs_ordered_extent *ordered;
1486
1487 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1488 cached_state);
1489 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1490 last_pos - start_pos + 1);
1491 if (ordered &&
1492 ordered->file_offset + ordered->num_bytes > start_pos &&
1493 ordered->file_offset <= last_pos) {
1494 unlock_extent_cached(&inode->io_tree, start_pos,
1495 last_pos, cached_state);
1496 for (i = 0; i < num_pages; i++) {
1497 unlock_page(pages[i]);
1498 put_page(pages[i]);
1499 }
1500 btrfs_start_ordered_extent(&inode->vfs_inode,
1501 ordered, 1);
1502 btrfs_put_ordered_extent(ordered);
1503 return -EAGAIN;
1504 }
1505 if (ordered)
1506 btrfs_put_ordered_extent(ordered);
1507
1508 *lockstart = start_pos;
1509 *lockend = last_pos;
1510 ret = 1;
1511 }
1512
1513 /*
1514 * It's possible the pages are dirty right now, but we don't want
1515 * to clean them yet because copy_from_user may catch a page fault
1516 * and we might have to fall back to one page at a time. If that
1517 * happens, we'll unlock these pages and we'd have a window where
1518 * reclaim could sneak in and drop the once-dirty page on the floor
1519 * without writing it.
1520 *
1521 * We have the pages locked and the extent range locked, so there's
1522 * no way someone can start IO on any dirty pages in this range.
1523 *
1524 * We'll call btrfs_dirty_pages() later on, and that will flip around
1525 * delalloc bits and dirty the pages as required.
1526 */
1527 for (i = 0; i < num_pages; i++) {
1528 set_page_extent_mapped(pages[i]);
1529 WARN_ON(!PageLocked(pages[i]));
1530 }
1531
1532 return ret;
1533}
1534
1535static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1536 size_t *write_bytes, bool nowait)
1537{
1538 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1539 struct btrfs_root *root = inode->root;
1540 u64 lockstart, lockend;
1541 u64 num_bytes;
1542 int ret;
1543
1544 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1545 return 0;
1546
1547 if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1548 return -EAGAIN;
1549
1550 lockstart = round_down(pos, fs_info->sectorsize);
1551 lockend = round_up(pos + *write_bytes,
1552 fs_info->sectorsize) - 1;
1553 num_bytes = lockend - lockstart + 1;
1554
1555 if (nowait) {
1556 struct btrfs_ordered_extent *ordered;
1557
1558 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1559 return -EAGAIN;
1560
1561 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1562 num_bytes);
1563 if (ordered) {
1564 btrfs_put_ordered_extent(ordered);
1565 ret = -EAGAIN;
1566 goto out_unlock;
1567 }
1568 } else {
1569 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1570 lockend, NULL);
1571 }
1572
1573 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1574 NULL, NULL, NULL, false);
1575 if (ret <= 0) {
1576 ret = 0;
1577 if (!nowait)
1578 btrfs_drew_write_unlock(&root->snapshot_lock);
1579 } else {
1580 *write_bytes = min_t(size_t, *write_bytes ,
1581 num_bytes - pos + lockstart);
1582 }
1583out_unlock:
1584 unlock_extent(&inode->io_tree, lockstart, lockend);
1585
1586 return ret;
1587}
1588
1589static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1590 size_t *write_bytes)
1591{
1592 return check_can_nocow(inode, pos, write_bytes, true);
1593}
1594
1595/*
1596 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1597 *
1598 * @pos: File offset
1599 * @write_bytes: The length to write, will be updated to the nocow writeable
1600 * range
1601 *
1602 * This function will flush ordered extents in the range to ensure proper
1603 * nocow checks.
1604 *
1605 * Return:
1606 * >0 and update @write_bytes if we can do nocow write
1607 * 0 if we can't do nocow write
1608 * -EAGAIN if we can't get the needed lock or there are ordered extents
1609 * for * (nowait == true) case
1610 * <0 if other error happened
1611 *
1612 * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1613 */
1614int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1615 size_t *write_bytes)
1616{
1617 return check_can_nocow(inode, pos, write_bytes, false);
1618}
1619
1620void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1621{
1622 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1623}
1624
1625static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1626 struct iov_iter *i)
1627{
1628 struct file *file = iocb->ki_filp;
1629 loff_t pos = iocb->ki_pos;
1630 struct inode *inode = file_inode(file);
1631 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1632 struct page **pages = NULL;
1633 struct extent_changeset *data_reserved = NULL;
1634 u64 release_bytes = 0;
1635 u64 lockstart;
1636 u64 lockend;
1637 size_t num_written = 0;
1638 int nrptrs;
1639 int ret = 0;
1640 bool only_release_metadata = false;
1641 bool force_page_uptodate = false;
1642
1643 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1644 PAGE_SIZE / (sizeof(struct page *)));
1645 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1646 nrptrs = max(nrptrs, 8);
1647 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1648 if (!pages)
1649 return -ENOMEM;
1650
1651 while (iov_iter_count(i) > 0) {
1652 struct extent_state *cached_state = NULL;
1653 size_t offset = offset_in_page(pos);
1654 size_t sector_offset;
1655 size_t write_bytes = min(iov_iter_count(i),
1656 nrptrs * (size_t)PAGE_SIZE -
1657 offset);
1658 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1659 PAGE_SIZE);
1660 size_t reserve_bytes;
1661 size_t dirty_pages;
1662 size_t copied;
1663 size_t dirty_sectors;
1664 size_t num_sectors;
1665 int extents_locked;
1666
1667 WARN_ON(num_pages > nrptrs);
1668
1669 /*
1670 * Fault pages before locking them in prepare_pages
1671 * to avoid recursive lock
1672 */
1673 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1674 ret = -EFAULT;
1675 break;
1676 }
1677
1678 only_release_metadata = false;
1679 sector_offset = pos & (fs_info->sectorsize - 1);
1680 reserve_bytes = round_up(write_bytes + sector_offset,
1681 fs_info->sectorsize);
1682
1683 extent_changeset_release(data_reserved);
1684 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1685 &data_reserved, pos,
1686 write_bytes);
1687 if (ret < 0) {
1688 if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1689 &write_bytes) > 0) {
1690 /*
1691 * For nodata cow case, no need to reserve
1692 * data space.
1693 */
1694 only_release_metadata = true;
1695 /*
1696 * our prealloc extent may be smaller than
1697 * write_bytes, so scale down.
1698 */
1699 num_pages = DIV_ROUND_UP(write_bytes + offset,
1700 PAGE_SIZE);
1701 reserve_bytes = round_up(write_bytes +
1702 sector_offset,
1703 fs_info->sectorsize);
1704 } else {
1705 break;
1706 }
1707 }
1708
1709 WARN_ON(reserve_bytes == 0);
1710 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1711 reserve_bytes);
1712 if (ret) {
1713 if (!only_release_metadata)
1714 btrfs_free_reserved_data_space(BTRFS_I(inode),
1715 data_reserved, pos,
1716 write_bytes);
1717 else
1718 btrfs_check_nocow_unlock(BTRFS_I(inode));
1719 break;
1720 }
1721
1722 release_bytes = reserve_bytes;
1723again:
1724 /*
1725 * This is going to setup the pages array with the number of
1726 * pages we want, so we don't really need to worry about the
1727 * contents of pages from loop to loop
1728 */
1729 ret = prepare_pages(inode, pages, num_pages,
1730 pos, write_bytes,
1731 force_page_uptodate);
1732 if (ret) {
1733 btrfs_delalloc_release_extents(BTRFS_I(inode),
1734 reserve_bytes);
1735 break;
1736 }
1737
1738 extents_locked = lock_and_cleanup_extent_if_need(
1739 BTRFS_I(inode), pages,
1740 num_pages, pos, write_bytes, &lockstart,
1741 &lockend, &cached_state);
1742 if (extents_locked < 0) {
1743 if (extents_locked == -EAGAIN)
1744 goto again;
1745 btrfs_delalloc_release_extents(BTRFS_I(inode),
1746 reserve_bytes);
1747 ret = extents_locked;
1748 break;
1749 }
1750
1751 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1752
1753 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1754 dirty_sectors = round_up(copied + sector_offset,
1755 fs_info->sectorsize);
1756 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1757
1758 /*
1759 * if we have trouble faulting in the pages, fall
1760 * back to one page at a time
1761 */
1762 if (copied < write_bytes)
1763 nrptrs = 1;
1764
1765 if (copied == 0) {
1766 force_page_uptodate = true;
1767 dirty_sectors = 0;
1768 dirty_pages = 0;
1769 } else {
1770 force_page_uptodate = false;
1771 dirty_pages = DIV_ROUND_UP(copied + offset,
1772 PAGE_SIZE);
1773 }
1774
1775 if (num_sectors > dirty_sectors) {
1776 /* release everything except the sectors we dirtied */
1777 release_bytes -= dirty_sectors <<
1778 fs_info->sb->s_blocksize_bits;
1779 if (only_release_metadata) {
1780 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1781 release_bytes, true);
1782 } else {
1783 u64 __pos;
1784
1785 __pos = round_down(pos,
1786 fs_info->sectorsize) +
1787 (dirty_pages << PAGE_SHIFT);
1788 btrfs_delalloc_release_space(BTRFS_I(inode),
1789 data_reserved, __pos,
1790 release_bytes, true);
1791 }
1792 }
1793
1794 release_bytes = round_up(copied + sector_offset,
1795 fs_info->sectorsize);
1796
1797 if (copied > 0)
1798 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1799 dirty_pages, pos, copied,
1800 &cached_state);
1801
1802 /*
1803 * If we have not locked the extent range, because the range's
1804 * start offset is >= i_size, we might still have a non-NULL
1805 * cached extent state, acquired while marking the extent range
1806 * as delalloc through btrfs_dirty_pages(). Therefore free any
1807 * possible cached extent state to avoid a memory leak.
1808 */
1809 if (extents_locked)
1810 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1811 lockstart, lockend, &cached_state);
1812 else
1813 free_extent_state(cached_state);
1814
1815 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1816 if (ret) {
1817 btrfs_drop_pages(pages, num_pages);
1818 break;
1819 }
1820
1821 release_bytes = 0;
1822 if (only_release_metadata)
1823 btrfs_check_nocow_unlock(BTRFS_I(inode));
1824
1825 if (only_release_metadata && copied > 0) {
1826 lockstart = round_down(pos,
1827 fs_info->sectorsize);
1828 lockend = round_up(pos + copied,
1829 fs_info->sectorsize) - 1;
1830
1831 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1832 lockend, EXTENT_NORESERVE, NULL,
1833 NULL, GFP_NOFS);
1834 }
1835
1836 btrfs_drop_pages(pages, num_pages);
1837
1838 cond_resched();
1839
1840 balance_dirty_pages_ratelimited(inode->i_mapping);
1841
1842 pos += copied;
1843 num_written += copied;
1844 }
1845
1846 kfree(pages);
1847
1848 if (release_bytes) {
1849 if (only_release_metadata) {
1850 btrfs_check_nocow_unlock(BTRFS_I(inode));
1851 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1852 release_bytes, true);
1853 } else {
1854 btrfs_delalloc_release_space(BTRFS_I(inode),
1855 data_reserved,
1856 round_down(pos, fs_info->sectorsize),
1857 release_bytes, true);
1858 }
1859 }
1860
1861 extent_changeset_free(data_reserved);
1862 return num_written ? num_written : ret;
1863}
1864
1865static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1866{
1867 struct file *file = iocb->ki_filp;
1868 struct inode *inode = file_inode(file);
1869 loff_t pos;
1870 ssize_t written;
1871 ssize_t written_buffered;
1872 loff_t endbyte;
1873 int err;
1874
1875 written = generic_file_direct_write(iocb, from);
1876
1877 if (written < 0 || !iov_iter_count(from))
1878 return written;
1879
1880 pos = iocb->ki_pos;
1881 written_buffered = btrfs_buffered_write(iocb, from);
1882 if (written_buffered < 0) {
1883 err = written_buffered;
1884 goto out;
1885 }
1886 /*
1887 * Ensure all data is persisted. We want the next direct IO read to be
1888 * able to read what was just written.
1889 */
1890 endbyte = pos + written_buffered - 1;
1891 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1892 if (err)
1893 goto out;
1894 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1895 if (err)
1896 goto out;
1897 written += written_buffered;
1898 iocb->ki_pos = pos + written_buffered;
1899 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1900 endbyte >> PAGE_SHIFT);
1901out:
1902 return written ? written : err;
1903}
1904
1905static void update_time_for_write(struct inode *inode)
1906{
1907 struct timespec64 now;
1908
1909 if (IS_NOCMTIME(inode))
1910 return;
1911
1912 now = current_time(inode);
1913 if (!timespec64_equal(&inode->i_mtime, &now))
1914 inode->i_mtime = now;
1915
1916 if (!timespec64_equal(&inode->i_ctime, &now))
1917 inode->i_ctime = now;
1918
1919 if (IS_I_VERSION(inode))
1920 inode_inc_iversion(inode);
1921}
1922
1923static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1924 struct iov_iter *from)
1925{
1926 struct file *file = iocb->ki_filp;
1927 struct inode *inode = file_inode(file);
1928 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1929 struct btrfs_root *root = BTRFS_I(inode)->root;
1930 u64 start_pos;
1931 u64 end_pos;
1932 ssize_t num_written = 0;
1933 const bool sync = iocb->ki_flags & IOCB_DSYNC;
1934 ssize_t err;
1935 loff_t pos;
1936 size_t count;
1937 loff_t oldsize;
1938 int clean_page = 0;
1939
1940 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1941 (iocb->ki_flags & IOCB_NOWAIT))
1942 return -EOPNOTSUPP;
1943
1944 if (iocb->ki_flags & IOCB_NOWAIT) {
1945 if (!inode_trylock(inode))
1946 return -EAGAIN;
1947 } else {
1948 inode_lock(inode);
1949 }
1950
1951 err = generic_write_checks(iocb, from);
1952 if (err <= 0) {
1953 inode_unlock(inode);
1954 return err;
1955 }
1956
1957 pos = iocb->ki_pos;
1958 count = iov_iter_count(from);
1959 if (iocb->ki_flags & IOCB_NOWAIT) {
1960 size_t nocow_bytes = count;
1961
1962 /*
1963 * We will allocate space in case nodatacow is not set,
1964 * so bail
1965 */
1966 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes)
1967 <= 0) {
1968 inode_unlock(inode);
1969 return -EAGAIN;
1970 }
1971 /*
1972 * There are holes in the range or parts of the range that must
1973 * be COWed (shared extents, RO block groups, etc), so just bail
1974 * out.
1975 */
1976 if (nocow_bytes < count) {
1977 inode_unlock(inode);
1978 return -EAGAIN;
1979 }
1980 }
1981
1982 current->backing_dev_info = inode_to_bdi(inode);
1983 err = file_remove_privs(file);
1984 if (err) {
1985 inode_unlock(inode);
1986 goto out;
1987 }
1988
1989 /*
1990 * If BTRFS flips readonly due to some impossible error
1991 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1992 * although we have opened a file as writable, we have
1993 * to stop this write operation to ensure FS consistency.
1994 */
1995 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1996 inode_unlock(inode);
1997 err = -EROFS;
1998 goto out;
1999 }
2000
2001 /*
2002 * We reserve space for updating the inode when we reserve space for the
2003 * extent we are going to write, so we will enospc out there. We don't
2004 * need to start yet another transaction to update the inode as we will
2005 * update the inode when we finish writing whatever data we write.
2006 */
2007 update_time_for_write(inode);
2008
2009 start_pos = round_down(pos, fs_info->sectorsize);
2010 oldsize = i_size_read(inode);
2011 if (start_pos > oldsize) {
2012 /* Expand hole size to cover write data, preventing empty gap */
2013 end_pos = round_up(pos + count,
2014 fs_info->sectorsize);
2015 err = btrfs_cont_expand(inode, oldsize, end_pos);
2016 if (err) {
2017 inode_unlock(inode);
2018 goto out;
2019 }
2020 if (start_pos > round_up(oldsize, fs_info->sectorsize))
2021 clean_page = 1;
2022 }
2023
2024 if (sync)
2025 atomic_inc(&BTRFS_I(inode)->sync_writers);
2026
2027 if (iocb->ki_flags & IOCB_DIRECT) {
2028 num_written = __btrfs_direct_write(iocb, from);
2029 } else {
2030 num_written = btrfs_buffered_write(iocb, from);
2031 if (num_written > 0)
2032 iocb->ki_pos = pos + num_written;
2033 if (clean_page)
2034 pagecache_isize_extended(inode, oldsize,
2035 i_size_read(inode));
2036 }
2037
2038 inode_unlock(inode);
2039
2040 /*
2041 * We also have to set last_sub_trans to the current log transid,
2042 * otherwise subsequent syncs to a file that's been synced in this
2043 * transaction will appear to have already occurred.
2044 */
2045 spin_lock(&BTRFS_I(inode)->lock);
2046 BTRFS_I(inode)->last_sub_trans = root->log_transid;
2047 spin_unlock(&BTRFS_I(inode)->lock);
2048 if (num_written > 0)
2049 num_written = generic_write_sync(iocb, num_written);
2050
2051 if (sync)
2052 atomic_dec(&BTRFS_I(inode)->sync_writers);
2053out:
2054 current->backing_dev_info = NULL;
2055 return num_written ? num_written : err;
2056}
2057
2058int btrfs_release_file(struct inode *inode, struct file *filp)
2059{
2060 struct btrfs_file_private *private = filp->private_data;
2061
2062 if (private && private->filldir_buf)
2063 kfree(private->filldir_buf);
2064 kfree(private);
2065 filp->private_data = NULL;
2066
2067 /*
2068 * ordered_data_close is set by setattr when we are about to truncate
2069 * a file from a non-zero size to a zero size. This tries to
2070 * flush down new bytes that may have been written if the
2071 * application were using truncate to replace a file in place.
2072 */
2073 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2074 &BTRFS_I(inode)->runtime_flags))
2075 filemap_flush(inode->i_mapping);
2076 return 0;
2077}
2078
2079static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2080{
2081 int ret;
2082 struct blk_plug plug;
2083
2084 /*
2085 * This is only called in fsync, which would do synchronous writes, so
2086 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2087 * multiple disks using raid profile, a large IO can be split to
2088 * several segments of stripe length (currently 64K).
2089 */
2090 blk_start_plug(&plug);
2091 atomic_inc(&BTRFS_I(inode)->sync_writers);
2092 ret = btrfs_fdatawrite_range(inode, start, end);
2093 atomic_dec(&BTRFS_I(inode)->sync_writers);
2094 blk_finish_plug(&plug);
2095
2096 return ret;
2097}
2098
2099/*
2100 * fsync call for both files and directories. This logs the inode into
2101 * the tree log instead of forcing full commits whenever possible.
2102 *
2103 * It needs to call filemap_fdatawait so that all ordered extent updates are
2104 * in the metadata btree are up to date for copying to the log.
2105 *
2106 * It drops the inode mutex before doing the tree log commit. This is an
2107 * important optimization for directories because holding the mutex prevents
2108 * new operations on the dir while we write to disk.
2109 */
2110int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2111{
2112 struct dentry *dentry = file_dentry(file);
2113 struct inode *inode = d_inode(dentry);
2114 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2115 struct btrfs_root *root = BTRFS_I(inode)->root;
2116 struct btrfs_trans_handle *trans;
2117 struct btrfs_log_ctx ctx;
2118 int ret = 0, err;
2119
2120 trace_btrfs_sync_file(file, datasync);
2121
2122 btrfs_init_log_ctx(&ctx, inode);
2123
2124 /*
2125 * Set the range to full if the NO_HOLES feature is not enabled.
2126 * This is to avoid missing file extent items representing holes after
2127 * replaying the log.
2128 */
2129 if (!btrfs_fs_incompat(fs_info, NO_HOLES)) {
2130 start = 0;
2131 end = LLONG_MAX;
2132 }
2133
2134 /*
2135 * We write the dirty pages in the range and wait until they complete
2136 * out of the ->i_mutex. If so, we can flush the dirty pages by
2137 * multi-task, and make the performance up. See
2138 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2139 */
2140 ret = start_ordered_ops(inode, start, end);
2141 if (ret)
2142 goto out;
2143
2144 inode_lock(inode);
2145
2146 /*
2147 * We take the dio_sem here because the tree log stuff can race with
2148 * lockless dio writes and get an extent map logged for an extent we
2149 * never waited on. We need it this high up for lockdep reasons.
2150 */
2151 down_write(&BTRFS_I(inode)->dio_sem);
2152
2153 atomic_inc(&root->log_batch);
2154
2155 /*
2156 * If the inode needs a full sync, make sure we use a full range to
2157 * avoid log tree corruption, due to hole detection racing with ordered
2158 * extent completion for adjacent ranges and races between logging and
2159 * completion of ordered extents for adjancent ranges - both races
2160 * could lead to file extent items in the log with overlapping ranges.
2161 * Do this while holding the inode lock, to avoid races with other
2162 * tasks.
2163 */
2164 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2165 &BTRFS_I(inode)->runtime_flags)) {
2166 start = 0;
2167 end = LLONG_MAX;
2168 }
2169
2170 /*
2171 * Before we acquired the inode's lock, someone may have dirtied more
2172 * pages in the target range. We need to make sure that writeback for
2173 * any such pages does not start while we are logging the inode, because
2174 * if it does, any of the following might happen when we are not doing a
2175 * full inode sync:
2176 *
2177 * 1) We log an extent after its writeback finishes but before its
2178 * checksums are added to the csum tree, leading to -EIO errors
2179 * when attempting to read the extent after a log replay.
2180 *
2181 * 2) We can end up logging an extent before its writeback finishes.
2182 * Therefore after the log replay we will have a file extent item
2183 * pointing to an unwritten extent (and no data checksums as well).
2184 *
2185 * So trigger writeback for any eventual new dirty pages and then we
2186 * wait for all ordered extents to complete below.
2187 */
2188 ret = start_ordered_ops(inode, start, end);
2189 if (ret) {
2190 up_write(&BTRFS_I(inode)->dio_sem);
2191 inode_unlock(inode);
2192 goto out;
2193 }
2194
2195 /*
2196 * We have to do this here to avoid the priority inversion of waiting on
2197 * IO of a lower priority task while holding a transaction open.
2198 *
2199 * Also, the range length can be represented by u64, we have to do the
2200 * typecasts to avoid signed overflow if it's [0, LLONG_MAX].
2201 */
2202 ret = btrfs_wait_ordered_range(inode, start, (u64)end - (u64)start + 1);
2203 if (ret) {
2204 up_write(&BTRFS_I(inode)->dio_sem);
2205 inode_unlock(inode);
2206 goto out;
2207 }
2208 atomic_inc(&root->log_batch);
2209
2210 smp_mb();
2211 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2212 BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2213 /*
2214 * We've had everything committed since the last time we were
2215 * modified so clear this flag in case it was set for whatever
2216 * reason, it's no longer relevant.
2217 */
2218 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2219 &BTRFS_I(inode)->runtime_flags);
2220 /*
2221 * An ordered extent might have started before and completed
2222 * already with io errors, in which case the inode was not
2223 * updated and we end up here. So check the inode's mapping
2224 * for any errors that might have happened since we last
2225 * checked called fsync.
2226 */
2227 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2228 up_write(&BTRFS_I(inode)->dio_sem);
2229 inode_unlock(inode);
2230 goto out;
2231 }
2232
2233 /*
2234 * We use start here because we will need to wait on the IO to complete
2235 * in btrfs_sync_log, which could require joining a transaction (for
2236 * example checking cross references in the nocow path). If we use join
2237 * here we could get into a situation where we're waiting on IO to
2238 * happen that is blocked on a transaction trying to commit. With start
2239 * we inc the extwriter counter, so we wait for all extwriters to exit
2240 * before we start blocking joiners. This comment is to keep somebody
2241 * from thinking they are super smart and changing this to
2242 * btrfs_join_transaction *cough*Josef*cough*.
2243 */
2244 trans = btrfs_start_transaction(root, 0);
2245 if (IS_ERR(trans)) {
2246 ret = PTR_ERR(trans);
2247 up_write(&BTRFS_I(inode)->dio_sem);
2248 inode_unlock(inode);
2249 goto out;
2250 }
2251
2252 ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2253 if (ret < 0) {
2254 /* Fallthrough and commit/free transaction. */
2255 ret = 1;
2256 }
2257
2258 /* we've logged all the items and now have a consistent
2259 * version of the file in the log. It is possible that
2260 * someone will come in and modify the file, but that's
2261 * fine because the log is consistent on disk, and we
2262 * have references to all of the file's extents
2263 *
2264 * It is possible that someone will come in and log the
2265 * file again, but that will end up using the synchronization
2266 * inside btrfs_sync_log to keep things safe.
2267 */
2268 up_write(&BTRFS_I(inode)->dio_sem);
2269 inode_unlock(inode);
2270
2271 if (ret != BTRFS_NO_LOG_SYNC) {
2272 if (!ret) {
2273 ret = btrfs_sync_log(trans, root, &ctx);
2274 if (!ret) {
2275 ret = btrfs_end_transaction(trans);
2276 goto out;
2277 }
2278 }
2279 ret = btrfs_commit_transaction(trans);
2280 } else {
2281 ret = btrfs_end_transaction(trans);
2282 }
2283out:
2284 ASSERT(list_empty(&ctx.list));
2285 err = file_check_and_advance_wb_err(file);
2286 if (!ret)
2287 ret = err;
2288 return ret > 0 ? -EIO : ret;
2289}
2290
2291static const struct vm_operations_struct btrfs_file_vm_ops = {
2292 .fault = filemap_fault,
2293 .map_pages = filemap_map_pages,
2294 .page_mkwrite = btrfs_page_mkwrite,
2295};
2296
2297static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2298{
2299 struct address_space *mapping = filp->f_mapping;
2300
2301 if (!mapping->a_ops->readpage)
2302 return -ENOEXEC;
2303
2304 file_accessed(filp);
2305 vma->vm_ops = &btrfs_file_vm_ops;
2306
2307 return 0;
2308}
2309
2310static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2311 int slot, u64 start, u64 end)
2312{
2313 struct btrfs_file_extent_item *fi;
2314 struct btrfs_key key;
2315
2316 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2317 return 0;
2318
2319 btrfs_item_key_to_cpu(leaf, &key, slot);
2320 if (key.objectid != btrfs_ino(inode) ||
2321 key.type != BTRFS_EXTENT_DATA_KEY)
2322 return 0;
2323
2324 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2325
2326 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2327 return 0;
2328
2329 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2330 return 0;
2331
2332 if (key.offset == end)
2333 return 1;
2334 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2335 return 1;
2336 return 0;
2337}
2338
2339static int fill_holes(struct btrfs_trans_handle *trans,
2340 struct btrfs_inode *inode,
2341 struct btrfs_path *path, u64 offset, u64 end)
2342{
2343 struct btrfs_fs_info *fs_info = trans->fs_info;
2344 struct btrfs_root *root = inode->root;
2345 struct extent_buffer *leaf;
2346 struct btrfs_file_extent_item *fi;
2347 struct extent_map *hole_em;
2348 struct extent_map_tree *em_tree = &inode->extent_tree;
2349 struct btrfs_key key;
2350 int ret;
2351
2352 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2353 goto out;
2354
2355 key.objectid = btrfs_ino(inode);
2356 key.type = BTRFS_EXTENT_DATA_KEY;
2357 key.offset = offset;
2358
2359 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2360 if (ret <= 0) {
2361 /*
2362 * We should have dropped this offset, so if we find it then
2363 * something has gone horribly wrong.
2364 */
2365 if (ret == 0)
2366 ret = -EINVAL;
2367 return ret;
2368 }
2369
2370 leaf = path->nodes[0];
2371 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2372 u64 num_bytes;
2373
2374 path->slots[0]--;
2375 fi = btrfs_item_ptr(leaf, path->slots[0],
2376 struct btrfs_file_extent_item);
2377 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2378 end - offset;
2379 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2380 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2381 btrfs_set_file_extent_offset(leaf, fi, 0);
2382 btrfs_mark_buffer_dirty(leaf);
2383 goto out;
2384 }
2385
2386 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2387 u64 num_bytes;
2388
2389 key.offset = offset;
2390 btrfs_set_item_key_safe(fs_info, path, &key);
2391 fi = btrfs_item_ptr(leaf, path->slots[0],
2392 struct btrfs_file_extent_item);
2393 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2394 offset;
2395 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2396 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2397 btrfs_set_file_extent_offset(leaf, fi, 0);
2398 btrfs_mark_buffer_dirty(leaf);
2399 goto out;
2400 }
2401 btrfs_release_path(path);
2402
2403 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2404 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2405 if (ret)
2406 return ret;
2407
2408out:
2409 btrfs_release_path(path);
2410
2411 hole_em = alloc_extent_map();
2412 if (!hole_em) {
2413 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2414 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2415 } else {
2416 hole_em->start = offset;
2417 hole_em->len = end - offset;
2418 hole_em->ram_bytes = hole_em->len;
2419 hole_em->orig_start = offset;
2420
2421 hole_em->block_start = EXTENT_MAP_HOLE;
2422 hole_em->block_len = 0;
2423 hole_em->orig_block_len = 0;
2424 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2425 hole_em->generation = trans->transid;
2426
2427 do {
2428 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2429 write_lock(&em_tree->lock);
2430 ret = add_extent_mapping(em_tree, hole_em, 1);
2431 write_unlock(&em_tree->lock);
2432 } while (ret == -EEXIST);
2433 free_extent_map(hole_em);
2434 if (ret)
2435 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2436 &inode->runtime_flags);
2437 }
2438
2439 return 0;
2440}
2441
2442/*
2443 * Find a hole extent on given inode and change start/len to the end of hole
2444 * extent.(hole/vacuum extent whose em->start <= start &&
2445 * em->start + em->len > start)
2446 * When a hole extent is found, return 1 and modify start/len.
2447 */
2448static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2449{
2450 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2451 struct extent_map *em;
2452 int ret = 0;
2453
2454 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2455 round_down(*start, fs_info->sectorsize),
2456 round_up(*len, fs_info->sectorsize));
2457 if (IS_ERR(em))
2458 return PTR_ERR(em);
2459
2460 /* Hole or vacuum extent(only exists in no-hole mode) */
2461 if (em->block_start == EXTENT_MAP_HOLE) {
2462 ret = 1;
2463 *len = em->start + em->len > *start + *len ?
2464 0 : *start + *len - em->start - em->len;
2465 *start = em->start + em->len;
2466 }
2467 free_extent_map(em);
2468 return ret;
2469}
2470
2471static int btrfs_punch_hole_lock_range(struct inode *inode,
2472 const u64 lockstart,
2473 const u64 lockend,
2474 struct extent_state **cached_state)
2475{
2476 while (1) {
2477 struct btrfs_ordered_extent *ordered;
2478 int ret;
2479
2480 truncate_pagecache_range(inode, lockstart, lockend);
2481
2482 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2483 cached_state);
2484 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2485
2486 /*
2487 * We need to make sure we have no ordered extents in this range
2488 * and nobody raced in and read a page in this range, if we did
2489 * we need to try again.
2490 */
2491 if ((!ordered ||
2492 (ordered->file_offset + ordered->num_bytes <= lockstart ||
2493 ordered->file_offset > lockend)) &&
2494 !filemap_range_has_page(inode->i_mapping,
2495 lockstart, lockend)) {
2496 if (ordered)
2497 btrfs_put_ordered_extent(ordered);
2498 break;
2499 }
2500 if (ordered)
2501 btrfs_put_ordered_extent(ordered);
2502 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2503 lockend, cached_state);
2504 ret = btrfs_wait_ordered_range(inode, lockstart,
2505 lockend - lockstart + 1);
2506 if (ret)
2507 return ret;
2508 }
2509 return 0;
2510}
2511
2512static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2513 struct inode *inode,
2514 struct btrfs_path *path,
2515 struct btrfs_clone_extent_info *clone_info,
2516 const u64 clone_len)
2517{
2518 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2519 struct btrfs_root *root = BTRFS_I(inode)->root;
2520 struct btrfs_file_extent_item *extent;
2521 struct extent_buffer *leaf;
2522 struct btrfs_key key;
2523 int slot;
2524 struct btrfs_ref ref = { 0 };
2525 u64 ref_offset;
2526 int ret;
2527
2528 if (clone_len == 0)
2529 return 0;
2530
2531 if (clone_info->disk_offset == 0 &&
2532 btrfs_fs_incompat(fs_info, NO_HOLES))
2533 return 0;
2534
2535 key.objectid = btrfs_ino(BTRFS_I(inode));
2536 key.type = BTRFS_EXTENT_DATA_KEY;
2537 key.offset = clone_info->file_offset;
2538 ret = btrfs_insert_empty_item(trans, root, path, &key,
2539 clone_info->item_size);
2540 if (ret)
2541 return ret;
2542 leaf = path->nodes[0];
2543 slot = path->slots[0];
2544 write_extent_buffer(leaf, clone_info->extent_buf,
2545 btrfs_item_ptr_offset(leaf, slot),
2546 clone_info->item_size);
2547 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2548 btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2549 btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2550 btrfs_mark_buffer_dirty(leaf);
2551 btrfs_release_path(path);
2552
2553 ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
2554 clone_info->file_offset, clone_len);
2555 if (ret)
2556 return ret;
2557
2558 /* If it's a hole, nothing more needs to be done. */
2559 if (clone_info->disk_offset == 0)
2560 return 0;
2561
2562 inode_add_bytes(inode, clone_len);
2563 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2564 clone_info->disk_offset,
2565 clone_info->disk_len, 0);
2566 ref_offset = clone_info->file_offset - clone_info->data_offset;
2567 btrfs_init_data_ref(&ref, root->root_key.objectid,
2568 btrfs_ino(BTRFS_I(inode)), ref_offset);
2569 ret = btrfs_inc_extent_ref(trans, &ref);
2570
2571 return ret;
2572}
2573
2574/*
2575 * The respective range must have been previously locked, as well as the inode.
2576 * The end offset is inclusive (last byte of the range).
2577 * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2578 * cloning.
2579 * When cloning, we don't want to end up in a state where we dropped extents
2580 * without inserting a new one, so we must abort the transaction to avoid a
2581 * corruption.
2582 */
2583int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2584 const u64 start, const u64 end,
2585 struct btrfs_clone_extent_info *clone_info,
2586 struct btrfs_trans_handle **trans_out)
2587{
2588 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2589 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2590 u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2591 struct btrfs_root *root = BTRFS_I(inode)->root;
2592 struct btrfs_trans_handle *trans = NULL;
2593 struct btrfs_block_rsv *rsv;
2594 unsigned int rsv_count;
2595 u64 cur_offset;
2596 u64 drop_end;
2597 u64 len = end - start;
2598 int ret = 0;
2599
2600 if (end <= start)
2601 return -EINVAL;
2602
2603 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2604 if (!rsv) {
2605 ret = -ENOMEM;
2606 goto out;
2607 }
2608 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2609 rsv->failfast = 1;
2610
2611 /*
2612 * 1 - update the inode
2613 * 1 - removing the extents in the range
2614 * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2615 * an extent
2616 */
2617 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2618 rsv_count = 3;
2619 else
2620 rsv_count = 2;
2621
2622 trans = btrfs_start_transaction(root, rsv_count);
2623 if (IS_ERR(trans)) {
2624 ret = PTR_ERR(trans);
2625 trans = NULL;
2626 goto out_free;
2627 }
2628
2629 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2630 min_size, false);
2631 BUG_ON(ret);
2632 trans->block_rsv = rsv;
2633
2634 cur_offset = start;
2635 while (cur_offset < end) {
2636 ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path,
2637 cur_offset, end + 1, &drop_end,
2638 1, 0, 0, NULL);
2639 if (ret != -ENOSPC) {
2640 /*
2641 * When cloning we want to avoid transaction aborts when
2642 * nothing was done and we are attempting to clone parts
2643 * of inline extents, in such cases -EOPNOTSUPP is
2644 * returned by __btrfs_drop_extents() without having
2645 * changed anything in the file.
2646 */
2647 if (clone_info && ret && ret != -EOPNOTSUPP)
2648 btrfs_abort_transaction(trans, ret);
2649 break;
2650 }
2651
2652 trans->block_rsv = &fs_info->trans_block_rsv;
2653
2654 if (!clone_info && cur_offset < drop_end &&
2655 cur_offset < ino_size) {
2656 ret = fill_holes(trans, BTRFS_I(inode), path,
2657 cur_offset, drop_end);
2658 if (ret) {
2659 /*
2660 * If we failed then we didn't insert our hole
2661 * entries for the area we dropped, so now the
2662 * fs is corrupted, so we must abort the
2663 * transaction.
2664 */
2665 btrfs_abort_transaction(trans, ret);
2666 break;
2667 }
2668 } else if (!clone_info && cur_offset < drop_end) {
2669 /*
2670 * We are past the i_size here, but since we didn't
2671 * insert holes we need to clear the mapped area so we
2672 * know to not set disk_i_size in this area until a new
2673 * file extent is inserted here.
2674 */
2675 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2676 cur_offset, drop_end - cur_offset);
2677 if (ret) {
2678 /*
2679 * We couldn't clear our area, so we could
2680 * presumably adjust up and corrupt the fs, so
2681 * we need to abort.
2682 */
2683 btrfs_abort_transaction(trans, ret);
2684 break;
2685 }
2686 }
2687
2688 if (clone_info && drop_end > clone_info->file_offset) {
2689 u64 clone_len = drop_end - clone_info->file_offset;
2690
2691 ret = btrfs_insert_clone_extent(trans, inode, path,
2692 clone_info, clone_len);
2693 if (ret) {
2694 btrfs_abort_transaction(trans, ret);
2695 break;
2696 }
2697 clone_info->data_len -= clone_len;
2698 clone_info->data_offset += clone_len;
2699 clone_info->file_offset += clone_len;
2700 }
2701
2702 cur_offset = drop_end;
2703
2704 ret = btrfs_update_inode(trans, root, inode);
2705 if (ret)
2706 break;
2707
2708 btrfs_end_transaction(trans);
2709 btrfs_btree_balance_dirty(fs_info);
2710
2711 trans = btrfs_start_transaction(root, rsv_count);
2712 if (IS_ERR(trans)) {
2713 ret = PTR_ERR(trans);
2714 trans = NULL;
2715 break;
2716 }
2717
2718 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2719 rsv, min_size, false);
2720 BUG_ON(ret); /* shouldn't happen */
2721 trans->block_rsv = rsv;
2722
2723 if (!clone_info) {
2724 ret = find_first_non_hole(inode, &cur_offset, &len);
2725 if (unlikely(ret < 0))
2726 break;
2727 if (ret && !len) {
2728 ret = 0;
2729 break;
2730 }
2731 }
2732 }
2733
2734 /*
2735 * If we were cloning, force the next fsync to be a full one since we
2736 * we replaced (or just dropped in the case of cloning holes when
2737 * NO_HOLES is enabled) extents and extent maps.
2738 * This is for the sake of simplicity, and cloning into files larger
2739 * than 16Mb would force the full fsync any way (when
2740 * try_release_extent_mapping() is invoked during page cache truncation.
2741 */
2742 if (clone_info)
2743 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2744 &BTRFS_I(inode)->runtime_flags);
2745
2746 if (ret)
2747 goto out_trans;
2748
2749 trans->block_rsv = &fs_info->trans_block_rsv;
2750 /*
2751 * If we are using the NO_HOLES feature we might have had already an
2752 * hole that overlaps a part of the region [lockstart, lockend] and
2753 * ends at (or beyond) lockend. Since we have no file extent items to
2754 * represent holes, drop_end can be less than lockend and so we must
2755 * make sure we have an extent map representing the existing hole (the
2756 * call to __btrfs_drop_extents() might have dropped the existing extent
2757 * map representing the existing hole), otherwise the fast fsync path
2758 * will not record the existence of the hole region
2759 * [existing_hole_start, lockend].
2760 */
2761 if (drop_end <= end)
2762 drop_end = end + 1;
2763 /*
2764 * Don't insert file hole extent item if it's for a range beyond eof
2765 * (because it's useless) or if it represents a 0 bytes range (when
2766 * cur_offset == drop_end).
2767 */
2768 if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2769 ret = fill_holes(trans, BTRFS_I(inode), path,
2770 cur_offset, drop_end);
2771 if (ret) {
2772 /* Same comment as above. */
2773 btrfs_abort_transaction(trans, ret);
2774 goto out_trans;
2775 }
2776 } else if (!clone_info && cur_offset < drop_end) {
2777 /* See the comment in the loop above for the reasoning here. */
2778 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2779 cur_offset, drop_end - cur_offset);
2780 if (ret) {
2781 btrfs_abort_transaction(trans, ret);
2782 goto out_trans;
2783 }
2784
2785 }
2786 if (clone_info) {
2787 ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2788 clone_info->data_len);
2789 if (ret) {
2790 btrfs_abort_transaction(trans, ret);
2791 goto out_trans;
2792 }
2793 }
2794
2795out_trans:
2796 if (!trans)
2797 goto out_free;
2798
2799 trans->block_rsv = &fs_info->trans_block_rsv;
2800 if (ret)
2801 btrfs_end_transaction(trans);
2802 else
2803 *trans_out = trans;
2804out_free:
2805 btrfs_free_block_rsv(fs_info, rsv);
2806out:
2807 return ret;
2808}
2809
2810static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2811{
2812 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2813 struct btrfs_root *root = BTRFS_I(inode)->root;
2814 struct extent_state *cached_state = NULL;
2815 struct btrfs_path *path;
2816 struct btrfs_trans_handle *trans = NULL;
2817 u64 lockstart;
2818 u64 lockend;
2819 u64 tail_start;
2820 u64 tail_len;
2821 u64 orig_start = offset;
2822 int ret = 0;
2823 bool same_block;
2824 u64 ino_size;
2825 bool truncated_block = false;
2826 bool updated_inode = false;
2827
2828 ret = btrfs_wait_ordered_range(inode, offset, len);
2829 if (ret)
2830 return ret;
2831
2832 inode_lock(inode);
2833 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2834 ret = find_first_non_hole(inode, &offset, &len);
2835 if (ret < 0)
2836 goto out_only_mutex;
2837 if (ret && !len) {
2838 /* Already in a large hole */
2839 ret = 0;
2840 goto out_only_mutex;
2841 }
2842
2843 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2844 lockend = round_down(offset + len,
2845 btrfs_inode_sectorsize(inode)) - 1;
2846 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2847 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2848 /*
2849 * We needn't truncate any block which is beyond the end of the file
2850 * because we are sure there is no data there.
2851 */
2852 /*
2853 * Only do this if we are in the same block and we aren't doing the
2854 * entire block.
2855 */
2856 if (same_block && len < fs_info->sectorsize) {
2857 if (offset < ino_size) {
2858 truncated_block = true;
2859 ret = btrfs_truncate_block(inode, offset, len, 0);
2860 } else {
2861 ret = 0;
2862 }
2863 goto out_only_mutex;
2864 }
2865
2866 /* zero back part of the first block */
2867 if (offset < ino_size) {
2868 truncated_block = true;
2869 ret = btrfs_truncate_block(inode, offset, 0, 0);
2870 if (ret) {
2871 inode_unlock(inode);
2872 return ret;
2873 }
2874 }
2875
2876 /* Check the aligned pages after the first unaligned page,
2877 * if offset != orig_start, which means the first unaligned page
2878 * including several following pages are already in holes,
2879 * the extra check can be skipped */
2880 if (offset == orig_start) {
2881 /* after truncate page, check hole again */
2882 len = offset + len - lockstart;
2883 offset = lockstart;
2884 ret = find_first_non_hole(inode, &offset, &len);
2885 if (ret < 0)
2886 goto out_only_mutex;
2887 if (ret && !len) {
2888 ret = 0;
2889 goto out_only_mutex;
2890 }
2891 lockstart = offset;
2892 }
2893
2894 /* Check the tail unaligned part is in a hole */
2895 tail_start = lockend + 1;
2896 tail_len = offset + len - tail_start;
2897 if (tail_len) {
2898 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2899 if (unlikely(ret < 0))
2900 goto out_only_mutex;
2901 if (!ret) {
2902 /* zero the front end of the last page */
2903 if (tail_start + tail_len < ino_size) {
2904 truncated_block = true;
2905 ret = btrfs_truncate_block(inode,
2906 tail_start + tail_len,
2907 0, 1);
2908 if (ret)
2909 goto out_only_mutex;
2910 }
2911 }
2912 }
2913
2914 if (lockend < lockstart) {
2915 ret = 0;
2916 goto out_only_mutex;
2917 }
2918
2919 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2920 &cached_state);
2921 if (ret)
2922 goto out_only_mutex;
2923
2924 path = btrfs_alloc_path();
2925 if (!path) {
2926 ret = -ENOMEM;
2927 goto out;
2928 }
2929
2930 ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2931 &trans);
2932 btrfs_free_path(path);
2933 if (ret)
2934 goto out;
2935
2936 ASSERT(trans != NULL);
2937 inode_inc_iversion(inode);
2938 inode->i_mtime = inode->i_ctime = current_time(inode);
2939 ret = btrfs_update_inode(trans, root, inode);
2940 updated_inode = true;
2941 btrfs_end_transaction(trans);
2942 btrfs_btree_balance_dirty(fs_info);
2943out:
2944 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2945 &cached_state);
2946out_only_mutex:
2947 if (!updated_inode && truncated_block && !ret) {
2948 /*
2949 * If we only end up zeroing part of a page, we still need to
2950 * update the inode item, so that all the time fields are
2951 * updated as well as the necessary btrfs inode in memory fields
2952 * for detecting, at fsync time, if the inode isn't yet in the
2953 * log tree or it's there but not up to date.
2954 */
2955 struct timespec64 now = current_time(inode);
2956
2957 inode_inc_iversion(inode);
2958 inode->i_mtime = now;
2959 inode->i_ctime = now;
2960 trans = btrfs_start_transaction(root, 1);
2961 if (IS_ERR(trans)) {
2962 ret = PTR_ERR(trans);
2963 } else {
2964 int ret2;
2965
2966 ret = btrfs_update_inode(trans, root, inode);
2967 ret2 = btrfs_end_transaction(trans);
2968 if (!ret)
2969 ret = ret2;
2970 }
2971 }
2972 inode_unlock(inode);
2973 return ret;
2974}
2975
2976/* Helper structure to record which range is already reserved */
2977struct falloc_range {
2978 struct list_head list;
2979 u64 start;
2980 u64 len;
2981};
2982
2983/*
2984 * Helper function to add falloc range
2985 *
2986 * Caller should have locked the larger range of extent containing
2987 * [start, len)
2988 */
2989static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2990{
2991 struct falloc_range *prev = NULL;
2992 struct falloc_range *range = NULL;
2993
2994 if (list_empty(head))
2995 goto insert;
2996
2997 /*
2998 * As fallocate iterate by bytenr order, we only need to check
2999 * the last range.
3000 */
3001 prev = list_entry(head->prev, struct falloc_range, list);
3002 if (prev->start + prev->len == start) {
3003 prev->len += len;
3004 return 0;
3005 }
3006insert:
3007 range = kmalloc(sizeof(*range), GFP_KERNEL);
3008 if (!range)
3009 return -ENOMEM;
3010 range->start = start;
3011 range->len = len;
3012 list_add_tail(&range->list, head);
3013 return 0;
3014}
3015
3016static int btrfs_fallocate_update_isize(struct inode *inode,
3017 const u64 end,
3018 const int mode)
3019{
3020 struct btrfs_trans_handle *trans;
3021 struct btrfs_root *root = BTRFS_I(inode)->root;
3022 int ret;
3023 int ret2;
3024
3025 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3026 return 0;
3027
3028 trans = btrfs_start_transaction(root, 1);
3029 if (IS_ERR(trans))
3030 return PTR_ERR(trans);
3031
3032 inode->i_ctime = current_time(inode);
3033 i_size_write(inode, end);
3034 btrfs_inode_safe_disk_i_size_write(inode, 0);
3035 ret = btrfs_update_inode(trans, root, inode);
3036 ret2 = btrfs_end_transaction(trans);
3037
3038 return ret ? ret : ret2;
3039}
3040
3041enum {
3042 RANGE_BOUNDARY_WRITTEN_EXTENT,
3043 RANGE_BOUNDARY_PREALLOC_EXTENT,
3044 RANGE_BOUNDARY_HOLE,
3045};
3046
3047static int btrfs_zero_range_check_range_boundary(struct inode *inode,
3048 u64 offset)
3049{
3050 const u64 sectorsize = btrfs_inode_sectorsize(inode);
3051 struct extent_map *em;
3052 int ret;
3053
3054 offset = round_down(offset, sectorsize);
3055 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize);
3056 if (IS_ERR(em))
3057 return PTR_ERR(em);
3058
3059 if (em->block_start == EXTENT_MAP_HOLE)
3060 ret = RANGE_BOUNDARY_HOLE;
3061 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3062 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3063 else
3064 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3065
3066 free_extent_map(em);
3067 return ret;
3068}
3069
3070static int btrfs_zero_range(struct inode *inode,
3071 loff_t offset,
3072 loff_t len,
3073 const int mode)
3074{
3075 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3076 struct extent_map *em;
3077 struct extent_changeset *data_reserved = NULL;
3078 int ret;
3079 u64 alloc_hint = 0;
3080 const u64 sectorsize = btrfs_inode_sectorsize(inode);
3081 u64 alloc_start = round_down(offset, sectorsize);
3082 u64 alloc_end = round_up(offset + len, sectorsize);
3083 u64 bytes_to_reserve = 0;
3084 bool space_reserved = false;
3085
3086 inode_dio_wait(inode);
3087
3088 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3089 alloc_end - alloc_start);
3090 if (IS_ERR(em)) {
3091 ret = PTR_ERR(em);
3092 goto out;
3093 }
3094
3095 /*
3096 * Avoid hole punching and extent allocation for some cases. More cases
3097 * could be considered, but these are unlikely common and we keep things
3098 * as simple as possible for now. Also, intentionally, if the target
3099 * range contains one or more prealloc extents together with regular
3100 * extents and holes, we drop all the existing extents and allocate a
3101 * new prealloc extent, so that we get a larger contiguous disk extent.
3102 */
3103 if (em->start <= alloc_start &&
3104 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3105 const u64 em_end = em->start + em->len;
3106
3107 if (em_end >= offset + len) {
3108 /*
3109 * The whole range is already a prealloc extent,
3110 * do nothing except updating the inode's i_size if
3111 * needed.
3112 */
3113 free_extent_map(em);
3114 ret = btrfs_fallocate_update_isize(inode, offset + len,
3115 mode);
3116 goto out;
3117 }
3118 /*
3119 * Part of the range is already a prealloc extent, so operate
3120 * only on the remaining part of the range.
3121 */
3122 alloc_start = em_end;
3123 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3124 len = offset + len - alloc_start;
3125 offset = alloc_start;
3126 alloc_hint = em->block_start + em->len;
3127 }
3128 free_extent_map(em);
3129
3130 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3131 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3132 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3133 sectorsize);
3134 if (IS_ERR(em)) {
3135 ret = PTR_ERR(em);
3136 goto out;
3137 }
3138
3139 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3140 free_extent_map(em);
3141 ret = btrfs_fallocate_update_isize(inode, offset + len,
3142 mode);
3143 goto out;
3144 }
3145 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3146 free_extent_map(em);
3147 ret = btrfs_truncate_block(inode, offset, len, 0);
3148 if (!ret)
3149 ret = btrfs_fallocate_update_isize(inode,
3150 offset + len,
3151 mode);
3152 return ret;
3153 }
3154 free_extent_map(em);
3155 alloc_start = round_down(offset, sectorsize);
3156 alloc_end = alloc_start + sectorsize;
3157 goto reserve_space;
3158 }
3159
3160 alloc_start = round_up(offset, sectorsize);
3161 alloc_end = round_down(offset + len, sectorsize);
3162
3163 /*
3164 * For unaligned ranges, check the pages at the boundaries, they might
3165 * map to an extent, in which case we need to partially zero them, or
3166 * they might map to a hole, in which case we need our allocation range
3167 * to cover them.
3168 */
3169 if (!IS_ALIGNED(offset, sectorsize)) {
3170 ret = btrfs_zero_range_check_range_boundary(inode, offset);
3171 if (ret < 0)
3172 goto out;
3173 if (ret == RANGE_BOUNDARY_HOLE) {
3174 alloc_start = round_down(offset, sectorsize);
3175 ret = 0;
3176 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3177 ret = btrfs_truncate_block(inode, offset, 0, 0);
3178 if (ret)
3179 goto out;
3180 } else {
3181 ret = 0;
3182 }
3183 }
3184
3185 if (!IS_ALIGNED(offset + len, sectorsize)) {
3186 ret = btrfs_zero_range_check_range_boundary(inode,
3187 offset + len);
3188 if (ret < 0)
3189 goto out;
3190 if (ret == RANGE_BOUNDARY_HOLE) {
3191 alloc_end = round_up(offset + len, sectorsize);
3192 ret = 0;
3193 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3194 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3195 if (ret)
3196 goto out;
3197 } else {
3198 ret = 0;
3199 }
3200 }
3201
3202reserve_space:
3203 if (alloc_start < alloc_end) {
3204 struct extent_state *cached_state = NULL;
3205 const u64 lockstart = alloc_start;
3206 const u64 lockend = alloc_end - 1;
3207
3208 bytes_to_reserve = alloc_end - alloc_start;
3209 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3210 bytes_to_reserve);
3211 if (ret < 0)
3212 goto out;
3213 space_reserved = true;
3214 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3215 &cached_state);
3216 if (ret)
3217 goto out;
3218 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3219 alloc_start, bytes_to_reserve);
3220 if (ret)
3221 goto out;
3222 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3223 alloc_end - alloc_start,
3224 i_blocksize(inode),
3225 offset + len, &alloc_hint);
3226 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3227 lockend, &cached_state);
3228 /* btrfs_prealloc_file_range releases reserved space on error */
3229 if (ret) {
3230 space_reserved = false;
3231 goto out;
3232 }
3233 }
3234 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3235 out:
3236 if (ret && space_reserved)
3237 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3238 alloc_start, bytes_to_reserve);
3239 extent_changeset_free(data_reserved);
3240
3241 return ret;
3242}
3243
3244static long btrfs_fallocate(struct file *file, int mode,
3245 loff_t offset, loff_t len)
3246{
3247 struct inode *inode = file_inode(file);
3248 struct extent_state *cached_state = NULL;
3249 struct extent_changeset *data_reserved = NULL;
3250 struct falloc_range *range;
3251 struct falloc_range *tmp;
3252 struct list_head reserve_list;
3253 u64 cur_offset;
3254 u64 last_byte;
3255 u64 alloc_start;
3256 u64 alloc_end;
3257 u64 alloc_hint = 0;
3258 u64 locked_end;
3259 u64 actual_end = 0;
3260 struct extent_map *em;
3261 int blocksize = btrfs_inode_sectorsize(inode);
3262 int ret;
3263
3264 alloc_start = round_down(offset, blocksize);
3265 alloc_end = round_up(offset + len, blocksize);
3266 cur_offset = alloc_start;
3267
3268 /* Make sure we aren't being give some crap mode */
3269 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3270 FALLOC_FL_ZERO_RANGE))
3271 return -EOPNOTSUPP;
3272
3273 if (mode & FALLOC_FL_PUNCH_HOLE)
3274 return btrfs_punch_hole(inode, offset, len);
3275
3276 /*
3277 * Only trigger disk allocation, don't trigger qgroup reserve
3278 *
3279 * For qgroup space, it will be checked later.
3280 */
3281 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3282 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3283 alloc_end - alloc_start);
3284 if (ret < 0)
3285 return ret;
3286 }
3287
3288 inode_lock(inode);
3289
3290 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3291 ret = inode_newsize_ok(inode, offset + len);
3292 if (ret)
3293 goto out;
3294 }
3295
3296 /*
3297 * TODO: Move these two operations after we have checked
3298 * accurate reserved space, or fallocate can still fail but
3299 * with page truncated or size expanded.
3300 *
3301 * But that's a minor problem and won't do much harm BTW.
3302 */
3303 if (alloc_start > inode->i_size) {
3304 ret = btrfs_cont_expand(inode, i_size_read(inode),
3305 alloc_start);
3306 if (ret)
3307 goto out;
3308 } else if (offset + len > inode->i_size) {
3309 /*
3310 * If we are fallocating from the end of the file onward we
3311 * need to zero out the end of the block if i_size lands in the
3312 * middle of a block.
3313 */
3314 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3315 if (ret)
3316 goto out;
3317 }
3318
3319 /*
3320 * wait for ordered IO before we have any locks. We'll loop again
3321 * below with the locks held.
3322 */
3323 ret = btrfs_wait_ordered_range(inode, alloc_start,
3324 alloc_end - alloc_start);
3325 if (ret)
3326 goto out;
3327
3328 if (mode & FALLOC_FL_ZERO_RANGE) {
3329 ret = btrfs_zero_range(inode, offset, len, mode);
3330 inode_unlock(inode);
3331 return ret;
3332 }
3333
3334 locked_end = alloc_end - 1;
3335 while (1) {
3336 struct btrfs_ordered_extent *ordered;
3337
3338 /* the extent lock is ordered inside the running
3339 * transaction
3340 */
3341 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3342 locked_end, &cached_state);
3343 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3344
3345 if (ordered &&
3346 ordered->file_offset + ordered->num_bytes > alloc_start &&
3347 ordered->file_offset < alloc_end) {
3348 btrfs_put_ordered_extent(ordered);
3349 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3350 alloc_start, locked_end,
3351 &cached_state);
3352 /*
3353 * we can't wait on the range with the transaction
3354 * running or with the extent lock held
3355 */
3356 ret = btrfs_wait_ordered_range(inode, alloc_start,
3357 alloc_end - alloc_start);
3358 if (ret)
3359 goto out;
3360 } else {
3361 if (ordered)
3362 btrfs_put_ordered_extent(ordered);
3363 break;
3364 }
3365 }
3366
3367 /* First, check if we exceed the qgroup limit */
3368 INIT_LIST_HEAD(&reserve_list);
3369 while (cur_offset < alloc_end) {
3370 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3371 alloc_end - cur_offset);
3372 if (IS_ERR(em)) {
3373 ret = PTR_ERR(em);
3374 break;
3375 }
3376 last_byte = min(extent_map_end(em), alloc_end);
3377 actual_end = min_t(u64, extent_map_end(em), offset + len);
3378 last_byte = ALIGN(last_byte, blocksize);
3379 if (em->block_start == EXTENT_MAP_HOLE ||
3380 (cur_offset >= inode->i_size &&
3381 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3382 ret = add_falloc_range(&reserve_list, cur_offset,
3383 last_byte - cur_offset);
3384 if (ret < 0) {
3385 free_extent_map(em);
3386 break;
3387 }
3388 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3389 &data_reserved, cur_offset,
3390 last_byte - cur_offset);
3391 if (ret < 0) {
3392 cur_offset = last_byte;
3393 free_extent_map(em);
3394 break;
3395 }
3396 } else {
3397 /*
3398 * Do not need to reserve unwritten extent for this
3399 * range, free reserved data space first, otherwise
3400 * it'll result in false ENOSPC error.
3401 */
3402 btrfs_free_reserved_data_space(BTRFS_I(inode),
3403 data_reserved, cur_offset,
3404 last_byte - cur_offset);
3405 }
3406 free_extent_map(em);
3407 cur_offset = last_byte;
3408 }
3409
3410 /*
3411 * If ret is still 0, means we're OK to fallocate.
3412 * Or just cleanup the list and exit.
3413 */
3414 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3415 if (!ret)
3416 ret = btrfs_prealloc_file_range(inode, mode,
3417 range->start,
3418 range->len, i_blocksize(inode),
3419 offset + len, &alloc_hint);
3420 else
3421 btrfs_free_reserved_data_space(BTRFS_I(inode),
3422 data_reserved, range->start,
3423 range->len);
3424 list_del(&range->list);
3425 kfree(range);
3426 }
3427 if (ret < 0)
3428 goto out_unlock;
3429
3430 /*
3431 * We didn't need to allocate any more space, but we still extended the
3432 * size of the file so we need to update i_size and the inode item.
3433 */
3434 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3435out_unlock:
3436 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3437 &cached_state);
3438out:
3439 inode_unlock(inode);
3440 /* Let go of our reservation. */
3441 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3442 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3443 cur_offset, alloc_end - cur_offset);
3444 extent_changeset_free(data_reserved);
3445 return ret;
3446}
3447
3448static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3449 int whence)
3450{
3451 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3452 struct extent_map *em = NULL;
3453 struct extent_state *cached_state = NULL;
3454 loff_t i_size = inode->i_size;
3455 u64 lockstart;
3456 u64 lockend;
3457 u64 start;
3458 u64 len;
3459 int ret = 0;
3460
3461 if (i_size == 0 || offset >= i_size)
3462 return -ENXIO;
3463
3464 /*
3465 * offset can be negative, in this case we start finding DATA/HOLE from
3466 * the very start of the file.
3467 */
3468 start = max_t(loff_t, 0, offset);
3469
3470 lockstart = round_down(start, fs_info->sectorsize);
3471 lockend = round_up(i_size, fs_info->sectorsize);
3472 if (lockend <= lockstart)
3473 lockend = lockstart + fs_info->sectorsize;
3474 lockend--;
3475 len = lockend - lockstart + 1;
3476
3477 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3478 &cached_state);
3479
3480 while (start < i_size) {
3481 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3482 if (IS_ERR(em)) {
3483 ret = PTR_ERR(em);
3484 em = NULL;
3485 break;
3486 }
3487
3488 if (whence == SEEK_HOLE &&
3489 (em->block_start == EXTENT_MAP_HOLE ||
3490 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3491 break;
3492 else if (whence == SEEK_DATA &&
3493 (em->block_start != EXTENT_MAP_HOLE &&
3494 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3495 break;
3496
3497 start = em->start + em->len;
3498 free_extent_map(em);
3499 em = NULL;
3500 cond_resched();
3501 }
3502 free_extent_map(em);
3503 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3504 &cached_state);
3505 if (ret) {
3506 offset = ret;
3507 } else {
3508 if (whence == SEEK_DATA && start >= i_size)
3509 offset = -ENXIO;
3510 else
3511 offset = min_t(loff_t, start, i_size);
3512 }
3513
3514 return offset;
3515}
3516
3517static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3518{
3519 struct inode *inode = file->f_mapping->host;
3520
3521 switch (whence) {
3522 default:
3523 return generic_file_llseek(file, offset, whence);
3524 case SEEK_DATA:
3525 case SEEK_HOLE:
3526 inode_lock_shared(inode);
3527 offset = find_desired_extent(inode, offset, whence);
3528 inode_unlock_shared(inode);
3529 break;
3530 }
3531
3532 if (offset < 0)
3533 return offset;
3534
3535 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3536}
3537
3538static int btrfs_file_open(struct inode *inode, struct file *filp)
3539{
3540 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3541 return generic_file_open(inode, filp);
3542}
3543
3544const struct file_operations btrfs_file_operations = {
3545 .llseek = btrfs_file_llseek,
3546 .read_iter = generic_file_read_iter,
3547 .splice_read = generic_file_splice_read,
3548 .write_iter = btrfs_file_write_iter,
3549 .splice_write = iter_file_splice_write,
3550 .mmap = btrfs_file_mmap,
3551 .open = btrfs_file_open,
3552 .release = btrfs_release_file,
3553 .fsync = btrfs_sync_file,
3554 .fallocate = btrfs_fallocate,
3555 .unlocked_ioctl = btrfs_ioctl,
3556#ifdef CONFIG_COMPAT
3557 .compat_ioctl = btrfs_compat_ioctl,
3558#endif
3559 .remap_file_range = btrfs_remap_file_range,
3560};
3561
3562void __cold btrfs_auto_defrag_exit(void)
3563{
3564 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3565}
3566
3567int __init btrfs_auto_defrag_init(void)
3568{
3569 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3570 sizeof(struct inode_defrag), 0,
3571 SLAB_MEM_SPREAD,
3572 NULL);
3573 if (!btrfs_inode_defrag_cachep)
3574 return -ENOMEM;
3575
3576 return 0;
3577}
3578
3579int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3580{
3581 int ret;
3582
3583 /*
3584 * So with compression we will find and lock a dirty page and clear the
3585 * first one as dirty, setup an async extent, and immediately return
3586 * with the entire range locked but with nobody actually marked with
3587 * writeback. So we can't just filemap_write_and_wait_range() and
3588 * expect it to work since it will just kick off a thread to do the
3589 * actual work. So we need to call filemap_fdatawrite_range _again_
3590 * since it will wait on the page lock, which won't be unlocked until
3591 * after the pages have been marked as writeback and so we're good to go
3592 * from there. We have to do this otherwise we'll miss the ordered
3593 * extents and that results in badness. Please Josef, do not think you
3594 * know better and pull this out at some point in the future, it is
3595 * right and you are wrong.
3596 */
3597 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3598 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3599 &BTRFS_I(inode)->runtime_flags))
3600 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3601
3602 return ret;
3603}