Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6#include <linux/fs.h>
7#include <linux/pagemap.h>
8#include <linux/time.h>
9#include <linux/init.h>
10#include <linux/string.h>
11#include <linux/backing-dev.h>
12#include <linux/falloc.h>
13#include <linux/writeback.h>
14#include <linux/compat.h>
15#include <linux/slab.h>
16#include <linux/btrfs.h>
17#include <linux/uio.h>
18#include <linux/iversion.h>
19#include <linux/fsverity.h>
20#include <linux/iomap.h>
21#include "ctree.h"
22#include "disk-io.h"
23#include "transaction.h"
24#include "btrfs_inode.h"
25#include "tree-log.h"
26#include "locking.h"
27#include "qgroup.h"
28#include "compression.h"
29#include "delalloc-space.h"
30#include "reflink.h"
31#include "subpage.h"
32#include "fs.h"
33#include "accessors.h"
34#include "extent-tree.h"
35#include "file-item.h"
36#include "ioctl.h"
37#include "file.h"
38#include "super.h"
39
40/* simple helper to fault in pages and copy. This should go away
41 * and be replaced with calls into generic code.
42 */
43static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
44 struct page **prepared_pages,
45 struct iov_iter *i)
46{
47 size_t copied = 0;
48 size_t total_copied = 0;
49 int pg = 0;
50 int offset = offset_in_page(pos);
51
52 while (write_bytes > 0) {
53 size_t count = min_t(size_t,
54 PAGE_SIZE - offset, write_bytes);
55 struct page *page = prepared_pages[pg];
56 /*
57 * Copy data from userspace to the current page
58 */
59 copied = copy_page_from_iter_atomic(page, offset, count, i);
60
61 /* Flush processor's dcache for this page */
62 flush_dcache_page(page);
63
64 /*
65 * if we get a partial write, we can end up with
66 * partially up to date pages. These add
67 * a lot of complexity, so make sure they don't
68 * happen by forcing this copy to be retried.
69 *
70 * The rest of the btrfs_file_write code will fall
71 * back to page at a time copies after we return 0.
72 */
73 if (unlikely(copied < count)) {
74 if (!PageUptodate(page)) {
75 iov_iter_revert(i, copied);
76 copied = 0;
77 }
78 if (!copied)
79 break;
80 }
81
82 write_bytes -= copied;
83 total_copied += copied;
84 offset += copied;
85 if (offset == PAGE_SIZE) {
86 pg++;
87 offset = 0;
88 }
89 }
90 return total_copied;
91}
92
93/*
94 * unlocks pages after btrfs_file_write is done with them
95 */
96static void btrfs_drop_pages(struct btrfs_fs_info *fs_info,
97 struct page **pages, size_t num_pages,
98 u64 pos, u64 copied)
99{
100 size_t i;
101 u64 block_start = round_down(pos, fs_info->sectorsize);
102 u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
103
104 ASSERT(block_len <= U32_MAX);
105 for (i = 0; i < num_pages; i++) {
106 /* page checked is some magic around finding pages that
107 * have been modified without going through btrfs_set_page_dirty
108 * clear it here. There should be no need to mark the pages
109 * accessed as prepare_pages should have marked them accessed
110 * in prepare_pages via find_or_create_page()
111 */
112 btrfs_folio_clamp_clear_checked(fs_info, page_folio(pages[i]),
113 block_start, block_len);
114 unlock_page(pages[i]);
115 put_page(pages[i]);
116 }
117}
118
119/*
120 * After btrfs_copy_from_user(), update the following things for delalloc:
121 * - Mark newly dirtied pages as DELALLOC in the io tree.
122 * Used to advise which range is to be written back.
123 * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
124 * - Update inode size for past EOF write
125 */
126int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
127 size_t num_pages, loff_t pos, size_t write_bytes,
128 struct extent_state **cached, bool noreserve)
129{
130 struct btrfs_fs_info *fs_info = inode->root->fs_info;
131 int err = 0;
132 int i;
133 u64 num_bytes;
134 u64 start_pos;
135 u64 end_of_last_block;
136 u64 end_pos = pos + write_bytes;
137 loff_t isize = i_size_read(&inode->vfs_inode);
138 unsigned int extra_bits = 0;
139
140 if (write_bytes == 0)
141 return 0;
142
143 if (noreserve)
144 extra_bits |= EXTENT_NORESERVE;
145
146 start_pos = round_down(pos, fs_info->sectorsize);
147 num_bytes = round_up(write_bytes + pos - start_pos,
148 fs_info->sectorsize);
149 ASSERT(num_bytes <= U32_MAX);
150
151 end_of_last_block = start_pos + num_bytes - 1;
152
153 /*
154 * The pages may have already been dirty, clear out old accounting so
155 * we can set things up properly
156 */
157 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
158 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
159 cached);
160
161 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
162 extra_bits, cached);
163 if (err)
164 return err;
165
166 for (i = 0; i < num_pages; i++) {
167 struct page *p = pages[i];
168
169 btrfs_folio_clamp_set_uptodate(fs_info, page_folio(p),
170 start_pos, num_bytes);
171 btrfs_folio_clamp_clear_checked(fs_info, page_folio(p),
172 start_pos, num_bytes);
173 btrfs_folio_clamp_set_dirty(fs_info, page_folio(p),
174 start_pos, num_bytes);
175 }
176
177 /*
178 * we've only changed i_size in ram, and we haven't updated
179 * the disk i_size. There is no need to log the inode
180 * at this time.
181 */
182 if (end_pos > isize)
183 i_size_write(&inode->vfs_inode, end_pos);
184 return 0;
185}
186
187/*
188 * this is very complex, but the basic idea is to drop all extents
189 * in the range start - end. hint_block is filled in with a block number
190 * that would be a good hint to the block allocator for this file.
191 *
192 * If an extent intersects the range but is not entirely inside the range
193 * it is either truncated or split. Anything entirely inside the range
194 * is deleted from the tree.
195 *
196 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
197 * to deal with that. We set the field 'bytes_found' of the arguments structure
198 * with the number of allocated bytes found in the target range, so that the
199 * caller can update the inode's number of bytes in an atomic way when
200 * replacing extents in a range to avoid races with stat(2).
201 */
202int btrfs_drop_extents(struct btrfs_trans_handle *trans,
203 struct btrfs_root *root, struct btrfs_inode *inode,
204 struct btrfs_drop_extents_args *args)
205{
206 struct btrfs_fs_info *fs_info = root->fs_info;
207 struct extent_buffer *leaf;
208 struct btrfs_file_extent_item *fi;
209 struct btrfs_ref ref = { 0 };
210 struct btrfs_key key;
211 struct btrfs_key new_key;
212 u64 ino = btrfs_ino(inode);
213 u64 search_start = args->start;
214 u64 disk_bytenr = 0;
215 u64 num_bytes = 0;
216 u64 extent_offset = 0;
217 u64 extent_end = 0;
218 u64 last_end = args->start;
219 int del_nr = 0;
220 int del_slot = 0;
221 int extent_type;
222 int recow;
223 int ret;
224 int modify_tree = -1;
225 int update_refs;
226 int found = 0;
227 struct btrfs_path *path = args->path;
228
229 args->bytes_found = 0;
230 args->extent_inserted = false;
231
232 /* Must always have a path if ->replace_extent is true */
233 ASSERT(!(args->replace_extent && !args->path));
234
235 if (!path) {
236 path = btrfs_alloc_path();
237 if (!path) {
238 ret = -ENOMEM;
239 goto out;
240 }
241 }
242
243 if (args->drop_cache)
244 btrfs_drop_extent_map_range(inode, args->start, args->end - 1, false);
245
246 if (args->start >= inode->disk_i_size && !args->replace_extent)
247 modify_tree = 0;
248
249 update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
250 while (1) {
251 recow = 0;
252 ret = btrfs_lookup_file_extent(trans, root, path, ino,
253 search_start, modify_tree);
254 if (ret < 0)
255 break;
256 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
257 leaf = path->nodes[0];
258 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
259 if (key.objectid == ino &&
260 key.type == BTRFS_EXTENT_DATA_KEY)
261 path->slots[0]--;
262 }
263 ret = 0;
264next_slot:
265 leaf = path->nodes[0];
266 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
267 BUG_ON(del_nr > 0);
268 ret = btrfs_next_leaf(root, path);
269 if (ret < 0)
270 break;
271 if (ret > 0) {
272 ret = 0;
273 break;
274 }
275 leaf = path->nodes[0];
276 recow = 1;
277 }
278
279 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
280
281 if (key.objectid > ino)
282 break;
283 if (WARN_ON_ONCE(key.objectid < ino) ||
284 key.type < BTRFS_EXTENT_DATA_KEY) {
285 ASSERT(del_nr == 0);
286 path->slots[0]++;
287 goto next_slot;
288 }
289 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
290 break;
291
292 fi = btrfs_item_ptr(leaf, path->slots[0],
293 struct btrfs_file_extent_item);
294 extent_type = btrfs_file_extent_type(leaf, fi);
295
296 if (extent_type == BTRFS_FILE_EXTENT_REG ||
297 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
298 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
299 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
300 extent_offset = btrfs_file_extent_offset(leaf, fi);
301 extent_end = key.offset +
302 btrfs_file_extent_num_bytes(leaf, fi);
303 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
304 extent_end = key.offset +
305 btrfs_file_extent_ram_bytes(leaf, fi);
306 } else {
307 /* can't happen */
308 BUG();
309 }
310
311 /*
312 * Don't skip extent items representing 0 byte lengths. They
313 * used to be created (bug) if while punching holes we hit
314 * -ENOSPC condition. So if we find one here, just ensure we
315 * delete it, otherwise we would insert a new file extent item
316 * with the same key (offset) as that 0 bytes length file
317 * extent item in the call to setup_items_for_insert() later
318 * in this function.
319 */
320 if (extent_end == key.offset && extent_end >= search_start) {
321 last_end = extent_end;
322 goto delete_extent_item;
323 }
324
325 if (extent_end <= search_start) {
326 path->slots[0]++;
327 goto next_slot;
328 }
329
330 found = 1;
331 search_start = max(key.offset, args->start);
332 if (recow || !modify_tree) {
333 modify_tree = -1;
334 btrfs_release_path(path);
335 continue;
336 }
337
338 /*
339 * | - range to drop - |
340 * | -------- extent -------- |
341 */
342 if (args->start > key.offset && args->end < extent_end) {
343 BUG_ON(del_nr > 0);
344 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
345 ret = -EOPNOTSUPP;
346 break;
347 }
348
349 memcpy(&new_key, &key, sizeof(new_key));
350 new_key.offset = args->start;
351 ret = btrfs_duplicate_item(trans, root, path,
352 &new_key);
353 if (ret == -EAGAIN) {
354 btrfs_release_path(path);
355 continue;
356 }
357 if (ret < 0)
358 break;
359
360 leaf = path->nodes[0];
361 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
362 struct btrfs_file_extent_item);
363 btrfs_set_file_extent_num_bytes(leaf, fi,
364 args->start - key.offset);
365
366 fi = btrfs_item_ptr(leaf, path->slots[0],
367 struct btrfs_file_extent_item);
368
369 extent_offset += args->start - key.offset;
370 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
371 btrfs_set_file_extent_num_bytes(leaf, fi,
372 extent_end - args->start);
373 btrfs_mark_buffer_dirty(trans, leaf);
374
375 if (update_refs && disk_bytenr > 0) {
376 btrfs_init_generic_ref(&ref,
377 BTRFS_ADD_DELAYED_REF,
378 disk_bytenr, num_bytes, 0,
379 root->root_key.objectid);
380 btrfs_init_data_ref(&ref,
381 root->root_key.objectid,
382 new_key.objectid,
383 args->start - extent_offset,
384 0, false);
385 ret = btrfs_inc_extent_ref(trans, &ref);
386 if (ret) {
387 btrfs_abort_transaction(trans, ret);
388 break;
389 }
390 }
391 key.offset = args->start;
392 }
393 /*
394 * From here on out we will have actually dropped something, so
395 * last_end can be updated.
396 */
397 last_end = extent_end;
398
399 /*
400 * | ---- range to drop ----- |
401 * | -------- extent -------- |
402 */
403 if (args->start <= key.offset && args->end < extent_end) {
404 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
405 ret = -EOPNOTSUPP;
406 break;
407 }
408
409 memcpy(&new_key, &key, sizeof(new_key));
410 new_key.offset = args->end;
411 btrfs_set_item_key_safe(trans, path, &new_key);
412
413 extent_offset += args->end - key.offset;
414 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
415 btrfs_set_file_extent_num_bytes(leaf, fi,
416 extent_end - args->end);
417 btrfs_mark_buffer_dirty(trans, leaf);
418 if (update_refs && disk_bytenr > 0)
419 args->bytes_found += args->end - key.offset;
420 break;
421 }
422
423 search_start = extent_end;
424 /*
425 * | ---- range to drop ----- |
426 * | -------- extent -------- |
427 */
428 if (args->start > key.offset && args->end >= extent_end) {
429 BUG_ON(del_nr > 0);
430 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
431 ret = -EOPNOTSUPP;
432 break;
433 }
434
435 btrfs_set_file_extent_num_bytes(leaf, fi,
436 args->start - key.offset);
437 btrfs_mark_buffer_dirty(trans, leaf);
438 if (update_refs && disk_bytenr > 0)
439 args->bytes_found += extent_end - args->start;
440 if (args->end == extent_end)
441 break;
442
443 path->slots[0]++;
444 goto next_slot;
445 }
446
447 /*
448 * | ---- range to drop ----- |
449 * | ------ extent ------ |
450 */
451 if (args->start <= key.offset && args->end >= extent_end) {
452delete_extent_item:
453 if (del_nr == 0) {
454 del_slot = path->slots[0];
455 del_nr = 1;
456 } else {
457 BUG_ON(del_slot + del_nr != path->slots[0]);
458 del_nr++;
459 }
460
461 if (update_refs &&
462 extent_type == BTRFS_FILE_EXTENT_INLINE) {
463 args->bytes_found += extent_end - key.offset;
464 extent_end = ALIGN(extent_end,
465 fs_info->sectorsize);
466 } else if (update_refs && disk_bytenr > 0) {
467 btrfs_init_generic_ref(&ref,
468 BTRFS_DROP_DELAYED_REF,
469 disk_bytenr, num_bytes, 0,
470 root->root_key.objectid);
471 btrfs_init_data_ref(&ref,
472 root->root_key.objectid,
473 key.objectid,
474 key.offset - extent_offset, 0,
475 false);
476 ret = btrfs_free_extent(trans, &ref);
477 if (ret) {
478 btrfs_abort_transaction(trans, ret);
479 break;
480 }
481 args->bytes_found += extent_end - key.offset;
482 }
483
484 if (args->end == extent_end)
485 break;
486
487 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
488 path->slots[0]++;
489 goto next_slot;
490 }
491
492 ret = btrfs_del_items(trans, root, path, del_slot,
493 del_nr);
494 if (ret) {
495 btrfs_abort_transaction(trans, ret);
496 break;
497 }
498
499 del_nr = 0;
500 del_slot = 0;
501
502 btrfs_release_path(path);
503 continue;
504 }
505
506 BUG();
507 }
508
509 if (!ret && del_nr > 0) {
510 /*
511 * Set path->slots[0] to first slot, so that after the delete
512 * if items are move off from our leaf to its immediate left or
513 * right neighbor leafs, we end up with a correct and adjusted
514 * path->slots[0] for our insertion (if args->replace_extent).
515 */
516 path->slots[0] = del_slot;
517 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
518 if (ret)
519 btrfs_abort_transaction(trans, ret);
520 }
521
522 leaf = path->nodes[0];
523 /*
524 * If btrfs_del_items() was called, it might have deleted a leaf, in
525 * which case it unlocked our path, so check path->locks[0] matches a
526 * write lock.
527 */
528 if (!ret && args->replace_extent &&
529 path->locks[0] == BTRFS_WRITE_LOCK &&
530 btrfs_leaf_free_space(leaf) >=
531 sizeof(struct btrfs_item) + args->extent_item_size) {
532
533 key.objectid = ino;
534 key.type = BTRFS_EXTENT_DATA_KEY;
535 key.offset = args->start;
536 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
537 struct btrfs_key slot_key;
538
539 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
540 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
541 path->slots[0]++;
542 }
543 btrfs_setup_item_for_insert(trans, root, path, &key,
544 args->extent_item_size);
545 args->extent_inserted = true;
546 }
547
548 if (!args->path)
549 btrfs_free_path(path);
550 else if (!args->extent_inserted)
551 btrfs_release_path(path);
552out:
553 args->drop_end = found ? min(args->end, last_end) : args->end;
554
555 return ret;
556}
557
558static int extent_mergeable(struct extent_buffer *leaf, int slot,
559 u64 objectid, u64 bytenr, u64 orig_offset,
560 u64 *start, u64 *end)
561{
562 struct btrfs_file_extent_item *fi;
563 struct btrfs_key key;
564 u64 extent_end;
565
566 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
567 return 0;
568
569 btrfs_item_key_to_cpu(leaf, &key, slot);
570 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
571 return 0;
572
573 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
574 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
575 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
576 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
577 btrfs_file_extent_compression(leaf, fi) ||
578 btrfs_file_extent_encryption(leaf, fi) ||
579 btrfs_file_extent_other_encoding(leaf, fi))
580 return 0;
581
582 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
583 if ((*start && *start != key.offset) || (*end && *end != extent_end))
584 return 0;
585
586 *start = key.offset;
587 *end = extent_end;
588 return 1;
589}
590
591/*
592 * Mark extent in the range start - end as written.
593 *
594 * This changes extent type from 'pre-allocated' to 'regular'. If only
595 * part of extent is marked as written, the extent will be split into
596 * two or three.
597 */
598int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
599 struct btrfs_inode *inode, u64 start, u64 end)
600{
601 struct btrfs_root *root = inode->root;
602 struct extent_buffer *leaf;
603 struct btrfs_path *path;
604 struct btrfs_file_extent_item *fi;
605 struct btrfs_ref ref = { 0 };
606 struct btrfs_key key;
607 struct btrfs_key new_key;
608 u64 bytenr;
609 u64 num_bytes;
610 u64 extent_end;
611 u64 orig_offset;
612 u64 other_start;
613 u64 other_end;
614 u64 split;
615 int del_nr = 0;
616 int del_slot = 0;
617 int recow;
618 int ret = 0;
619 u64 ino = btrfs_ino(inode);
620
621 path = btrfs_alloc_path();
622 if (!path)
623 return -ENOMEM;
624again:
625 recow = 0;
626 split = start;
627 key.objectid = ino;
628 key.type = BTRFS_EXTENT_DATA_KEY;
629 key.offset = split;
630
631 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
632 if (ret < 0)
633 goto out;
634 if (ret > 0 && path->slots[0] > 0)
635 path->slots[0]--;
636
637 leaf = path->nodes[0];
638 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
639 if (key.objectid != ino ||
640 key.type != BTRFS_EXTENT_DATA_KEY) {
641 ret = -EINVAL;
642 btrfs_abort_transaction(trans, ret);
643 goto out;
644 }
645 fi = btrfs_item_ptr(leaf, path->slots[0],
646 struct btrfs_file_extent_item);
647 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
648 ret = -EINVAL;
649 btrfs_abort_transaction(trans, ret);
650 goto out;
651 }
652 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
653 if (key.offset > start || extent_end < end) {
654 ret = -EINVAL;
655 btrfs_abort_transaction(trans, ret);
656 goto out;
657 }
658
659 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
660 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
661 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
662 memcpy(&new_key, &key, sizeof(new_key));
663
664 if (start == key.offset && end < extent_end) {
665 other_start = 0;
666 other_end = start;
667 if (extent_mergeable(leaf, path->slots[0] - 1,
668 ino, bytenr, orig_offset,
669 &other_start, &other_end)) {
670 new_key.offset = end;
671 btrfs_set_item_key_safe(trans, path, &new_key);
672 fi = btrfs_item_ptr(leaf, path->slots[0],
673 struct btrfs_file_extent_item);
674 btrfs_set_file_extent_generation(leaf, fi,
675 trans->transid);
676 btrfs_set_file_extent_num_bytes(leaf, fi,
677 extent_end - end);
678 btrfs_set_file_extent_offset(leaf, fi,
679 end - orig_offset);
680 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
681 struct btrfs_file_extent_item);
682 btrfs_set_file_extent_generation(leaf, fi,
683 trans->transid);
684 btrfs_set_file_extent_num_bytes(leaf, fi,
685 end - other_start);
686 btrfs_mark_buffer_dirty(trans, leaf);
687 goto out;
688 }
689 }
690
691 if (start > key.offset && end == extent_end) {
692 other_start = end;
693 other_end = 0;
694 if (extent_mergeable(leaf, path->slots[0] + 1,
695 ino, bytenr, orig_offset,
696 &other_start, &other_end)) {
697 fi = btrfs_item_ptr(leaf, path->slots[0],
698 struct btrfs_file_extent_item);
699 btrfs_set_file_extent_num_bytes(leaf, fi,
700 start - key.offset);
701 btrfs_set_file_extent_generation(leaf, fi,
702 trans->transid);
703 path->slots[0]++;
704 new_key.offset = start;
705 btrfs_set_item_key_safe(trans, path, &new_key);
706
707 fi = btrfs_item_ptr(leaf, path->slots[0],
708 struct btrfs_file_extent_item);
709 btrfs_set_file_extent_generation(leaf, fi,
710 trans->transid);
711 btrfs_set_file_extent_num_bytes(leaf, fi,
712 other_end - start);
713 btrfs_set_file_extent_offset(leaf, fi,
714 start - orig_offset);
715 btrfs_mark_buffer_dirty(trans, leaf);
716 goto out;
717 }
718 }
719
720 while (start > key.offset || end < extent_end) {
721 if (key.offset == start)
722 split = end;
723
724 new_key.offset = split;
725 ret = btrfs_duplicate_item(trans, root, path, &new_key);
726 if (ret == -EAGAIN) {
727 btrfs_release_path(path);
728 goto again;
729 }
730 if (ret < 0) {
731 btrfs_abort_transaction(trans, ret);
732 goto out;
733 }
734
735 leaf = path->nodes[0];
736 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
737 struct btrfs_file_extent_item);
738 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
739 btrfs_set_file_extent_num_bytes(leaf, fi,
740 split - key.offset);
741
742 fi = btrfs_item_ptr(leaf, path->slots[0],
743 struct btrfs_file_extent_item);
744
745 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
746 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
747 btrfs_set_file_extent_num_bytes(leaf, fi,
748 extent_end - split);
749 btrfs_mark_buffer_dirty(trans, leaf);
750
751 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
752 num_bytes, 0, root->root_key.objectid);
753 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
754 orig_offset, 0, false);
755 ret = btrfs_inc_extent_ref(trans, &ref);
756 if (ret) {
757 btrfs_abort_transaction(trans, ret);
758 goto out;
759 }
760
761 if (split == start) {
762 key.offset = start;
763 } else {
764 if (start != key.offset) {
765 ret = -EINVAL;
766 btrfs_abort_transaction(trans, ret);
767 goto out;
768 }
769 path->slots[0]--;
770 extent_end = end;
771 }
772 recow = 1;
773 }
774
775 other_start = end;
776 other_end = 0;
777 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
778 num_bytes, 0, root->root_key.objectid);
779 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset,
780 0, false);
781 if (extent_mergeable(leaf, path->slots[0] + 1,
782 ino, bytenr, orig_offset,
783 &other_start, &other_end)) {
784 if (recow) {
785 btrfs_release_path(path);
786 goto again;
787 }
788 extent_end = other_end;
789 del_slot = path->slots[0] + 1;
790 del_nr++;
791 ret = btrfs_free_extent(trans, &ref);
792 if (ret) {
793 btrfs_abort_transaction(trans, ret);
794 goto out;
795 }
796 }
797 other_start = 0;
798 other_end = start;
799 if (extent_mergeable(leaf, path->slots[0] - 1,
800 ino, bytenr, orig_offset,
801 &other_start, &other_end)) {
802 if (recow) {
803 btrfs_release_path(path);
804 goto again;
805 }
806 key.offset = other_start;
807 del_slot = path->slots[0];
808 del_nr++;
809 ret = btrfs_free_extent(trans, &ref);
810 if (ret) {
811 btrfs_abort_transaction(trans, ret);
812 goto out;
813 }
814 }
815 if (del_nr == 0) {
816 fi = btrfs_item_ptr(leaf, path->slots[0],
817 struct btrfs_file_extent_item);
818 btrfs_set_file_extent_type(leaf, fi,
819 BTRFS_FILE_EXTENT_REG);
820 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
821 btrfs_mark_buffer_dirty(trans, leaf);
822 } else {
823 fi = btrfs_item_ptr(leaf, del_slot - 1,
824 struct btrfs_file_extent_item);
825 btrfs_set_file_extent_type(leaf, fi,
826 BTRFS_FILE_EXTENT_REG);
827 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
828 btrfs_set_file_extent_num_bytes(leaf, fi,
829 extent_end - key.offset);
830 btrfs_mark_buffer_dirty(trans, leaf);
831
832 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
833 if (ret < 0) {
834 btrfs_abort_transaction(trans, ret);
835 goto out;
836 }
837 }
838out:
839 btrfs_free_path(path);
840 return ret;
841}
842
843/*
844 * on error we return an unlocked page and the error value
845 * on success we return a locked page and 0
846 */
847static int prepare_uptodate_page(struct inode *inode,
848 struct page *page, u64 pos,
849 bool force_uptodate)
850{
851 struct folio *folio = page_folio(page);
852 int ret = 0;
853
854 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
855 !PageUptodate(page)) {
856 ret = btrfs_read_folio(NULL, folio);
857 if (ret)
858 return ret;
859 lock_page(page);
860 if (!PageUptodate(page)) {
861 unlock_page(page);
862 return -EIO;
863 }
864
865 /*
866 * Since btrfs_read_folio() will unlock the folio before it
867 * returns, there is a window where btrfs_release_folio() can be
868 * called to release the page. Here we check both inode
869 * mapping and PagePrivate() to make sure the page was not
870 * released.
871 *
872 * The private flag check is essential for subpage as we need
873 * to store extra bitmap using folio private.
874 */
875 if (page->mapping != inode->i_mapping || !folio_test_private(folio)) {
876 unlock_page(page);
877 return -EAGAIN;
878 }
879 }
880 return 0;
881}
882
883static fgf_t get_prepare_fgp_flags(bool nowait)
884{
885 fgf_t fgp_flags = FGP_LOCK | FGP_ACCESSED | FGP_CREAT;
886
887 if (nowait)
888 fgp_flags |= FGP_NOWAIT;
889
890 return fgp_flags;
891}
892
893static gfp_t get_prepare_gfp_flags(struct inode *inode, bool nowait)
894{
895 gfp_t gfp;
896
897 gfp = btrfs_alloc_write_mask(inode->i_mapping);
898 if (nowait) {
899 gfp &= ~__GFP_DIRECT_RECLAIM;
900 gfp |= GFP_NOWAIT;
901 }
902
903 return gfp;
904}
905
906/*
907 * this just gets pages into the page cache and locks them down.
908 */
909static noinline int prepare_pages(struct inode *inode, struct page **pages,
910 size_t num_pages, loff_t pos,
911 size_t write_bytes, bool force_uptodate,
912 bool nowait)
913{
914 int i;
915 unsigned long index = pos >> PAGE_SHIFT;
916 gfp_t mask = get_prepare_gfp_flags(inode, nowait);
917 fgf_t fgp_flags = get_prepare_fgp_flags(nowait);
918 int err = 0;
919 int faili;
920
921 for (i = 0; i < num_pages; i++) {
922again:
923 pages[i] = pagecache_get_page(inode->i_mapping, index + i,
924 fgp_flags, mask | __GFP_WRITE);
925 if (!pages[i]) {
926 faili = i - 1;
927 if (nowait)
928 err = -EAGAIN;
929 else
930 err = -ENOMEM;
931 goto fail;
932 }
933
934 err = set_page_extent_mapped(pages[i]);
935 if (err < 0) {
936 faili = i;
937 goto fail;
938 }
939
940 if (i == 0)
941 err = prepare_uptodate_page(inode, pages[i], pos,
942 force_uptodate);
943 if (!err && i == num_pages - 1)
944 err = prepare_uptodate_page(inode, pages[i],
945 pos + write_bytes, false);
946 if (err) {
947 put_page(pages[i]);
948 if (!nowait && err == -EAGAIN) {
949 err = 0;
950 goto again;
951 }
952 faili = i - 1;
953 goto fail;
954 }
955 wait_on_page_writeback(pages[i]);
956 }
957
958 return 0;
959fail:
960 while (faili >= 0) {
961 unlock_page(pages[faili]);
962 put_page(pages[faili]);
963 faili--;
964 }
965 return err;
966
967}
968
969/*
970 * This function locks the extent and properly waits for data=ordered extents
971 * to finish before allowing the pages to be modified if need.
972 *
973 * The return value:
974 * 1 - the extent is locked
975 * 0 - the extent is not locked, and everything is OK
976 * -EAGAIN - need re-prepare the pages
977 * the other < 0 number - Something wrong happens
978 */
979static noinline int
980lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
981 size_t num_pages, loff_t pos,
982 size_t write_bytes,
983 u64 *lockstart, u64 *lockend, bool nowait,
984 struct extent_state **cached_state)
985{
986 struct btrfs_fs_info *fs_info = inode->root->fs_info;
987 u64 start_pos;
988 u64 last_pos;
989 int i;
990 int ret = 0;
991
992 start_pos = round_down(pos, fs_info->sectorsize);
993 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
994
995 if (start_pos < inode->vfs_inode.i_size) {
996 struct btrfs_ordered_extent *ordered;
997
998 if (nowait) {
999 if (!try_lock_extent(&inode->io_tree, start_pos, last_pos,
1000 cached_state)) {
1001 for (i = 0; i < num_pages; i++) {
1002 unlock_page(pages[i]);
1003 put_page(pages[i]);
1004 pages[i] = NULL;
1005 }
1006
1007 return -EAGAIN;
1008 }
1009 } else {
1010 lock_extent(&inode->io_tree, start_pos, last_pos, cached_state);
1011 }
1012
1013 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1014 last_pos - start_pos + 1);
1015 if (ordered &&
1016 ordered->file_offset + ordered->num_bytes > start_pos &&
1017 ordered->file_offset <= last_pos) {
1018 unlock_extent(&inode->io_tree, start_pos, last_pos,
1019 cached_state);
1020 for (i = 0; i < num_pages; i++) {
1021 unlock_page(pages[i]);
1022 put_page(pages[i]);
1023 }
1024 btrfs_start_ordered_extent(ordered);
1025 btrfs_put_ordered_extent(ordered);
1026 return -EAGAIN;
1027 }
1028 if (ordered)
1029 btrfs_put_ordered_extent(ordered);
1030
1031 *lockstart = start_pos;
1032 *lockend = last_pos;
1033 ret = 1;
1034 }
1035
1036 /*
1037 * We should be called after prepare_pages() which should have locked
1038 * all pages in the range.
1039 */
1040 for (i = 0; i < num_pages; i++)
1041 WARN_ON(!PageLocked(pages[i]));
1042
1043 return ret;
1044}
1045
1046/*
1047 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1048 *
1049 * @pos: File offset.
1050 * @write_bytes: The length to write, will be updated to the nocow writeable
1051 * range.
1052 *
1053 * This function will flush ordered extents in the range to ensure proper
1054 * nocow checks.
1055 *
1056 * Return:
1057 * > 0 If we can nocow, and updates @write_bytes.
1058 * 0 If we can't do a nocow write.
1059 * -EAGAIN If we can't do a nocow write because snapshoting of the inode's
1060 * root is in progress.
1061 * < 0 If an error happened.
1062 *
1063 * NOTE: Callers need to call btrfs_check_nocow_unlock() if we return > 0.
1064 */
1065int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1066 size_t *write_bytes, bool nowait)
1067{
1068 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1069 struct btrfs_root *root = inode->root;
1070 struct extent_state *cached_state = NULL;
1071 u64 lockstart, lockend;
1072 u64 num_bytes;
1073 int ret;
1074
1075 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1076 return 0;
1077
1078 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1079 return -EAGAIN;
1080
1081 lockstart = round_down(pos, fs_info->sectorsize);
1082 lockend = round_up(pos + *write_bytes,
1083 fs_info->sectorsize) - 1;
1084 num_bytes = lockend - lockstart + 1;
1085
1086 if (nowait) {
1087 if (!btrfs_try_lock_ordered_range(inode, lockstart, lockend,
1088 &cached_state)) {
1089 btrfs_drew_write_unlock(&root->snapshot_lock);
1090 return -EAGAIN;
1091 }
1092 } else {
1093 btrfs_lock_and_flush_ordered_range(inode, lockstart, lockend,
1094 &cached_state);
1095 }
1096 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1097 NULL, NULL, NULL, nowait, false);
1098 if (ret <= 0)
1099 btrfs_drew_write_unlock(&root->snapshot_lock);
1100 else
1101 *write_bytes = min_t(size_t, *write_bytes ,
1102 num_bytes - pos + lockstart);
1103 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
1104
1105 return ret;
1106}
1107
1108void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1109{
1110 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1111}
1112
1113static void update_time_for_write(struct inode *inode)
1114{
1115 struct timespec64 now, ts;
1116
1117 if (IS_NOCMTIME(inode))
1118 return;
1119
1120 now = current_time(inode);
1121 ts = inode_get_mtime(inode);
1122 if (!timespec64_equal(&ts, &now))
1123 inode_set_mtime_to_ts(inode, now);
1124
1125 ts = inode_get_ctime(inode);
1126 if (!timespec64_equal(&ts, &now))
1127 inode_set_ctime_to_ts(inode, now);
1128
1129 if (IS_I_VERSION(inode))
1130 inode_inc_iversion(inode);
1131}
1132
1133static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1134 size_t count)
1135{
1136 struct file *file = iocb->ki_filp;
1137 struct inode *inode = file_inode(file);
1138 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1139 loff_t pos = iocb->ki_pos;
1140 int ret;
1141 loff_t oldsize;
1142 loff_t start_pos;
1143
1144 /*
1145 * Quickly bail out on NOWAIT writes if we don't have the nodatacow or
1146 * prealloc flags, as without those flags we always have to COW. We will
1147 * later check if we can really COW into the target range (using
1148 * can_nocow_extent() at btrfs_get_blocks_direct_write()).
1149 */
1150 if ((iocb->ki_flags & IOCB_NOWAIT) &&
1151 !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1152 return -EAGAIN;
1153
1154 ret = file_remove_privs(file);
1155 if (ret)
1156 return ret;
1157
1158 /*
1159 * We reserve space for updating the inode when we reserve space for the
1160 * extent we are going to write, so we will enospc out there. We don't
1161 * need to start yet another transaction to update the inode as we will
1162 * update the inode when we finish writing whatever data we write.
1163 */
1164 update_time_for_write(inode);
1165
1166 start_pos = round_down(pos, fs_info->sectorsize);
1167 oldsize = i_size_read(inode);
1168 if (start_pos > oldsize) {
1169 /* Expand hole size to cover write data, preventing empty gap */
1170 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1171
1172 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1173 if (ret)
1174 return ret;
1175 }
1176
1177 return 0;
1178}
1179
1180static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1181 struct iov_iter *i)
1182{
1183 struct file *file = iocb->ki_filp;
1184 loff_t pos;
1185 struct inode *inode = file_inode(file);
1186 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1187 struct page **pages = NULL;
1188 struct extent_changeset *data_reserved = NULL;
1189 u64 release_bytes = 0;
1190 u64 lockstart;
1191 u64 lockend;
1192 size_t num_written = 0;
1193 int nrptrs;
1194 ssize_t ret;
1195 bool only_release_metadata = false;
1196 bool force_page_uptodate = false;
1197 loff_t old_isize = i_size_read(inode);
1198 unsigned int ilock_flags = 0;
1199 const bool nowait = (iocb->ki_flags & IOCB_NOWAIT);
1200 unsigned int bdp_flags = (nowait ? BDP_ASYNC : 0);
1201
1202 if (nowait)
1203 ilock_flags |= BTRFS_ILOCK_TRY;
1204
1205 ret = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1206 if (ret < 0)
1207 return ret;
1208
1209 ret = generic_write_checks(iocb, i);
1210 if (ret <= 0)
1211 goto out;
1212
1213 ret = btrfs_write_check(iocb, i, ret);
1214 if (ret < 0)
1215 goto out;
1216
1217 pos = iocb->ki_pos;
1218 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1219 PAGE_SIZE / (sizeof(struct page *)));
1220 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1221 nrptrs = max(nrptrs, 8);
1222 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1223 if (!pages) {
1224 ret = -ENOMEM;
1225 goto out;
1226 }
1227
1228 while (iov_iter_count(i) > 0) {
1229 struct extent_state *cached_state = NULL;
1230 size_t offset = offset_in_page(pos);
1231 size_t sector_offset;
1232 size_t write_bytes = min(iov_iter_count(i),
1233 nrptrs * (size_t)PAGE_SIZE -
1234 offset);
1235 size_t num_pages;
1236 size_t reserve_bytes;
1237 size_t dirty_pages;
1238 size_t copied;
1239 size_t dirty_sectors;
1240 size_t num_sectors;
1241 int extents_locked;
1242
1243 /*
1244 * Fault pages before locking them in prepare_pages
1245 * to avoid recursive lock
1246 */
1247 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1248 ret = -EFAULT;
1249 break;
1250 }
1251
1252 only_release_metadata = false;
1253 sector_offset = pos & (fs_info->sectorsize - 1);
1254
1255 extent_changeset_release(data_reserved);
1256 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1257 &data_reserved, pos,
1258 write_bytes, nowait);
1259 if (ret < 0) {
1260 int can_nocow;
1261
1262 if (nowait && (ret == -ENOSPC || ret == -EAGAIN)) {
1263 ret = -EAGAIN;
1264 break;
1265 }
1266
1267 /*
1268 * If we don't have to COW at the offset, reserve
1269 * metadata only. write_bytes may get smaller than
1270 * requested here.
1271 */
1272 can_nocow = btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1273 &write_bytes, nowait);
1274 if (can_nocow < 0)
1275 ret = can_nocow;
1276 if (can_nocow > 0)
1277 ret = 0;
1278 if (ret)
1279 break;
1280 only_release_metadata = true;
1281 }
1282
1283 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1284 WARN_ON(num_pages > nrptrs);
1285 reserve_bytes = round_up(write_bytes + sector_offset,
1286 fs_info->sectorsize);
1287 WARN_ON(reserve_bytes == 0);
1288 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1289 reserve_bytes,
1290 reserve_bytes, nowait);
1291 if (ret) {
1292 if (!only_release_metadata)
1293 btrfs_free_reserved_data_space(BTRFS_I(inode),
1294 data_reserved, pos,
1295 write_bytes);
1296 else
1297 btrfs_check_nocow_unlock(BTRFS_I(inode));
1298
1299 if (nowait && ret == -ENOSPC)
1300 ret = -EAGAIN;
1301 break;
1302 }
1303
1304 release_bytes = reserve_bytes;
1305again:
1306 ret = balance_dirty_pages_ratelimited_flags(inode->i_mapping, bdp_flags);
1307 if (ret) {
1308 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1309 break;
1310 }
1311
1312 /*
1313 * This is going to setup the pages array with the number of
1314 * pages we want, so we don't really need to worry about the
1315 * contents of pages from loop to loop
1316 */
1317 ret = prepare_pages(inode, pages, num_pages,
1318 pos, write_bytes, force_page_uptodate, false);
1319 if (ret) {
1320 btrfs_delalloc_release_extents(BTRFS_I(inode),
1321 reserve_bytes);
1322 break;
1323 }
1324
1325 extents_locked = lock_and_cleanup_extent_if_need(
1326 BTRFS_I(inode), pages,
1327 num_pages, pos, write_bytes, &lockstart,
1328 &lockend, nowait, &cached_state);
1329 if (extents_locked < 0) {
1330 if (!nowait && extents_locked == -EAGAIN)
1331 goto again;
1332
1333 btrfs_delalloc_release_extents(BTRFS_I(inode),
1334 reserve_bytes);
1335 ret = extents_locked;
1336 break;
1337 }
1338
1339 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1340
1341 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1342 dirty_sectors = round_up(copied + sector_offset,
1343 fs_info->sectorsize);
1344 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1345
1346 /*
1347 * if we have trouble faulting in the pages, fall
1348 * back to one page at a time
1349 */
1350 if (copied < write_bytes)
1351 nrptrs = 1;
1352
1353 if (copied == 0) {
1354 force_page_uptodate = true;
1355 dirty_sectors = 0;
1356 dirty_pages = 0;
1357 } else {
1358 force_page_uptodate = false;
1359 dirty_pages = DIV_ROUND_UP(copied + offset,
1360 PAGE_SIZE);
1361 }
1362
1363 if (num_sectors > dirty_sectors) {
1364 /* release everything except the sectors we dirtied */
1365 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1366 if (only_release_metadata) {
1367 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1368 release_bytes, true);
1369 } else {
1370 u64 __pos;
1371
1372 __pos = round_down(pos,
1373 fs_info->sectorsize) +
1374 (dirty_pages << PAGE_SHIFT);
1375 btrfs_delalloc_release_space(BTRFS_I(inode),
1376 data_reserved, __pos,
1377 release_bytes, true);
1378 }
1379 }
1380
1381 release_bytes = round_up(copied + sector_offset,
1382 fs_info->sectorsize);
1383
1384 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1385 dirty_pages, pos, copied,
1386 &cached_state, only_release_metadata);
1387
1388 /*
1389 * If we have not locked the extent range, because the range's
1390 * start offset is >= i_size, we might still have a non-NULL
1391 * cached extent state, acquired while marking the extent range
1392 * as delalloc through btrfs_dirty_pages(). Therefore free any
1393 * possible cached extent state to avoid a memory leak.
1394 */
1395 if (extents_locked)
1396 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
1397 lockend, &cached_state);
1398 else
1399 free_extent_state(cached_state);
1400
1401 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1402 if (ret) {
1403 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1404 break;
1405 }
1406
1407 release_bytes = 0;
1408 if (only_release_metadata)
1409 btrfs_check_nocow_unlock(BTRFS_I(inode));
1410
1411 btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1412
1413 cond_resched();
1414
1415 pos += copied;
1416 num_written += copied;
1417 }
1418
1419 kfree(pages);
1420
1421 if (release_bytes) {
1422 if (only_release_metadata) {
1423 btrfs_check_nocow_unlock(BTRFS_I(inode));
1424 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1425 release_bytes, true);
1426 } else {
1427 btrfs_delalloc_release_space(BTRFS_I(inode),
1428 data_reserved,
1429 round_down(pos, fs_info->sectorsize),
1430 release_bytes, true);
1431 }
1432 }
1433
1434 extent_changeset_free(data_reserved);
1435 if (num_written > 0) {
1436 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1437 iocb->ki_pos += num_written;
1438 }
1439out:
1440 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1441 return num_written ? num_written : ret;
1442}
1443
1444static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1445 const struct iov_iter *iter, loff_t offset)
1446{
1447 const u32 blocksize_mask = fs_info->sectorsize - 1;
1448
1449 if (offset & blocksize_mask)
1450 return -EINVAL;
1451
1452 if (iov_iter_alignment(iter) & blocksize_mask)
1453 return -EINVAL;
1454
1455 return 0;
1456}
1457
1458static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1459{
1460 struct file *file = iocb->ki_filp;
1461 struct inode *inode = file_inode(file);
1462 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1463 loff_t pos;
1464 ssize_t written = 0;
1465 ssize_t written_buffered;
1466 size_t prev_left = 0;
1467 loff_t endbyte;
1468 ssize_t err;
1469 unsigned int ilock_flags = 0;
1470 struct iomap_dio *dio;
1471
1472 if (iocb->ki_flags & IOCB_NOWAIT)
1473 ilock_flags |= BTRFS_ILOCK_TRY;
1474
1475 /*
1476 * If the write DIO is within EOF, use a shared lock and also only if
1477 * security bits will likely not be dropped by file_remove_privs() called
1478 * from btrfs_write_check(). Either will need to be rechecked after the
1479 * lock was acquired.
1480 */
1481 if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode) && IS_NOSEC(inode))
1482 ilock_flags |= BTRFS_ILOCK_SHARED;
1483
1484relock:
1485 err = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1486 if (err < 0)
1487 return err;
1488
1489 /* Shared lock cannot be used with security bits set. */
1490 if ((ilock_flags & BTRFS_ILOCK_SHARED) && !IS_NOSEC(inode)) {
1491 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1492 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1493 goto relock;
1494 }
1495
1496 err = generic_write_checks(iocb, from);
1497 if (err <= 0) {
1498 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1499 return err;
1500 }
1501
1502 err = btrfs_write_check(iocb, from, err);
1503 if (err < 0) {
1504 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1505 goto out;
1506 }
1507
1508 pos = iocb->ki_pos;
1509 /*
1510 * Re-check since file size may have changed just before taking the
1511 * lock or pos may have changed because of O_APPEND in generic_write_check()
1512 */
1513 if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1514 pos + iov_iter_count(from) > i_size_read(inode)) {
1515 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1516 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1517 goto relock;
1518 }
1519
1520 if (check_direct_IO(fs_info, from, pos)) {
1521 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1522 goto buffered;
1523 }
1524
1525 /*
1526 * The iov_iter can be mapped to the same file range we are writing to.
1527 * If that's the case, then we will deadlock in the iomap code, because
1528 * it first calls our callback btrfs_dio_iomap_begin(), which will create
1529 * an ordered extent, and after that it will fault in the pages that the
1530 * iov_iter refers to. During the fault in we end up in the readahead
1531 * pages code (starting at btrfs_readahead()), which will lock the range,
1532 * find that ordered extent and then wait for it to complete (at
1533 * btrfs_lock_and_flush_ordered_range()), resulting in a deadlock since
1534 * obviously the ordered extent can never complete as we didn't submit
1535 * yet the respective bio(s). This always happens when the buffer is
1536 * memory mapped to the same file range, since the iomap DIO code always
1537 * invalidates pages in the target file range (after starting and waiting
1538 * for any writeback).
1539 *
1540 * So here we disable page faults in the iov_iter and then retry if we
1541 * got -EFAULT, faulting in the pages before the retry.
1542 */
1543 from->nofault = true;
1544 dio = btrfs_dio_write(iocb, from, written);
1545 from->nofault = false;
1546
1547 /*
1548 * iomap_dio_complete() will call btrfs_sync_file() if we have a dsync
1549 * iocb, and that needs to lock the inode. So unlock it before calling
1550 * iomap_dio_complete() to avoid a deadlock.
1551 */
1552 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1553
1554 if (IS_ERR_OR_NULL(dio))
1555 err = PTR_ERR_OR_ZERO(dio);
1556 else
1557 err = iomap_dio_complete(dio);
1558
1559 /* No increment (+=) because iomap returns a cumulative value. */
1560 if (err > 0)
1561 written = err;
1562
1563 if (iov_iter_count(from) > 0 && (err == -EFAULT || err > 0)) {
1564 const size_t left = iov_iter_count(from);
1565 /*
1566 * We have more data left to write. Try to fault in as many as
1567 * possible of the remainder pages and retry. We do this without
1568 * releasing and locking again the inode, to prevent races with
1569 * truncate.
1570 *
1571 * Also, in case the iov refers to pages in the file range of the
1572 * file we want to write to (due to a mmap), we could enter an
1573 * infinite loop if we retry after faulting the pages in, since
1574 * iomap will invalidate any pages in the range early on, before
1575 * it tries to fault in the pages of the iov. So we keep track of
1576 * how much was left of iov in the previous EFAULT and fallback
1577 * to buffered IO in case we haven't made any progress.
1578 */
1579 if (left == prev_left) {
1580 err = -ENOTBLK;
1581 } else {
1582 fault_in_iov_iter_readable(from, left);
1583 prev_left = left;
1584 goto relock;
1585 }
1586 }
1587
1588 /*
1589 * If 'err' is -ENOTBLK or we have not written all data, then it means
1590 * we must fallback to buffered IO.
1591 */
1592 if ((err < 0 && err != -ENOTBLK) || !iov_iter_count(from))
1593 goto out;
1594
1595buffered:
1596 /*
1597 * If we are in a NOWAIT context, then return -EAGAIN to signal the caller
1598 * it must retry the operation in a context where blocking is acceptable,
1599 * because even if we end up not blocking during the buffered IO attempt
1600 * below, we will block when flushing and waiting for the IO.
1601 */
1602 if (iocb->ki_flags & IOCB_NOWAIT) {
1603 err = -EAGAIN;
1604 goto out;
1605 }
1606
1607 pos = iocb->ki_pos;
1608 written_buffered = btrfs_buffered_write(iocb, from);
1609 if (written_buffered < 0) {
1610 err = written_buffered;
1611 goto out;
1612 }
1613 /*
1614 * Ensure all data is persisted. We want the next direct IO read to be
1615 * able to read what was just written.
1616 */
1617 endbyte = pos + written_buffered - 1;
1618 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1619 if (err)
1620 goto out;
1621 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1622 if (err)
1623 goto out;
1624 written += written_buffered;
1625 iocb->ki_pos = pos + written_buffered;
1626 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1627 endbyte >> PAGE_SHIFT);
1628out:
1629 return err < 0 ? err : written;
1630}
1631
1632static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
1633 const struct btrfs_ioctl_encoded_io_args *encoded)
1634{
1635 struct file *file = iocb->ki_filp;
1636 struct inode *inode = file_inode(file);
1637 loff_t count;
1638 ssize_t ret;
1639
1640 btrfs_inode_lock(BTRFS_I(inode), 0);
1641 count = encoded->len;
1642 ret = generic_write_checks_count(iocb, &count);
1643 if (ret == 0 && count != encoded->len) {
1644 /*
1645 * The write got truncated by generic_write_checks_count(). We
1646 * can't do a partial encoded write.
1647 */
1648 ret = -EFBIG;
1649 }
1650 if (ret || encoded->len == 0)
1651 goto out;
1652
1653 ret = btrfs_write_check(iocb, from, encoded->len);
1654 if (ret < 0)
1655 goto out;
1656
1657 ret = btrfs_do_encoded_write(iocb, from, encoded);
1658out:
1659 btrfs_inode_unlock(BTRFS_I(inode), 0);
1660 return ret;
1661}
1662
1663ssize_t btrfs_do_write_iter(struct kiocb *iocb, struct iov_iter *from,
1664 const struct btrfs_ioctl_encoded_io_args *encoded)
1665{
1666 struct file *file = iocb->ki_filp;
1667 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1668 ssize_t num_written, num_sync;
1669
1670 /*
1671 * If the fs flips readonly due to some impossible error, although we
1672 * have opened a file as writable, we have to stop this write operation
1673 * to ensure consistency.
1674 */
1675 if (BTRFS_FS_ERROR(inode->root->fs_info))
1676 return -EROFS;
1677
1678 if (encoded && (iocb->ki_flags & IOCB_NOWAIT))
1679 return -EOPNOTSUPP;
1680
1681 if (encoded) {
1682 num_written = btrfs_encoded_write(iocb, from, encoded);
1683 num_sync = encoded->len;
1684 } else if (iocb->ki_flags & IOCB_DIRECT) {
1685 num_written = btrfs_direct_write(iocb, from);
1686 num_sync = num_written;
1687 } else {
1688 num_written = btrfs_buffered_write(iocb, from);
1689 num_sync = num_written;
1690 }
1691
1692 btrfs_set_inode_last_sub_trans(inode);
1693
1694 if (num_sync > 0) {
1695 num_sync = generic_write_sync(iocb, num_sync);
1696 if (num_sync < 0)
1697 num_written = num_sync;
1698 }
1699
1700 return num_written;
1701}
1702
1703static ssize_t btrfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1704{
1705 return btrfs_do_write_iter(iocb, from, NULL);
1706}
1707
1708int btrfs_release_file(struct inode *inode, struct file *filp)
1709{
1710 struct btrfs_file_private *private = filp->private_data;
1711
1712 if (private) {
1713 kfree(private->filldir_buf);
1714 free_extent_state(private->llseek_cached_state);
1715 kfree(private);
1716 filp->private_data = NULL;
1717 }
1718
1719 /*
1720 * Set by setattr when we are about to truncate a file from a non-zero
1721 * size to a zero size. This tries to flush down new bytes that may
1722 * have been written if the application were using truncate to replace
1723 * a file in place.
1724 */
1725 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
1726 &BTRFS_I(inode)->runtime_flags))
1727 filemap_flush(inode->i_mapping);
1728 return 0;
1729}
1730
1731static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1732{
1733 int ret;
1734 struct blk_plug plug;
1735
1736 /*
1737 * This is only called in fsync, which would do synchronous writes, so
1738 * a plug can merge adjacent IOs as much as possible. Esp. in case of
1739 * multiple disks using raid profile, a large IO can be split to
1740 * several segments of stripe length (currently 64K).
1741 */
1742 blk_start_plug(&plug);
1743 ret = btrfs_fdatawrite_range(inode, start, end);
1744 blk_finish_plug(&plug);
1745
1746 return ret;
1747}
1748
1749static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
1750{
1751 struct btrfs_inode *inode = BTRFS_I(ctx->inode);
1752 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1753
1754 if (btrfs_inode_in_log(inode, btrfs_get_fs_generation(fs_info)) &&
1755 list_empty(&ctx->ordered_extents))
1756 return true;
1757
1758 /*
1759 * If we are doing a fast fsync we can not bail out if the inode's
1760 * last_trans is <= then the last committed transaction, because we only
1761 * update the last_trans of the inode during ordered extent completion,
1762 * and for a fast fsync we don't wait for that, we only wait for the
1763 * writeback to complete.
1764 */
1765 if (inode->last_trans <= btrfs_get_last_trans_committed(fs_info) &&
1766 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
1767 list_empty(&ctx->ordered_extents)))
1768 return true;
1769
1770 return false;
1771}
1772
1773/*
1774 * fsync call for both files and directories. This logs the inode into
1775 * the tree log instead of forcing full commits whenever possible.
1776 *
1777 * It needs to call filemap_fdatawait so that all ordered extent updates are
1778 * in the metadata btree are up to date for copying to the log.
1779 *
1780 * It drops the inode mutex before doing the tree log commit. This is an
1781 * important optimization for directories because holding the mutex prevents
1782 * new operations on the dir while we write to disk.
1783 */
1784int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1785{
1786 struct dentry *dentry = file_dentry(file);
1787 struct inode *inode = d_inode(dentry);
1788 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1789 struct btrfs_root *root = BTRFS_I(inode)->root;
1790 struct btrfs_trans_handle *trans;
1791 struct btrfs_log_ctx ctx;
1792 int ret = 0, err;
1793 u64 len;
1794 bool full_sync;
1795
1796 trace_btrfs_sync_file(file, datasync);
1797
1798 btrfs_init_log_ctx(&ctx, inode);
1799
1800 /*
1801 * Always set the range to a full range, otherwise we can get into
1802 * several problems, from missing file extent items to represent holes
1803 * when not using the NO_HOLES feature, to log tree corruption due to
1804 * races between hole detection during logging and completion of ordered
1805 * extents outside the range, to missing checksums due to ordered extents
1806 * for which we flushed only a subset of their pages.
1807 */
1808 start = 0;
1809 end = LLONG_MAX;
1810 len = (u64)LLONG_MAX + 1;
1811
1812 /*
1813 * We write the dirty pages in the range and wait until they complete
1814 * out of the ->i_mutex. If so, we can flush the dirty pages by
1815 * multi-task, and make the performance up. See
1816 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1817 */
1818 ret = start_ordered_ops(inode, start, end);
1819 if (ret)
1820 goto out;
1821
1822 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1823
1824 atomic_inc(&root->log_batch);
1825
1826 /*
1827 * Before we acquired the inode's lock and the mmap lock, someone may
1828 * have dirtied more pages in the target range. We need to make sure
1829 * that writeback for any such pages does not start while we are logging
1830 * the inode, because if it does, any of the following might happen when
1831 * we are not doing a full inode sync:
1832 *
1833 * 1) We log an extent after its writeback finishes but before its
1834 * checksums are added to the csum tree, leading to -EIO errors
1835 * when attempting to read the extent after a log replay.
1836 *
1837 * 2) We can end up logging an extent before its writeback finishes.
1838 * Therefore after the log replay we will have a file extent item
1839 * pointing to an unwritten extent (and no data checksums as well).
1840 *
1841 * So trigger writeback for any eventual new dirty pages and then we
1842 * wait for all ordered extents to complete below.
1843 */
1844 ret = start_ordered_ops(inode, start, end);
1845 if (ret) {
1846 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1847 goto out;
1848 }
1849
1850 /*
1851 * Always check for the full sync flag while holding the inode's lock,
1852 * to avoid races with other tasks. The flag must be either set all the
1853 * time during logging or always off all the time while logging.
1854 * We check the flag here after starting delalloc above, because when
1855 * running delalloc the full sync flag may be set if we need to drop
1856 * extra extent map ranges due to temporary memory allocation failures.
1857 */
1858 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1859 &BTRFS_I(inode)->runtime_flags);
1860
1861 /*
1862 * We have to do this here to avoid the priority inversion of waiting on
1863 * IO of a lower priority task while holding a transaction open.
1864 *
1865 * For a full fsync we wait for the ordered extents to complete while
1866 * for a fast fsync we wait just for writeback to complete, and then
1867 * attach the ordered extents to the transaction so that a transaction
1868 * commit waits for their completion, to avoid data loss if we fsync,
1869 * the current transaction commits before the ordered extents complete
1870 * and a power failure happens right after that.
1871 *
1872 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
1873 * logical address recorded in the ordered extent may change. We need
1874 * to wait for the IO to stabilize the logical address.
1875 */
1876 if (full_sync || btrfs_is_zoned(fs_info)) {
1877 ret = btrfs_wait_ordered_range(inode, start, len);
1878 } else {
1879 /*
1880 * Get our ordered extents as soon as possible to avoid doing
1881 * checksum lookups in the csum tree, and use instead the
1882 * checksums attached to the ordered extents.
1883 */
1884 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
1885 &ctx.ordered_extents);
1886 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
1887 }
1888
1889 if (ret)
1890 goto out_release_extents;
1891
1892 atomic_inc(&root->log_batch);
1893
1894 if (skip_inode_logging(&ctx)) {
1895 /*
1896 * We've had everything committed since the last time we were
1897 * modified so clear this flag in case it was set for whatever
1898 * reason, it's no longer relevant.
1899 */
1900 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1901 &BTRFS_I(inode)->runtime_flags);
1902 /*
1903 * An ordered extent might have started before and completed
1904 * already with io errors, in which case the inode was not
1905 * updated and we end up here. So check the inode's mapping
1906 * for any errors that might have happened since we last
1907 * checked called fsync.
1908 */
1909 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
1910 goto out_release_extents;
1911 }
1912
1913 btrfs_init_log_ctx_scratch_eb(&ctx);
1914
1915 /*
1916 * We use start here because we will need to wait on the IO to complete
1917 * in btrfs_sync_log, which could require joining a transaction (for
1918 * example checking cross references in the nocow path). If we use join
1919 * here we could get into a situation where we're waiting on IO to
1920 * happen that is blocked on a transaction trying to commit. With start
1921 * we inc the extwriter counter, so we wait for all extwriters to exit
1922 * before we start blocking joiners. This comment is to keep somebody
1923 * from thinking they are super smart and changing this to
1924 * btrfs_join_transaction *cough*Josef*cough*.
1925 */
1926 trans = btrfs_start_transaction(root, 0);
1927 if (IS_ERR(trans)) {
1928 ret = PTR_ERR(trans);
1929 goto out_release_extents;
1930 }
1931 trans->in_fsync = true;
1932
1933 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
1934 /*
1935 * Scratch eb no longer needed, release before syncing log or commit
1936 * transaction, to avoid holding unnecessary memory during such long
1937 * operations.
1938 */
1939 if (ctx.scratch_eb) {
1940 free_extent_buffer(ctx.scratch_eb);
1941 ctx.scratch_eb = NULL;
1942 }
1943 btrfs_release_log_ctx_extents(&ctx);
1944 if (ret < 0) {
1945 /* Fallthrough and commit/free transaction. */
1946 ret = BTRFS_LOG_FORCE_COMMIT;
1947 }
1948
1949 /* we've logged all the items and now have a consistent
1950 * version of the file in the log. It is possible that
1951 * someone will come in and modify the file, but that's
1952 * fine because the log is consistent on disk, and we
1953 * have references to all of the file's extents
1954 *
1955 * It is possible that someone will come in and log the
1956 * file again, but that will end up using the synchronization
1957 * inside btrfs_sync_log to keep things safe.
1958 */
1959 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
1960
1961 if (ret == BTRFS_NO_LOG_SYNC) {
1962 ret = btrfs_end_transaction(trans);
1963 goto out;
1964 }
1965
1966 /* We successfully logged the inode, attempt to sync the log. */
1967 if (!ret) {
1968 ret = btrfs_sync_log(trans, root, &ctx);
1969 if (!ret) {
1970 ret = btrfs_end_transaction(trans);
1971 goto out;
1972 }
1973 }
1974
1975 /*
1976 * At this point we need to commit the transaction because we had
1977 * btrfs_need_log_full_commit() or some other error.
1978 *
1979 * If we didn't do a full sync we have to stop the trans handle, wait on
1980 * the ordered extents, start it again and commit the transaction. If
1981 * we attempt to wait on the ordered extents here we could deadlock with
1982 * something like fallocate() that is holding the extent lock trying to
1983 * start a transaction while some other thread is trying to commit the
1984 * transaction while we (fsync) are currently holding the transaction
1985 * open.
1986 */
1987 if (!full_sync) {
1988 ret = btrfs_end_transaction(trans);
1989 if (ret)
1990 goto out;
1991 ret = btrfs_wait_ordered_range(inode, start, len);
1992 if (ret)
1993 goto out;
1994
1995 /*
1996 * This is safe to use here because we're only interested in
1997 * making sure the transaction that had the ordered extents is
1998 * committed. We aren't waiting on anything past this point,
1999 * we're purely getting the transaction and committing it.
2000 */
2001 trans = btrfs_attach_transaction_barrier(root);
2002 if (IS_ERR(trans)) {
2003 ret = PTR_ERR(trans);
2004
2005 /*
2006 * We committed the transaction and there's no currently
2007 * running transaction, this means everything we care
2008 * about made it to disk and we are done.
2009 */
2010 if (ret == -ENOENT)
2011 ret = 0;
2012 goto out;
2013 }
2014 }
2015
2016 ret = btrfs_commit_transaction(trans);
2017out:
2018 free_extent_buffer(ctx.scratch_eb);
2019 ASSERT(list_empty(&ctx.list));
2020 ASSERT(list_empty(&ctx.conflict_inodes));
2021 err = file_check_and_advance_wb_err(file);
2022 if (!ret)
2023 ret = err;
2024 return ret > 0 ? -EIO : ret;
2025
2026out_release_extents:
2027 btrfs_release_log_ctx_extents(&ctx);
2028 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2029 goto out;
2030}
2031
2032static const struct vm_operations_struct btrfs_file_vm_ops = {
2033 .fault = filemap_fault,
2034 .map_pages = filemap_map_pages,
2035 .page_mkwrite = btrfs_page_mkwrite,
2036};
2037
2038static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2039{
2040 struct address_space *mapping = filp->f_mapping;
2041
2042 if (!mapping->a_ops->read_folio)
2043 return -ENOEXEC;
2044
2045 file_accessed(filp);
2046 vma->vm_ops = &btrfs_file_vm_ops;
2047
2048 return 0;
2049}
2050
2051static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2052 int slot, u64 start, u64 end)
2053{
2054 struct btrfs_file_extent_item *fi;
2055 struct btrfs_key key;
2056
2057 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2058 return 0;
2059
2060 btrfs_item_key_to_cpu(leaf, &key, slot);
2061 if (key.objectid != btrfs_ino(inode) ||
2062 key.type != BTRFS_EXTENT_DATA_KEY)
2063 return 0;
2064
2065 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2066
2067 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2068 return 0;
2069
2070 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2071 return 0;
2072
2073 if (key.offset == end)
2074 return 1;
2075 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2076 return 1;
2077 return 0;
2078}
2079
2080static int fill_holes(struct btrfs_trans_handle *trans,
2081 struct btrfs_inode *inode,
2082 struct btrfs_path *path, u64 offset, u64 end)
2083{
2084 struct btrfs_fs_info *fs_info = trans->fs_info;
2085 struct btrfs_root *root = inode->root;
2086 struct extent_buffer *leaf;
2087 struct btrfs_file_extent_item *fi;
2088 struct extent_map *hole_em;
2089 struct btrfs_key key;
2090 int ret;
2091
2092 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2093 goto out;
2094
2095 key.objectid = btrfs_ino(inode);
2096 key.type = BTRFS_EXTENT_DATA_KEY;
2097 key.offset = offset;
2098
2099 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2100 if (ret <= 0) {
2101 /*
2102 * We should have dropped this offset, so if we find it then
2103 * something has gone horribly wrong.
2104 */
2105 if (ret == 0)
2106 ret = -EINVAL;
2107 return ret;
2108 }
2109
2110 leaf = path->nodes[0];
2111 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2112 u64 num_bytes;
2113
2114 path->slots[0]--;
2115 fi = btrfs_item_ptr(leaf, path->slots[0],
2116 struct btrfs_file_extent_item);
2117 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2118 end - offset;
2119 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2120 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2121 btrfs_set_file_extent_offset(leaf, fi, 0);
2122 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2123 btrfs_mark_buffer_dirty(trans, leaf);
2124 goto out;
2125 }
2126
2127 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2128 u64 num_bytes;
2129
2130 key.offset = offset;
2131 btrfs_set_item_key_safe(trans, path, &key);
2132 fi = btrfs_item_ptr(leaf, path->slots[0],
2133 struct btrfs_file_extent_item);
2134 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2135 offset;
2136 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2137 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2138 btrfs_set_file_extent_offset(leaf, fi, 0);
2139 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2140 btrfs_mark_buffer_dirty(trans, leaf);
2141 goto out;
2142 }
2143 btrfs_release_path(path);
2144
2145 ret = btrfs_insert_hole_extent(trans, root, btrfs_ino(inode), offset,
2146 end - offset);
2147 if (ret)
2148 return ret;
2149
2150out:
2151 btrfs_release_path(path);
2152
2153 hole_em = alloc_extent_map();
2154 if (!hole_em) {
2155 btrfs_drop_extent_map_range(inode, offset, end - 1, false);
2156 btrfs_set_inode_full_sync(inode);
2157 } else {
2158 hole_em->start = offset;
2159 hole_em->len = end - offset;
2160 hole_em->ram_bytes = hole_em->len;
2161 hole_em->orig_start = offset;
2162
2163 hole_em->block_start = EXTENT_MAP_HOLE;
2164 hole_em->block_len = 0;
2165 hole_em->orig_block_len = 0;
2166 hole_em->generation = trans->transid;
2167
2168 ret = btrfs_replace_extent_map_range(inode, hole_em, true);
2169 free_extent_map(hole_em);
2170 if (ret)
2171 btrfs_set_inode_full_sync(inode);
2172 }
2173
2174 return 0;
2175}
2176
2177/*
2178 * Find a hole extent on given inode and change start/len to the end of hole
2179 * extent.(hole/vacuum extent whose em->start <= start &&
2180 * em->start + em->len > start)
2181 * When a hole extent is found, return 1 and modify start/len.
2182 */
2183static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2184{
2185 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2186 struct extent_map *em;
2187 int ret = 0;
2188
2189 em = btrfs_get_extent(inode, NULL,
2190 round_down(*start, fs_info->sectorsize),
2191 round_up(*len, fs_info->sectorsize));
2192 if (IS_ERR(em))
2193 return PTR_ERR(em);
2194
2195 /* Hole or vacuum extent(only exists in no-hole mode) */
2196 if (em->block_start == EXTENT_MAP_HOLE) {
2197 ret = 1;
2198 *len = em->start + em->len > *start + *len ?
2199 0 : *start + *len - em->start - em->len;
2200 *start = em->start + em->len;
2201 }
2202 free_extent_map(em);
2203 return ret;
2204}
2205
2206static void btrfs_punch_hole_lock_range(struct inode *inode,
2207 const u64 lockstart,
2208 const u64 lockend,
2209 struct extent_state **cached_state)
2210{
2211 /*
2212 * For subpage case, if the range is not at page boundary, we could
2213 * have pages at the leading/tailing part of the range.
2214 * This could lead to dead loop since filemap_range_has_page()
2215 * will always return true.
2216 * So here we need to do extra page alignment for
2217 * filemap_range_has_page().
2218 */
2219 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2220 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2221
2222 while (1) {
2223 truncate_pagecache_range(inode, lockstart, lockend);
2224
2225 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2226 cached_state);
2227 /*
2228 * We can't have ordered extents in the range, nor dirty/writeback
2229 * pages, because we have locked the inode's VFS lock in exclusive
2230 * mode, we have locked the inode's i_mmap_lock in exclusive mode,
2231 * we have flushed all delalloc in the range and we have waited
2232 * for any ordered extents in the range to complete.
2233 * We can race with anyone reading pages from this range, so after
2234 * locking the range check if we have pages in the range, and if
2235 * we do, unlock the range and retry.
2236 */
2237 if (!filemap_range_has_page(inode->i_mapping, page_lockstart,
2238 page_lockend))
2239 break;
2240
2241 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2242 cached_state);
2243 }
2244
2245 btrfs_assert_inode_range_clean(BTRFS_I(inode), lockstart, lockend);
2246}
2247
2248static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2249 struct btrfs_inode *inode,
2250 struct btrfs_path *path,
2251 struct btrfs_replace_extent_info *extent_info,
2252 const u64 replace_len,
2253 const u64 bytes_to_drop)
2254{
2255 struct btrfs_fs_info *fs_info = trans->fs_info;
2256 struct btrfs_root *root = inode->root;
2257 struct btrfs_file_extent_item *extent;
2258 struct extent_buffer *leaf;
2259 struct btrfs_key key;
2260 int slot;
2261 struct btrfs_ref ref = { 0 };
2262 int ret;
2263
2264 if (replace_len == 0)
2265 return 0;
2266
2267 if (extent_info->disk_offset == 0 &&
2268 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2269 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2270 return 0;
2271 }
2272
2273 key.objectid = btrfs_ino(inode);
2274 key.type = BTRFS_EXTENT_DATA_KEY;
2275 key.offset = extent_info->file_offset;
2276 ret = btrfs_insert_empty_item(trans, root, path, &key,
2277 sizeof(struct btrfs_file_extent_item));
2278 if (ret)
2279 return ret;
2280 leaf = path->nodes[0];
2281 slot = path->slots[0];
2282 write_extent_buffer(leaf, extent_info->extent_buf,
2283 btrfs_item_ptr_offset(leaf, slot),
2284 sizeof(struct btrfs_file_extent_item));
2285 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2286 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2287 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2288 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2289 if (extent_info->is_new_extent)
2290 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2291 btrfs_mark_buffer_dirty(trans, leaf);
2292 btrfs_release_path(path);
2293
2294 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2295 replace_len);
2296 if (ret)
2297 return ret;
2298
2299 /* If it's a hole, nothing more needs to be done. */
2300 if (extent_info->disk_offset == 0) {
2301 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2302 return 0;
2303 }
2304
2305 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2306
2307 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2308 key.objectid = extent_info->disk_offset;
2309 key.type = BTRFS_EXTENT_ITEM_KEY;
2310 key.offset = extent_info->disk_len;
2311 ret = btrfs_alloc_reserved_file_extent(trans, root,
2312 btrfs_ino(inode),
2313 extent_info->file_offset,
2314 extent_info->qgroup_reserved,
2315 &key);
2316 } else {
2317 u64 ref_offset;
2318
2319 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2320 extent_info->disk_offset,
2321 extent_info->disk_len, 0,
2322 root->root_key.objectid);
2323 ref_offset = extent_info->file_offset - extent_info->data_offset;
2324 btrfs_init_data_ref(&ref, root->root_key.objectid,
2325 btrfs_ino(inode), ref_offset, 0, false);
2326 ret = btrfs_inc_extent_ref(trans, &ref);
2327 }
2328
2329 extent_info->insertions++;
2330
2331 return ret;
2332}
2333
2334/*
2335 * The respective range must have been previously locked, as well as the inode.
2336 * The end offset is inclusive (last byte of the range).
2337 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2338 * the file range with an extent.
2339 * When not punching a hole, we don't want to end up in a state where we dropped
2340 * extents without inserting a new one, so we must abort the transaction to avoid
2341 * a corruption.
2342 */
2343int btrfs_replace_file_extents(struct btrfs_inode *inode,
2344 struct btrfs_path *path, const u64 start,
2345 const u64 end,
2346 struct btrfs_replace_extent_info *extent_info,
2347 struct btrfs_trans_handle **trans_out)
2348{
2349 struct btrfs_drop_extents_args drop_args = { 0 };
2350 struct btrfs_root *root = inode->root;
2351 struct btrfs_fs_info *fs_info = root->fs_info;
2352 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2353 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2354 struct btrfs_trans_handle *trans = NULL;
2355 struct btrfs_block_rsv *rsv;
2356 unsigned int rsv_count;
2357 u64 cur_offset;
2358 u64 len = end - start;
2359 int ret = 0;
2360
2361 if (end <= start)
2362 return -EINVAL;
2363
2364 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2365 if (!rsv) {
2366 ret = -ENOMEM;
2367 goto out;
2368 }
2369 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2370 rsv->failfast = true;
2371
2372 /*
2373 * 1 - update the inode
2374 * 1 - removing the extents in the range
2375 * 1 - adding the hole extent if no_holes isn't set or if we are
2376 * replacing the range with a new extent
2377 */
2378 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2379 rsv_count = 3;
2380 else
2381 rsv_count = 2;
2382
2383 trans = btrfs_start_transaction(root, rsv_count);
2384 if (IS_ERR(trans)) {
2385 ret = PTR_ERR(trans);
2386 trans = NULL;
2387 goto out_free;
2388 }
2389
2390 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2391 min_size, false);
2392 if (WARN_ON(ret))
2393 goto out_trans;
2394 trans->block_rsv = rsv;
2395
2396 cur_offset = start;
2397 drop_args.path = path;
2398 drop_args.end = end + 1;
2399 drop_args.drop_cache = true;
2400 while (cur_offset < end) {
2401 drop_args.start = cur_offset;
2402 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2403 /* If we are punching a hole decrement the inode's byte count */
2404 if (!extent_info)
2405 btrfs_update_inode_bytes(inode, 0,
2406 drop_args.bytes_found);
2407 if (ret != -ENOSPC) {
2408 /*
2409 * The only time we don't want to abort is if we are
2410 * attempting to clone a partial inline extent, in which
2411 * case we'll get EOPNOTSUPP. However if we aren't
2412 * clone we need to abort no matter what, because if we
2413 * got EOPNOTSUPP via prealloc then we messed up and
2414 * need to abort.
2415 */
2416 if (ret &&
2417 (ret != -EOPNOTSUPP ||
2418 (extent_info && extent_info->is_new_extent)))
2419 btrfs_abort_transaction(trans, ret);
2420 break;
2421 }
2422
2423 trans->block_rsv = &fs_info->trans_block_rsv;
2424
2425 if (!extent_info && cur_offset < drop_args.drop_end &&
2426 cur_offset < ino_size) {
2427 ret = fill_holes(trans, inode, path, cur_offset,
2428 drop_args.drop_end);
2429 if (ret) {
2430 /*
2431 * If we failed then we didn't insert our hole
2432 * entries for the area we dropped, so now the
2433 * fs is corrupted, so we must abort the
2434 * transaction.
2435 */
2436 btrfs_abort_transaction(trans, ret);
2437 break;
2438 }
2439 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2440 /*
2441 * We are past the i_size here, but since we didn't
2442 * insert holes we need to clear the mapped area so we
2443 * know to not set disk_i_size in this area until a new
2444 * file extent is inserted here.
2445 */
2446 ret = btrfs_inode_clear_file_extent_range(inode,
2447 cur_offset,
2448 drop_args.drop_end - cur_offset);
2449 if (ret) {
2450 /*
2451 * We couldn't clear our area, so we could
2452 * presumably adjust up and corrupt the fs, so
2453 * we need to abort.
2454 */
2455 btrfs_abort_transaction(trans, ret);
2456 break;
2457 }
2458 }
2459
2460 if (extent_info &&
2461 drop_args.drop_end > extent_info->file_offset) {
2462 u64 replace_len = drop_args.drop_end -
2463 extent_info->file_offset;
2464
2465 ret = btrfs_insert_replace_extent(trans, inode, path,
2466 extent_info, replace_len,
2467 drop_args.bytes_found);
2468 if (ret) {
2469 btrfs_abort_transaction(trans, ret);
2470 break;
2471 }
2472 extent_info->data_len -= replace_len;
2473 extent_info->data_offset += replace_len;
2474 extent_info->file_offset += replace_len;
2475 }
2476
2477 /*
2478 * We are releasing our handle on the transaction, balance the
2479 * dirty pages of the btree inode and flush delayed items, and
2480 * then get a new transaction handle, which may now point to a
2481 * new transaction in case someone else may have committed the
2482 * transaction we used to replace/drop file extent items. So
2483 * bump the inode's iversion and update mtime and ctime except
2484 * if we are called from a dedupe context. This is because a
2485 * power failure/crash may happen after the transaction is
2486 * committed and before we finish replacing/dropping all the
2487 * file extent items we need.
2488 */
2489 inode_inc_iversion(&inode->vfs_inode);
2490
2491 if (!extent_info || extent_info->update_times)
2492 inode_set_mtime_to_ts(&inode->vfs_inode,
2493 inode_set_ctime_current(&inode->vfs_inode));
2494
2495 ret = btrfs_update_inode(trans, inode);
2496 if (ret)
2497 break;
2498
2499 btrfs_end_transaction(trans);
2500 btrfs_btree_balance_dirty(fs_info);
2501
2502 trans = btrfs_start_transaction(root, rsv_count);
2503 if (IS_ERR(trans)) {
2504 ret = PTR_ERR(trans);
2505 trans = NULL;
2506 break;
2507 }
2508
2509 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2510 rsv, min_size, false);
2511 if (WARN_ON(ret))
2512 break;
2513 trans->block_rsv = rsv;
2514
2515 cur_offset = drop_args.drop_end;
2516 len = end - cur_offset;
2517 if (!extent_info && len) {
2518 ret = find_first_non_hole(inode, &cur_offset, &len);
2519 if (unlikely(ret < 0))
2520 break;
2521 if (ret && !len) {
2522 ret = 0;
2523 break;
2524 }
2525 }
2526 }
2527
2528 /*
2529 * If we were cloning, force the next fsync to be a full one since we
2530 * we replaced (or just dropped in the case of cloning holes when
2531 * NO_HOLES is enabled) file extent items and did not setup new extent
2532 * maps for the replacement extents (or holes).
2533 */
2534 if (extent_info && !extent_info->is_new_extent)
2535 btrfs_set_inode_full_sync(inode);
2536
2537 if (ret)
2538 goto out_trans;
2539
2540 trans->block_rsv = &fs_info->trans_block_rsv;
2541 /*
2542 * If we are using the NO_HOLES feature we might have had already an
2543 * hole that overlaps a part of the region [lockstart, lockend] and
2544 * ends at (or beyond) lockend. Since we have no file extent items to
2545 * represent holes, drop_end can be less than lockend and so we must
2546 * make sure we have an extent map representing the existing hole (the
2547 * call to __btrfs_drop_extents() might have dropped the existing extent
2548 * map representing the existing hole), otherwise the fast fsync path
2549 * will not record the existence of the hole region
2550 * [existing_hole_start, lockend].
2551 */
2552 if (drop_args.drop_end <= end)
2553 drop_args.drop_end = end + 1;
2554 /*
2555 * Don't insert file hole extent item if it's for a range beyond eof
2556 * (because it's useless) or if it represents a 0 bytes range (when
2557 * cur_offset == drop_end).
2558 */
2559 if (!extent_info && cur_offset < ino_size &&
2560 cur_offset < drop_args.drop_end) {
2561 ret = fill_holes(trans, inode, path, cur_offset,
2562 drop_args.drop_end);
2563 if (ret) {
2564 /* Same comment as above. */
2565 btrfs_abort_transaction(trans, ret);
2566 goto out_trans;
2567 }
2568 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2569 /* See the comment in the loop above for the reasoning here. */
2570 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2571 drop_args.drop_end - cur_offset);
2572 if (ret) {
2573 btrfs_abort_transaction(trans, ret);
2574 goto out_trans;
2575 }
2576
2577 }
2578 if (extent_info) {
2579 ret = btrfs_insert_replace_extent(trans, inode, path,
2580 extent_info, extent_info->data_len,
2581 drop_args.bytes_found);
2582 if (ret) {
2583 btrfs_abort_transaction(trans, ret);
2584 goto out_trans;
2585 }
2586 }
2587
2588out_trans:
2589 if (!trans)
2590 goto out_free;
2591
2592 trans->block_rsv = &fs_info->trans_block_rsv;
2593 if (ret)
2594 btrfs_end_transaction(trans);
2595 else
2596 *trans_out = trans;
2597out_free:
2598 btrfs_free_block_rsv(fs_info, rsv);
2599out:
2600 return ret;
2601}
2602
2603static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2604{
2605 struct inode *inode = file_inode(file);
2606 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2607 struct btrfs_root *root = BTRFS_I(inode)->root;
2608 struct extent_state *cached_state = NULL;
2609 struct btrfs_path *path;
2610 struct btrfs_trans_handle *trans = NULL;
2611 u64 lockstart;
2612 u64 lockend;
2613 u64 tail_start;
2614 u64 tail_len;
2615 u64 orig_start = offset;
2616 int ret = 0;
2617 bool same_block;
2618 u64 ino_size;
2619 bool truncated_block = false;
2620 bool updated_inode = false;
2621
2622 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2623
2624 ret = btrfs_wait_ordered_range(inode, offset, len);
2625 if (ret)
2626 goto out_only_mutex;
2627
2628 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2629 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2630 if (ret < 0)
2631 goto out_only_mutex;
2632 if (ret && !len) {
2633 /* Already in a large hole */
2634 ret = 0;
2635 goto out_only_mutex;
2636 }
2637
2638 ret = file_modified(file);
2639 if (ret)
2640 goto out_only_mutex;
2641
2642 lockstart = round_up(offset, fs_info->sectorsize);
2643 lockend = round_down(offset + len, fs_info->sectorsize) - 1;
2644 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2645 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2646 /*
2647 * We needn't truncate any block which is beyond the end of the file
2648 * because we are sure there is no data there.
2649 */
2650 /*
2651 * Only do this if we are in the same block and we aren't doing the
2652 * entire block.
2653 */
2654 if (same_block && len < fs_info->sectorsize) {
2655 if (offset < ino_size) {
2656 truncated_block = true;
2657 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2658 0);
2659 } else {
2660 ret = 0;
2661 }
2662 goto out_only_mutex;
2663 }
2664
2665 /* zero back part of the first block */
2666 if (offset < ino_size) {
2667 truncated_block = true;
2668 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2669 if (ret) {
2670 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2671 return ret;
2672 }
2673 }
2674
2675 /* Check the aligned pages after the first unaligned page,
2676 * if offset != orig_start, which means the first unaligned page
2677 * including several following pages are already in holes,
2678 * the extra check can be skipped */
2679 if (offset == orig_start) {
2680 /* after truncate page, check hole again */
2681 len = offset + len - lockstart;
2682 offset = lockstart;
2683 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2684 if (ret < 0)
2685 goto out_only_mutex;
2686 if (ret && !len) {
2687 ret = 0;
2688 goto out_only_mutex;
2689 }
2690 lockstart = offset;
2691 }
2692
2693 /* Check the tail unaligned part is in a hole */
2694 tail_start = lockend + 1;
2695 tail_len = offset + len - tail_start;
2696 if (tail_len) {
2697 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2698 if (unlikely(ret < 0))
2699 goto out_only_mutex;
2700 if (!ret) {
2701 /* zero the front end of the last page */
2702 if (tail_start + tail_len < ino_size) {
2703 truncated_block = true;
2704 ret = btrfs_truncate_block(BTRFS_I(inode),
2705 tail_start + tail_len,
2706 0, 1);
2707 if (ret)
2708 goto out_only_mutex;
2709 }
2710 }
2711 }
2712
2713 if (lockend < lockstart) {
2714 ret = 0;
2715 goto out_only_mutex;
2716 }
2717
2718 btrfs_punch_hole_lock_range(inode, lockstart, lockend, &cached_state);
2719
2720 path = btrfs_alloc_path();
2721 if (!path) {
2722 ret = -ENOMEM;
2723 goto out;
2724 }
2725
2726 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
2727 lockend, NULL, &trans);
2728 btrfs_free_path(path);
2729 if (ret)
2730 goto out;
2731
2732 ASSERT(trans != NULL);
2733 inode_inc_iversion(inode);
2734 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
2735 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2736 updated_inode = true;
2737 btrfs_end_transaction(trans);
2738 btrfs_btree_balance_dirty(fs_info);
2739out:
2740 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2741 &cached_state);
2742out_only_mutex:
2743 if (!updated_inode && truncated_block && !ret) {
2744 /*
2745 * If we only end up zeroing part of a page, we still need to
2746 * update the inode item, so that all the time fields are
2747 * updated as well as the necessary btrfs inode in memory fields
2748 * for detecting, at fsync time, if the inode isn't yet in the
2749 * log tree or it's there but not up to date.
2750 */
2751 struct timespec64 now = inode_set_ctime_current(inode);
2752
2753 inode_inc_iversion(inode);
2754 inode_set_mtime_to_ts(inode, now);
2755 trans = btrfs_start_transaction(root, 1);
2756 if (IS_ERR(trans)) {
2757 ret = PTR_ERR(trans);
2758 } else {
2759 int ret2;
2760
2761 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2762 ret2 = btrfs_end_transaction(trans);
2763 if (!ret)
2764 ret = ret2;
2765 }
2766 }
2767 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2768 return ret;
2769}
2770
2771/* Helper structure to record which range is already reserved */
2772struct falloc_range {
2773 struct list_head list;
2774 u64 start;
2775 u64 len;
2776};
2777
2778/*
2779 * Helper function to add falloc range
2780 *
2781 * Caller should have locked the larger range of extent containing
2782 * [start, len)
2783 */
2784static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2785{
2786 struct falloc_range *range = NULL;
2787
2788 if (!list_empty(head)) {
2789 /*
2790 * As fallocate iterates by bytenr order, we only need to check
2791 * the last range.
2792 */
2793 range = list_last_entry(head, struct falloc_range, list);
2794 if (range->start + range->len == start) {
2795 range->len += len;
2796 return 0;
2797 }
2798 }
2799
2800 range = kmalloc(sizeof(*range), GFP_KERNEL);
2801 if (!range)
2802 return -ENOMEM;
2803 range->start = start;
2804 range->len = len;
2805 list_add_tail(&range->list, head);
2806 return 0;
2807}
2808
2809static int btrfs_fallocate_update_isize(struct inode *inode,
2810 const u64 end,
2811 const int mode)
2812{
2813 struct btrfs_trans_handle *trans;
2814 struct btrfs_root *root = BTRFS_I(inode)->root;
2815 int ret;
2816 int ret2;
2817
2818 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2819 return 0;
2820
2821 trans = btrfs_start_transaction(root, 1);
2822 if (IS_ERR(trans))
2823 return PTR_ERR(trans);
2824
2825 inode_set_ctime_current(inode);
2826 i_size_write(inode, end);
2827 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
2828 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2829 ret2 = btrfs_end_transaction(trans);
2830
2831 return ret ? ret : ret2;
2832}
2833
2834enum {
2835 RANGE_BOUNDARY_WRITTEN_EXTENT,
2836 RANGE_BOUNDARY_PREALLOC_EXTENT,
2837 RANGE_BOUNDARY_HOLE,
2838};
2839
2840static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
2841 u64 offset)
2842{
2843 const u64 sectorsize = inode->root->fs_info->sectorsize;
2844 struct extent_map *em;
2845 int ret;
2846
2847 offset = round_down(offset, sectorsize);
2848 em = btrfs_get_extent(inode, NULL, offset, sectorsize);
2849 if (IS_ERR(em))
2850 return PTR_ERR(em);
2851
2852 if (em->block_start == EXTENT_MAP_HOLE)
2853 ret = RANGE_BOUNDARY_HOLE;
2854 else if (em->flags & EXTENT_FLAG_PREALLOC)
2855 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2856 else
2857 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2858
2859 free_extent_map(em);
2860 return ret;
2861}
2862
2863static int btrfs_zero_range(struct inode *inode,
2864 loff_t offset,
2865 loff_t len,
2866 const int mode)
2867{
2868 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2869 struct extent_map *em;
2870 struct extent_changeset *data_reserved = NULL;
2871 int ret;
2872 u64 alloc_hint = 0;
2873 const u64 sectorsize = fs_info->sectorsize;
2874 u64 alloc_start = round_down(offset, sectorsize);
2875 u64 alloc_end = round_up(offset + len, sectorsize);
2876 u64 bytes_to_reserve = 0;
2877 bool space_reserved = false;
2878
2879 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start,
2880 alloc_end - alloc_start);
2881 if (IS_ERR(em)) {
2882 ret = PTR_ERR(em);
2883 goto out;
2884 }
2885
2886 /*
2887 * Avoid hole punching and extent allocation for some cases. More cases
2888 * could be considered, but these are unlikely common and we keep things
2889 * as simple as possible for now. Also, intentionally, if the target
2890 * range contains one or more prealloc extents together with regular
2891 * extents and holes, we drop all the existing extents and allocate a
2892 * new prealloc extent, so that we get a larger contiguous disk extent.
2893 */
2894 if (em->start <= alloc_start && (em->flags & EXTENT_FLAG_PREALLOC)) {
2895 const u64 em_end = em->start + em->len;
2896
2897 if (em_end >= offset + len) {
2898 /*
2899 * The whole range is already a prealloc extent,
2900 * do nothing except updating the inode's i_size if
2901 * needed.
2902 */
2903 free_extent_map(em);
2904 ret = btrfs_fallocate_update_isize(inode, offset + len,
2905 mode);
2906 goto out;
2907 }
2908 /*
2909 * Part of the range is already a prealloc extent, so operate
2910 * only on the remaining part of the range.
2911 */
2912 alloc_start = em_end;
2913 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2914 len = offset + len - alloc_start;
2915 offset = alloc_start;
2916 alloc_hint = em->block_start + em->len;
2917 }
2918 free_extent_map(em);
2919
2920 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2921 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2922 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start, sectorsize);
2923 if (IS_ERR(em)) {
2924 ret = PTR_ERR(em);
2925 goto out;
2926 }
2927
2928 if (em->flags & EXTENT_FLAG_PREALLOC) {
2929 free_extent_map(em);
2930 ret = btrfs_fallocate_update_isize(inode, offset + len,
2931 mode);
2932 goto out;
2933 }
2934 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2935 free_extent_map(em);
2936 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2937 0);
2938 if (!ret)
2939 ret = btrfs_fallocate_update_isize(inode,
2940 offset + len,
2941 mode);
2942 return ret;
2943 }
2944 free_extent_map(em);
2945 alloc_start = round_down(offset, sectorsize);
2946 alloc_end = alloc_start + sectorsize;
2947 goto reserve_space;
2948 }
2949
2950 alloc_start = round_up(offset, sectorsize);
2951 alloc_end = round_down(offset + len, sectorsize);
2952
2953 /*
2954 * For unaligned ranges, check the pages at the boundaries, they might
2955 * map to an extent, in which case we need to partially zero them, or
2956 * they might map to a hole, in which case we need our allocation range
2957 * to cover them.
2958 */
2959 if (!IS_ALIGNED(offset, sectorsize)) {
2960 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2961 offset);
2962 if (ret < 0)
2963 goto out;
2964 if (ret == RANGE_BOUNDARY_HOLE) {
2965 alloc_start = round_down(offset, sectorsize);
2966 ret = 0;
2967 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2968 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2969 if (ret)
2970 goto out;
2971 } else {
2972 ret = 0;
2973 }
2974 }
2975
2976 if (!IS_ALIGNED(offset + len, sectorsize)) {
2977 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2978 offset + len);
2979 if (ret < 0)
2980 goto out;
2981 if (ret == RANGE_BOUNDARY_HOLE) {
2982 alloc_end = round_up(offset + len, sectorsize);
2983 ret = 0;
2984 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2985 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
2986 0, 1);
2987 if (ret)
2988 goto out;
2989 } else {
2990 ret = 0;
2991 }
2992 }
2993
2994reserve_space:
2995 if (alloc_start < alloc_end) {
2996 struct extent_state *cached_state = NULL;
2997 const u64 lockstart = alloc_start;
2998 const u64 lockend = alloc_end - 1;
2999
3000 bytes_to_reserve = alloc_end - alloc_start;
3001 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3002 bytes_to_reserve);
3003 if (ret < 0)
3004 goto out;
3005 space_reserved = true;
3006 btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3007 &cached_state);
3008 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3009 alloc_start, bytes_to_reserve);
3010 if (ret) {
3011 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
3012 lockend, &cached_state);
3013 goto out;
3014 }
3015 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3016 alloc_end - alloc_start,
3017 fs_info->sectorsize,
3018 offset + len, &alloc_hint);
3019 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3020 &cached_state);
3021 /* btrfs_prealloc_file_range releases reserved space on error */
3022 if (ret) {
3023 space_reserved = false;
3024 goto out;
3025 }
3026 }
3027 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3028 out:
3029 if (ret && space_reserved)
3030 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3031 alloc_start, bytes_to_reserve);
3032 extent_changeset_free(data_reserved);
3033
3034 return ret;
3035}
3036
3037static long btrfs_fallocate(struct file *file, int mode,
3038 loff_t offset, loff_t len)
3039{
3040 struct inode *inode = file_inode(file);
3041 struct extent_state *cached_state = NULL;
3042 struct extent_changeset *data_reserved = NULL;
3043 struct falloc_range *range;
3044 struct falloc_range *tmp;
3045 LIST_HEAD(reserve_list);
3046 u64 cur_offset;
3047 u64 last_byte;
3048 u64 alloc_start;
3049 u64 alloc_end;
3050 u64 alloc_hint = 0;
3051 u64 locked_end;
3052 u64 actual_end = 0;
3053 u64 data_space_needed = 0;
3054 u64 data_space_reserved = 0;
3055 u64 qgroup_reserved = 0;
3056 struct extent_map *em;
3057 int blocksize = BTRFS_I(inode)->root->fs_info->sectorsize;
3058 int ret;
3059
3060 /* Do not allow fallocate in ZONED mode */
3061 if (btrfs_is_zoned(inode_to_fs_info(inode)))
3062 return -EOPNOTSUPP;
3063
3064 alloc_start = round_down(offset, blocksize);
3065 alloc_end = round_up(offset + len, blocksize);
3066 cur_offset = alloc_start;
3067
3068 /* Make sure we aren't being give some crap mode */
3069 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3070 FALLOC_FL_ZERO_RANGE))
3071 return -EOPNOTSUPP;
3072
3073 if (mode & FALLOC_FL_PUNCH_HOLE)
3074 return btrfs_punch_hole(file, offset, len);
3075
3076 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3077
3078 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3079 ret = inode_newsize_ok(inode, offset + len);
3080 if (ret)
3081 goto out;
3082 }
3083
3084 ret = file_modified(file);
3085 if (ret)
3086 goto out;
3087
3088 /*
3089 * TODO: Move these two operations after we have checked
3090 * accurate reserved space, or fallocate can still fail but
3091 * with page truncated or size expanded.
3092 *
3093 * But that's a minor problem and won't do much harm BTW.
3094 */
3095 if (alloc_start > inode->i_size) {
3096 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3097 alloc_start);
3098 if (ret)
3099 goto out;
3100 } else if (offset + len > inode->i_size) {
3101 /*
3102 * If we are fallocating from the end of the file onward we
3103 * need to zero out the end of the block if i_size lands in the
3104 * middle of a block.
3105 */
3106 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3107 if (ret)
3108 goto out;
3109 }
3110
3111 /*
3112 * We have locked the inode at the VFS level (in exclusive mode) and we
3113 * have locked the i_mmap_lock lock (in exclusive mode). Now before
3114 * locking the file range, flush all dealloc in the range and wait for
3115 * all ordered extents in the range to complete. After this we can lock
3116 * the file range and, due to the previous locking we did, we know there
3117 * can't be more delalloc or ordered extents in the range.
3118 */
3119 ret = btrfs_wait_ordered_range(inode, alloc_start,
3120 alloc_end - alloc_start);
3121 if (ret)
3122 goto out;
3123
3124 if (mode & FALLOC_FL_ZERO_RANGE) {
3125 ret = btrfs_zero_range(inode, offset, len, mode);
3126 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3127 return ret;
3128 }
3129
3130 locked_end = alloc_end - 1;
3131 lock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3132 &cached_state);
3133
3134 btrfs_assert_inode_range_clean(BTRFS_I(inode), alloc_start, locked_end);
3135
3136 /* First, check if we exceed the qgroup limit */
3137 while (cur_offset < alloc_end) {
3138 em = btrfs_get_extent(BTRFS_I(inode), NULL, cur_offset,
3139 alloc_end - cur_offset);
3140 if (IS_ERR(em)) {
3141 ret = PTR_ERR(em);
3142 break;
3143 }
3144 last_byte = min(extent_map_end(em), alloc_end);
3145 actual_end = min_t(u64, extent_map_end(em), offset + len);
3146 last_byte = ALIGN(last_byte, blocksize);
3147 if (em->block_start == EXTENT_MAP_HOLE ||
3148 (cur_offset >= inode->i_size &&
3149 !(em->flags & EXTENT_FLAG_PREALLOC))) {
3150 const u64 range_len = last_byte - cur_offset;
3151
3152 ret = add_falloc_range(&reserve_list, cur_offset, range_len);
3153 if (ret < 0) {
3154 free_extent_map(em);
3155 break;
3156 }
3157 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3158 &data_reserved, cur_offset, range_len);
3159 if (ret < 0) {
3160 free_extent_map(em);
3161 break;
3162 }
3163 qgroup_reserved += range_len;
3164 data_space_needed += range_len;
3165 }
3166 free_extent_map(em);
3167 cur_offset = last_byte;
3168 }
3169
3170 if (!ret && data_space_needed > 0) {
3171 /*
3172 * We are safe to reserve space here as we can't have delalloc
3173 * in the range, see above.
3174 */
3175 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3176 data_space_needed);
3177 if (!ret)
3178 data_space_reserved = data_space_needed;
3179 }
3180
3181 /*
3182 * If ret is still 0, means we're OK to fallocate.
3183 * Or just cleanup the list and exit.
3184 */
3185 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3186 if (!ret) {
3187 ret = btrfs_prealloc_file_range(inode, mode,
3188 range->start,
3189 range->len, blocksize,
3190 offset + len, &alloc_hint);
3191 /*
3192 * btrfs_prealloc_file_range() releases space even
3193 * if it returns an error.
3194 */
3195 data_space_reserved -= range->len;
3196 qgroup_reserved -= range->len;
3197 } else if (data_space_reserved > 0) {
3198 btrfs_free_reserved_data_space(BTRFS_I(inode),
3199 data_reserved, range->start,
3200 range->len);
3201 data_space_reserved -= range->len;
3202 qgroup_reserved -= range->len;
3203 } else if (qgroup_reserved > 0) {
3204 btrfs_qgroup_free_data(BTRFS_I(inode), data_reserved,
3205 range->start, range->len, NULL);
3206 qgroup_reserved -= range->len;
3207 }
3208 list_del(&range->list);
3209 kfree(range);
3210 }
3211 if (ret < 0)
3212 goto out_unlock;
3213
3214 /*
3215 * We didn't need to allocate any more space, but we still extended the
3216 * size of the file so we need to update i_size and the inode item.
3217 */
3218 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3219out_unlock:
3220 unlock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3221 &cached_state);
3222out:
3223 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3224 extent_changeset_free(data_reserved);
3225 return ret;
3226}
3227
3228/*
3229 * Helper for btrfs_find_delalloc_in_range(). Find a subrange in a given range
3230 * that has unflushed and/or flushing delalloc. There might be other adjacent
3231 * subranges after the one it found, so btrfs_find_delalloc_in_range() keeps
3232 * looping while it gets adjacent subranges, and merging them together.
3233 */
3234static bool find_delalloc_subrange(struct btrfs_inode *inode, u64 start, u64 end,
3235 struct extent_state **cached_state,
3236 bool *search_io_tree,
3237 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3238{
3239 u64 len = end + 1 - start;
3240 u64 delalloc_len = 0;
3241 struct btrfs_ordered_extent *oe;
3242 u64 oe_start;
3243 u64 oe_end;
3244
3245 /*
3246 * Search the io tree first for EXTENT_DELALLOC. If we find any, it
3247 * means we have delalloc (dirty pages) for which writeback has not
3248 * started yet.
3249 */
3250 if (*search_io_tree) {
3251 spin_lock(&inode->lock);
3252 if (inode->delalloc_bytes > 0) {
3253 spin_unlock(&inode->lock);
3254 *delalloc_start_ret = start;
3255 delalloc_len = count_range_bits(&inode->io_tree,
3256 delalloc_start_ret, end,
3257 len, EXTENT_DELALLOC, 1,
3258 cached_state);
3259 } else {
3260 spin_unlock(&inode->lock);
3261 }
3262 }
3263
3264 if (delalloc_len > 0) {
3265 /*
3266 * If delalloc was found then *delalloc_start_ret has a sector size
3267 * aligned value (rounded down).
3268 */
3269 *delalloc_end_ret = *delalloc_start_ret + delalloc_len - 1;
3270
3271 if (*delalloc_start_ret == start) {
3272 /* Delalloc for the whole range, nothing more to do. */
3273 if (*delalloc_end_ret == end)
3274 return true;
3275 /* Else trim our search range for ordered extents. */
3276 start = *delalloc_end_ret + 1;
3277 len = end + 1 - start;
3278 }
3279 } else {
3280 /* No delalloc, future calls don't need to search again. */
3281 *search_io_tree = false;
3282 }
3283
3284 /*
3285 * Now also check if there's any ordered extent in the range.
3286 * We do this because:
3287 *
3288 * 1) When delalloc is flushed, the file range is locked, we clear the
3289 * EXTENT_DELALLOC bit from the io tree and create an extent map and
3290 * an ordered extent for the write. So we might just have been called
3291 * after delalloc is flushed and before the ordered extent completes
3292 * and inserts the new file extent item in the subvolume's btree;
3293 *
3294 * 2) We may have an ordered extent created by flushing delalloc for a
3295 * subrange that starts before the subrange we found marked with
3296 * EXTENT_DELALLOC in the io tree.
3297 *
3298 * We could also use the extent map tree to find such delalloc that is
3299 * being flushed, but using the ordered extents tree is more efficient
3300 * because it's usually much smaller as ordered extents are removed from
3301 * the tree once they complete. With the extent maps, we mau have them
3302 * in the extent map tree for a very long time, and they were either
3303 * created by previous writes or loaded by read operations.
3304 */
3305 oe = btrfs_lookup_first_ordered_range(inode, start, len);
3306 if (!oe)
3307 return (delalloc_len > 0);
3308
3309 /* The ordered extent may span beyond our search range. */
3310 oe_start = max(oe->file_offset, start);
3311 oe_end = min(oe->file_offset + oe->num_bytes - 1, end);
3312
3313 btrfs_put_ordered_extent(oe);
3314
3315 /* Don't have unflushed delalloc, return the ordered extent range. */
3316 if (delalloc_len == 0) {
3317 *delalloc_start_ret = oe_start;
3318 *delalloc_end_ret = oe_end;
3319 return true;
3320 }
3321
3322 /*
3323 * We have both unflushed delalloc (io_tree) and an ordered extent.
3324 * If the ranges are adjacent returned a combined range, otherwise
3325 * return the leftmost range.
3326 */
3327 if (oe_start < *delalloc_start_ret) {
3328 if (oe_end < *delalloc_start_ret)
3329 *delalloc_end_ret = oe_end;
3330 *delalloc_start_ret = oe_start;
3331 } else if (*delalloc_end_ret + 1 == oe_start) {
3332 *delalloc_end_ret = oe_end;
3333 }
3334
3335 return true;
3336}
3337
3338/*
3339 * Check if there's delalloc in a given range.
3340 *
3341 * @inode: The inode.
3342 * @start: The start offset of the range. It does not need to be
3343 * sector size aligned.
3344 * @end: The end offset (inclusive value) of the search range.
3345 * It does not need to be sector size aligned.
3346 * @cached_state: Extent state record used for speeding up delalloc
3347 * searches in the inode's io_tree. Can be NULL.
3348 * @delalloc_start_ret: Output argument, set to the start offset of the
3349 * subrange found with delalloc (may not be sector size
3350 * aligned).
3351 * @delalloc_end_ret: Output argument, set to he end offset (inclusive value)
3352 * of the subrange found with delalloc.
3353 *
3354 * Returns true if a subrange with delalloc is found within the given range, and
3355 * if so it sets @delalloc_start_ret and @delalloc_end_ret with the start and
3356 * end offsets of the subrange.
3357 */
3358bool btrfs_find_delalloc_in_range(struct btrfs_inode *inode, u64 start, u64 end,
3359 struct extent_state **cached_state,
3360 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3361{
3362 u64 cur_offset = round_down(start, inode->root->fs_info->sectorsize);
3363 u64 prev_delalloc_end = 0;
3364 bool search_io_tree = true;
3365 bool ret = false;
3366
3367 while (cur_offset <= end) {
3368 u64 delalloc_start;
3369 u64 delalloc_end;
3370 bool delalloc;
3371
3372 delalloc = find_delalloc_subrange(inode, cur_offset, end,
3373 cached_state, &search_io_tree,
3374 &delalloc_start,
3375 &delalloc_end);
3376 if (!delalloc)
3377 break;
3378
3379 if (prev_delalloc_end == 0) {
3380 /* First subrange found. */
3381 *delalloc_start_ret = max(delalloc_start, start);
3382 *delalloc_end_ret = delalloc_end;
3383 ret = true;
3384 } else if (delalloc_start == prev_delalloc_end + 1) {
3385 /* Subrange adjacent to the previous one, merge them. */
3386 *delalloc_end_ret = delalloc_end;
3387 } else {
3388 /* Subrange not adjacent to the previous one, exit. */
3389 break;
3390 }
3391
3392 prev_delalloc_end = delalloc_end;
3393 cur_offset = delalloc_end + 1;
3394 cond_resched();
3395 }
3396
3397 return ret;
3398}
3399
3400/*
3401 * Check if there's a hole or delalloc range in a range representing a hole (or
3402 * prealloc extent) found in the inode's subvolume btree.
3403 *
3404 * @inode: The inode.
3405 * @whence: Seek mode (SEEK_DATA or SEEK_HOLE).
3406 * @start: Start offset of the hole region. It does not need to be sector
3407 * size aligned.
3408 * @end: End offset (inclusive value) of the hole region. It does not
3409 * need to be sector size aligned.
3410 * @start_ret: Return parameter, used to set the start of the subrange in the
3411 * hole that matches the search criteria (seek mode), if such
3412 * subrange is found (return value of the function is true).
3413 * The value returned here may not be sector size aligned.
3414 *
3415 * Returns true if a subrange matching the given seek mode is found, and if one
3416 * is found, it updates @start_ret with the start of the subrange.
3417 */
3418static bool find_desired_extent_in_hole(struct btrfs_inode *inode, int whence,
3419 struct extent_state **cached_state,
3420 u64 start, u64 end, u64 *start_ret)
3421{
3422 u64 delalloc_start;
3423 u64 delalloc_end;
3424 bool delalloc;
3425
3426 delalloc = btrfs_find_delalloc_in_range(inode, start, end, cached_state,
3427 &delalloc_start, &delalloc_end);
3428 if (delalloc && whence == SEEK_DATA) {
3429 *start_ret = delalloc_start;
3430 return true;
3431 }
3432
3433 if (delalloc && whence == SEEK_HOLE) {
3434 /*
3435 * We found delalloc but it starts after out start offset. So we
3436 * have a hole between our start offset and the delalloc start.
3437 */
3438 if (start < delalloc_start) {
3439 *start_ret = start;
3440 return true;
3441 }
3442 /*
3443 * Delalloc range starts at our start offset.
3444 * If the delalloc range's length is smaller than our range,
3445 * then it means we have a hole that starts where the delalloc
3446 * subrange ends.
3447 */
3448 if (delalloc_end < end) {
3449 *start_ret = delalloc_end + 1;
3450 return true;
3451 }
3452
3453 /* There's delalloc for the whole range. */
3454 return false;
3455 }
3456
3457 if (!delalloc && whence == SEEK_HOLE) {
3458 *start_ret = start;
3459 return true;
3460 }
3461
3462 /*
3463 * No delalloc in the range and we are seeking for data. The caller has
3464 * to iterate to the next extent item in the subvolume btree.
3465 */
3466 return false;
3467}
3468
3469static loff_t find_desired_extent(struct file *file, loff_t offset, int whence)
3470{
3471 struct btrfs_inode *inode = BTRFS_I(file->f_mapping->host);
3472 struct btrfs_file_private *private = file->private_data;
3473 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3474 struct extent_state *cached_state = NULL;
3475 struct extent_state **delalloc_cached_state;
3476 const loff_t i_size = i_size_read(&inode->vfs_inode);
3477 const u64 ino = btrfs_ino(inode);
3478 struct btrfs_root *root = inode->root;
3479 struct btrfs_path *path;
3480 struct btrfs_key key;
3481 u64 last_extent_end;
3482 u64 lockstart;
3483 u64 lockend;
3484 u64 start;
3485 int ret;
3486 bool found = false;
3487
3488 if (i_size == 0 || offset >= i_size)
3489 return -ENXIO;
3490
3491 /*
3492 * Quick path. If the inode has no prealloc extents and its number of
3493 * bytes used matches its i_size, then it can not have holes.
3494 */
3495 if (whence == SEEK_HOLE &&
3496 !(inode->flags & BTRFS_INODE_PREALLOC) &&
3497 inode_get_bytes(&inode->vfs_inode) == i_size)
3498 return i_size;
3499
3500 if (!private) {
3501 private = kzalloc(sizeof(*private), GFP_KERNEL);
3502 /*
3503 * No worries if memory allocation failed.
3504 * The private structure is used only for speeding up multiple
3505 * lseek SEEK_HOLE/DATA calls to a file when there's delalloc,
3506 * so everything will still be correct.
3507 */
3508 file->private_data = private;
3509 }
3510
3511 if (private)
3512 delalloc_cached_state = &private->llseek_cached_state;
3513 else
3514 delalloc_cached_state = NULL;
3515
3516 /*
3517 * offset can be negative, in this case we start finding DATA/HOLE from
3518 * the very start of the file.
3519 */
3520 start = max_t(loff_t, 0, offset);
3521
3522 lockstart = round_down(start, fs_info->sectorsize);
3523 lockend = round_up(i_size, fs_info->sectorsize);
3524 if (lockend <= lockstart)
3525 lockend = lockstart + fs_info->sectorsize;
3526 lockend--;
3527
3528 path = btrfs_alloc_path();
3529 if (!path)
3530 return -ENOMEM;
3531 path->reada = READA_FORWARD;
3532
3533 key.objectid = ino;
3534 key.type = BTRFS_EXTENT_DATA_KEY;
3535 key.offset = start;
3536
3537 last_extent_end = lockstart;
3538
3539 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3540
3541 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3542 if (ret < 0) {
3543 goto out;
3544 } else if (ret > 0 && path->slots[0] > 0) {
3545 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
3546 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
3547 path->slots[0]--;
3548 }
3549
3550 while (start < i_size) {
3551 struct extent_buffer *leaf = path->nodes[0];
3552 struct btrfs_file_extent_item *extent;
3553 u64 extent_end;
3554 u8 type;
3555
3556 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3557 ret = btrfs_next_leaf(root, path);
3558 if (ret < 0)
3559 goto out;
3560 else if (ret > 0)
3561 break;
3562
3563 leaf = path->nodes[0];
3564 }
3565
3566 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3567 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3568 break;
3569
3570 extent_end = btrfs_file_extent_end(path);
3571
3572 /*
3573 * In the first iteration we may have a slot that points to an
3574 * extent that ends before our start offset, so skip it.
3575 */
3576 if (extent_end <= start) {
3577 path->slots[0]++;
3578 continue;
3579 }
3580
3581 /* We have an implicit hole, NO_HOLES feature is likely set. */
3582 if (last_extent_end < key.offset) {
3583 u64 search_start = last_extent_end;
3584 u64 found_start;
3585
3586 /*
3587 * First iteration, @start matches @offset and it's
3588 * within the hole.
3589 */
3590 if (start == offset)
3591 search_start = offset;
3592
3593 found = find_desired_extent_in_hole(inode, whence,
3594 delalloc_cached_state,
3595 search_start,
3596 key.offset - 1,
3597 &found_start);
3598 if (found) {
3599 start = found_start;
3600 break;
3601 }
3602 /*
3603 * Didn't find data or a hole (due to delalloc) in the
3604 * implicit hole range, so need to analyze the extent.
3605 */
3606 }
3607
3608 extent = btrfs_item_ptr(leaf, path->slots[0],
3609 struct btrfs_file_extent_item);
3610 type = btrfs_file_extent_type(leaf, extent);
3611
3612 /*
3613 * Can't access the extent's disk_bytenr field if this is an
3614 * inline extent, since at that offset, it's where the extent
3615 * data starts.
3616 */
3617 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
3618 (type == BTRFS_FILE_EXTENT_REG &&
3619 btrfs_file_extent_disk_bytenr(leaf, extent) == 0)) {
3620 /*
3621 * Explicit hole or prealloc extent, search for delalloc.
3622 * A prealloc extent is treated like a hole.
3623 */
3624 u64 search_start = key.offset;
3625 u64 found_start;
3626
3627 /*
3628 * First iteration, @start matches @offset and it's
3629 * within the hole.
3630 */
3631 if (start == offset)
3632 search_start = offset;
3633
3634 found = find_desired_extent_in_hole(inode, whence,
3635 delalloc_cached_state,
3636 search_start,
3637 extent_end - 1,
3638 &found_start);
3639 if (found) {
3640 start = found_start;
3641 break;
3642 }
3643 /*
3644 * Didn't find data or a hole (due to delalloc) in the
3645 * implicit hole range, so need to analyze the next
3646 * extent item.
3647 */
3648 } else {
3649 /*
3650 * Found a regular or inline extent.
3651 * If we are seeking for data, adjust the start offset
3652 * and stop, we're done.
3653 */
3654 if (whence == SEEK_DATA) {
3655 start = max_t(u64, key.offset, offset);
3656 found = true;
3657 break;
3658 }
3659 /*
3660 * Else, we are seeking for a hole, check the next file
3661 * extent item.
3662 */
3663 }
3664
3665 start = extent_end;
3666 last_extent_end = extent_end;
3667 path->slots[0]++;
3668 if (fatal_signal_pending(current)) {
3669 ret = -EINTR;
3670 goto out;
3671 }
3672 cond_resched();
3673 }
3674
3675 /* We have an implicit hole from the last extent found up to i_size. */
3676 if (!found && start < i_size) {
3677 found = find_desired_extent_in_hole(inode, whence,
3678 delalloc_cached_state, start,
3679 i_size - 1, &start);
3680 if (!found)
3681 start = i_size;
3682 }
3683
3684out:
3685 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3686 btrfs_free_path(path);
3687
3688 if (ret < 0)
3689 return ret;
3690
3691 if (whence == SEEK_DATA && start >= i_size)
3692 return -ENXIO;
3693
3694 return min_t(loff_t, start, i_size);
3695}
3696
3697static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3698{
3699 struct inode *inode = file->f_mapping->host;
3700
3701 switch (whence) {
3702 default:
3703 return generic_file_llseek(file, offset, whence);
3704 case SEEK_DATA:
3705 case SEEK_HOLE:
3706 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3707 offset = find_desired_extent(file, offset, whence);
3708 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3709 break;
3710 }
3711
3712 if (offset < 0)
3713 return offset;
3714
3715 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3716}
3717
3718static int btrfs_file_open(struct inode *inode, struct file *filp)
3719{
3720 int ret;
3721
3722 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC |
3723 FMODE_CAN_ODIRECT;
3724
3725 ret = fsverity_file_open(inode, filp);
3726 if (ret)
3727 return ret;
3728 return generic_file_open(inode, filp);
3729}
3730
3731static int check_direct_read(struct btrfs_fs_info *fs_info,
3732 const struct iov_iter *iter, loff_t offset)
3733{
3734 int ret;
3735 int i, seg;
3736
3737 ret = check_direct_IO(fs_info, iter, offset);
3738 if (ret < 0)
3739 return ret;
3740
3741 if (!iter_is_iovec(iter))
3742 return 0;
3743
3744 for (seg = 0; seg < iter->nr_segs; seg++) {
3745 for (i = seg + 1; i < iter->nr_segs; i++) {
3746 const struct iovec *iov1 = iter_iov(iter) + seg;
3747 const struct iovec *iov2 = iter_iov(iter) + i;
3748
3749 if (iov1->iov_base == iov2->iov_base)
3750 return -EINVAL;
3751 }
3752 }
3753 return 0;
3754}
3755
3756static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3757{
3758 struct inode *inode = file_inode(iocb->ki_filp);
3759 size_t prev_left = 0;
3760 ssize_t read = 0;
3761 ssize_t ret;
3762
3763 if (fsverity_active(inode))
3764 return 0;
3765
3766 if (check_direct_read(inode_to_fs_info(inode), to, iocb->ki_pos))
3767 return 0;
3768
3769 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3770again:
3771 /*
3772 * This is similar to what we do for direct IO writes, see the comment
3773 * at btrfs_direct_write(), but we also disable page faults in addition
3774 * to disabling them only at the iov_iter level. This is because when
3775 * reading from a hole or prealloc extent, iomap calls iov_iter_zero(),
3776 * which can still trigger page fault ins despite having set ->nofault
3777 * to true of our 'to' iov_iter.
3778 *
3779 * The difference to direct IO writes is that we deadlock when trying
3780 * to lock the extent range in the inode's tree during he page reads
3781 * triggered by the fault in (while for writes it is due to waiting for
3782 * our own ordered extent). This is because for direct IO reads,
3783 * btrfs_dio_iomap_begin() returns with the extent range locked, which
3784 * is only unlocked in the endio callback (end_bio_extent_readpage()).
3785 */
3786 pagefault_disable();
3787 to->nofault = true;
3788 ret = btrfs_dio_read(iocb, to, read);
3789 to->nofault = false;
3790 pagefault_enable();
3791
3792 /* No increment (+=) because iomap returns a cumulative value. */
3793 if (ret > 0)
3794 read = ret;
3795
3796 if (iov_iter_count(to) > 0 && (ret == -EFAULT || ret > 0)) {
3797 const size_t left = iov_iter_count(to);
3798
3799 if (left == prev_left) {
3800 /*
3801 * We didn't make any progress since the last attempt,
3802 * fallback to a buffered read for the remainder of the
3803 * range. This is just to avoid any possibility of looping
3804 * for too long.
3805 */
3806 ret = read;
3807 } else {
3808 /*
3809 * We made some progress since the last retry or this is
3810 * the first time we are retrying. Fault in as many pages
3811 * as possible and retry.
3812 */
3813 fault_in_iov_iter_writeable(to, left);
3814 prev_left = left;
3815 goto again;
3816 }
3817 }
3818 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3819 return ret < 0 ? ret : read;
3820}
3821
3822static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3823{
3824 ssize_t ret = 0;
3825
3826 if (iocb->ki_flags & IOCB_DIRECT) {
3827 ret = btrfs_direct_read(iocb, to);
3828 if (ret < 0 || !iov_iter_count(to) ||
3829 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3830 return ret;
3831 }
3832
3833 return filemap_read(iocb, to, ret);
3834}
3835
3836const struct file_operations btrfs_file_operations = {
3837 .llseek = btrfs_file_llseek,
3838 .read_iter = btrfs_file_read_iter,
3839 .splice_read = filemap_splice_read,
3840 .write_iter = btrfs_file_write_iter,
3841 .splice_write = iter_file_splice_write,
3842 .mmap = btrfs_file_mmap,
3843 .open = btrfs_file_open,
3844 .release = btrfs_release_file,
3845 .get_unmapped_area = thp_get_unmapped_area,
3846 .fsync = btrfs_sync_file,
3847 .fallocate = btrfs_fallocate,
3848 .unlocked_ioctl = btrfs_ioctl,
3849#ifdef CONFIG_COMPAT
3850 .compat_ioctl = btrfs_compat_ioctl,
3851#endif
3852 .remap_file_range = btrfs_remap_file_range,
3853};
3854
3855int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3856{
3857 int ret;
3858
3859 /*
3860 * So with compression we will find and lock a dirty page and clear the
3861 * first one as dirty, setup an async extent, and immediately return
3862 * with the entire range locked but with nobody actually marked with
3863 * writeback. So we can't just filemap_write_and_wait_range() and
3864 * expect it to work since it will just kick off a thread to do the
3865 * actual work. So we need to call filemap_fdatawrite_range _again_
3866 * since it will wait on the page lock, which won't be unlocked until
3867 * after the pages have been marked as writeback and so we're good to go
3868 * from there. We have to do this otherwise we'll miss the ordered
3869 * extents and that results in badness. Please Josef, do not think you
3870 * know better and pull this out at some point in the future, it is
3871 * right and you are wrong.
3872 */
3873 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3874 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3875 &BTRFS_I(inode)->runtime_flags))
3876 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3877
3878 return ret;
3879}
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#include "subpage.h"
32
33static struct kmem_cache *btrfs_inode_defrag_cachep;
34/*
35 * when auto defrag is enabled we
36 * queue up these defrag structs to remember which
37 * inodes need defragging passes
38 */
39struct inode_defrag {
40 struct rb_node rb_node;
41 /* objectid */
42 u64 ino;
43 /*
44 * transid where the defrag was added, we search for
45 * extents newer than this
46 */
47 u64 transid;
48
49 /* root objectid */
50 u64 root;
51
52 /* last offset we were able to defrag */
53 u64 last_offset;
54
55 /* if we've wrapped around back to zero once already */
56 int cycled;
57};
58
59static int __compare_inode_defrag(struct inode_defrag *defrag1,
60 struct inode_defrag *defrag2)
61{
62 if (defrag1->root > defrag2->root)
63 return 1;
64 else if (defrag1->root < defrag2->root)
65 return -1;
66 else if (defrag1->ino > defrag2->ino)
67 return 1;
68 else if (defrag1->ino < defrag2->ino)
69 return -1;
70 else
71 return 0;
72}
73
74/* pop a record for an inode into the defrag tree. The lock
75 * must be held already
76 *
77 * If you're inserting a record for an older transid than an
78 * existing record, the transid already in the tree is lowered
79 *
80 * If an existing record is found the defrag item you
81 * pass in is freed
82 */
83static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
84 struct inode_defrag *defrag)
85{
86 struct btrfs_fs_info *fs_info = inode->root->fs_info;
87 struct inode_defrag *entry;
88 struct rb_node **p;
89 struct rb_node *parent = NULL;
90 int ret;
91
92 p = &fs_info->defrag_inodes.rb_node;
93 while (*p) {
94 parent = *p;
95 entry = rb_entry(parent, struct inode_defrag, rb_node);
96
97 ret = __compare_inode_defrag(defrag, entry);
98 if (ret < 0)
99 p = &parent->rb_left;
100 else if (ret > 0)
101 p = &parent->rb_right;
102 else {
103 /* if we're reinserting an entry for
104 * an old defrag run, make sure to
105 * lower the transid of our existing record
106 */
107 if (defrag->transid < entry->transid)
108 entry->transid = defrag->transid;
109 if (defrag->last_offset > entry->last_offset)
110 entry->last_offset = defrag->last_offset;
111 return -EEXIST;
112 }
113 }
114 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
115 rb_link_node(&defrag->rb_node, parent, p);
116 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
117 return 0;
118}
119
120static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
121{
122 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
123 return 0;
124
125 if (btrfs_fs_closing(fs_info))
126 return 0;
127
128 return 1;
129}
130
131/*
132 * insert a defrag record for this inode if auto defrag is
133 * enabled
134 */
135int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
136 struct btrfs_inode *inode)
137{
138 struct btrfs_root *root = inode->root;
139 struct btrfs_fs_info *fs_info = root->fs_info;
140 struct inode_defrag *defrag;
141 u64 transid;
142 int ret;
143
144 if (!__need_auto_defrag(fs_info))
145 return 0;
146
147 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
148 return 0;
149
150 if (trans)
151 transid = trans->transid;
152 else
153 transid = inode->root->last_trans;
154
155 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
156 if (!defrag)
157 return -ENOMEM;
158
159 defrag->ino = btrfs_ino(inode);
160 defrag->transid = transid;
161 defrag->root = root->root_key.objectid;
162
163 spin_lock(&fs_info->defrag_inodes_lock);
164 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
165 /*
166 * If we set IN_DEFRAG flag and evict the inode from memory,
167 * and then re-read this inode, this new inode doesn't have
168 * IN_DEFRAG flag. At the case, we may find the existed defrag.
169 */
170 ret = __btrfs_add_inode_defrag(inode, defrag);
171 if (ret)
172 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
173 } else {
174 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
175 }
176 spin_unlock(&fs_info->defrag_inodes_lock);
177 return 0;
178}
179
180/*
181 * Requeue the defrag object. If there is a defrag object that points to
182 * the same inode in the tree, we will merge them together (by
183 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
184 */
185static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
186 struct inode_defrag *defrag)
187{
188 struct btrfs_fs_info *fs_info = inode->root->fs_info;
189 int ret;
190
191 if (!__need_auto_defrag(fs_info))
192 goto out;
193
194 /*
195 * Here we don't check the IN_DEFRAG flag, because we need merge
196 * them together.
197 */
198 spin_lock(&fs_info->defrag_inodes_lock);
199 ret = __btrfs_add_inode_defrag(inode, defrag);
200 spin_unlock(&fs_info->defrag_inodes_lock);
201 if (ret)
202 goto out;
203 return;
204out:
205 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
206}
207
208/*
209 * pick the defragable inode that we want, if it doesn't exist, we will get
210 * the next one.
211 */
212static struct inode_defrag *
213btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
214{
215 struct inode_defrag *entry = NULL;
216 struct inode_defrag tmp;
217 struct rb_node *p;
218 struct rb_node *parent = NULL;
219 int ret;
220
221 tmp.ino = ino;
222 tmp.root = root;
223
224 spin_lock(&fs_info->defrag_inodes_lock);
225 p = fs_info->defrag_inodes.rb_node;
226 while (p) {
227 parent = p;
228 entry = rb_entry(parent, struct inode_defrag, rb_node);
229
230 ret = __compare_inode_defrag(&tmp, entry);
231 if (ret < 0)
232 p = parent->rb_left;
233 else if (ret > 0)
234 p = parent->rb_right;
235 else
236 goto out;
237 }
238
239 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
240 parent = rb_next(parent);
241 if (parent)
242 entry = rb_entry(parent, struct inode_defrag, rb_node);
243 else
244 entry = NULL;
245 }
246out:
247 if (entry)
248 rb_erase(parent, &fs_info->defrag_inodes);
249 spin_unlock(&fs_info->defrag_inodes_lock);
250 return entry;
251}
252
253void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
254{
255 struct inode_defrag *defrag;
256 struct rb_node *node;
257
258 spin_lock(&fs_info->defrag_inodes_lock);
259 node = rb_first(&fs_info->defrag_inodes);
260 while (node) {
261 rb_erase(node, &fs_info->defrag_inodes);
262 defrag = rb_entry(node, struct inode_defrag, rb_node);
263 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
264
265 cond_resched_lock(&fs_info->defrag_inodes_lock);
266
267 node = rb_first(&fs_info->defrag_inodes);
268 }
269 spin_unlock(&fs_info->defrag_inodes_lock);
270}
271
272#define BTRFS_DEFRAG_BATCH 1024
273
274static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
275 struct inode_defrag *defrag)
276{
277 struct btrfs_root *inode_root;
278 struct inode *inode;
279 struct btrfs_ioctl_defrag_range_args range;
280 int num_defrag;
281 int ret;
282
283 /* get the inode */
284 inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
285 if (IS_ERR(inode_root)) {
286 ret = PTR_ERR(inode_root);
287 goto cleanup;
288 }
289
290 inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
291 btrfs_put_root(inode_root);
292 if (IS_ERR(inode)) {
293 ret = PTR_ERR(inode);
294 goto cleanup;
295 }
296
297 /* do a chunk of defrag */
298 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
299 memset(&range, 0, sizeof(range));
300 range.len = (u64)-1;
301 range.start = defrag->last_offset;
302
303 sb_start_write(fs_info->sb);
304 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
305 BTRFS_DEFRAG_BATCH);
306 sb_end_write(fs_info->sb);
307 /*
308 * if we filled the whole defrag batch, there
309 * must be more work to do. Queue this defrag
310 * again
311 */
312 if (num_defrag == BTRFS_DEFRAG_BATCH) {
313 defrag->last_offset = range.start;
314 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
315 } else if (defrag->last_offset && !defrag->cycled) {
316 /*
317 * we didn't fill our defrag batch, but
318 * we didn't start at zero. Make sure we loop
319 * around to the start of the file.
320 */
321 defrag->last_offset = 0;
322 defrag->cycled = 1;
323 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
324 } else {
325 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
326 }
327
328 iput(inode);
329 return 0;
330cleanup:
331 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
332 return ret;
333}
334
335/*
336 * run through the list of inodes in the FS that need
337 * defragging
338 */
339int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
340{
341 struct inode_defrag *defrag;
342 u64 first_ino = 0;
343 u64 root_objectid = 0;
344
345 atomic_inc(&fs_info->defrag_running);
346 while (1) {
347 /* Pause the auto defragger. */
348 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
349 &fs_info->fs_state))
350 break;
351
352 if (!__need_auto_defrag(fs_info))
353 break;
354
355 /* find an inode to defrag */
356 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
357 first_ino);
358 if (!defrag) {
359 if (root_objectid || first_ino) {
360 root_objectid = 0;
361 first_ino = 0;
362 continue;
363 } else {
364 break;
365 }
366 }
367
368 first_ino = defrag->ino + 1;
369 root_objectid = defrag->root;
370
371 __btrfs_run_defrag_inode(fs_info, defrag);
372 }
373 atomic_dec(&fs_info->defrag_running);
374
375 /*
376 * during unmount, we use the transaction_wait queue to
377 * wait for the defragger to stop
378 */
379 wake_up(&fs_info->transaction_wait);
380 return 0;
381}
382
383/* simple helper to fault in pages and copy. This should go away
384 * and be replaced with calls into generic code.
385 */
386static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
387 struct page **prepared_pages,
388 struct iov_iter *i)
389{
390 size_t copied = 0;
391 size_t total_copied = 0;
392 int pg = 0;
393 int offset = offset_in_page(pos);
394
395 while (write_bytes > 0) {
396 size_t count = min_t(size_t,
397 PAGE_SIZE - offset, write_bytes);
398 struct page *page = prepared_pages[pg];
399 /*
400 * Copy data from userspace to the current page
401 */
402 copied = copy_page_from_iter_atomic(page, offset, count, i);
403
404 /* Flush processor's dcache for this page */
405 flush_dcache_page(page);
406
407 /*
408 * if we get a partial write, we can end up with
409 * partially up to date pages. These add
410 * a lot of complexity, so make sure they don't
411 * happen by forcing this copy to be retried.
412 *
413 * The rest of the btrfs_file_write code will fall
414 * back to page at a time copies after we return 0.
415 */
416 if (unlikely(copied < count)) {
417 if (!PageUptodate(page)) {
418 iov_iter_revert(i, copied);
419 copied = 0;
420 }
421 if (!copied)
422 break;
423 }
424
425 write_bytes -= copied;
426 total_copied += copied;
427 offset += copied;
428 if (offset == PAGE_SIZE) {
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
455/*
456 * After btrfs_copy_from_user(), update the following things for delalloc:
457 * - Mark newly dirtied pages as DELALLOC in the io tree.
458 * Used to advise which range is to be written back.
459 * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
460 * - Update inode size for past EOF write
461 */
462int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
463 size_t num_pages, loff_t pos, size_t write_bytes,
464 struct extent_state **cached, bool noreserve)
465{
466 struct btrfs_fs_info *fs_info = inode->root->fs_info;
467 int err = 0;
468 int i;
469 u64 num_bytes;
470 u64 start_pos;
471 u64 end_of_last_block;
472 u64 end_pos = pos + write_bytes;
473 loff_t isize = i_size_read(&inode->vfs_inode);
474 unsigned int extra_bits = 0;
475
476 if (write_bytes == 0)
477 return 0;
478
479 if (noreserve)
480 extra_bits |= EXTENT_NORESERVE;
481
482 start_pos = round_down(pos, fs_info->sectorsize);
483 num_bytes = round_up(write_bytes + pos - start_pos,
484 fs_info->sectorsize);
485 ASSERT(num_bytes <= U32_MAX);
486
487 end_of_last_block = start_pos + num_bytes - 1;
488
489 /*
490 * The pages may have already been dirty, clear out old accounting so
491 * we can set things up properly
492 */
493 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
494 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
495 0, 0, cached);
496
497 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
498 extra_bits, cached);
499 if (err)
500 return err;
501
502 for (i = 0; i < num_pages; i++) {
503 struct page *p = pages[i];
504
505 btrfs_page_clamp_set_uptodate(fs_info, p, start_pos, num_bytes);
506 ClearPageChecked(p);
507 btrfs_page_clamp_set_dirty(fs_info, p, start_pos, num_bytes);
508 }
509
510 /*
511 * we've only changed i_size in ram, and we haven't updated
512 * the disk i_size. There is no need to log the inode
513 * at this time.
514 */
515 if (end_pos > isize)
516 i_size_write(&inode->vfs_inode, end_pos);
517 return 0;
518}
519
520/*
521 * this drops all the extents in the cache that intersect the range
522 * [start, end]. Existing extents are split as required.
523 */
524void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
525 int skip_pinned)
526{
527 struct extent_map *em;
528 struct extent_map *split = NULL;
529 struct extent_map *split2 = NULL;
530 struct extent_map_tree *em_tree = &inode->extent_tree;
531 u64 len = end - start + 1;
532 u64 gen;
533 int ret;
534 int testend = 1;
535 unsigned long flags;
536 int compressed = 0;
537 bool modified;
538
539 WARN_ON(end < start);
540 if (end == (u64)-1) {
541 len = (u64)-1;
542 testend = 0;
543 }
544 while (1) {
545 int no_splits = 0;
546
547 modified = false;
548 if (!split)
549 split = alloc_extent_map();
550 if (!split2)
551 split2 = alloc_extent_map();
552 if (!split || !split2)
553 no_splits = 1;
554
555 write_lock(&em_tree->lock);
556 em = lookup_extent_mapping(em_tree, start, len);
557 if (!em) {
558 write_unlock(&em_tree->lock);
559 break;
560 }
561 flags = em->flags;
562 gen = em->generation;
563 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
564 if (testend && em->start + em->len >= start + len) {
565 free_extent_map(em);
566 write_unlock(&em_tree->lock);
567 break;
568 }
569 start = em->start + em->len;
570 if (testend)
571 len = start + len - (em->start + em->len);
572 free_extent_map(em);
573 write_unlock(&em_tree->lock);
574 continue;
575 }
576 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
577 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
578 clear_bit(EXTENT_FLAG_LOGGING, &flags);
579 modified = !list_empty(&em->list);
580 if (no_splits)
581 goto next;
582
583 if (em->start < start) {
584 split->start = em->start;
585 split->len = start - em->start;
586
587 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
588 split->orig_start = em->orig_start;
589 split->block_start = em->block_start;
590
591 if (compressed)
592 split->block_len = em->block_len;
593 else
594 split->block_len = split->len;
595 split->orig_block_len = max(split->block_len,
596 em->orig_block_len);
597 split->ram_bytes = em->ram_bytes;
598 } else {
599 split->orig_start = split->start;
600 split->block_len = 0;
601 split->block_start = em->block_start;
602 split->orig_block_len = 0;
603 split->ram_bytes = split->len;
604 }
605
606 split->generation = gen;
607 split->flags = flags;
608 split->compress_type = em->compress_type;
609 replace_extent_mapping(em_tree, em, split, modified);
610 free_extent_map(split);
611 split = split2;
612 split2 = NULL;
613 }
614 if (testend && em->start + em->len > start + len) {
615 u64 diff = start + len - em->start;
616
617 split->start = start + len;
618 split->len = em->start + em->len - (start + len);
619 split->flags = flags;
620 split->compress_type = em->compress_type;
621 split->generation = gen;
622
623 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
624 split->orig_block_len = max(em->block_len,
625 em->orig_block_len);
626
627 split->ram_bytes = em->ram_bytes;
628 if (compressed) {
629 split->block_len = em->block_len;
630 split->block_start = em->block_start;
631 split->orig_start = em->orig_start;
632 } else {
633 split->block_len = split->len;
634 split->block_start = em->block_start
635 + diff;
636 split->orig_start = em->orig_start;
637 }
638 } else {
639 split->ram_bytes = split->len;
640 split->orig_start = split->start;
641 split->block_len = 0;
642 split->block_start = em->block_start;
643 split->orig_block_len = 0;
644 }
645
646 if (extent_map_in_tree(em)) {
647 replace_extent_mapping(em_tree, em, split,
648 modified);
649 } else {
650 ret = add_extent_mapping(em_tree, split,
651 modified);
652 ASSERT(ret == 0); /* Logic error */
653 }
654 free_extent_map(split);
655 split = NULL;
656 }
657next:
658 if (extent_map_in_tree(em))
659 remove_extent_mapping(em_tree, em);
660 write_unlock(&em_tree->lock);
661
662 /* once for us */
663 free_extent_map(em);
664 /* once for the tree*/
665 free_extent_map(em);
666 }
667 if (split)
668 free_extent_map(split);
669 if (split2)
670 free_extent_map(split2);
671}
672
673/*
674 * this is very complex, but the basic idea is to drop all extents
675 * in the range start - end. hint_block is filled in with a block number
676 * that would be a good hint to the block allocator for this file.
677 *
678 * If an extent intersects the range but is not entirely inside the range
679 * it is either truncated or split. Anything entirely inside the range
680 * is deleted from the tree.
681 *
682 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
683 * to deal with that. We set the field 'bytes_found' of the arguments structure
684 * with the number of allocated bytes found in the target range, so that the
685 * caller can update the inode's number of bytes in an atomic way when
686 * replacing extents in a range to avoid races with stat(2).
687 */
688int btrfs_drop_extents(struct btrfs_trans_handle *trans,
689 struct btrfs_root *root, struct btrfs_inode *inode,
690 struct btrfs_drop_extents_args *args)
691{
692 struct btrfs_fs_info *fs_info = root->fs_info;
693 struct extent_buffer *leaf;
694 struct btrfs_file_extent_item *fi;
695 struct btrfs_ref ref = { 0 };
696 struct btrfs_key key;
697 struct btrfs_key new_key;
698 u64 ino = btrfs_ino(inode);
699 u64 search_start = args->start;
700 u64 disk_bytenr = 0;
701 u64 num_bytes = 0;
702 u64 extent_offset = 0;
703 u64 extent_end = 0;
704 u64 last_end = args->start;
705 int del_nr = 0;
706 int del_slot = 0;
707 int extent_type;
708 int recow;
709 int ret;
710 int modify_tree = -1;
711 int update_refs;
712 int found = 0;
713 int leafs_visited = 0;
714 struct btrfs_path *path = args->path;
715
716 args->bytes_found = 0;
717 args->extent_inserted = false;
718
719 /* Must always have a path if ->replace_extent is true */
720 ASSERT(!(args->replace_extent && !args->path));
721
722 if (!path) {
723 path = btrfs_alloc_path();
724 if (!path) {
725 ret = -ENOMEM;
726 goto out;
727 }
728 }
729
730 if (args->drop_cache)
731 btrfs_drop_extent_cache(inode, args->start, args->end - 1, 0);
732
733 if (args->start >= inode->disk_i_size && !args->replace_extent)
734 modify_tree = 0;
735
736 update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
737 while (1) {
738 recow = 0;
739 ret = btrfs_lookup_file_extent(trans, root, path, ino,
740 search_start, modify_tree);
741 if (ret < 0)
742 break;
743 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
744 leaf = path->nodes[0];
745 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
746 if (key.objectid == ino &&
747 key.type == BTRFS_EXTENT_DATA_KEY)
748 path->slots[0]--;
749 }
750 ret = 0;
751 leafs_visited++;
752next_slot:
753 leaf = path->nodes[0];
754 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
755 BUG_ON(del_nr > 0);
756 ret = btrfs_next_leaf(root, path);
757 if (ret < 0)
758 break;
759 if (ret > 0) {
760 ret = 0;
761 break;
762 }
763 leafs_visited++;
764 leaf = path->nodes[0];
765 recow = 1;
766 }
767
768 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
769
770 if (key.objectid > ino)
771 break;
772 if (WARN_ON_ONCE(key.objectid < ino) ||
773 key.type < BTRFS_EXTENT_DATA_KEY) {
774 ASSERT(del_nr == 0);
775 path->slots[0]++;
776 goto next_slot;
777 }
778 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
779 break;
780
781 fi = btrfs_item_ptr(leaf, path->slots[0],
782 struct btrfs_file_extent_item);
783 extent_type = btrfs_file_extent_type(leaf, fi);
784
785 if (extent_type == BTRFS_FILE_EXTENT_REG ||
786 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
787 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
788 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
789 extent_offset = btrfs_file_extent_offset(leaf, fi);
790 extent_end = key.offset +
791 btrfs_file_extent_num_bytes(leaf, fi);
792 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
793 extent_end = key.offset +
794 btrfs_file_extent_ram_bytes(leaf, fi);
795 } else {
796 /* can't happen */
797 BUG();
798 }
799
800 /*
801 * Don't skip extent items representing 0 byte lengths. They
802 * used to be created (bug) if while punching holes we hit
803 * -ENOSPC condition. So if we find one here, just ensure we
804 * delete it, otherwise we would insert a new file extent item
805 * with the same key (offset) as that 0 bytes length file
806 * extent item in the call to setup_items_for_insert() later
807 * in this function.
808 */
809 if (extent_end == key.offset && extent_end >= search_start) {
810 last_end = extent_end;
811 goto delete_extent_item;
812 }
813
814 if (extent_end <= search_start) {
815 path->slots[0]++;
816 goto next_slot;
817 }
818
819 found = 1;
820 search_start = max(key.offset, args->start);
821 if (recow || !modify_tree) {
822 modify_tree = -1;
823 btrfs_release_path(path);
824 continue;
825 }
826
827 /*
828 * | - range to drop - |
829 * | -------- extent -------- |
830 */
831 if (args->start > key.offset && args->end < extent_end) {
832 BUG_ON(del_nr > 0);
833 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
834 ret = -EOPNOTSUPP;
835 break;
836 }
837
838 memcpy(&new_key, &key, sizeof(new_key));
839 new_key.offset = args->start;
840 ret = btrfs_duplicate_item(trans, root, path,
841 &new_key);
842 if (ret == -EAGAIN) {
843 btrfs_release_path(path);
844 continue;
845 }
846 if (ret < 0)
847 break;
848
849 leaf = path->nodes[0];
850 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
851 struct btrfs_file_extent_item);
852 btrfs_set_file_extent_num_bytes(leaf, fi,
853 args->start - key.offset);
854
855 fi = btrfs_item_ptr(leaf, path->slots[0],
856 struct btrfs_file_extent_item);
857
858 extent_offset += args->start - key.offset;
859 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
860 btrfs_set_file_extent_num_bytes(leaf, fi,
861 extent_end - args->start);
862 btrfs_mark_buffer_dirty(leaf);
863
864 if (update_refs && disk_bytenr > 0) {
865 btrfs_init_generic_ref(&ref,
866 BTRFS_ADD_DELAYED_REF,
867 disk_bytenr, num_bytes, 0);
868 btrfs_init_data_ref(&ref,
869 root->root_key.objectid,
870 new_key.objectid,
871 args->start - extent_offset);
872 ret = btrfs_inc_extent_ref(trans, &ref);
873 BUG_ON(ret); /* -ENOMEM */
874 }
875 key.offset = args->start;
876 }
877 /*
878 * From here on out we will have actually dropped something, so
879 * last_end can be updated.
880 */
881 last_end = extent_end;
882
883 /*
884 * | ---- range to drop ----- |
885 * | -------- extent -------- |
886 */
887 if (args->start <= key.offset && args->end < extent_end) {
888 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
889 ret = -EOPNOTSUPP;
890 break;
891 }
892
893 memcpy(&new_key, &key, sizeof(new_key));
894 new_key.offset = args->end;
895 btrfs_set_item_key_safe(fs_info, path, &new_key);
896
897 extent_offset += args->end - key.offset;
898 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
899 btrfs_set_file_extent_num_bytes(leaf, fi,
900 extent_end - args->end);
901 btrfs_mark_buffer_dirty(leaf);
902 if (update_refs && disk_bytenr > 0)
903 args->bytes_found += args->end - key.offset;
904 break;
905 }
906
907 search_start = extent_end;
908 /*
909 * | ---- range to drop ----- |
910 * | -------- extent -------- |
911 */
912 if (args->start > key.offset && args->end >= extent_end) {
913 BUG_ON(del_nr > 0);
914 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
915 ret = -EOPNOTSUPP;
916 break;
917 }
918
919 btrfs_set_file_extent_num_bytes(leaf, fi,
920 args->start - key.offset);
921 btrfs_mark_buffer_dirty(leaf);
922 if (update_refs && disk_bytenr > 0)
923 args->bytes_found += extent_end - args->start;
924 if (args->end == extent_end)
925 break;
926
927 path->slots[0]++;
928 goto next_slot;
929 }
930
931 /*
932 * | ---- range to drop ----- |
933 * | ------ extent ------ |
934 */
935 if (args->start <= key.offset && args->end >= extent_end) {
936delete_extent_item:
937 if (del_nr == 0) {
938 del_slot = path->slots[0];
939 del_nr = 1;
940 } else {
941 BUG_ON(del_slot + del_nr != path->slots[0]);
942 del_nr++;
943 }
944
945 if (update_refs &&
946 extent_type == BTRFS_FILE_EXTENT_INLINE) {
947 args->bytes_found += extent_end - key.offset;
948 extent_end = ALIGN(extent_end,
949 fs_info->sectorsize);
950 } else if (update_refs && disk_bytenr > 0) {
951 btrfs_init_generic_ref(&ref,
952 BTRFS_DROP_DELAYED_REF,
953 disk_bytenr, num_bytes, 0);
954 btrfs_init_data_ref(&ref,
955 root->root_key.objectid,
956 key.objectid,
957 key.offset - extent_offset);
958 ret = btrfs_free_extent(trans, &ref);
959 BUG_ON(ret); /* -ENOMEM */
960 args->bytes_found += extent_end - key.offset;
961 }
962
963 if (args->end == extent_end)
964 break;
965
966 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
967 path->slots[0]++;
968 goto next_slot;
969 }
970
971 ret = btrfs_del_items(trans, root, path, del_slot,
972 del_nr);
973 if (ret) {
974 btrfs_abort_transaction(trans, ret);
975 break;
976 }
977
978 del_nr = 0;
979 del_slot = 0;
980
981 btrfs_release_path(path);
982 continue;
983 }
984
985 BUG();
986 }
987
988 if (!ret && del_nr > 0) {
989 /*
990 * Set path->slots[0] to first slot, so that after the delete
991 * if items are move off from our leaf to its immediate left or
992 * right neighbor leafs, we end up with a correct and adjusted
993 * path->slots[0] for our insertion (if args->replace_extent).
994 */
995 path->slots[0] = del_slot;
996 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
997 if (ret)
998 btrfs_abort_transaction(trans, ret);
999 }
1000
1001 leaf = path->nodes[0];
1002 /*
1003 * If btrfs_del_items() was called, it might have deleted a leaf, in
1004 * which case it unlocked our path, so check path->locks[0] matches a
1005 * write lock.
1006 */
1007 if (!ret && args->replace_extent && leafs_visited == 1 &&
1008 path->locks[0] == BTRFS_WRITE_LOCK &&
1009 btrfs_leaf_free_space(leaf) >=
1010 sizeof(struct btrfs_item) + args->extent_item_size) {
1011
1012 key.objectid = ino;
1013 key.type = BTRFS_EXTENT_DATA_KEY;
1014 key.offset = args->start;
1015 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1016 struct btrfs_key slot_key;
1017
1018 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1019 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1020 path->slots[0]++;
1021 }
1022 setup_items_for_insert(root, path, &key,
1023 &args->extent_item_size, 1);
1024 args->extent_inserted = true;
1025 }
1026
1027 if (!args->path)
1028 btrfs_free_path(path);
1029 else if (!args->extent_inserted)
1030 btrfs_release_path(path);
1031out:
1032 args->drop_end = found ? min(args->end, last_end) : args->end;
1033
1034 return ret;
1035}
1036
1037static int extent_mergeable(struct extent_buffer *leaf, int slot,
1038 u64 objectid, u64 bytenr, u64 orig_offset,
1039 u64 *start, u64 *end)
1040{
1041 struct btrfs_file_extent_item *fi;
1042 struct btrfs_key key;
1043 u64 extent_end;
1044
1045 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1046 return 0;
1047
1048 btrfs_item_key_to_cpu(leaf, &key, slot);
1049 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1050 return 0;
1051
1052 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1053 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1054 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1055 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1056 btrfs_file_extent_compression(leaf, fi) ||
1057 btrfs_file_extent_encryption(leaf, fi) ||
1058 btrfs_file_extent_other_encoding(leaf, fi))
1059 return 0;
1060
1061 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1062 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1063 return 0;
1064
1065 *start = key.offset;
1066 *end = extent_end;
1067 return 1;
1068}
1069
1070/*
1071 * Mark extent in the range start - end as written.
1072 *
1073 * This changes extent type from 'pre-allocated' to 'regular'. If only
1074 * part of extent is marked as written, the extent will be split into
1075 * two or three.
1076 */
1077int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1078 struct btrfs_inode *inode, u64 start, u64 end)
1079{
1080 struct btrfs_fs_info *fs_info = trans->fs_info;
1081 struct btrfs_root *root = inode->root;
1082 struct extent_buffer *leaf;
1083 struct btrfs_path *path;
1084 struct btrfs_file_extent_item *fi;
1085 struct btrfs_ref ref = { 0 };
1086 struct btrfs_key key;
1087 struct btrfs_key new_key;
1088 u64 bytenr;
1089 u64 num_bytes;
1090 u64 extent_end;
1091 u64 orig_offset;
1092 u64 other_start;
1093 u64 other_end;
1094 u64 split;
1095 int del_nr = 0;
1096 int del_slot = 0;
1097 int recow;
1098 int ret = 0;
1099 u64 ino = btrfs_ino(inode);
1100
1101 path = btrfs_alloc_path();
1102 if (!path)
1103 return -ENOMEM;
1104again:
1105 recow = 0;
1106 split = start;
1107 key.objectid = ino;
1108 key.type = BTRFS_EXTENT_DATA_KEY;
1109 key.offset = split;
1110
1111 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1112 if (ret < 0)
1113 goto out;
1114 if (ret > 0 && path->slots[0] > 0)
1115 path->slots[0]--;
1116
1117 leaf = path->nodes[0];
1118 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1119 if (key.objectid != ino ||
1120 key.type != BTRFS_EXTENT_DATA_KEY) {
1121 ret = -EINVAL;
1122 btrfs_abort_transaction(trans, ret);
1123 goto out;
1124 }
1125 fi = btrfs_item_ptr(leaf, path->slots[0],
1126 struct btrfs_file_extent_item);
1127 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1128 ret = -EINVAL;
1129 btrfs_abort_transaction(trans, ret);
1130 goto out;
1131 }
1132 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1133 if (key.offset > start || extent_end < end) {
1134 ret = -EINVAL;
1135 btrfs_abort_transaction(trans, ret);
1136 goto out;
1137 }
1138
1139 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1140 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1141 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1142 memcpy(&new_key, &key, sizeof(new_key));
1143
1144 if (start == key.offset && end < extent_end) {
1145 other_start = 0;
1146 other_end = start;
1147 if (extent_mergeable(leaf, path->slots[0] - 1,
1148 ino, bytenr, orig_offset,
1149 &other_start, &other_end)) {
1150 new_key.offset = end;
1151 btrfs_set_item_key_safe(fs_info, path, &new_key);
1152 fi = btrfs_item_ptr(leaf, path->slots[0],
1153 struct btrfs_file_extent_item);
1154 btrfs_set_file_extent_generation(leaf, fi,
1155 trans->transid);
1156 btrfs_set_file_extent_num_bytes(leaf, fi,
1157 extent_end - end);
1158 btrfs_set_file_extent_offset(leaf, fi,
1159 end - orig_offset);
1160 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1161 struct btrfs_file_extent_item);
1162 btrfs_set_file_extent_generation(leaf, fi,
1163 trans->transid);
1164 btrfs_set_file_extent_num_bytes(leaf, fi,
1165 end - other_start);
1166 btrfs_mark_buffer_dirty(leaf);
1167 goto out;
1168 }
1169 }
1170
1171 if (start > key.offset && end == extent_end) {
1172 other_start = end;
1173 other_end = 0;
1174 if (extent_mergeable(leaf, path->slots[0] + 1,
1175 ino, bytenr, orig_offset,
1176 &other_start, &other_end)) {
1177 fi = btrfs_item_ptr(leaf, path->slots[0],
1178 struct btrfs_file_extent_item);
1179 btrfs_set_file_extent_num_bytes(leaf, fi,
1180 start - key.offset);
1181 btrfs_set_file_extent_generation(leaf, fi,
1182 trans->transid);
1183 path->slots[0]++;
1184 new_key.offset = start;
1185 btrfs_set_item_key_safe(fs_info, path, &new_key);
1186
1187 fi = btrfs_item_ptr(leaf, path->slots[0],
1188 struct btrfs_file_extent_item);
1189 btrfs_set_file_extent_generation(leaf, fi,
1190 trans->transid);
1191 btrfs_set_file_extent_num_bytes(leaf, fi,
1192 other_end - start);
1193 btrfs_set_file_extent_offset(leaf, fi,
1194 start - orig_offset);
1195 btrfs_mark_buffer_dirty(leaf);
1196 goto out;
1197 }
1198 }
1199
1200 while (start > key.offset || end < extent_end) {
1201 if (key.offset == start)
1202 split = end;
1203
1204 new_key.offset = split;
1205 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1206 if (ret == -EAGAIN) {
1207 btrfs_release_path(path);
1208 goto again;
1209 }
1210 if (ret < 0) {
1211 btrfs_abort_transaction(trans, ret);
1212 goto out;
1213 }
1214
1215 leaf = path->nodes[0];
1216 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1217 struct btrfs_file_extent_item);
1218 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1219 btrfs_set_file_extent_num_bytes(leaf, fi,
1220 split - key.offset);
1221
1222 fi = btrfs_item_ptr(leaf, path->slots[0],
1223 struct btrfs_file_extent_item);
1224
1225 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1226 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1227 btrfs_set_file_extent_num_bytes(leaf, fi,
1228 extent_end - split);
1229 btrfs_mark_buffer_dirty(leaf);
1230
1231 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1232 num_bytes, 0);
1233 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1234 orig_offset);
1235 ret = btrfs_inc_extent_ref(trans, &ref);
1236 if (ret) {
1237 btrfs_abort_transaction(trans, ret);
1238 goto out;
1239 }
1240
1241 if (split == start) {
1242 key.offset = start;
1243 } else {
1244 if (start != key.offset) {
1245 ret = -EINVAL;
1246 btrfs_abort_transaction(trans, ret);
1247 goto out;
1248 }
1249 path->slots[0]--;
1250 extent_end = end;
1251 }
1252 recow = 1;
1253 }
1254
1255 other_start = end;
1256 other_end = 0;
1257 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1258 num_bytes, 0);
1259 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1260 if (extent_mergeable(leaf, path->slots[0] + 1,
1261 ino, bytenr, orig_offset,
1262 &other_start, &other_end)) {
1263 if (recow) {
1264 btrfs_release_path(path);
1265 goto again;
1266 }
1267 extent_end = other_end;
1268 del_slot = path->slots[0] + 1;
1269 del_nr++;
1270 ret = btrfs_free_extent(trans, &ref);
1271 if (ret) {
1272 btrfs_abort_transaction(trans, ret);
1273 goto out;
1274 }
1275 }
1276 other_start = 0;
1277 other_end = start;
1278 if (extent_mergeable(leaf, path->slots[0] - 1,
1279 ino, bytenr, orig_offset,
1280 &other_start, &other_end)) {
1281 if (recow) {
1282 btrfs_release_path(path);
1283 goto again;
1284 }
1285 key.offset = other_start;
1286 del_slot = path->slots[0];
1287 del_nr++;
1288 ret = btrfs_free_extent(trans, &ref);
1289 if (ret) {
1290 btrfs_abort_transaction(trans, ret);
1291 goto out;
1292 }
1293 }
1294 if (del_nr == 0) {
1295 fi = btrfs_item_ptr(leaf, path->slots[0],
1296 struct btrfs_file_extent_item);
1297 btrfs_set_file_extent_type(leaf, fi,
1298 BTRFS_FILE_EXTENT_REG);
1299 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1300 btrfs_mark_buffer_dirty(leaf);
1301 } else {
1302 fi = btrfs_item_ptr(leaf, del_slot - 1,
1303 struct btrfs_file_extent_item);
1304 btrfs_set_file_extent_type(leaf, fi,
1305 BTRFS_FILE_EXTENT_REG);
1306 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1307 btrfs_set_file_extent_num_bytes(leaf, fi,
1308 extent_end - key.offset);
1309 btrfs_mark_buffer_dirty(leaf);
1310
1311 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1312 if (ret < 0) {
1313 btrfs_abort_transaction(trans, ret);
1314 goto out;
1315 }
1316 }
1317out:
1318 btrfs_free_path(path);
1319 return ret;
1320}
1321
1322/*
1323 * on error we return an unlocked page and the error value
1324 * on success we return a locked page and 0
1325 */
1326static int prepare_uptodate_page(struct inode *inode,
1327 struct page *page, u64 pos,
1328 bool force_uptodate)
1329{
1330 int ret = 0;
1331
1332 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1333 !PageUptodate(page)) {
1334 ret = btrfs_readpage(NULL, page);
1335 if (ret)
1336 return ret;
1337 lock_page(page);
1338 if (!PageUptodate(page)) {
1339 unlock_page(page);
1340 return -EIO;
1341 }
1342 if (page->mapping != inode->i_mapping) {
1343 unlock_page(page);
1344 return -EAGAIN;
1345 }
1346 }
1347 return 0;
1348}
1349
1350/*
1351 * this just gets pages into the page cache and locks them down.
1352 */
1353static noinline int prepare_pages(struct inode *inode, struct page **pages,
1354 size_t num_pages, loff_t pos,
1355 size_t write_bytes, bool force_uptodate)
1356{
1357 int i;
1358 unsigned long index = pos >> PAGE_SHIFT;
1359 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1360 int err = 0;
1361 int faili;
1362
1363 for (i = 0; i < num_pages; i++) {
1364again:
1365 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1366 mask | __GFP_WRITE);
1367 if (!pages[i]) {
1368 faili = i - 1;
1369 err = -ENOMEM;
1370 goto fail;
1371 }
1372
1373 err = set_page_extent_mapped(pages[i]);
1374 if (err < 0) {
1375 faili = i;
1376 goto fail;
1377 }
1378
1379 if (i == 0)
1380 err = prepare_uptodate_page(inode, pages[i], pos,
1381 force_uptodate);
1382 if (!err && i == num_pages - 1)
1383 err = prepare_uptodate_page(inode, pages[i],
1384 pos + write_bytes, false);
1385 if (err) {
1386 put_page(pages[i]);
1387 if (err == -EAGAIN) {
1388 err = 0;
1389 goto again;
1390 }
1391 faili = i - 1;
1392 goto fail;
1393 }
1394 wait_on_page_writeback(pages[i]);
1395 }
1396
1397 return 0;
1398fail:
1399 while (faili >= 0) {
1400 unlock_page(pages[faili]);
1401 put_page(pages[faili]);
1402 faili--;
1403 }
1404 return err;
1405
1406}
1407
1408/*
1409 * This function locks the extent and properly waits for data=ordered extents
1410 * to finish before allowing the pages to be modified if need.
1411 *
1412 * The return value:
1413 * 1 - the extent is locked
1414 * 0 - the extent is not locked, and everything is OK
1415 * -EAGAIN - need re-prepare the pages
1416 * the other < 0 number - Something wrong happens
1417 */
1418static noinline int
1419lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1420 size_t num_pages, loff_t pos,
1421 size_t write_bytes,
1422 u64 *lockstart, u64 *lockend,
1423 struct extent_state **cached_state)
1424{
1425 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1426 u64 start_pos;
1427 u64 last_pos;
1428 int i;
1429 int ret = 0;
1430
1431 start_pos = round_down(pos, fs_info->sectorsize);
1432 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1433
1434 if (start_pos < inode->vfs_inode.i_size) {
1435 struct btrfs_ordered_extent *ordered;
1436
1437 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1438 cached_state);
1439 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1440 last_pos - start_pos + 1);
1441 if (ordered &&
1442 ordered->file_offset + ordered->num_bytes > start_pos &&
1443 ordered->file_offset <= last_pos) {
1444 unlock_extent_cached(&inode->io_tree, start_pos,
1445 last_pos, cached_state);
1446 for (i = 0; i < num_pages; i++) {
1447 unlock_page(pages[i]);
1448 put_page(pages[i]);
1449 }
1450 btrfs_start_ordered_extent(ordered, 1);
1451 btrfs_put_ordered_extent(ordered);
1452 return -EAGAIN;
1453 }
1454 if (ordered)
1455 btrfs_put_ordered_extent(ordered);
1456
1457 *lockstart = start_pos;
1458 *lockend = last_pos;
1459 ret = 1;
1460 }
1461
1462 /*
1463 * We should be called after prepare_pages() which should have locked
1464 * all pages in the range.
1465 */
1466 for (i = 0; i < num_pages; i++)
1467 WARN_ON(!PageLocked(pages[i]));
1468
1469 return ret;
1470}
1471
1472static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1473 size_t *write_bytes, bool nowait)
1474{
1475 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1476 struct btrfs_root *root = inode->root;
1477 u64 lockstart, lockend;
1478 u64 num_bytes;
1479 int ret;
1480
1481 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1482 return 0;
1483
1484 if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1485 return -EAGAIN;
1486
1487 lockstart = round_down(pos, fs_info->sectorsize);
1488 lockend = round_up(pos + *write_bytes,
1489 fs_info->sectorsize) - 1;
1490 num_bytes = lockend - lockstart + 1;
1491
1492 if (nowait) {
1493 struct btrfs_ordered_extent *ordered;
1494
1495 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1496 return -EAGAIN;
1497
1498 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1499 num_bytes);
1500 if (ordered) {
1501 btrfs_put_ordered_extent(ordered);
1502 ret = -EAGAIN;
1503 goto out_unlock;
1504 }
1505 } else {
1506 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1507 lockend, NULL);
1508 }
1509
1510 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1511 NULL, NULL, NULL, false);
1512 if (ret <= 0) {
1513 ret = 0;
1514 if (!nowait)
1515 btrfs_drew_write_unlock(&root->snapshot_lock);
1516 } else {
1517 *write_bytes = min_t(size_t, *write_bytes ,
1518 num_bytes - pos + lockstart);
1519 }
1520out_unlock:
1521 unlock_extent(&inode->io_tree, lockstart, lockend);
1522
1523 return ret;
1524}
1525
1526static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1527 size_t *write_bytes)
1528{
1529 return check_can_nocow(inode, pos, write_bytes, true);
1530}
1531
1532/*
1533 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1534 *
1535 * @pos: File offset
1536 * @write_bytes: The length to write, will be updated to the nocow writeable
1537 * range
1538 *
1539 * This function will flush ordered extents in the range to ensure proper
1540 * nocow checks.
1541 *
1542 * Return:
1543 * >0 and update @write_bytes if we can do nocow write
1544 * 0 if we can't do nocow write
1545 * -EAGAIN if we can't get the needed lock or there are ordered extents
1546 * for * (nowait == true) case
1547 * <0 if other error happened
1548 *
1549 * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1550 */
1551int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1552 size_t *write_bytes)
1553{
1554 return check_can_nocow(inode, pos, write_bytes, false);
1555}
1556
1557void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1558{
1559 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1560}
1561
1562static void update_time_for_write(struct inode *inode)
1563{
1564 struct timespec64 now;
1565
1566 if (IS_NOCMTIME(inode))
1567 return;
1568
1569 now = current_time(inode);
1570 if (!timespec64_equal(&inode->i_mtime, &now))
1571 inode->i_mtime = now;
1572
1573 if (!timespec64_equal(&inode->i_ctime, &now))
1574 inode->i_ctime = now;
1575
1576 if (IS_I_VERSION(inode))
1577 inode_inc_iversion(inode);
1578}
1579
1580static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1581 size_t count)
1582{
1583 struct file *file = iocb->ki_filp;
1584 struct inode *inode = file_inode(file);
1585 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1586 loff_t pos = iocb->ki_pos;
1587 int ret;
1588 loff_t oldsize;
1589 loff_t start_pos;
1590
1591 if (iocb->ki_flags & IOCB_NOWAIT) {
1592 size_t nocow_bytes = count;
1593
1594 /* We will allocate space in case nodatacow is not set, so bail */
1595 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes) <= 0)
1596 return -EAGAIN;
1597 /*
1598 * There are holes in the range or parts of the range that must
1599 * be COWed (shared extents, RO block groups, etc), so just bail
1600 * out.
1601 */
1602 if (nocow_bytes < count)
1603 return -EAGAIN;
1604 }
1605
1606 current->backing_dev_info = inode_to_bdi(inode);
1607 ret = file_remove_privs(file);
1608 if (ret)
1609 return ret;
1610
1611 /*
1612 * We reserve space for updating the inode when we reserve space for the
1613 * extent we are going to write, so we will enospc out there. We don't
1614 * need to start yet another transaction to update the inode as we will
1615 * update the inode when we finish writing whatever data we write.
1616 */
1617 update_time_for_write(inode);
1618
1619 start_pos = round_down(pos, fs_info->sectorsize);
1620 oldsize = i_size_read(inode);
1621 if (start_pos > oldsize) {
1622 /* Expand hole size to cover write data, preventing empty gap */
1623 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1624
1625 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1626 if (ret) {
1627 current->backing_dev_info = NULL;
1628 return ret;
1629 }
1630 }
1631
1632 return 0;
1633}
1634
1635static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1636 struct iov_iter *i)
1637{
1638 struct file *file = iocb->ki_filp;
1639 loff_t pos;
1640 struct inode *inode = file_inode(file);
1641 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1642 struct page **pages = NULL;
1643 struct extent_changeset *data_reserved = NULL;
1644 u64 release_bytes = 0;
1645 u64 lockstart;
1646 u64 lockend;
1647 size_t num_written = 0;
1648 int nrptrs;
1649 ssize_t ret;
1650 bool only_release_metadata = false;
1651 bool force_page_uptodate = false;
1652 loff_t old_isize = i_size_read(inode);
1653 unsigned int ilock_flags = 0;
1654
1655 if (iocb->ki_flags & IOCB_NOWAIT)
1656 ilock_flags |= BTRFS_ILOCK_TRY;
1657
1658 ret = btrfs_inode_lock(inode, ilock_flags);
1659 if (ret < 0)
1660 return ret;
1661
1662 ret = generic_write_checks(iocb, i);
1663 if (ret <= 0)
1664 goto out;
1665
1666 ret = btrfs_write_check(iocb, i, ret);
1667 if (ret < 0)
1668 goto out;
1669
1670 pos = iocb->ki_pos;
1671 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1672 PAGE_SIZE / (sizeof(struct page *)));
1673 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1674 nrptrs = max(nrptrs, 8);
1675 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1676 if (!pages) {
1677 ret = -ENOMEM;
1678 goto out;
1679 }
1680
1681 while (iov_iter_count(i) > 0) {
1682 struct extent_state *cached_state = NULL;
1683 size_t offset = offset_in_page(pos);
1684 size_t sector_offset;
1685 size_t write_bytes = min(iov_iter_count(i),
1686 nrptrs * (size_t)PAGE_SIZE -
1687 offset);
1688 size_t num_pages;
1689 size_t reserve_bytes;
1690 size_t dirty_pages;
1691 size_t copied;
1692 size_t dirty_sectors;
1693 size_t num_sectors;
1694 int extents_locked;
1695
1696 /*
1697 * Fault pages before locking them in prepare_pages
1698 * to avoid recursive lock
1699 */
1700 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1701 ret = -EFAULT;
1702 break;
1703 }
1704
1705 only_release_metadata = false;
1706 sector_offset = pos & (fs_info->sectorsize - 1);
1707
1708 extent_changeset_release(data_reserved);
1709 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1710 &data_reserved, pos,
1711 write_bytes);
1712 if (ret < 0) {
1713 /*
1714 * If we don't have to COW at the offset, reserve
1715 * metadata only. write_bytes may get smaller than
1716 * requested here.
1717 */
1718 if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1719 &write_bytes) > 0)
1720 only_release_metadata = true;
1721 else
1722 break;
1723 }
1724
1725 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1726 WARN_ON(num_pages > nrptrs);
1727 reserve_bytes = round_up(write_bytes + sector_offset,
1728 fs_info->sectorsize);
1729 WARN_ON(reserve_bytes == 0);
1730 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1731 reserve_bytes);
1732 if (ret) {
1733 if (!only_release_metadata)
1734 btrfs_free_reserved_data_space(BTRFS_I(inode),
1735 data_reserved, pos,
1736 write_bytes);
1737 else
1738 btrfs_check_nocow_unlock(BTRFS_I(inode));
1739 break;
1740 }
1741
1742 release_bytes = reserve_bytes;
1743again:
1744 /*
1745 * This is going to setup the pages array with the number of
1746 * pages we want, so we don't really need to worry about the
1747 * contents of pages from loop to loop
1748 */
1749 ret = prepare_pages(inode, pages, num_pages,
1750 pos, write_bytes,
1751 force_page_uptodate);
1752 if (ret) {
1753 btrfs_delalloc_release_extents(BTRFS_I(inode),
1754 reserve_bytes);
1755 break;
1756 }
1757
1758 extents_locked = lock_and_cleanup_extent_if_need(
1759 BTRFS_I(inode), pages,
1760 num_pages, pos, write_bytes, &lockstart,
1761 &lockend, &cached_state);
1762 if (extents_locked < 0) {
1763 if (extents_locked == -EAGAIN)
1764 goto again;
1765 btrfs_delalloc_release_extents(BTRFS_I(inode),
1766 reserve_bytes);
1767 ret = extents_locked;
1768 break;
1769 }
1770
1771 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1772
1773 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1774 dirty_sectors = round_up(copied + sector_offset,
1775 fs_info->sectorsize);
1776 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1777
1778 /*
1779 * if we have trouble faulting in the pages, fall
1780 * back to one page at a time
1781 */
1782 if (copied < write_bytes)
1783 nrptrs = 1;
1784
1785 if (copied == 0) {
1786 force_page_uptodate = true;
1787 dirty_sectors = 0;
1788 dirty_pages = 0;
1789 } else {
1790 force_page_uptodate = false;
1791 dirty_pages = DIV_ROUND_UP(copied + offset,
1792 PAGE_SIZE);
1793 }
1794
1795 if (num_sectors > dirty_sectors) {
1796 /* release everything except the sectors we dirtied */
1797 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1798 if (only_release_metadata) {
1799 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1800 release_bytes, true);
1801 } else {
1802 u64 __pos;
1803
1804 __pos = round_down(pos,
1805 fs_info->sectorsize) +
1806 (dirty_pages << PAGE_SHIFT);
1807 btrfs_delalloc_release_space(BTRFS_I(inode),
1808 data_reserved, __pos,
1809 release_bytes, true);
1810 }
1811 }
1812
1813 release_bytes = round_up(copied + sector_offset,
1814 fs_info->sectorsize);
1815
1816 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1817 dirty_pages, pos, copied,
1818 &cached_state, only_release_metadata);
1819
1820 /*
1821 * If we have not locked the extent range, because the range's
1822 * start offset is >= i_size, we might still have a non-NULL
1823 * cached extent state, acquired while marking the extent range
1824 * as delalloc through btrfs_dirty_pages(). Therefore free any
1825 * possible cached extent state to avoid a memory leak.
1826 */
1827 if (extents_locked)
1828 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1829 lockstart, lockend, &cached_state);
1830 else
1831 free_extent_state(cached_state);
1832
1833 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1834 if (ret) {
1835 btrfs_drop_pages(pages, num_pages);
1836 break;
1837 }
1838
1839 release_bytes = 0;
1840 if (only_release_metadata)
1841 btrfs_check_nocow_unlock(BTRFS_I(inode));
1842
1843 btrfs_drop_pages(pages, num_pages);
1844
1845 cond_resched();
1846
1847 balance_dirty_pages_ratelimited(inode->i_mapping);
1848
1849 pos += copied;
1850 num_written += copied;
1851 }
1852
1853 kfree(pages);
1854
1855 if (release_bytes) {
1856 if (only_release_metadata) {
1857 btrfs_check_nocow_unlock(BTRFS_I(inode));
1858 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1859 release_bytes, true);
1860 } else {
1861 btrfs_delalloc_release_space(BTRFS_I(inode),
1862 data_reserved,
1863 round_down(pos, fs_info->sectorsize),
1864 release_bytes, true);
1865 }
1866 }
1867
1868 extent_changeset_free(data_reserved);
1869 if (num_written > 0) {
1870 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1871 iocb->ki_pos += num_written;
1872 }
1873out:
1874 btrfs_inode_unlock(inode, ilock_flags);
1875 return num_written ? num_written : ret;
1876}
1877
1878static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1879 const struct iov_iter *iter, loff_t offset)
1880{
1881 const u32 blocksize_mask = fs_info->sectorsize - 1;
1882
1883 if (offset & blocksize_mask)
1884 return -EINVAL;
1885
1886 if (iov_iter_alignment(iter) & blocksize_mask)
1887 return -EINVAL;
1888
1889 return 0;
1890}
1891
1892static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1893{
1894 struct file *file = iocb->ki_filp;
1895 struct inode *inode = file_inode(file);
1896 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1897 loff_t pos;
1898 ssize_t written = 0;
1899 ssize_t written_buffered;
1900 loff_t endbyte;
1901 ssize_t err;
1902 unsigned int ilock_flags = 0;
1903 struct iomap_dio *dio = NULL;
1904
1905 if (iocb->ki_flags & IOCB_NOWAIT)
1906 ilock_flags |= BTRFS_ILOCK_TRY;
1907
1908 /* If the write DIO is within EOF, use a shared lock */
1909 if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1910 ilock_flags |= BTRFS_ILOCK_SHARED;
1911
1912relock:
1913 err = btrfs_inode_lock(inode, ilock_flags);
1914 if (err < 0)
1915 return err;
1916
1917 err = generic_write_checks(iocb, from);
1918 if (err <= 0) {
1919 btrfs_inode_unlock(inode, ilock_flags);
1920 return err;
1921 }
1922
1923 err = btrfs_write_check(iocb, from, err);
1924 if (err < 0) {
1925 btrfs_inode_unlock(inode, ilock_flags);
1926 goto out;
1927 }
1928
1929 pos = iocb->ki_pos;
1930 /*
1931 * Re-check since file size may have changed just before taking the
1932 * lock or pos may have changed because of O_APPEND in generic_write_check()
1933 */
1934 if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1935 pos + iov_iter_count(from) > i_size_read(inode)) {
1936 btrfs_inode_unlock(inode, ilock_flags);
1937 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1938 goto relock;
1939 }
1940
1941 if (check_direct_IO(fs_info, from, pos)) {
1942 btrfs_inode_unlock(inode, ilock_flags);
1943 goto buffered;
1944 }
1945
1946 dio = __iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
1947 0);
1948
1949 btrfs_inode_unlock(inode, ilock_flags);
1950
1951 if (IS_ERR_OR_NULL(dio)) {
1952 err = PTR_ERR_OR_ZERO(dio);
1953 if (err < 0 && err != -ENOTBLK)
1954 goto out;
1955 } else {
1956 written = iomap_dio_complete(dio);
1957 }
1958
1959 if (written < 0 || !iov_iter_count(from)) {
1960 err = written;
1961 goto out;
1962 }
1963
1964buffered:
1965 pos = iocb->ki_pos;
1966 written_buffered = btrfs_buffered_write(iocb, from);
1967 if (written_buffered < 0) {
1968 err = written_buffered;
1969 goto out;
1970 }
1971 /*
1972 * Ensure all data is persisted. We want the next direct IO read to be
1973 * able to read what was just written.
1974 */
1975 endbyte = pos + written_buffered - 1;
1976 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1977 if (err)
1978 goto out;
1979 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1980 if (err)
1981 goto out;
1982 written += written_buffered;
1983 iocb->ki_pos = pos + written_buffered;
1984 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1985 endbyte >> PAGE_SHIFT);
1986out:
1987 return written ? written : err;
1988}
1989
1990static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1991 struct iov_iter *from)
1992{
1993 struct file *file = iocb->ki_filp;
1994 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1995 ssize_t num_written = 0;
1996 const bool sync = iocb->ki_flags & IOCB_DSYNC;
1997
1998 /*
1999 * If the fs flips readonly due to some impossible error, although we
2000 * have opened a file as writable, we have to stop this write operation
2001 * to ensure consistency.
2002 */
2003 if (test_bit(BTRFS_FS_STATE_ERROR, &inode->root->fs_info->fs_state))
2004 return -EROFS;
2005
2006 if (!(iocb->ki_flags & IOCB_DIRECT) &&
2007 (iocb->ki_flags & IOCB_NOWAIT))
2008 return -EOPNOTSUPP;
2009
2010 if (sync)
2011 atomic_inc(&inode->sync_writers);
2012
2013 if (iocb->ki_flags & IOCB_DIRECT)
2014 num_written = btrfs_direct_write(iocb, from);
2015 else
2016 num_written = btrfs_buffered_write(iocb, from);
2017
2018 btrfs_set_inode_last_sub_trans(inode);
2019
2020 if (num_written > 0)
2021 num_written = generic_write_sync(iocb, num_written);
2022
2023 if (sync)
2024 atomic_dec(&inode->sync_writers);
2025
2026 current->backing_dev_info = NULL;
2027 return num_written;
2028}
2029
2030int btrfs_release_file(struct inode *inode, struct file *filp)
2031{
2032 struct btrfs_file_private *private = filp->private_data;
2033
2034 if (private && private->filldir_buf)
2035 kfree(private->filldir_buf);
2036 kfree(private);
2037 filp->private_data = NULL;
2038
2039 /*
2040 * Set by setattr when we are about to truncate a file from a non-zero
2041 * size to a zero size. This tries to flush down new bytes that may
2042 * have been written if the application were using truncate to replace
2043 * a file in place.
2044 */
2045 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2046 &BTRFS_I(inode)->runtime_flags))
2047 filemap_flush(inode->i_mapping);
2048 return 0;
2049}
2050
2051static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2052{
2053 int ret;
2054 struct blk_plug plug;
2055
2056 /*
2057 * This is only called in fsync, which would do synchronous writes, so
2058 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2059 * multiple disks using raid profile, a large IO can be split to
2060 * several segments of stripe length (currently 64K).
2061 */
2062 blk_start_plug(&plug);
2063 atomic_inc(&BTRFS_I(inode)->sync_writers);
2064 ret = btrfs_fdatawrite_range(inode, start, end);
2065 atomic_dec(&BTRFS_I(inode)->sync_writers);
2066 blk_finish_plug(&plug);
2067
2068 return ret;
2069}
2070
2071static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
2072{
2073 struct btrfs_inode *inode = BTRFS_I(ctx->inode);
2074 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2075
2076 if (btrfs_inode_in_log(inode, fs_info->generation) &&
2077 list_empty(&ctx->ordered_extents))
2078 return true;
2079
2080 /*
2081 * If we are doing a fast fsync we can not bail out if the inode's
2082 * last_trans is <= then the last committed transaction, because we only
2083 * update the last_trans of the inode during ordered extent completion,
2084 * and for a fast fsync we don't wait for that, we only wait for the
2085 * writeback to complete.
2086 */
2087 if (inode->last_trans <= fs_info->last_trans_committed &&
2088 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
2089 list_empty(&ctx->ordered_extents)))
2090 return true;
2091
2092 return false;
2093}
2094
2095/*
2096 * fsync call for both files and directories. This logs the inode into
2097 * the tree log instead of forcing full commits whenever possible.
2098 *
2099 * It needs to call filemap_fdatawait so that all ordered extent updates are
2100 * in the metadata btree are up to date for copying to the log.
2101 *
2102 * It drops the inode mutex before doing the tree log commit. This is an
2103 * important optimization for directories because holding the mutex prevents
2104 * new operations on the dir while we write to disk.
2105 */
2106int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2107{
2108 struct dentry *dentry = file_dentry(file);
2109 struct inode *inode = d_inode(dentry);
2110 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2111 struct btrfs_root *root = BTRFS_I(inode)->root;
2112 struct btrfs_trans_handle *trans;
2113 struct btrfs_log_ctx ctx;
2114 int ret = 0, err;
2115 u64 len;
2116 bool full_sync;
2117
2118 trace_btrfs_sync_file(file, datasync);
2119
2120 btrfs_init_log_ctx(&ctx, inode);
2121
2122 /*
2123 * Always set the range to a full range, otherwise we can get into
2124 * several problems, from missing file extent items to represent holes
2125 * when not using the NO_HOLES feature, to log tree corruption due to
2126 * races between hole detection during logging and completion of ordered
2127 * extents outside the range, to missing checksums due to ordered extents
2128 * for which we flushed only a subset of their pages.
2129 */
2130 start = 0;
2131 end = LLONG_MAX;
2132 len = (u64)LLONG_MAX + 1;
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 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2145
2146 atomic_inc(&root->log_batch);
2147
2148 /*
2149 * Always check for the full sync flag while holding the inode's lock,
2150 * to avoid races with other tasks. The flag must be either set all the
2151 * time during logging or always off all the time while logging.
2152 */
2153 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2154 &BTRFS_I(inode)->runtime_flags);
2155
2156 /*
2157 * Before we acquired the inode's lock and the mmap lock, someone may
2158 * have dirtied more pages in the target range. We need to make sure
2159 * that writeback for any such pages does not start while we are logging
2160 * the inode, because if it does, any of the following might happen when
2161 * we are not doing a full inode sync:
2162 *
2163 * 1) We log an extent after its writeback finishes but before its
2164 * checksums are added to the csum tree, leading to -EIO errors
2165 * when attempting to read the extent after a log replay.
2166 *
2167 * 2) We can end up logging an extent before its writeback finishes.
2168 * Therefore after the log replay we will have a file extent item
2169 * pointing to an unwritten extent (and no data checksums as well).
2170 *
2171 * So trigger writeback for any eventual new dirty pages and then we
2172 * wait for all ordered extents to complete below.
2173 */
2174 ret = start_ordered_ops(inode, start, end);
2175 if (ret) {
2176 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2177 goto out;
2178 }
2179
2180 /*
2181 * We have to do this here to avoid the priority inversion of waiting on
2182 * IO of a lower priority task while holding a transaction open.
2183 *
2184 * For a full fsync we wait for the ordered extents to complete while
2185 * for a fast fsync we wait just for writeback to complete, and then
2186 * attach the ordered extents to the transaction so that a transaction
2187 * commit waits for their completion, to avoid data loss if we fsync,
2188 * the current transaction commits before the ordered extents complete
2189 * and a power failure happens right after that.
2190 *
2191 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
2192 * logical address recorded in the ordered extent may change. We need
2193 * to wait for the IO to stabilize the logical address.
2194 */
2195 if (full_sync || btrfs_is_zoned(fs_info)) {
2196 ret = btrfs_wait_ordered_range(inode, start, len);
2197 } else {
2198 /*
2199 * Get our ordered extents as soon as possible to avoid doing
2200 * checksum lookups in the csum tree, and use instead the
2201 * checksums attached to the ordered extents.
2202 */
2203 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2204 &ctx.ordered_extents);
2205 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2206 }
2207
2208 if (ret)
2209 goto out_release_extents;
2210
2211 atomic_inc(&root->log_batch);
2212
2213 smp_mb();
2214 if (skip_inode_logging(&ctx)) {
2215 /*
2216 * We've had everything committed since the last time we were
2217 * modified so clear this flag in case it was set for whatever
2218 * reason, it's no longer relevant.
2219 */
2220 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2221 &BTRFS_I(inode)->runtime_flags);
2222 /*
2223 * An ordered extent might have started before and completed
2224 * already with io errors, in which case the inode was not
2225 * updated and we end up here. So check the inode's mapping
2226 * for any errors that might have happened since we last
2227 * checked called fsync.
2228 */
2229 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2230 goto out_release_extents;
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 goto out_release_extents;
2248 }
2249 trans->in_fsync = true;
2250
2251 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2252 btrfs_release_log_ctx_extents(&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 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2269
2270 if (ret != BTRFS_NO_LOG_SYNC) {
2271 if (!ret) {
2272 ret = btrfs_sync_log(trans, root, &ctx);
2273 if (!ret) {
2274 ret = btrfs_end_transaction(trans);
2275 goto out;
2276 }
2277 }
2278 if (!full_sync) {
2279 ret = btrfs_wait_ordered_range(inode, start, len);
2280 if (ret) {
2281 btrfs_end_transaction(trans);
2282 goto out;
2283 }
2284 }
2285 ret = btrfs_commit_transaction(trans);
2286 } else {
2287 ret = btrfs_end_transaction(trans);
2288 }
2289out:
2290 ASSERT(list_empty(&ctx.list));
2291 err = file_check_and_advance_wb_err(file);
2292 if (!ret)
2293 ret = err;
2294 return ret > 0 ? -EIO : ret;
2295
2296out_release_extents:
2297 btrfs_release_log_ctx_extents(&ctx);
2298 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2299 goto out;
2300}
2301
2302static const struct vm_operations_struct btrfs_file_vm_ops = {
2303 .fault = filemap_fault,
2304 .map_pages = filemap_map_pages,
2305 .page_mkwrite = btrfs_page_mkwrite,
2306};
2307
2308static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2309{
2310 struct address_space *mapping = filp->f_mapping;
2311
2312 if (!mapping->a_ops->readpage)
2313 return -ENOEXEC;
2314
2315 file_accessed(filp);
2316 vma->vm_ops = &btrfs_file_vm_ops;
2317
2318 return 0;
2319}
2320
2321static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2322 int slot, u64 start, u64 end)
2323{
2324 struct btrfs_file_extent_item *fi;
2325 struct btrfs_key key;
2326
2327 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2328 return 0;
2329
2330 btrfs_item_key_to_cpu(leaf, &key, slot);
2331 if (key.objectid != btrfs_ino(inode) ||
2332 key.type != BTRFS_EXTENT_DATA_KEY)
2333 return 0;
2334
2335 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2336
2337 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2338 return 0;
2339
2340 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2341 return 0;
2342
2343 if (key.offset == end)
2344 return 1;
2345 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2346 return 1;
2347 return 0;
2348}
2349
2350static int fill_holes(struct btrfs_trans_handle *trans,
2351 struct btrfs_inode *inode,
2352 struct btrfs_path *path, u64 offset, u64 end)
2353{
2354 struct btrfs_fs_info *fs_info = trans->fs_info;
2355 struct btrfs_root *root = inode->root;
2356 struct extent_buffer *leaf;
2357 struct btrfs_file_extent_item *fi;
2358 struct extent_map *hole_em;
2359 struct extent_map_tree *em_tree = &inode->extent_tree;
2360 struct btrfs_key key;
2361 int ret;
2362
2363 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2364 goto out;
2365
2366 key.objectid = btrfs_ino(inode);
2367 key.type = BTRFS_EXTENT_DATA_KEY;
2368 key.offset = offset;
2369
2370 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2371 if (ret <= 0) {
2372 /*
2373 * We should have dropped this offset, so if we find it then
2374 * something has gone horribly wrong.
2375 */
2376 if (ret == 0)
2377 ret = -EINVAL;
2378 return ret;
2379 }
2380
2381 leaf = path->nodes[0];
2382 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2383 u64 num_bytes;
2384
2385 path->slots[0]--;
2386 fi = btrfs_item_ptr(leaf, path->slots[0],
2387 struct btrfs_file_extent_item);
2388 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2389 end - offset;
2390 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2391 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2392 btrfs_set_file_extent_offset(leaf, fi, 0);
2393 btrfs_mark_buffer_dirty(leaf);
2394 goto out;
2395 }
2396
2397 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2398 u64 num_bytes;
2399
2400 key.offset = offset;
2401 btrfs_set_item_key_safe(fs_info, path, &key);
2402 fi = btrfs_item_ptr(leaf, path->slots[0],
2403 struct btrfs_file_extent_item);
2404 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2405 offset;
2406 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2407 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2408 btrfs_set_file_extent_offset(leaf, fi, 0);
2409 btrfs_mark_buffer_dirty(leaf);
2410 goto out;
2411 }
2412 btrfs_release_path(path);
2413
2414 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2415 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2416 if (ret)
2417 return ret;
2418
2419out:
2420 btrfs_release_path(path);
2421
2422 hole_em = alloc_extent_map();
2423 if (!hole_em) {
2424 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2425 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2426 } else {
2427 hole_em->start = offset;
2428 hole_em->len = end - offset;
2429 hole_em->ram_bytes = hole_em->len;
2430 hole_em->orig_start = offset;
2431
2432 hole_em->block_start = EXTENT_MAP_HOLE;
2433 hole_em->block_len = 0;
2434 hole_em->orig_block_len = 0;
2435 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2436 hole_em->generation = trans->transid;
2437
2438 do {
2439 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2440 write_lock(&em_tree->lock);
2441 ret = add_extent_mapping(em_tree, hole_em, 1);
2442 write_unlock(&em_tree->lock);
2443 } while (ret == -EEXIST);
2444 free_extent_map(hole_em);
2445 if (ret)
2446 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2447 &inode->runtime_flags);
2448 }
2449
2450 return 0;
2451}
2452
2453/*
2454 * Find a hole extent on given inode and change start/len to the end of hole
2455 * extent.(hole/vacuum extent whose em->start <= start &&
2456 * em->start + em->len > start)
2457 * When a hole extent is found, return 1 and modify start/len.
2458 */
2459static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2460{
2461 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2462 struct extent_map *em;
2463 int ret = 0;
2464
2465 em = btrfs_get_extent(inode, NULL, 0,
2466 round_down(*start, fs_info->sectorsize),
2467 round_up(*len, fs_info->sectorsize));
2468 if (IS_ERR(em))
2469 return PTR_ERR(em);
2470
2471 /* Hole or vacuum extent(only exists in no-hole mode) */
2472 if (em->block_start == EXTENT_MAP_HOLE) {
2473 ret = 1;
2474 *len = em->start + em->len > *start + *len ?
2475 0 : *start + *len - em->start - em->len;
2476 *start = em->start + em->len;
2477 }
2478 free_extent_map(em);
2479 return ret;
2480}
2481
2482static int btrfs_punch_hole_lock_range(struct inode *inode,
2483 const u64 lockstart,
2484 const u64 lockend,
2485 struct extent_state **cached_state)
2486{
2487 /*
2488 * For subpage case, if the range is not at page boundary, we could
2489 * have pages at the leading/tailing part of the range.
2490 * This could lead to dead loop since filemap_range_has_page()
2491 * will always return true.
2492 * So here we need to do extra page alignment for
2493 * filemap_range_has_page().
2494 */
2495 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2496 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2497
2498 while (1) {
2499 struct btrfs_ordered_extent *ordered;
2500 int ret;
2501
2502 truncate_pagecache_range(inode, lockstart, lockend);
2503
2504 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2505 cached_state);
2506 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2507 lockend);
2508
2509 /*
2510 * We need to make sure we have no ordered extents in this range
2511 * and nobody raced in and read a page in this range, if we did
2512 * we need to try again.
2513 */
2514 if ((!ordered ||
2515 (ordered->file_offset + ordered->num_bytes <= lockstart ||
2516 ordered->file_offset > lockend)) &&
2517 !filemap_range_has_page(inode->i_mapping,
2518 page_lockstart, page_lockend)) {
2519 if (ordered)
2520 btrfs_put_ordered_extent(ordered);
2521 break;
2522 }
2523 if (ordered)
2524 btrfs_put_ordered_extent(ordered);
2525 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2526 lockend, cached_state);
2527 ret = btrfs_wait_ordered_range(inode, lockstart,
2528 lockend - lockstart + 1);
2529 if (ret)
2530 return ret;
2531 }
2532 return 0;
2533}
2534
2535static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2536 struct btrfs_inode *inode,
2537 struct btrfs_path *path,
2538 struct btrfs_replace_extent_info *extent_info,
2539 const u64 replace_len,
2540 const u64 bytes_to_drop)
2541{
2542 struct btrfs_fs_info *fs_info = trans->fs_info;
2543 struct btrfs_root *root = inode->root;
2544 struct btrfs_file_extent_item *extent;
2545 struct extent_buffer *leaf;
2546 struct btrfs_key key;
2547 int slot;
2548 struct btrfs_ref ref = { 0 };
2549 int ret;
2550
2551 if (replace_len == 0)
2552 return 0;
2553
2554 if (extent_info->disk_offset == 0 &&
2555 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2556 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2557 return 0;
2558 }
2559
2560 key.objectid = btrfs_ino(inode);
2561 key.type = BTRFS_EXTENT_DATA_KEY;
2562 key.offset = extent_info->file_offset;
2563 ret = btrfs_insert_empty_item(trans, root, path, &key,
2564 sizeof(struct btrfs_file_extent_item));
2565 if (ret)
2566 return ret;
2567 leaf = path->nodes[0];
2568 slot = path->slots[0];
2569 write_extent_buffer(leaf, extent_info->extent_buf,
2570 btrfs_item_ptr_offset(leaf, slot),
2571 sizeof(struct btrfs_file_extent_item));
2572 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2573 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2574 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2575 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2576 if (extent_info->is_new_extent)
2577 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2578 btrfs_mark_buffer_dirty(leaf);
2579 btrfs_release_path(path);
2580
2581 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2582 replace_len);
2583 if (ret)
2584 return ret;
2585
2586 /* If it's a hole, nothing more needs to be done. */
2587 if (extent_info->disk_offset == 0) {
2588 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2589 return 0;
2590 }
2591
2592 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2593
2594 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2595 key.objectid = extent_info->disk_offset;
2596 key.type = BTRFS_EXTENT_ITEM_KEY;
2597 key.offset = extent_info->disk_len;
2598 ret = btrfs_alloc_reserved_file_extent(trans, root,
2599 btrfs_ino(inode),
2600 extent_info->file_offset,
2601 extent_info->qgroup_reserved,
2602 &key);
2603 } else {
2604 u64 ref_offset;
2605
2606 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2607 extent_info->disk_offset,
2608 extent_info->disk_len, 0);
2609 ref_offset = extent_info->file_offset - extent_info->data_offset;
2610 btrfs_init_data_ref(&ref, root->root_key.objectid,
2611 btrfs_ino(inode), ref_offset);
2612 ret = btrfs_inc_extent_ref(trans, &ref);
2613 }
2614
2615 extent_info->insertions++;
2616
2617 return ret;
2618}
2619
2620/*
2621 * The respective range must have been previously locked, as well as the inode.
2622 * The end offset is inclusive (last byte of the range).
2623 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2624 * the file range with an extent.
2625 * When not punching a hole, we don't want to end up in a state where we dropped
2626 * extents without inserting a new one, so we must abort the transaction to avoid
2627 * a corruption.
2628 */
2629int btrfs_replace_file_extents(struct btrfs_inode *inode,
2630 struct btrfs_path *path, const u64 start,
2631 const u64 end,
2632 struct btrfs_replace_extent_info *extent_info,
2633 struct btrfs_trans_handle **trans_out)
2634{
2635 struct btrfs_drop_extents_args drop_args = { 0 };
2636 struct btrfs_root *root = inode->root;
2637 struct btrfs_fs_info *fs_info = root->fs_info;
2638 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2639 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2640 struct btrfs_trans_handle *trans = NULL;
2641 struct btrfs_block_rsv *rsv;
2642 unsigned int rsv_count;
2643 u64 cur_offset;
2644 u64 len = end - start;
2645 int ret = 0;
2646
2647 if (end <= start)
2648 return -EINVAL;
2649
2650 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2651 if (!rsv) {
2652 ret = -ENOMEM;
2653 goto out;
2654 }
2655 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2656 rsv->failfast = 1;
2657
2658 /*
2659 * 1 - update the inode
2660 * 1 - removing the extents in the range
2661 * 1 - adding the hole extent if no_holes isn't set or if we are
2662 * replacing the range with a new extent
2663 */
2664 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2665 rsv_count = 3;
2666 else
2667 rsv_count = 2;
2668
2669 trans = btrfs_start_transaction(root, rsv_count);
2670 if (IS_ERR(trans)) {
2671 ret = PTR_ERR(trans);
2672 trans = NULL;
2673 goto out_free;
2674 }
2675
2676 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2677 min_size, false);
2678 BUG_ON(ret);
2679 trans->block_rsv = rsv;
2680
2681 cur_offset = start;
2682 drop_args.path = path;
2683 drop_args.end = end + 1;
2684 drop_args.drop_cache = true;
2685 while (cur_offset < end) {
2686 drop_args.start = cur_offset;
2687 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2688 /* If we are punching a hole decrement the inode's byte count */
2689 if (!extent_info)
2690 btrfs_update_inode_bytes(inode, 0,
2691 drop_args.bytes_found);
2692 if (ret != -ENOSPC) {
2693 /*
2694 * The only time we don't want to abort is if we are
2695 * attempting to clone a partial inline extent, in which
2696 * case we'll get EOPNOTSUPP. However if we aren't
2697 * clone we need to abort no matter what, because if we
2698 * got EOPNOTSUPP via prealloc then we messed up and
2699 * need to abort.
2700 */
2701 if (ret &&
2702 (ret != -EOPNOTSUPP ||
2703 (extent_info && extent_info->is_new_extent)))
2704 btrfs_abort_transaction(trans, ret);
2705 break;
2706 }
2707
2708 trans->block_rsv = &fs_info->trans_block_rsv;
2709
2710 if (!extent_info && cur_offset < drop_args.drop_end &&
2711 cur_offset < ino_size) {
2712 ret = fill_holes(trans, inode, path, cur_offset,
2713 drop_args.drop_end);
2714 if (ret) {
2715 /*
2716 * If we failed then we didn't insert our hole
2717 * entries for the area we dropped, so now the
2718 * fs is corrupted, so we must abort the
2719 * transaction.
2720 */
2721 btrfs_abort_transaction(trans, ret);
2722 break;
2723 }
2724 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2725 /*
2726 * We are past the i_size here, but since we didn't
2727 * insert holes we need to clear the mapped area so we
2728 * know to not set disk_i_size in this area until a new
2729 * file extent is inserted here.
2730 */
2731 ret = btrfs_inode_clear_file_extent_range(inode,
2732 cur_offset,
2733 drop_args.drop_end - cur_offset);
2734 if (ret) {
2735 /*
2736 * We couldn't clear our area, so we could
2737 * presumably adjust up and corrupt the fs, so
2738 * we need to abort.
2739 */
2740 btrfs_abort_transaction(trans, ret);
2741 break;
2742 }
2743 }
2744
2745 if (extent_info &&
2746 drop_args.drop_end > extent_info->file_offset) {
2747 u64 replace_len = drop_args.drop_end -
2748 extent_info->file_offset;
2749
2750 ret = btrfs_insert_replace_extent(trans, inode, path,
2751 extent_info, replace_len,
2752 drop_args.bytes_found);
2753 if (ret) {
2754 btrfs_abort_transaction(trans, ret);
2755 break;
2756 }
2757 extent_info->data_len -= replace_len;
2758 extent_info->data_offset += replace_len;
2759 extent_info->file_offset += replace_len;
2760 }
2761
2762 ret = btrfs_update_inode(trans, root, inode);
2763 if (ret)
2764 break;
2765
2766 btrfs_end_transaction(trans);
2767 btrfs_btree_balance_dirty(fs_info);
2768
2769 trans = btrfs_start_transaction(root, rsv_count);
2770 if (IS_ERR(trans)) {
2771 ret = PTR_ERR(trans);
2772 trans = NULL;
2773 break;
2774 }
2775
2776 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2777 rsv, min_size, false);
2778 BUG_ON(ret); /* shouldn't happen */
2779 trans->block_rsv = rsv;
2780
2781 cur_offset = drop_args.drop_end;
2782 len = end - cur_offset;
2783 if (!extent_info && len) {
2784 ret = find_first_non_hole(inode, &cur_offset, &len);
2785 if (unlikely(ret < 0))
2786 break;
2787 if (ret && !len) {
2788 ret = 0;
2789 break;
2790 }
2791 }
2792 }
2793
2794 /*
2795 * If we were cloning, force the next fsync to be a full one since we
2796 * we replaced (or just dropped in the case of cloning holes when
2797 * NO_HOLES is enabled) file extent items and did not setup new extent
2798 * maps for the replacement extents (or holes).
2799 */
2800 if (extent_info && !extent_info->is_new_extent)
2801 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2802
2803 if (ret)
2804 goto out_trans;
2805
2806 trans->block_rsv = &fs_info->trans_block_rsv;
2807 /*
2808 * If we are using the NO_HOLES feature we might have had already an
2809 * hole that overlaps a part of the region [lockstart, lockend] and
2810 * ends at (or beyond) lockend. Since we have no file extent items to
2811 * represent holes, drop_end can be less than lockend and so we must
2812 * make sure we have an extent map representing the existing hole (the
2813 * call to __btrfs_drop_extents() might have dropped the existing extent
2814 * map representing the existing hole), otherwise the fast fsync path
2815 * will not record the existence of the hole region
2816 * [existing_hole_start, lockend].
2817 */
2818 if (drop_args.drop_end <= end)
2819 drop_args.drop_end = end + 1;
2820 /*
2821 * Don't insert file hole extent item if it's for a range beyond eof
2822 * (because it's useless) or if it represents a 0 bytes range (when
2823 * cur_offset == drop_end).
2824 */
2825 if (!extent_info && cur_offset < ino_size &&
2826 cur_offset < drop_args.drop_end) {
2827 ret = fill_holes(trans, inode, path, cur_offset,
2828 drop_args.drop_end);
2829 if (ret) {
2830 /* Same comment as above. */
2831 btrfs_abort_transaction(trans, ret);
2832 goto out_trans;
2833 }
2834 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2835 /* See the comment in the loop above for the reasoning here. */
2836 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2837 drop_args.drop_end - cur_offset);
2838 if (ret) {
2839 btrfs_abort_transaction(trans, ret);
2840 goto out_trans;
2841 }
2842
2843 }
2844 if (extent_info) {
2845 ret = btrfs_insert_replace_extent(trans, inode, path,
2846 extent_info, extent_info->data_len,
2847 drop_args.bytes_found);
2848 if (ret) {
2849 btrfs_abort_transaction(trans, ret);
2850 goto out_trans;
2851 }
2852 }
2853
2854out_trans:
2855 if (!trans)
2856 goto out_free;
2857
2858 trans->block_rsv = &fs_info->trans_block_rsv;
2859 if (ret)
2860 btrfs_end_transaction(trans);
2861 else
2862 *trans_out = trans;
2863out_free:
2864 btrfs_free_block_rsv(fs_info, rsv);
2865out:
2866 return ret;
2867}
2868
2869static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2870{
2871 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2872 struct btrfs_root *root = BTRFS_I(inode)->root;
2873 struct extent_state *cached_state = NULL;
2874 struct btrfs_path *path;
2875 struct btrfs_trans_handle *trans = NULL;
2876 u64 lockstart;
2877 u64 lockend;
2878 u64 tail_start;
2879 u64 tail_len;
2880 u64 orig_start = offset;
2881 int ret = 0;
2882 bool same_block;
2883 u64 ino_size;
2884 bool truncated_block = false;
2885 bool updated_inode = false;
2886
2887 ret = btrfs_wait_ordered_range(inode, offset, len);
2888 if (ret)
2889 return ret;
2890
2891 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2892 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2893 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2894 if (ret < 0)
2895 goto out_only_mutex;
2896 if (ret && !len) {
2897 /* Already in a large hole */
2898 ret = 0;
2899 goto out_only_mutex;
2900 }
2901
2902 lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2903 lockend = round_down(offset + len,
2904 btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2905 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2906 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2907 /*
2908 * We needn't truncate any block which is beyond the end of the file
2909 * because we are sure there is no data there.
2910 */
2911 /*
2912 * Only do this if we are in the same block and we aren't doing the
2913 * entire block.
2914 */
2915 if (same_block && len < fs_info->sectorsize) {
2916 if (offset < ino_size) {
2917 truncated_block = true;
2918 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2919 0);
2920 } else {
2921 ret = 0;
2922 }
2923 goto out_only_mutex;
2924 }
2925
2926 /* zero back part of the first block */
2927 if (offset < ino_size) {
2928 truncated_block = true;
2929 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2930 if (ret) {
2931 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2932 return ret;
2933 }
2934 }
2935
2936 /* Check the aligned pages after the first unaligned page,
2937 * if offset != orig_start, which means the first unaligned page
2938 * including several following pages are already in holes,
2939 * the extra check can be skipped */
2940 if (offset == orig_start) {
2941 /* after truncate page, check hole again */
2942 len = offset + len - lockstart;
2943 offset = lockstart;
2944 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2945 if (ret < 0)
2946 goto out_only_mutex;
2947 if (ret && !len) {
2948 ret = 0;
2949 goto out_only_mutex;
2950 }
2951 lockstart = offset;
2952 }
2953
2954 /* Check the tail unaligned part is in a hole */
2955 tail_start = lockend + 1;
2956 tail_len = offset + len - tail_start;
2957 if (tail_len) {
2958 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2959 if (unlikely(ret < 0))
2960 goto out_only_mutex;
2961 if (!ret) {
2962 /* zero the front end of the last page */
2963 if (tail_start + tail_len < ino_size) {
2964 truncated_block = true;
2965 ret = btrfs_truncate_block(BTRFS_I(inode),
2966 tail_start + tail_len,
2967 0, 1);
2968 if (ret)
2969 goto out_only_mutex;
2970 }
2971 }
2972 }
2973
2974 if (lockend < lockstart) {
2975 ret = 0;
2976 goto out_only_mutex;
2977 }
2978
2979 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2980 &cached_state);
2981 if (ret)
2982 goto out_only_mutex;
2983
2984 path = btrfs_alloc_path();
2985 if (!path) {
2986 ret = -ENOMEM;
2987 goto out;
2988 }
2989
2990 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
2991 lockend, NULL, &trans);
2992 btrfs_free_path(path);
2993 if (ret)
2994 goto out;
2995
2996 ASSERT(trans != NULL);
2997 inode_inc_iversion(inode);
2998 inode->i_mtime = inode->i_ctime = current_time(inode);
2999 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3000 updated_inode = true;
3001 btrfs_end_transaction(trans);
3002 btrfs_btree_balance_dirty(fs_info);
3003out:
3004 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3005 &cached_state);
3006out_only_mutex:
3007 if (!updated_inode && truncated_block && !ret) {
3008 /*
3009 * If we only end up zeroing part of a page, we still need to
3010 * update the inode item, so that all the time fields are
3011 * updated as well as the necessary btrfs inode in memory fields
3012 * for detecting, at fsync time, if the inode isn't yet in the
3013 * log tree or it's there but not up to date.
3014 */
3015 struct timespec64 now = current_time(inode);
3016
3017 inode_inc_iversion(inode);
3018 inode->i_mtime = now;
3019 inode->i_ctime = now;
3020 trans = btrfs_start_transaction(root, 1);
3021 if (IS_ERR(trans)) {
3022 ret = PTR_ERR(trans);
3023 } else {
3024 int ret2;
3025
3026 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3027 ret2 = btrfs_end_transaction(trans);
3028 if (!ret)
3029 ret = ret2;
3030 }
3031 }
3032 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3033 return ret;
3034}
3035
3036/* Helper structure to record which range is already reserved */
3037struct falloc_range {
3038 struct list_head list;
3039 u64 start;
3040 u64 len;
3041};
3042
3043/*
3044 * Helper function to add falloc range
3045 *
3046 * Caller should have locked the larger range of extent containing
3047 * [start, len)
3048 */
3049static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3050{
3051 struct falloc_range *range = NULL;
3052
3053 if (!list_empty(head)) {
3054 /*
3055 * As fallocate iterates by bytenr order, we only need to check
3056 * the last range.
3057 */
3058 range = list_last_entry(head, struct falloc_range, list);
3059 if (range->start + range->len == start) {
3060 range->len += len;
3061 return 0;
3062 }
3063 }
3064
3065 range = kmalloc(sizeof(*range), GFP_KERNEL);
3066 if (!range)
3067 return -ENOMEM;
3068 range->start = start;
3069 range->len = len;
3070 list_add_tail(&range->list, head);
3071 return 0;
3072}
3073
3074static int btrfs_fallocate_update_isize(struct inode *inode,
3075 const u64 end,
3076 const int mode)
3077{
3078 struct btrfs_trans_handle *trans;
3079 struct btrfs_root *root = BTRFS_I(inode)->root;
3080 int ret;
3081 int ret2;
3082
3083 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3084 return 0;
3085
3086 trans = btrfs_start_transaction(root, 1);
3087 if (IS_ERR(trans))
3088 return PTR_ERR(trans);
3089
3090 inode->i_ctime = current_time(inode);
3091 i_size_write(inode, end);
3092 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
3093 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3094 ret2 = btrfs_end_transaction(trans);
3095
3096 return ret ? ret : ret2;
3097}
3098
3099enum {
3100 RANGE_BOUNDARY_WRITTEN_EXTENT,
3101 RANGE_BOUNDARY_PREALLOC_EXTENT,
3102 RANGE_BOUNDARY_HOLE,
3103};
3104
3105static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3106 u64 offset)
3107{
3108 const u64 sectorsize = btrfs_inode_sectorsize(inode);
3109 struct extent_map *em;
3110 int ret;
3111
3112 offset = round_down(offset, sectorsize);
3113 em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3114 if (IS_ERR(em))
3115 return PTR_ERR(em);
3116
3117 if (em->block_start == EXTENT_MAP_HOLE)
3118 ret = RANGE_BOUNDARY_HOLE;
3119 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3120 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3121 else
3122 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3123
3124 free_extent_map(em);
3125 return ret;
3126}
3127
3128static int btrfs_zero_range(struct inode *inode,
3129 loff_t offset,
3130 loff_t len,
3131 const int mode)
3132{
3133 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3134 struct extent_map *em;
3135 struct extent_changeset *data_reserved = NULL;
3136 int ret;
3137 u64 alloc_hint = 0;
3138 const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3139 u64 alloc_start = round_down(offset, sectorsize);
3140 u64 alloc_end = round_up(offset + len, sectorsize);
3141 u64 bytes_to_reserve = 0;
3142 bool space_reserved = false;
3143
3144 inode_dio_wait(inode);
3145
3146 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3147 alloc_end - alloc_start);
3148 if (IS_ERR(em)) {
3149 ret = PTR_ERR(em);
3150 goto out;
3151 }
3152
3153 /*
3154 * Avoid hole punching and extent allocation for some cases. More cases
3155 * could be considered, but these are unlikely common and we keep things
3156 * as simple as possible for now. Also, intentionally, if the target
3157 * range contains one or more prealloc extents together with regular
3158 * extents and holes, we drop all the existing extents and allocate a
3159 * new prealloc extent, so that we get a larger contiguous disk extent.
3160 */
3161 if (em->start <= alloc_start &&
3162 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3163 const u64 em_end = em->start + em->len;
3164
3165 if (em_end >= offset + len) {
3166 /*
3167 * The whole range is already a prealloc extent,
3168 * do nothing except updating the inode's i_size if
3169 * needed.
3170 */
3171 free_extent_map(em);
3172 ret = btrfs_fallocate_update_isize(inode, offset + len,
3173 mode);
3174 goto out;
3175 }
3176 /*
3177 * Part of the range is already a prealloc extent, so operate
3178 * only on the remaining part of the range.
3179 */
3180 alloc_start = em_end;
3181 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3182 len = offset + len - alloc_start;
3183 offset = alloc_start;
3184 alloc_hint = em->block_start + em->len;
3185 }
3186 free_extent_map(em);
3187
3188 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3189 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3190 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3191 sectorsize);
3192 if (IS_ERR(em)) {
3193 ret = PTR_ERR(em);
3194 goto out;
3195 }
3196
3197 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3198 free_extent_map(em);
3199 ret = btrfs_fallocate_update_isize(inode, offset + len,
3200 mode);
3201 goto out;
3202 }
3203 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3204 free_extent_map(em);
3205 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3206 0);
3207 if (!ret)
3208 ret = btrfs_fallocate_update_isize(inode,
3209 offset + len,
3210 mode);
3211 return ret;
3212 }
3213 free_extent_map(em);
3214 alloc_start = round_down(offset, sectorsize);
3215 alloc_end = alloc_start + sectorsize;
3216 goto reserve_space;
3217 }
3218
3219 alloc_start = round_up(offset, sectorsize);
3220 alloc_end = round_down(offset + len, sectorsize);
3221
3222 /*
3223 * For unaligned ranges, check the pages at the boundaries, they might
3224 * map to an extent, in which case we need to partially zero them, or
3225 * they might map to a hole, in which case we need our allocation range
3226 * to cover them.
3227 */
3228 if (!IS_ALIGNED(offset, sectorsize)) {
3229 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3230 offset);
3231 if (ret < 0)
3232 goto out;
3233 if (ret == RANGE_BOUNDARY_HOLE) {
3234 alloc_start = round_down(offset, sectorsize);
3235 ret = 0;
3236 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3237 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3238 if (ret)
3239 goto out;
3240 } else {
3241 ret = 0;
3242 }
3243 }
3244
3245 if (!IS_ALIGNED(offset + len, sectorsize)) {
3246 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3247 offset + len);
3248 if (ret < 0)
3249 goto out;
3250 if (ret == RANGE_BOUNDARY_HOLE) {
3251 alloc_end = round_up(offset + len, sectorsize);
3252 ret = 0;
3253 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3254 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
3255 0, 1);
3256 if (ret)
3257 goto out;
3258 } else {
3259 ret = 0;
3260 }
3261 }
3262
3263reserve_space:
3264 if (alloc_start < alloc_end) {
3265 struct extent_state *cached_state = NULL;
3266 const u64 lockstart = alloc_start;
3267 const u64 lockend = alloc_end - 1;
3268
3269 bytes_to_reserve = alloc_end - alloc_start;
3270 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3271 bytes_to_reserve);
3272 if (ret < 0)
3273 goto out;
3274 space_reserved = true;
3275 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3276 &cached_state);
3277 if (ret)
3278 goto out;
3279 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3280 alloc_start, bytes_to_reserve);
3281 if (ret) {
3282 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3283 lockend, &cached_state);
3284 goto out;
3285 }
3286 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3287 alloc_end - alloc_start,
3288 i_blocksize(inode),
3289 offset + len, &alloc_hint);
3290 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3291 lockend, &cached_state);
3292 /* btrfs_prealloc_file_range releases reserved space on error */
3293 if (ret) {
3294 space_reserved = false;
3295 goto out;
3296 }
3297 }
3298 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3299 out:
3300 if (ret && space_reserved)
3301 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3302 alloc_start, bytes_to_reserve);
3303 extent_changeset_free(data_reserved);
3304
3305 return ret;
3306}
3307
3308static long btrfs_fallocate(struct file *file, int mode,
3309 loff_t offset, loff_t len)
3310{
3311 struct inode *inode = file_inode(file);
3312 struct extent_state *cached_state = NULL;
3313 struct extent_changeset *data_reserved = NULL;
3314 struct falloc_range *range;
3315 struct falloc_range *tmp;
3316 struct list_head reserve_list;
3317 u64 cur_offset;
3318 u64 last_byte;
3319 u64 alloc_start;
3320 u64 alloc_end;
3321 u64 alloc_hint = 0;
3322 u64 locked_end;
3323 u64 actual_end = 0;
3324 struct extent_map *em;
3325 int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3326 int ret;
3327
3328 /* Do not allow fallocate in ZONED mode */
3329 if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3330 return -EOPNOTSUPP;
3331
3332 alloc_start = round_down(offset, blocksize);
3333 alloc_end = round_up(offset + len, blocksize);
3334 cur_offset = alloc_start;
3335
3336 /* Make sure we aren't being give some crap mode */
3337 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3338 FALLOC_FL_ZERO_RANGE))
3339 return -EOPNOTSUPP;
3340
3341 if (mode & FALLOC_FL_PUNCH_HOLE)
3342 return btrfs_punch_hole(inode, offset, len);
3343
3344 /*
3345 * Only trigger disk allocation, don't trigger qgroup reserve
3346 *
3347 * For qgroup space, it will be checked later.
3348 */
3349 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3350 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3351 alloc_end - alloc_start);
3352 if (ret < 0)
3353 return ret;
3354 }
3355
3356 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
3357
3358 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3359 ret = inode_newsize_ok(inode, offset + len);
3360 if (ret)
3361 goto out;
3362 }
3363
3364 /*
3365 * TODO: Move these two operations after we have checked
3366 * accurate reserved space, or fallocate can still fail but
3367 * with page truncated or size expanded.
3368 *
3369 * But that's a minor problem and won't do much harm BTW.
3370 */
3371 if (alloc_start > inode->i_size) {
3372 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3373 alloc_start);
3374 if (ret)
3375 goto out;
3376 } else if (offset + len > inode->i_size) {
3377 /*
3378 * If we are fallocating from the end of the file onward we
3379 * need to zero out the end of the block if i_size lands in the
3380 * middle of a block.
3381 */
3382 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3383 if (ret)
3384 goto out;
3385 }
3386
3387 /*
3388 * wait for ordered IO before we have any locks. We'll loop again
3389 * below with the locks held.
3390 */
3391 ret = btrfs_wait_ordered_range(inode, alloc_start,
3392 alloc_end - alloc_start);
3393 if (ret)
3394 goto out;
3395
3396 if (mode & FALLOC_FL_ZERO_RANGE) {
3397 ret = btrfs_zero_range(inode, offset, len, mode);
3398 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3399 return ret;
3400 }
3401
3402 locked_end = alloc_end - 1;
3403 while (1) {
3404 struct btrfs_ordered_extent *ordered;
3405
3406 /* the extent lock is ordered inside the running
3407 * transaction
3408 */
3409 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3410 locked_end, &cached_state);
3411 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3412 locked_end);
3413
3414 if (ordered &&
3415 ordered->file_offset + ordered->num_bytes > alloc_start &&
3416 ordered->file_offset < alloc_end) {
3417 btrfs_put_ordered_extent(ordered);
3418 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3419 alloc_start, locked_end,
3420 &cached_state);
3421 /*
3422 * we can't wait on the range with the transaction
3423 * running or with the extent lock held
3424 */
3425 ret = btrfs_wait_ordered_range(inode, alloc_start,
3426 alloc_end - alloc_start);
3427 if (ret)
3428 goto out;
3429 } else {
3430 if (ordered)
3431 btrfs_put_ordered_extent(ordered);
3432 break;
3433 }
3434 }
3435
3436 /* First, check if we exceed the qgroup limit */
3437 INIT_LIST_HEAD(&reserve_list);
3438 while (cur_offset < alloc_end) {
3439 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3440 alloc_end - cur_offset);
3441 if (IS_ERR(em)) {
3442 ret = PTR_ERR(em);
3443 break;
3444 }
3445 last_byte = min(extent_map_end(em), alloc_end);
3446 actual_end = min_t(u64, extent_map_end(em), offset + len);
3447 last_byte = ALIGN(last_byte, blocksize);
3448 if (em->block_start == EXTENT_MAP_HOLE ||
3449 (cur_offset >= inode->i_size &&
3450 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3451 ret = add_falloc_range(&reserve_list, cur_offset,
3452 last_byte - cur_offset);
3453 if (ret < 0) {
3454 free_extent_map(em);
3455 break;
3456 }
3457 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3458 &data_reserved, cur_offset,
3459 last_byte - cur_offset);
3460 if (ret < 0) {
3461 cur_offset = last_byte;
3462 free_extent_map(em);
3463 break;
3464 }
3465 } else {
3466 /*
3467 * Do not need to reserve unwritten extent for this
3468 * range, free reserved data space first, otherwise
3469 * it'll result in false ENOSPC error.
3470 */
3471 btrfs_free_reserved_data_space(BTRFS_I(inode),
3472 data_reserved, cur_offset,
3473 last_byte - cur_offset);
3474 }
3475 free_extent_map(em);
3476 cur_offset = last_byte;
3477 }
3478
3479 /*
3480 * If ret is still 0, means we're OK to fallocate.
3481 * Or just cleanup the list and exit.
3482 */
3483 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3484 if (!ret)
3485 ret = btrfs_prealloc_file_range(inode, mode,
3486 range->start,
3487 range->len, i_blocksize(inode),
3488 offset + len, &alloc_hint);
3489 else
3490 btrfs_free_reserved_data_space(BTRFS_I(inode),
3491 data_reserved, range->start,
3492 range->len);
3493 list_del(&range->list);
3494 kfree(range);
3495 }
3496 if (ret < 0)
3497 goto out_unlock;
3498
3499 /*
3500 * We didn't need to allocate any more space, but we still extended the
3501 * size of the file so we need to update i_size and the inode item.
3502 */
3503 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3504out_unlock:
3505 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3506 &cached_state);
3507out:
3508 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3509 /* Let go of our reservation. */
3510 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3511 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3512 cur_offset, alloc_end - cur_offset);
3513 extent_changeset_free(data_reserved);
3514 return ret;
3515}
3516
3517static loff_t find_desired_extent(struct btrfs_inode *inode, loff_t offset,
3518 int whence)
3519{
3520 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3521 struct extent_map *em = NULL;
3522 struct extent_state *cached_state = NULL;
3523 loff_t i_size = inode->vfs_inode.i_size;
3524 u64 lockstart;
3525 u64 lockend;
3526 u64 start;
3527 u64 len;
3528 int ret = 0;
3529
3530 if (i_size == 0 || offset >= i_size)
3531 return -ENXIO;
3532
3533 /*
3534 * offset can be negative, in this case we start finding DATA/HOLE from
3535 * the very start of the file.
3536 */
3537 start = max_t(loff_t, 0, offset);
3538
3539 lockstart = round_down(start, fs_info->sectorsize);
3540 lockend = round_up(i_size, fs_info->sectorsize);
3541 if (lockend <= lockstart)
3542 lockend = lockstart + fs_info->sectorsize;
3543 lockend--;
3544 len = lockend - lockstart + 1;
3545
3546 lock_extent_bits(&inode->io_tree, lockstart, lockend, &cached_state);
3547
3548 while (start < i_size) {
3549 em = btrfs_get_extent_fiemap(inode, start, len);
3550 if (IS_ERR(em)) {
3551 ret = PTR_ERR(em);
3552 em = NULL;
3553 break;
3554 }
3555
3556 if (whence == SEEK_HOLE &&
3557 (em->block_start == EXTENT_MAP_HOLE ||
3558 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3559 break;
3560 else if (whence == SEEK_DATA &&
3561 (em->block_start != EXTENT_MAP_HOLE &&
3562 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3563 break;
3564
3565 start = em->start + em->len;
3566 free_extent_map(em);
3567 em = NULL;
3568 cond_resched();
3569 }
3570 free_extent_map(em);
3571 unlock_extent_cached(&inode->io_tree, lockstart, lockend,
3572 &cached_state);
3573 if (ret) {
3574 offset = ret;
3575 } else {
3576 if (whence == SEEK_DATA && start >= i_size)
3577 offset = -ENXIO;
3578 else
3579 offset = min_t(loff_t, start, i_size);
3580 }
3581
3582 return offset;
3583}
3584
3585static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3586{
3587 struct inode *inode = file->f_mapping->host;
3588
3589 switch (whence) {
3590 default:
3591 return generic_file_llseek(file, offset, whence);
3592 case SEEK_DATA:
3593 case SEEK_HOLE:
3594 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3595 offset = find_desired_extent(BTRFS_I(inode), offset, whence);
3596 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3597 break;
3598 }
3599
3600 if (offset < 0)
3601 return offset;
3602
3603 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3604}
3605
3606static int btrfs_file_open(struct inode *inode, struct file *filp)
3607{
3608 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3609 return generic_file_open(inode, filp);
3610}
3611
3612static int check_direct_read(struct btrfs_fs_info *fs_info,
3613 const struct iov_iter *iter, loff_t offset)
3614{
3615 int ret;
3616 int i, seg;
3617
3618 ret = check_direct_IO(fs_info, iter, offset);
3619 if (ret < 0)
3620 return ret;
3621
3622 if (!iter_is_iovec(iter))
3623 return 0;
3624
3625 for (seg = 0; seg < iter->nr_segs; seg++)
3626 for (i = seg + 1; i < iter->nr_segs; i++)
3627 if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3628 return -EINVAL;
3629 return 0;
3630}
3631
3632static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3633{
3634 struct inode *inode = file_inode(iocb->ki_filp);
3635 ssize_t ret;
3636
3637 if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3638 return 0;
3639
3640 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3641 ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops, 0);
3642 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3643 return ret;
3644}
3645
3646static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3647{
3648 ssize_t ret = 0;
3649
3650 if (iocb->ki_flags & IOCB_DIRECT) {
3651 ret = btrfs_direct_read(iocb, to);
3652 if (ret < 0 || !iov_iter_count(to) ||
3653 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3654 return ret;
3655 }
3656
3657 return filemap_read(iocb, to, ret);
3658}
3659
3660const struct file_operations btrfs_file_operations = {
3661 .llseek = btrfs_file_llseek,
3662 .read_iter = btrfs_file_read_iter,
3663 .splice_read = generic_file_splice_read,
3664 .write_iter = btrfs_file_write_iter,
3665 .splice_write = iter_file_splice_write,
3666 .mmap = btrfs_file_mmap,
3667 .open = btrfs_file_open,
3668 .release = btrfs_release_file,
3669 .fsync = btrfs_sync_file,
3670 .fallocate = btrfs_fallocate,
3671 .unlocked_ioctl = btrfs_ioctl,
3672#ifdef CONFIG_COMPAT
3673 .compat_ioctl = btrfs_compat_ioctl,
3674#endif
3675 .remap_file_range = btrfs_remap_file_range,
3676};
3677
3678void __cold btrfs_auto_defrag_exit(void)
3679{
3680 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3681}
3682
3683int __init btrfs_auto_defrag_init(void)
3684{
3685 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3686 sizeof(struct inode_defrag), 0,
3687 SLAB_MEM_SPREAD,
3688 NULL);
3689 if (!btrfs_inode_defrag_cachep)
3690 return -ENOMEM;
3691
3692 return 0;
3693}
3694
3695int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3696{
3697 int ret;
3698
3699 /*
3700 * So with compression we will find and lock a dirty page and clear the
3701 * first one as dirty, setup an async extent, and immediately return
3702 * with the entire range locked but with nobody actually marked with
3703 * writeback. So we can't just filemap_write_and_wait_range() and
3704 * expect it to work since it will just kick off a thread to do the
3705 * actual work. So we need to call filemap_fdatawrite_range _again_
3706 * since it will wait on the page lock, which won't be unlocked until
3707 * after the pages have been marked as writeback and so we're good to go
3708 * from there. We have to do this otherwise we'll miss the ordered
3709 * extents and that results in badness. Please Josef, do not think you
3710 * know better and pull this out at some point in the future, it is
3711 * right and you are wrong.
3712 */
3713 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3714 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3715 &BTRFS_I(inode)->runtime_flags))
3716 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3717
3718 return ret;
3719}