Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6#include <linux/fs.h>
7#include <linux/pagemap.h>
8#include <linux/time.h>
9#include <linux/init.h>
10#include <linux/string.h>
11#include <linux/backing-dev.h>
12#include <linux/falloc.h>
13#include <linux/writeback.h>
14#include <linux/compat.h>
15#include <linux/slab.h>
16#include <linux/btrfs.h>
17#include <linux/uio.h>
18#include <linux/iversion.h>
19#include <linux/fsverity.h>
20#include "ctree.h"
21#include "direct-io.h"
22#include "disk-io.h"
23#include "transaction.h"
24#include "btrfs_inode.h"
25#include "tree-log.h"
26#include "locking.h"
27#include "qgroup.h"
28#include "compression.h"
29#include "delalloc-space.h"
30#include "reflink.h"
31#include "subpage.h"
32#include "fs.h"
33#include "accessors.h"
34#include "extent-tree.h"
35#include "file-item.h"
36#include "ioctl.h"
37#include "file.h"
38#include "super.h"
39
40/*
41 * Helper to fault in page and copy. This should go away and be replaced with
42 * calls into generic code.
43 */
44static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
45 struct folio *folio, struct iov_iter *i)
46{
47 size_t copied = 0;
48 size_t total_copied = 0;
49 int offset = offset_in_page(pos);
50
51 while (write_bytes > 0) {
52 size_t count = min_t(size_t, PAGE_SIZE - offset, write_bytes);
53 /*
54 * Copy data from userspace to the current page
55 */
56 copied = copy_folio_from_iter_atomic(folio, offset, count, i);
57
58 /* Flush processor's dcache for this page */
59 flush_dcache_folio(folio);
60
61 /*
62 * if we get a partial write, we can end up with
63 * partially up to date page. These add
64 * a lot of complexity, so make sure they don't
65 * happen by forcing this copy to be retried.
66 *
67 * The rest of the btrfs_file_write code will fall
68 * back to page at a time copies after we return 0.
69 */
70 if (unlikely(copied < count)) {
71 if (!folio_test_uptodate(folio)) {
72 iov_iter_revert(i, copied);
73 copied = 0;
74 }
75 if (!copied)
76 break;
77 }
78
79 write_bytes -= copied;
80 total_copied += copied;
81 offset += copied;
82 }
83 return total_copied;
84}
85
86/*
87 * Unlock folio after btrfs_file_write() is done with it.
88 */
89static void btrfs_drop_folio(struct btrfs_fs_info *fs_info, struct folio *folio,
90 u64 pos, u64 copied)
91{
92 u64 block_start = round_down(pos, fs_info->sectorsize);
93 u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
94
95 ASSERT(block_len <= U32_MAX);
96 /*
97 * Folio checked is some magic around finding folios that have been
98 * modified without going through btrfs_dirty_folio(). Clear it here.
99 * There should be no need to mark the pages accessed as
100 * prepare_one_folio() should have marked them accessed in
101 * prepare_one_folio() via find_or_create_page()
102 */
103 btrfs_folio_clamp_clear_checked(fs_info, folio, block_start, block_len);
104 folio_unlock(folio);
105 folio_put(folio);
106}
107
108/*
109 * After btrfs_copy_from_user(), update the following things for delalloc:
110 * - Mark newly dirtied folio as DELALLOC in the io tree.
111 * Used to advise which range is to be written back.
112 * - Mark modified folio as Uptodate/Dirty and not needing COW fixup
113 * - Update inode size for past EOF write
114 */
115int btrfs_dirty_folio(struct btrfs_inode *inode, struct folio *folio, loff_t pos,
116 size_t write_bytes, struct extent_state **cached, bool noreserve)
117{
118 struct btrfs_fs_info *fs_info = inode->root->fs_info;
119 int ret = 0;
120 u64 num_bytes;
121 u64 start_pos;
122 u64 end_of_last_block;
123 u64 end_pos = pos + write_bytes;
124 loff_t isize = i_size_read(&inode->vfs_inode);
125 unsigned int extra_bits = 0;
126
127 if (write_bytes == 0)
128 return 0;
129
130 if (noreserve)
131 extra_bits |= EXTENT_NORESERVE;
132
133 start_pos = round_down(pos, fs_info->sectorsize);
134 num_bytes = round_up(write_bytes + pos - start_pos,
135 fs_info->sectorsize);
136 ASSERT(num_bytes <= U32_MAX);
137 ASSERT(folio_pos(folio) <= pos &&
138 folio_pos(folio) + folio_size(folio) >= pos + write_bytes);
139
140 end_of_last_block = start_pos + num_bytes - 1;
141
142 /*
143 * The pages may have already been dirty, clear out old accounting so
144 * we can set things up properly
145 */
146 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
147 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
148 cached);
149
150 ret = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
151 extra_bits, cached);
152 if (ret)
153 return ret;
154
155 btrfs_folio_clamp_set_uptodate(fs_info, folio, start_pos, num_bytes);
156 btrfs_folio_clamp_clear_checked(fs_info, folio, start_pos, num_bytes);
157 btrfs_folio_clamp_set_dirty(fs_info, folio, start_pos, num_bytes);
158
159 /*
160 * we've only changed i_size in ram, and we haven't updated
161 * the disk i_size. There is no need to log the inode
162 * at this time.
163 */
164 if (end_pos > isize)
165 i_size_write(&inode->vfs_inode, end_pos);
166 return 0;
167}
168
169/*
170 * this is very complex, but the basic idea is to drop all extents
171 * in the range start - end. hint_block is filled in with a block number
172 * that would be a good hint to the block allocator for this file.
173 *
174 * If an extent intersects the range but is not entirely inside the range
175 * it is either truncated or split. Anything entirely inside the range
176 * is deleted from the tree.
177 *
178 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
179 * to deal with that. We set the field 'bytes_found' of the arguments structure
180 * with the number of allocated bytes found in the target range, so that the
181 * caller can update the inode's number of bytes in an atomic way when
182 * replacing extents in a range to avoid races with stat(2).
183 */
184int btrfs_drop_extents(struct btrfs_trans_handle *trans,
185 struct btrfs_root *root, struct btrfs_inode *inode,
186 struct btrfs_drop_extents_args *args)
187{
188 struct btrfs_fs_info *fs_info = root->fs_info;
189 struct extent_buffer *leaf;
190 struct btrfs_file_extent_item *fi;
191 struct btrfs_key key;
192 struct btrfs_key new_key;
193 u64 ino = btrfs_ino(inode);
194 u64 search_start = args->start;
195 u64 disk_bytenr = 0;
196 u64 num_bytes = 0;
197 u64 extent_offset = 0;
198 u64 extent_end = 0;
199 u64 last_end = args->start;
200 int del_nr = 0;
201 int del_slot = 0;
202 int extent_type;
203 int recow;
204 int ret;
205 int modify_tree = -1;
206 int update_refs;
207 int found = 0;
208 struct btrfs_path *path = args->path;
209
210 args->bytes_found = 0;
211 args->extent_inserted = false;
212
213 /* Must always have a path if ->replace_extent is true */
214 ASSERT(!(args->replace_extent && !args->path));
215
216 if (!path) {
217 path = btrfs_alloc_path();
218 if (!path) {
219 ret = -ENOMEM;
220 goto out;
221 }
222 }
223
224 if (args->drop_cache)
225 btrfs_drop_extent_map_range(inode, args->start, args->end - 1, false);
226
227 if (data_race(args->start >= inode->disk_i_size) && !args->replace_extent)
228 modify_tree = 0;
229
230 update_refs = (btrfs_root_id(root) != BTRFS_TREE_LOG_OBJECTID);
231 while (1) {
232 recow = 0;
233 ret = btrfs_lookup_file_extent(trans, root, path, ino,
234 search_start, modify_tree);
235 if (ret < 0)
236 break;
237 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
238 leaf = path->nodes[0];
239 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
240 if (key.objectid == ino &&
241 key.type == BTRFS_EXTENT_DATA_KEY)
242 path->slots[0]--;
243 }
244 ret = 0;
245next_slot:
246 leaf = path->nodes[0];
247 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
248 BUG_ON(del_nr > 0);
249 ret = btrfs_next_leaf(root, path);
250 if (ret < 0)
251 break;
252 if (ret > 0) {
253 ret = 0;
254 break;
255 }
256 leaf = path->nodes[0];
257 recow = 1;
258 }
259
260 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
261
262 if (key.objectid > ino)
263 break;
264 if (WARN_ON_ONCE(key.objectid < ino) ||
265 key.type < BTRFS_EXTENT_DATA_KEY) {
266 ASSERT(del_nr == 0);
267 path->slots[0]++;
268 goto next_slot;
269 }
270 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
271 break;
272
273 fi = btrfs_item_ptr(leaf, path->slots[0],
274 struct btrfs_file_extent_item);
275 extent_type = btrfs_file_extent_type(leaf, fi);
276
277 if (extent_type == BTRFS_FILE_EXTENT_REG ||
278 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
279 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
280 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
281 extent_offset = btrfs_file_extent_offset(leaf, fi);
282 extent_end = key.offset +
283 btrfs_file_extent_num_bytes(leaf, fi);
284 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
285 extent_end = key.offset +
286 btrfs_file_extent_ram_bytes(leaf, fi);
287 } else {
288 /* can't happen */
289 BUG();
290 }
291
292 /*
293 * Don't skip extent items representing 0 byte lengths. They
294 * used to be created (bug) if while punching holes we hit
295 * -ENOSPC condition. So if we find one here, just ensure we
296 * delete it, otherwise we would insert a new file extent item
297 * with the same key (offset) as that 0 bytes length file
298 * extent item in the call to setup_items_for_insert() later
299 * in this function.
300 */
301 if (extent_end == key.offset && extent_end >= search_start) {
302 last_end = extent_end;
303 goto delete_extent_item;
304 }
305
306 if (extent_end <= search_start) {
307 path->slots[0]++;
308 goto next_slot;
309 }
310
311 found = 1;
312 search_start = max(key.offset, args->start);
313 if (recow || !modify_tree) {
314 modify_tree = -1;
315 btrfs_release_path(path);
316 continue;
317 }
318
319 /*
320 * | - range to drop - |
321 * | -------- extent -------- |
322 */
323 if (args->start > key.offset && args->end < extent_end) {
324 BUG_ON(del_nr > 0);
325 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
326 ret = -EOPNOTSUPP;
327 break;
328 }
329
330 memcpy(&new_key, &key, sizeof(new_key));
331 new_key.offset = args->start;
332 ret = btrfs_duplicate_item(trans, root, path,
333 &new_key);
334 if (ret == -EAGAIN) {
335 btrfs_release_path(path);
336 continue;
337 }
338 if (ret < 0)
339 break;
340
341 leaf = path->nodes[0];
342 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
343 struct btrfs_file_extent_item);
344 btrfs_set_file_extent_num_bytes(leaf, fi,
345 args->start - key.offset);
346
347 fi = btrfs_item_ptr(leaf, path->slots[0],
348 struct btrfs_file_extent_item);
349
350 extent_offset += args->start - key.offset;
351 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
352 btrfs_set_file_extent_num_bytes(leaf, fi,
353 extent_end - args->start);
354 btrfs_mark_buffer_dirty(trans, leaf);
355
356 if (update_refs && disk_bytenr > 0) {
357 struct btrfs_ref ref = {
358 .action = BTRFS_ADD_DELAYED_REF,
359 .bytenr = disk_bytenr,
360 .num_bytes = num_bytes,
361 .parent = 0,
362 .owning_root = btrfs_root_id(root),
363 .ref_root = btrfs_root_id(root),
364 };
365 btrfs_init_data_ref(&ref, new_key.objectid,
366 args->start - extent_offset,
367 0, false);
368 ret = btrfs_inc_extent_ref(trans, &ref);
369 if (ret) {
370 btrfs_abort_transaction(trans, ret);
371 break;
372 }
373 }
374 key.offset = args->start;
375 }
376 /*
377 * From here on out we will have actually dropped something, so
378 * last_end can be updated.
379 */
380 last_end = extent_end;
381
382 /*
383 * | ---- range to drop ----- |
384 * | -------- extent -------- |
385 */
386 if (args->start <= key.offset && args->end < extent_end) {
387 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
388 ret = -EOPNOTSUPP;
389 break;
390 }
391
392 memcpy(&new_key, &key, sizeof(new_key));
393 new_key.offset = args->end;
394 btrfs_set_item_key_safe(trans, path, &new_key);
395
396 extent_offset += args->end - key.offset;
397 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
398 btrfs_set_file_extent_num_bytes(leaf, fi,
399 extent_end - args->end);
400 btrfs_mark_buffer_dirty(trans, leaf);
401 if (update_refs && disk_bytenr > 0)
402 args->bytes_found += args->end - key.offset;
403 break;
404 }
405
406 search_start = extent_end;
407 /*
408 * | ---- range to drop ----- |
409 * | -------- extent -------- |
410 */
411 if (args->start > key.offset && args->end >= extent_end) {
412 BUG_ON(del_nr > 0);
413 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
414 ret = -EOPNOTSUPP;
415 break;
416 }
417
418 btrfs_set_file_extent_num_bytes(leaf, fi,
419 args->start - key.offset);
420 btrfs_mark_buffer_dirty(trans, leaf);
421 if (update_refs && disk_bytenr > 0)
422 args->bytes_found += extent_end - args->start;
423 if (args->end == extent_end)
424 break;
425
426 path->slots[0]++;
427 goto next_slot;
428 }
429
430 /*
431 * | ---- range to drop ----- |
432 * | ------ extent ------ |
433 */
434 if (args->start <= key.offset && args->end >= extent_end) {
435delete_extent_item:
436 if (del_nr == 0) {
437 del_slot = path->slots[0];
438 del_nr = 1;
439 } else {
440 BUG_ON(del_slot + del_nr != path->slots[0]);
441 del_nr++;
442 }
443
444 if (update_refs &&
445 extent_type == BTRFS_FILE_EXTENT_INLINE) {
446 args->bytes_found += extent_end - key.offset;
447 extent_end = ALIGN(extent_end,
448 fs_info->sectorsize);
449 } else if (update_refs && disk_bytenr > 0) {
450 struct btrfs_ref ref = {
451 .action = BTRFS_DROP_DELAYED_REF,
452 .bytenr = disk_bytenr,
453 .num_bytes = num_bytes,
454 .parent = 0,
455 .owning_root = btrfs_root_id(root),
456 .ref_root = btrfs_root_id(root),
457 };
458 btrfs_init_data_ref(&ref, key.objectid,
459 key.offset - extent_offset,
460 0, false);
461 ret = btrfs_free_extent(trans, &ref);
462 if (ret) {
463 btrfs_abort_transaction(trans, ret);
464 break;
465 }
466 args->bytes_found += extent_end - key.offset;
467 }
468
469 if (args->end == extent_end)
470 break;
471
472 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
473 path->slots[0]++;
474 goto next_slot;
475 }
476
477 ret = btrfs_del_items(trans, root, path, del_slot,
478 del_nr);
479 if (ret) {
480 btrfs_abort_transaction(trans, ret);
481 break;
482 }
483
484 del_nr = 0;
485 del_slot = 0;
486
487 btrfs_release_path(path);
488 continue;
489 }
490
491 BUG();
492 }
493
494 if (!ret && del_nr > 0) {
495 /*
496 * Set path->slots[0] to first slot, so that after the delete
497 * if items are move off from our leaf to its immediate left or
498 * right neighbor leafs, we end up with a correct and adjusted
499 * path->slots[0] for our insertion (if args->replace_extent).
500 */
501 path->slots[0] = del_slot;
502 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
503 if (ret)
504 btrfs_abort_transaction(trans, ret);
505 }
506
507 leaf = path->nodes[0];
508 /*
509 * If btrfs_del_items() was called, it might have deleted a leaf, in
510 * which case it unlocked our path, so check path->locks[0] matches a
511 * write lock.
512 */
513 if (!ret && args->replace_extent &&
514 path->locks[0] == BTRFS_WRITE_LOCK &&
515 btrfs_leaf_free_space(leaf) >=
516 sizeof(struct btrfs_item) + args->extent_item_size) {
517
518 key.objectid = ino;
519 key.type = BTRFS_EXTENT_DATA_KEY;
520 key.offset = args->start;
521 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
522 struct btrfs_key slot_key;
523
524 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
525 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
526 path->slots[0]++;
527 }
528 btrfs_setup_item_for_insert(trans, root, path, &key,
529 args->extent_item_size);
530 args->extent_inserted = true;
531 }
532
533 if (!args->path)
534 btrfs_free_path(path);
535 else if (!args->extent_inserted)
536 btrfs_release_path(path);
537out:
538 args->drop_end = found ? min(args->end, last_end) : args->end;
539
540 return ret;
541}
542
543static int extent_mergeable(struct extent_buffer *leaf, int slot,
544 u64 objectid, u64 bytenr, u64 orig_offset,
545 u64 *start, u64 *end)
546{
547 struct btrfs_file_extent_item *fi;
548 struct btrfs_key key;
549 u64 extent_end;
550
551 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
552 return 0;
553
554 btrfs_item_key_to_cpu(leaf, &key, slot);
555 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
556 return 0;
557
558 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
559 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
560 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
561 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
562 btrfs_file_extent_compression(leaf, fi) ||
563 btrfs_file_extent_encryption(leaf, fi) ||
564 btrfs_file_extent_other_encoding(leaf, fi))
565 return 0;
566
567 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
568 if ((*start && *start != key.offset) || (*end && *end != extent_end))
569 return 0;
570
571 *start = key.offset;
572 *end = extent_end;
573 return 1;
574}
575
576/*
577 * Mark extent in the range start - end as written.
578 *
579 * This changes extent type from 'pre-allocated' to 'regular'. If only
580 * part of extent is marked as written, the extent will be split into
581 * two or three.
582 */
583int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
584 struct btrfs_inode *inode, u64 start, u64 end)
585{
586 struct btrfs_root *root = inode->root;
587 struct extent_buffer *leaf;
588 struct btrfs_path *path;
589 struct btrfs_file_extent_item *fi;
590 struct btrfs_ref ref = { 0 };
591 struct btrfs_key key;
592 struct btrfs_key new_key;
593 u64 bytenr;
594 u64 num_bytes;
595 u64 extent_end;
596 u64 orig_offset;
597 u64 other_start;
598 u64 other_end;
599 u64 split;
600 int del_nr = 0;
601 int del_slot = 0;
602 int recow;
603 int ret = 0;
604 u64 ino = btrfs_ino(inode);
605
606 path = btrfs_alloc_path();
607 if (!path)
608 return -ENOMEM;
609again:
610 recow = 0;
611 split = start;
612 key.objectid = ino;
613 key.type = BTRFS_EXTENT_DATA_KEY;
614 key.offset = split;
615
616 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
617 if (ret < 0)
618 goto out;
619 if (ret > 0 && path->slots[0] > 0)
620 path->slots[0]--;
621
622 leaf = path->nodes[0];
623 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
624 if (key.objectid != ino ||
625 key.type != BTRFS_EXTENT_DATA_KEY) {
626 ret = -EINVAL;
627 btrfs_abort_transaction(trans, ret);
628 goto out;
629 }
630 fi = btrfs_item_ptr(leaf, path->slots[0],
631 struct btrfs_file_extent_item);
632 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
633 ret = -EINVAL;
634 btrfs_abort_transaction(trans, ret);
635 goto out;
636 }
637 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
638 if (key.offset > start || extent_end < end) {
639 ret = -EINVAL;
640 btrfs_abort_transaction(trans, ret);
641 goto out;
642 }
643
644 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
645 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
646 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
647 memcpy(&new_key, &key, sizeof(new_key));
648
649 if (start == key.offset && end < extent_end) {
650 other_start = 0;
651 other_end = start;
652 if (extent_mergeable(leaf, path->slots[0] - 1,
653 ino, bytenr, orig_offset,
654 &other_start, &other_end)) {
655 new_key.offset = end;
656 btrfs_set_item_key_safe(trans, path, &new_key);
657 fi = btrfs_item_ptr(leaf, path->slots[0],
658 struct btrfs_file_extent_item);
659 btrfs_set_file_extent_generation(leaf, fi,
660 trans->transid);
661 btrfs_set_file_extent_num_bytes(leaf, fi,
662 extent_end - end);
663 btrfs_set_file_extent_offset(leaf, fi,
664 end - orig_offset);
665 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
666 struct btrfs_file_extent_item);
667 btrfs_set_file_extent_generation(leaf, fi,
668 trans->transid);
669 btrfs_set_file_extent_num_bytes(leaf, fi,
670 end - other_start);
671 btrfs_mark_buffer_dirty(trans, leaf);
672 goto out;
673 }
674 }
675
676 if (start > key.offset && end == extent_end) {
677 other_start = end;
678 other_end = 0;
679 if (extent_mergeable(leaf, path->slots[0] + 1,
680 ino, bytenr, orig_offset,
681 &other_start, &other_end)) {
682 fi = btrfs_item_ptr(leaf, path->slots[0],
683 struct btrfs_file_extent_item);
684 btrfs_set_file_extent_num_bytes(leaf, fi,
685 start - key.offset);
686 btrfs_set_file_extent_generation(leaf, fi,
687 trans->transid);
688 path->slots[0]++;
689 new_key.offset = start;
690 btrfs_set_item_key_safe(trans, path, &new_key);
691
692 fi = btrfs_item_ptr(leaf, path->slots[0],
693 struct btrfs_file_extent_item);
694 btrfs_set_file_extent_generation(leaf, fi,
695 trans->transid);
696 btrfs_set_file_extent_num_bytes(leaf, fi,
697 other_end - start);
698 btrfs_set_file_extent_offset(leaf, fi,
699 start - orig_offset);
700 btrfs_mark_buffer_dirty(trans, leaf);
701 goto out;
702 }
703 }
704
705 while (start > key.offset || end < extent_end) {
706 if (key.offset == start)
707 split = end;
708
709 new_key.offset = split;
710 ret = btrfs_duplicate_item(trans, root, path, &new_key);
711 if (ret == -EAGAIN) {
712 btrfs_release_path(path);
713 goto again;
714 }
715 if (ret < 0) {
716 btrfs_abort_transaction(trans, ret);
717 goto out;
718 }
719
720 leaf = path->nodes[0];
721 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
722 struct btrfs_file_extent_item);
723 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
724 btrfs_set_file_extent_num_bytes(leaf, fi,
725 split - key.offset);
726
727 fi = btrfs_item_ptr(leaf, path->slots[0],
728 struct btrfs_file_extent_item);
729
730 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
731 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
732 btrfs_set_file_extent_num_bytes(leaf, fi,
733 extent_end - split);
734 btrfs_mark_buffer_dirty(trans, leaf);
735
736 ref.action = BTRFS_ADD_DELAYED_REF;
737 ref.bytenr = bytenr;
738 ref.num_bytes = num_bytes;
739 ref.parent = 0;
740 ref.owning_root = btrfs_root_id(root);
741 ref.ref_root = btrfs_root_id(root);
742 btrfs_init_data_ref(&ref, ino, orig_offset, 0, false);
743 ret = btrfs_inc_extent_ref(trans, &ref);
744 if (ret) {
745 btrfs_abort_transaction(trans, ret);
746 goto out;
747 }
748
749 if (split == start) {
750 key.offset = start;
751 } else {
752 if (start != key.offset) {
753 ret = -EINVAL;
754 btrfs_abort_transaction(trans, ret);
755 goto out;
756 }
757 path->slots[0]--;
758 extent_end = end;
759 }
760 recow = 1;
761 }
762
763 other_start = end;
764 other_end = 0;
765
766 ref.action = BTRFS_DROP_DELAYED_REF;
767 ref.bytenr = bytenr;
768 ref.num_bytes = num_bytes;
769 ref.parent = 0;
770 ref.owning_root = btrfs_root_id(root);
771 ref.ref_root = btrfs_root_id(root);
772 btrfs_init_data_ref(&ref, ino, orig_offset, 0, false);
773 if (extent_mergeable(leaf, path->slots[0] + 1,
774 ino, bytenr, orig_offset,
775 &other_start, &other_end)) {
776 if (recow) {
777 btrfs_release_path(path);
778 goto again;
779 }
780 extent_end = other_end;
781 del_slot = path->slots[0] + 1;
782 del_nr++;
783 ret = btrfs_free_extent(trans, &ref);
784 if (ret) {
785 btrfs_abort_transaction(trans, ret);
786 goto out;
787 }
788 }
789 other_start = 0;
790 other_end = start;
791 if (extent_mergeable(leaf, path->slots[0] - 1,
792 ino, bytenr, orig_offset,
793 &other_start, &other_end)) {
794 if (recow) {
795 btrfs_release_path(path);
796 goto again;
797 }
798 key.offset = other_start;
799 del_slot = path->slots[0];
800 del_nr++;
801 ret = btrfs_free_extent(trans, &ref);
802 if (ret) {
803 btrfs_abort_transaction(trans, ret);
804 goto out;
805 }
806 }
807 if (del_nr == 0) {
808 fi = btrfs_item_ptr(leaf, path->slots[0],
809 struct btrfs_file_extent_item);
810 btrfs_set_file_extent_type(leaf, fi,
811 BTRFS_FILE_EXTENT_REG);
812 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
813 btrfs_mark_buffer_dirty(trans, leaf);
814 } else {
815 fi = btrfs_item_ptr(leaf, del_slot - 1,
816 struct btrfs_file_extent_item);
817 btrfs_set_file_extent_type(leaf, fi,
818 BTRFS_FILE_EXTENT_REG);
819 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
820 btrfs_set_file_extent_num_bytes(leaf, fi,
821 extent_end - key.offset);
822 btrfs_mark_buffer_dirty(trans, leaf);
823
824 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
825 if (ret < 0) {
826 btrfs_abort_transaction(trans, ret);
827 goto out;
828 }
829 }
830out:
831 btrfs_free_path(path);
832 return ret;
833}
834
835/*
836 * On error return an unlocked folio and the error value
837 * On success return a locked folio and 0
838 */
839static int prepare_uptodate_folio(struct inode *inode, struct folio *folio, u64 pos,
840 u64 len, bool force_uptodate)
841{
842 u64 clamp_start = max_t(u64, pos, folio_pos(folio));
843 u64 clamp_end = min_t(u64, pos + len, folio_pos(folio) + folio_size(folio));
844 int ret = 0;
845
846 if (folio_test_uptodate(folio))
847 return 0;
848
849 if (!force_uptodate &&
850 IS_ALIGNED(clamp_start, PAGE_SIZE) &&
851 IS_ALIGNED(clamp_end, PAGE_SIZE))
852 return 0;
853
854 ret = btrfs_read_folio(NULL, folio);
855 if (ret)
856 return ret;
857 folio_lock(folio);
858 if (!folio_test_uptodate(folio)) {
859 folio_unlock(folio);
860 return -EIO;
861 }
862
863 /*
864 * Since btrfs_read_folio() will unlock the folio before it returns,
865 * there is a window where btrfs_release_folio() can be called to
866 * release the page. Here we check both inode mapping and page
867 * private to make sure the page was not released.
868 *
869 * The private flag check is essential for subpage as we need to store
870 * extra bitmap using folio private.
871 */
872 if (folio->mapping != inode->i_mapping || !folio_test_private(folio)) {
873 folio_unlock(folio);
874 return -EAGAIN;
875 }
876 return 0;
877}
878
879static gfp_t get_prepare_gfp_flags(struct inode *inode, bool nowait)
880{
881 gfp_t gfp;
882
883 gfp = btrfs_alloc_write_mask(inode->i_mapping);
884 if (nowait) {
885 gfp &= ~__GFP_DIRECT_RECLAIM;
886 gfp |= GFP_NOWAIT;
887 }
888
889 return gfp;
890}
891
892/*
893 * Get folio into the page cache and lock it.
894 */
895static noinline int prepare_one_folio(struct inode *inode, struct folio **folio_ret,
896 loff_t pos, size_t write_bytes,
897 bool force_uptodate, bool nowait)
898{
899 unsigned long index = pos >> PAGE_SHIFT;
900 gfp_t mask = get_prepare_gfp_flags(inode, nowait);
901 fgf_t fgp_flags = (nowait ? FGP_WRITEBEGIN | FGP_NOWAIT : FGP_WRITEBEGIN);
902 struct folio *folio;
903 int ret = 0;
904
905again:
906 folio = __filemap_get_folio(inode->i_mapping, index, fgp_flags, mask);
907 if (IS_ERR(folio)) {
908 if (nowait)
909 ret = -EAGAIN;
910 else
911 ret = PTR_ERR(folio);
912 return ret;
913 }
914 folio_wait_writeback(folio);
915 /* Only support page sized folio yet. */
916 ASSERT(folio_order(folio) == 0);
917 ret = set_folio_extent_mapped(folio);
918 if (ret < 0) {
919 folio_unlock(folio);
920 folio_put(folio);
921 return ret;
922 }
923 ret = prepare_uptodate_folio(inode, folio, pos, write_bytes, force_uptodate);
924 if (ret) {
925 /* The folio is already unlocked. */
926 folio_put(folio);
927 if (!nowait && ret == -EAGAIN) {
928 ret = 0;
929 goto again;
930 }
931 return ret;
932 }
933 *folio_ret = folio;
934 return 0;
935}
936
937/*
938 * Locks the extent and properly waits for data=ordered extents to finish
939 * before allowing the folios to be modified if need.
940 *
941 * Return:
942 * 1 - the extent is locked
943 * 0 - the extent is not locked, and everything is OK
944 * -EAGAIN - need to prepare the folios again
945 */
946static noinline int
947lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct folio *folio,
948 loff_t pos, size_t write_bytes,
949 u64 *lockstart, u64 *lockend, bool nowait,
950 struct extent_state **cached_state)
951{
952 struct btrfs_fs_info *fs_info = inode->root->fs_info;
953 u64 start_pos;
954 u64 last_pos;
955 int ret = 0;
956
957 start_pos = round_down(pos, fs_info->sectorsize);
958 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
959
960 if (start_pos < inode->vfs_inode.i_size) {
961 struct btrfs_ordered_extent *ordered;
962
963 if (nowait) {
964 if (!try_lock_extent(&inode->io_tree, start_pos, last_pos,
965 cached_state)) {
966 folio_unlock(folio);
967 folio_put(folio);
968 return -EAGAIN;
969 }
970 } else {
971 lock_extent(&inode->io_tree, start_pos, last_pos, cached_state);
972 }
973
974 ordered = btrfs_lookup_ordered_range(inode, start_pos,
975 last_pos - start_pos + 1);
976 if (ordered &&
977 ordered->file_offset + ordered->num_bytes > start_pos &&
978 ordered->file_offset <= last_pos) {
979 unlock_extent(&inode->io_tree, start_pos, last_pos,
980 cached_state);
981 folio_unlock(folio);
982 folio_put(folio);
983 btrfs_start_ordered_extent(ordered);
984 btrfs_put_ordered_extent(ordered);
985 return -EAGAIN;
986 }
987 if (ordered)
988 btrfs_put_ordered_extent(ordered);
989
990 *lockstart = start_pos;
991 *lockend = last_pos;
992 ret = 1;
993 }
994
995 /*
996 * We should be called after prepare_one_folio() which should have locked
997 * all pages in the range.
998 */
999 WARN_ON(!folio_test_locked(folio));
1000
1001 return ret;
1002}
1003
1004/*
1005 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1006 *
1007 * @pos: File offset.
1008 * @write_bytes: The length to write, will be updated to the nocow writeable
1009 * range.
1010 *
1011 * This function will flush ordered extents in the range to ensure proper
1012 * nocow checks.
1013 *
1014 * Return:
1015 * > 0 If we can nocow, and updates @write_bytes.
1016 * 0 If we can't do a nocow write.
1017 * -EAGAIN If we can't do a nocow write because snapshoting of the inode's
1018 * root is in progress.
1019 * < 0 If an error happened.
1020 *
1021 * NOTE: Callers need to call btrfs_check_nocow_unlock() if we return > 0.
1022 */
1023int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1024 size_t *write_bytes, bool nowait)
1025{
1026 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1027 struct btrfs_root *root = inode->root;
1028 struct extent_state *cached_state = NULL;
1029 u64 lockstart, lockend;
1030 u64 num_bytes;
1031 int ret;
1032
1033 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1034 return 0;
1035
1036 if (!btrfs_drew_try_write_lock(&root->snapshot_lock))
1037 return -EAGAIN;
1038
1039 lockstart = round_down(pos, fs_info->sectorsize);
1040 lockend = round_up(pos + *write_bytes,
1041 fs_info->sectorsize) - 1;
1042 num_bytes = lockend - lockstart + 1;
1043
1044 if (nowait) {
1045 if (!btrfs_try_lock_ordered_range(inode, lockstart, lockend,
1046 &cached_state)) {
1047 btrfs_drew_write_unlock(&root->snapshot_lock);
1048 return -EAGAIN;
1049 }
1050 } else {
1051 btrfs_lock_and_flush_ordered_range(inode, lockstart, lockend,
1052 &cached_state);
1053 }
1054 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1055 NULL, nowait, false);
1056 if (ret <= 0)
1057 btrfs_drew_write_unlock(&root->snapshot_lock);
1058 else
1059 *write_bytes = min_t(size_t, *write_bytes ,
1060 num_bytes - pos + lockstart);
1061 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
1062
1063 return ret;
1064}
1065
1066void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1067{
1068 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1069}
1070
1071int btrfs_write_check(struct kiocb *iocb, size_t count)
1072{
1073 struct file *file = iocb->ki_filp;
1074 struct inode *inode = file_inode(file);
1075 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1076 loff_t pos = iocb->ki_pos;
1077 int ret;
1078 loff_t oldsize;
1079
1080 /*
1081 * Quickly bail out on NOWAIT writes if we don't have the nodatacow or
1082 * prealloc flags, as without those flags we always have to COW. We will
1083 * later check if we can really COW into the target range (using
1084 * can_nocow_extent() at btrfs_get_blocks_direct_write()).
1085 */
1086 if ((iocb->ki_flags & IOCB_NOWAIT) &&
1087 !(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1088 return -EAGAIN;
1089
1090 ret = file_remove_privs(file);
1091 if (ret)
1092 return ret;
1093
1094 /*
1095 * We reserve space for updating the inode when we reserve space for the
1096 * extent we are going to write, so we will enospc out there. We don't
1097 * need to start yet another transaction to update the inode as we will
1098 * update the inode when we finish writing whatever data we write.
1099 */
1100 if (!IS_NOCMTIME(inode)) {
1101 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
1102 inode_inc_iversion(inode);
1103 }
1104
1105 oldsize = i_size_read(inode);
1106 if (pos > oldsize) {
1107 /* Expand hole size to cover write data, preventing empty gap */
1108 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1109
1110 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1111 if (ret)
1112 return ret;
1113 }
1114
1115 return 0;
1116}
1117
1118ssize_t btrfs_buffered_write(struct kiocb *iocb, struct iov_iter *i)
1119{
1120 struct file *file = iocb->ki_filp;
1121 loff_t pos;
1122 struct inode *inode = file_inode(file);
1123 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1124 struct extent_changeset *data_reserved = NULL;
1125 u64 release_bytes = 0;
1126 u64 lockstart;
1127 u64 lockend;
1128 size_t num_written = 0;
1129 ssize_t ret;
1130 loff_t old_isize;
1131 unsigned int ilock_flags = 0;
1132 const bool nowait = (iocb->ki_flags & IOCB_NOWAIT);
1133 unsigned int bdp_flags = (nowait ? BDP_ASYNC : 0);
1134 bool only_release_metadata = false;
1135
1136 if (nowait)
1137 ilock_flags |= BTRFS_ILOCK_TRY;
1138
1139 ret = btrfs_inode_lock(BTRFS_I(inode), ilock_flags);
1140 if (ret < 0)
1141 return ret;
1142
1143 /*
1144 * We can only trust the isize with inode lock held, or it can race with
1145 * other buffered writes and cause incorrect call of
1146 * pagecache_isize_extended() to overwrite existing data.
1147 */
1148 old_isize = i_size_read(inode);
1149
1150 ret = generic_write_checks(iocb, i);
1151 if (ret <= 0)
1152 goto out;
1153
1154 ret = btrfs_write_check(iocb, ret);
1155 if (ret < 0)
1156 goto out;
1157
1158 pos = iocb->ki_pos;
1159 while (iov_iter_count(i) > 0) {
1160 struct extent_state *cached_state = NULL;
1161 size_t offset = offset_in_page(pos);
1162 size_t sector_offset;
1163 size_t write_bytes = min(iov_iter_count(i), PAGE_SIZE - offset);
1164 size_t reserve_bytes;
1165 size_t copied;
1166 size_t dirty_sectors;
1167 size_t num_sectors;
1168 struct folio *folio = NULL;
1169 int extents_locked;
1170 bool force_page_uptodate = false;
1171
1172 /*
1173 * Fault pages before locking them in prepare_one_folio()
1174 * to avoid recursive lock
1175 */
1176 if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1177 ret = -EFAULT;
1178 break;
1179 }
1180
1181 only_release_metadata = false;
1182 sector_offset = pos & (fs_info->sectorsize - 1);
1183
1184 extent_changeset_release(data_reserved);
1185 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1186 &data_reserved, pos,
1187 write_bytes, nowait);
1188 if (ret < 0) {
1189 int can_nocow;
1190
1191 if (nowait && (ret == -ENOSPC || ret == -EAGAIN)) {
1192 ret = -EAGAIN;
1193 break;
1194 }
1195
1196 /*
1197 * If we don't have to COW at the offset, reserve
1198 * metadata only. write_bytes may get smaller than
1199 * requested here.
1200 */
1201 can_nocow = btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1202 &write_bytes, nowait);
1203 if (can_nocow < 0)
1204 ret = can_nocow;
1205 if (can_nocow > 0)
1206 ret = 0;
1207 if (ret)
1208 break;
1209 only_release_metadata = true;
1210 }
1211
1212 reserve_bytes = round_up(write_bytes + sector_offset,
1213 fs_info->sectorsize);
1214 WARN_ON(reserve_bytes == 0);
1215 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1216 reserve_bytes,
1217 reserve_bytes, nowait);
1218 if (ret) {
1219 if (!only_release_metadata)
1220 btrfs_free_reserved_data_space(BTRFS_I(inode),
1221 data_reserved, pos,
1222 write_bytes);
1223 else
1224 btrfs_check_nocow_unlock(BTRFS_I(inode));
1225
1226 if (nowait && ret == -ENOSPC)
1227 ret = -EAGAIN;
1228 break;
1229 }
1230
1231 release_bytes = reserve_bytes;
1232again:
1233 ret = balance_dirty_pages_ratelimited_flags(inode->i_mapping, bdp_flags);
1234 if (ret) {
1235 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1236 break;
1237 }
1238
1239 ret = prepare_one_folio(inode, &folio, pos, write_bytes,
1240 force_page_uptodate, false);
1241 if (ret) {
1242 btrfs_delalloc_release_extents(BTRFS_I(inode),
1243 reserve_bytes);
1244 break;
1245 }
1246
1247 extents_locked = lock_and_cleanup_extent_if_need(BTRFS_I(inode),
1248 folio, pos, write_bytes, &lockstart,
1249 &lockend, nowait, &cached_state);
1250 if (extents_locked < 0) {
1251 if (!nowait && extents_locked == -EAGAIN)
1252 goto again;
1253
1254 btrfs_delalloc_release_extents(BTRFS_I(inode),
1255 reserve_bytes);
1256 ret = extents_locked;
1257 break;
1258 }
1259
1260 copied = btrfs_copy_from_user(pos, write_bytes, folio, i);
1261
1262 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1263 dirty_sectors = round_up(copied + sector_offset,
1264 fs_info->sectorsize);
1265 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1266
1267 if (copied == 0) {
1268 force_page_uptodate = true;
1269 dirty_sectors = 0;
1270 } else {
1271 force_page_uptodate = false;
1272 }
1273
1274 if (num_sectors > dirty_sectors) {
1275 /* release everything except the sectors we dirtied */
1276 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1277 if (only_release_metadata) {
1278 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1279 release_bytes, true);
1280 } else {
1281 u64 release_start = round_up(pos + copied,
1282 fs_info->sectorsize);
1283 btrfs_delalloc_release_space(BTRFS_I(inode),
1284 data_reserved, release_start,
1285 release_bytes, true);
1286 }
1287 }
1288
1289 release_bytes = round_up(copied + sector_offset,
1290 fs_info->sectorsize);
1291
1292 ret = btrfs_dirty_folio(BTRFS_I(inode), folio, pos, copied,
1293 &cached_state, only_release_metadata);
1294
1295 /*
1296 * If we have not locked the extent range, because the range's
1297 * start offset is >= i_size, we might still have a non-NULL
1298 * cached extent state, acquired while marking the extent range
1299 * as delalloc through btrfs_dirty_page(). Therefore free any
1300 * possible cached extent state to avoid a memory leak.
1301 */
1302 if (extents_locked)
1303 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
1304 lockend, &cached_state);
1305 else
1306 free_extent_state(cached_state);
1307
1308 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1309 if (ret) {
1310 btrfs_drop_folio(fs_info, folio, pos, copied);
1311 break;
1312 }
1313
1314 release_bytes = 0;
1315 if (only_release_metadata)
1316 btrfs_check_nocow_unlock(BTRFS_I(inode));
1317
1318 btrfs_drop_folio(fs_info, folio, pos, copied);
1319
1320 cond_resched();
1321
1322 pos += copied;
1323 num_written += copied;
1324 }
1325
1326 if (release_bytes) {
1327 if (only_release_metadata) {
1328 btrfs_check_nocow_unlock(BTRFS_I(inode));
1329 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1330 release_bytes, true);
1331 } else {
1332 btrfs_delalloc_release_space(BTRFS_I(inode),
1333 data_reserved,
1334 round_down(pos, fs_info->sectorsize),
1335 release_bytes, true);
1336 }
1337 }
1338
1339 extent_changeset_free(data_reserved);
1340 if (num_written > 0) {
1341 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1342 iocb->ki_pos += num_written;
1343 }
1344out:
1345 btrfs_inode_unlock(BTRFS_I(inode), ilock_flags);
1346 return num_written ? num_written : ret;
1347}
1348
1349static ssize_t btrfs_encoded_write(struct kiocb *iocb, struct iov_iter *from,
1350 const struct btrfs_ioctl_encoded_io_args *encoded)
1351{
1352 struct file *file = iocb->ki_filp;
1353 struct inode *inode = file_inode(file);
1354 loff_t count;
1355 ssize_t ret;
1356
1357 btrfs_inode_lock(BTRFS_I(inode), 0);
1358 count = encoded->len;
1359 ret = generic_write_checks_count(iocb, &count);
1360 if (ret == 0 && count != encoded->len) {
1361 /*
1362 * The write got truncated by generic_write_checks_count(). We
1363 * can't do a partial encoded write.
1364 */
1365 ret = -EFBIG;
1366 }
1367 if (ret || encoded->len == 0)
1368 goto out;
1369
1370 ret = btrfs_write_check(iocb, encoded->len);
1371 if (ret < 0)
1372 goto out;
1373
1374 ret = btrfs_do_encoded_write(iocb, from, encoded);
1375out:
1376 btrfs_inode_unlock(BTRFS_I(inode), 0);
1377 return ret;
1378}
1379
1380ssize_t btrfs_do_write_iter(struct kiocb *iocb, struct iov_iter *from,
1381 const struct btrfs_ioctl_encoded_io_args *encoded)
1382{
1383 struct file *file = iocb->ki_filp;
1384 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1385 ssize_t num_written, num_sync;
1386
1387 /*
1388 * If the fs flips readonly due to some impossible error, although we
1389 * have opened a file as writable, we have to stop this write operation
1390 * to ensure consistency.
1391 */
1392 if (BTRFS_FS_ERROR(inode->root->fs_info))
1393 return -EROFS;
1394
1395 if (encoded && (iocb->ki_flags & IOCB_NOWAIT))
1396 return -EOPNOTSUPP;
1397
1398 if (encoded) {
1399 num_written = btrfs_encoded_write(iocb, from, encoded);
1400 num_sync = encoded->len;
1401 } else if (iocb->ki_flags & IOCB_DIRECT) {
1402 num_written = btrfs_direct_write(iocb, from);
1403 num_sync = num_written;
1404 } else {
1405 num_written = btrfs_buffered_write(iocb, from);
1406 num_sync = num_written;
1407 }
1408
1409 btrfs_set_inode_last_sub_trans(inode);
1410
1411 if (num_sync > 0) {
1412 num_sync = generic_write_sync(iocb, num_sync);
1413 if (num_sync < 0)
1414 num_written = num_sync;
1415 }
1416
1417 return num_written;
1418}
1419
1420static ssize_t btrfs_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
1421{
1422 return btrfs_do_write_iter(iocb, from, NULL);
1423}
1424
1425int btrfs_release_file(struct inode *inode, struct file *filp)
1426{
1427 struct btrfs_file_private *private = filp->private_data;
1428
1429 if (private) {
1430 kfree(private->filldir_buf);
1431 free_extent_state(private->llseek_cached_state);
1432 kfree(private);
1433 filp->private_data = NULL;
1434 }
1435
1436 /*
1437 * Set by setattr when we are about to truncate a file from a non-zero
1438 * size to a zero size. This tries to flush down new bytes that may
1439 * have been written if the application were using truncate to replace
1440 * a file in place.
1441 */
1442 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
1443 &BTRFS_I(inode)->runtime_flags))
1444 filemap_flush(inode->i_mapping);
1445 return 0;
1446}
1447
1448static int start_ordered_ops(struct btrfs_inode *inode, loff_t start, loff_t end)
1449{
1450 int ret;
1451 struct blk_plug plug;
1452
1453 /*
1454 * This is only called in fsync, which would do synchronous writes, so
1455 * a plug can merge adjacent IOs as much as possible. Esp. in case of
1456 * multiple disks using raid profile, a large IO can be split to
1457 * several segments of stripe length (currently 64K).
1458 */
1459 blk_start_plug(&plug);
1460 ret = btrfs_fdatawrite_range(inode, start, end);
1461 blk_finish_plug(&plug);
1462
1463 return ret;
1464}
1465
1466static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
1467{
1468 struct btrfs_inode *inode = ctx->inode;
1469 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1470
1471 if (btrfs_inode_in_log(inode, btrfs_get_fs_generation(fs_info)) &&
1472 list_empty(&ctx->ordered_extents))
1473 return true;
1474
1475 /*
1476 * If we are doing a fast fsync we can not bail out if the inode's
1477 * last_trans is <= then the last committed transaction, because we only
1478 * update the last_trans of the inode during ordered extent completion,
1479 * and for a fast fsync we don't wait for that, we only wait for the
1480 * writeback to complete.
1481 */
1482 if (inode->last_trans <= btrfs_get_last_trans_committed(fs_info) &&
1483 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
1484 list_empty(&ctx->ordered_extents)))
1485 return true;
1486
1487 return false;
1488}
1489
1490/*
1491 * fsync call for both files and directories. This logs the inode into
1492 * the tree log instead of forcing full commits whenever possible.
1493 *
1494 * It needs to call filemap_fdatawait so that all ordered extent updates are
1495 * in the metadata btree are up to date for copying to the log.
1496 *
1497 * It drops the inode mutex before doing the tree log commit. This is an
1498 * important optimization for directories because holding the mutex prevents
1499 * new operations on the dir while we write to disk.
1500 */
1501int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1502{
1503 struct dentry *dentry = file_dentry(file);
1504 struct btrfs_inode *inode = BTRFS_I(d_inode(dentry));
1505 struct btrfs_root *root = inode->root;
1506 struct btrfs_fs_info *fs_info = root->fs_info;
1507 struct btrfs_trans_handle *trans;
1508 struct btrfs_log_ctx ctx;
1509 int ret = 0, err;
1510 u64 len;
1511 bool full_sync;
1512 bool skip_ilock = false;
1513
1514 if (current->journal_info == BTRFS_TRANS_DIO_WRITE_STUB) {
1515 skip_ilock = true;
1516 current->journal_info = NULL;
1517 btrfs_assert_inode_locked(inode);
1518 }
1519
1520 trace_btrfs_sync_file(file, datasync);
1521
1522 btrfs_init_log_ctx(&ctx, inode);
1523
1524 /*
1525 * Always set the range to a full range, otherwise we can get into
1526 * several problems, from missing file extent items to represent holes
1527 * when not using the NO_HOLES feature, to log tree corruption due to
1528 * races between hole detection during logging and completion of ordered
1529 * extents outside the range, to missing checksums due to ordered extents
1530 * for which we flushed only a subset of their pages.
1531 */
1532 start = 0;
1533 end = LLONG_MAX;
1534 len = (u64)LLONG_MAX + 1;
1535
1536 /*
1537 * We write the dirty pages in the range and wait until they complete
1538 * out of the ->i_mutex. If so, we can flush the dirty pages by
1539 * multi-task, and make the performance up. See
1540 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1541 */
1542 ret = start_ordered_ops(inode, start, end);
1543 if (ret)
1544 goto out;
1545
1546 if (skip_ilock)
1547 down_write(&inode->i_mmap_lock);
1548 else
1549 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
1550
1551 atomic_inc(&root->log_batch);
1552
1553 /*
1554 * Before we acquired the inode's lock and the mmap lock, someone may
1555 * have dirtied more pages in the target range. We need to make sure
1556 * that writeback for any such pages does not start while we are logging
1557 * the inode, because if it does, any of the following might happen when
1558 * we are not doing a full inode sync:
1559 *
1560 * 1) We log an extent after its writeback finishes but before its
1561 * checksums are added to the csum tree, leading to -EIO errors
1562 * when attempting to read the extent after a log replay.
1563 *
1564 * 2) We can end up logging an extent before its writeback finishes.
1565 * Therefore after the log replay we will have a file extent item
1566 * pointing to an unwritten extent (and no data checksums as well).
1567 *
1568 * So trigger writeback for any eventual new dirty pages and then we
1569 * wait for all ordered extents to complete below.
1570 */
1571 ret = start_ordered_ops(inode, start, end);
1572 if (ret) {
1573 if (skip_ilock)
1574 up_write(&inode->i_mmap_lock);
1575 else
1576 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
1577 goto out;
1578 }
1579
1580 /*
1581 * Always check for the full sync flag while holding the inode's lock,
1582 * to avoid races with other tasks. The flag must be either set all the
1583 * time during logging or always off all the time while logging.
1584 * We check the flag here after starting delalloc above, because when
1585 * running delalloc the full sync flag may be set if we need to drop
1586 * extra extent map ranges due to temporary memory allocation failures.
1587 */
1588 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
1589
1590 /*
1591 * We have to do this here to avoid the priority inversion of waiting on
1592 * IO of a lower priority task while holding a transaction open.
1593 *
1594 * For a full fsync we wait for the ordered extents to complete while
1595 * for a fast fsync we wait just for writeback to complete, and then
1596 * attach the ordered extents to the transaction so that a transaction
1597 * commit waits for their completion, to avoid data loss if we fsync,
1598 * the current transaction commits before the ordered extents complete
1599 * and a power failure happens right after that.
1600 *
1601 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
1602 * logical address recorded in the ordered extent may change. We need
1603 * to wait for the IO to stabilize the logical address.
1604 */
1605 if (full_sync || btrfs_is_zoned(fs_info)) {
1606 ret = btrfs_wait_ordered_range(inode, start, len);
1607 clear_bit(BTRFS_INODE_COW_WRITE_ERROR, &inode->runtime_flags);
1608 } else {
1609 /*
1610 * Get our ordered extents as soon as possible to avoid doing
1611 * checksum lookups in the csum tree, and use instead the
1612 * checksums attached to the ordered extents.
1613 */
1614 btrfs_get_ordered_extents_for_logging(inode, &ctx.ordered_extents);
1615 ret = filemap_fdatawait_range(inode->vfs_inode.i_mapping, start, end);
1616 if (ret)
1617 goto out_release_extents;
1618
1619 /*
1620 * Check and clear the BTRFS_INODE_COW_WRITE_ERROR now after
1621 * starting and waiting for writeback, because for buffered IO
1622 * it may have been set during the end IO callback
1623 * (end_bbio_data_write() -> btrfs_finish_ordered_extent()) in
1624 * case an error happened and we need to wait for ordered
1625 * extents to complete so that any extent maps that point to
1626 * unwritten locations are dropped and we don't log them.
1627 */
1628 if (test_and_clear_bit(BTRFS_INODE_COW_WRITE_ERROR, &inode->runtime_flags))
1629 ret = btrfs_wait_ordered_range(inode, start, len);
1630 }
1631
1632 if (ret)
1633 goto out_release_extents;
1634
1635 atomic_inc(&root->log_batch);
1636
1637 if (skip_inode_logging(&ctx)) {
1638 /*
1639 * We've had everything committed since the last time we were
1640 * modified so clear this flag in case it was set for whatever
1641 * reason, it's no longer relevant.
1642 */
1643 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
1644 /*
1645 * An ordered extent might have started before and completed
1646 * already with io errors, in which case the inode was not
1647 * updated and we end up here. So check the inode's mapping
1648 * for any errors that might have happened since we last
1649 * checked called fsync.
1650 */
1651 ret = filemap_check_wb_err(inode->vfs_inode.i_mapping, file->f_wb_err);
1652 goto out_release_extents;
1653 }
1654
1655 btrfs_init_log_ctx_scratch_eb(&ctx);
1656
1657 /*
1658 * We use start here because we will need to wait on the IO to complete
1659 * in btrfs_sync_log, which could require joining a transaction (for
1660 * example checking cross references in the nocow path). If we use join
1661 * here we could get into a situation where we're waiting on IO to
1662 * happen that is blocked on a transaction trying to commit. With start
1663 * we inc the extwriter counter, so we wait for all extwriters to exit
1664 * before we start blocking joiners. This comment is to keep somebody
1665 * from thinking they are super smart and changing this to
1666 * btrfs_join_transaction *cough*Josef*cough*.
1667 */
1668 trans = btrfs_start_transaction(root, 0);
1669 if (IS_ERR(trans)) {
1670 ret = PTR_ERR(trans);
1671 goto out_release_extents;
1672 }
1673 trans->in_fsync = true;
1674
1675 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
1676 /*
1677 * Scratch eb no longer needed, release before syncing log or commit
1678 * transaction, to avoid holding unnecessary memory during such long
1679 * operations.
1680 */
1681 if (ctx.scratch_eb) {
1682 free_extent_buffer(ctx.scratch_eb);
1683 ctx.scratch_eb = NULL;
1684 }
1685 btrfs_release_log_ctx_extents(&ctx);
1686 if (ret < 0) {
1687 /* Fallthrough and commit/free transaction. */
1688 ret = BTRFS_LOG_FORCE_COMMIT;
1689 }
1690
1691 /* we've logged all the items and now have a consistent
1692 * version of the file in the log. It is possible that
1693 * someone will come in and modify the file, but that's
1694 * fine because the log is consistent on disk, and we
1695 * have references to all of the file's extents
1696 *
1697 * It is possible that someone will come in and log the
1698 * file again, but that will end up using the synchronization
1699 * inside btrfs_sync_log to keep things safe.
1700 */
1701 if (skip_ilock)
1702 up_write(&inode->i_mmap_lock);
1703 else
1704 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
1705
1706 if (ret == BTRFS_NO_LOG_SYNC) {
1707 ret = btrfs_end_transaction(trans);
1708 goto out;
1709 }
1710
1711 /* We successfully logged the inode, attempt to sync the log. */
1712 if (!ret) {
1713 ret = btrfs_sync_log(trans, root, &ctx);
1714 if (!ret) {
1715 ret = btrfs_end_transaction(trans);
1716 goto out;
1717 }
1718 }
1719
1720 /*
1721 * At this point we need to commit the transaction because we had
1722 * btrfs_need_log_full_commit() or some other error.
1723 *
1724 * If we didn't do a full sync we have to stop the trans handle, wait on
1725 * the ordered extents, start it again and commit the transaction. If
1726 * we attempt to wait on the ordered extents here we could deadlock with
1727 * something like fallocate() that is holding the extent lock trying to
1728 * start a transaction while some other thread is trying to commit the
1729 * transaction while we (fsync) are currently holding the transaction
1730 * open.
1731 */
1732 if (!full_sync) {
1733 ret = btrfs_end_transaction(trans);
1734 if (ret)
1735 goto out;
1736 ret = btrfs_wait_ordered_range(inode, start, len);
1737 if (ret)
1738 goto out;
1739
1740 /*
1741 * This is safe to use here because we're only interested in
1742 * making sure the transaction that had the ordered extents is
1743 * committed. We aren't waiting on anything past this point,
1744 * we're purely getting the transaction and committing it.
1745 */
1746 trans = btrfs_attach_transaction_barrier(root);
1747 if (IS_ERR(trans)) {
1748 ret = PTR_ERR(trans);
1749
1750 /*
1751 * We committed the transaction and there's no currently
1752 * running transaction, this means everything we care
1753 * about made it to disk and we are done.
1754 */
1755 if (ret == -ENOENT)
1756 ret = 0;
1757 goto out;
1758 }
1759 }
1760
1761 ret = btrfs_commit_transaction(trans);
1762out:
1763 free_extent_buffer(ctx.scratch_eb);
1764 ASSERT(list_empty(&ctx.list));
1765 ASSERT(list_empty(&ctx.conflict_inodes));
1766 err = file_check_and_advance_wb_err(file);
1767 if (!ret)
1768 ret = err;
1769 return ret > 0 ? -EIO : ret;
1770
1771out_release_extents:
1772 btrfs_release_log_ctx_extents(&ctx);
1773 if (skip_ilock)
1774 up_write(&inode->i_mmap_lock);
1775 else
1776 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
1777 goto out;
1778}
1779
1780/*
1781 * btrfs_page_mkwrite() is not allowed to change the file size as it gets
1782 * called from a page fault handler when a page is first dirtied. Hence we must
1783 * be careful to check for EOF conditions here. We set the page up correctly
1784 * for a written page which means we get ENOSPC checking when writing into
1785 * holes and correct delalloc and unwritten extent mapping on filesystems that
1786 * support these features.
1787 *
1788 * We are not allowed to take the i_mutex here so we have to play games to
1789 * protect against truncate races as the page could now be beyond EOF. Because
1790 * truncate_setsize() writes the inode size before removing pages, once we have
1791 * the page lock we can determine safely if the page is beyond EOF. If it is not
1792 * beyond EOF, then the page is guaranteed safe against truncation until we
1793 * unlock the page.
1794 */
1795static vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
1796{
1797 struct page *page = vmf->page;
1798 struct folio *folio = page_folio(page);
1799 struct inode *inode = file_inode(vmf->vma->vm_file);
1800 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
1801 struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1802 struct btrfs_ordered_extent *ordered;
1803 struct extent_state *cached_state = NULL;
1804 struct extent_changeset *data_reserved = NULL;
1805 unsigned long zero_start;
1806 loff_t size;
1807 vm_fault_t ret;
1808 int ret2;
1809 int reserved = 0;
1810 u64 reserved_space;
1811 u64 page_start;
1812 u64 page_end;
1813 u64 end;
1814
1815 ASSERT(folio_order(folio) == 0);
1816
1817 reserved_space = PAGE_SIZE;
1818
1819 sb_start_pagefault(inode->i_sb);
1820 page_start = folio_pos(folio);
1821 page_end = page_start + folio_size(folio) - 1;
1822 end = page_end;
1823
1824 /*
1825 * Reserving delalloc space after obtaining the page lock can lead to
1826 * deadlock. For example, if a dirty page is locked by this function
1827 * and the call to btrfs_delalloc_reserve_space() ends up triggering
1828 * dirty page write out, then the btrfs_writepages() function could
1829 * end up waiting indefinitely to get a lock on the page currently
1830 * being processed by btrfs_page_mkwrite() function.
1831 */
1832 ret2 = btrfs_delalloc_reserve_space(BTRFS_I(inode), &data_reserved,
1833 page_start, reserved_space);
1834 if (!ret2) {
1835 ret2 = file_update_time(vmf->vma->vm_file);
1836 reserved = 1;
1837 }
1838 if (ret2) {
1839 ret = vmf_error(ret2);
1840 if (reserved)
1841 goto out;
1842 goto out_noreserve;
1843 }
1844
1845 /* Make the VM retry the fault. */
1846 ret = VM_FAULT_NOPAGE;
1847again:
1848 down_read(&BTRFS_I(inode)->i_mmap_lock);
1849 folio_lock(folio);
1850 size = i_size_read(inode);
1851
1852 if ((folio->mapping != inode->i_mapping) ||
1853 (page_start >= size)) {
1854 /* Page got truncated out from underneath us. */
1855 goto out_unlock;
1856 }
1857 folio_wait_writeback(folio);
1858
1859 lock_extent(io_tree, page_start, page_end, &cached_state);
1860 ret2 = set_folio_extent_mapped(folio);
1861 if (ret2 < 0) {
1862 ret = vmf_error(ret2);
1863 unlock_extent(io_tree, page_start, page_end, &cached_state);
1864 goto out_unlock;
1865 }
1866
1867 /*
1868 * We can't set the delalloc bits if there are pending ordered
1869 * extents. Drop our locks and wait for them to finish.
1870 */
1871 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start, PAGE_SIZE);
1872 if (ordered) {
1873 unlock_extent(io_tree, page_start, page_end, &cached_state);
1874 folio_unlock(folio);
1875 up_read(&BTRFS_I(inode)->i_mmap_lock);
1876 btrfs_start_ordered_extent(ordered);
1877 btrfs_put_ordered_extent(ordered);
1878 goto again;
1879 }
1880
1881 if (folio->index == ((size - 1) >> PAGE_SHIFT)) {
1882 reserved_space = round_up(size - page_start, fs_info->sectorsize);
1883 if (reserved_space < PAGE_SIZE) {
1884 end = page_start + reserved_space - 1;
1885 btrfs_delalloc_release_space(BTRFS_I(inode),
1886 data_reserved, page_start,
1887 PAGE_SIZE - reserved_space, true);
1888 }
1889 }
1890
1891 /*
1892 * page_mkwrite gets called when the page is firstly dirtied after it's
1893 * faulted in, but write(2) could also dirty a page and set delalloc
1894 * bits, thus in this case for space account reason, we still need to
1895 * clear any delalloc bits within this page range since we have to
1896 * reserve data&meta space before lock_page() (see above comments).
1897 */
1898 clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
1899 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
1900 EXTENT_DEFRAG, &cached_state);
1901
1902 ret2 = btrfs_set_extent_delalloc(BTRFS_I(inode), page_start, end, 0,
1903 &cached_state);
1904 if (ret2) {
1905 unlock_extent(io_tree, page_start, page_end, &cached_state);
1906 ret = VM_FAULT_SIGBUS;
1907 goto out_unlock;
1908 }
1909
1910 /* Page is wholly or partially inside EOF. */
1911 if (page_start + folio_size(folio) > size)
1912 zero_start = offset_in_folio(folio, size);
1913 else
1914 zero_start = PAGE_SIZE;
1915
1916 if (zero_start != PAGE_SIZE)
1917 folio_zero_range(folio, zero_start, folio_size(folio) - zero_start);
1918
1919 btrfs_folio_clear_checked(fs_info, folio, page_start, PAGE_SIZE);
1920 btrfs_folio_set_dirty(fs_info, folio, page_start, end + 1 - page_start);
1921 btrfs_folio_set_uptodate(fs_info, folio, page_start, end + 1 - page_start);
1922
1923 btrfs_set_inode_last_sub_trans(BTRFS_I(inode));
1924
1925 unlock_extent(io_tree, page_start, page_end, &cached_state);
1926 up_read(&BTRFS_I(inode)->i_mmap_lock);
1927
1928 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
1929 sb_end_pagefault(inode->i_sb);
1930 extent_changeset_free(data_reserved);
1931 return VM_FAULT_LOCKED;
1932
1933out_unlock:
1934 folio_unlock(folio);
1935 up_read(&BTRFS_I(inode)->i_mmap_lock);
1936out:
1937 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
1938 btrfs_delalloc_release_space(BTRFS_I(inode), data_reserved, page_start,
1939 reserved_space, (ret != 0));
1940out_noreserve:
1941 sb_end_pagefault(inode->i_sb);
1942 extent_changeset_free(data_reserved);
1943 return ret;
1944}
1945
1946static const struct vm_operations_struct btrfs_file_vm_ops = {
1947 .fault = filemap_fault,
1948 .map_pages = filemap_map_pages,
1949 .page_mkwrite = btrfs_page_mkwrite,
1950};
1951
1952static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
1953{
1954 struct address_space *mapping = filp->f_mapping;
1955
1956 if (!mapping->a_ops->read_folio)
1957 return -ENOEXEC;
1958
1959 file_accessed(filp);
1960 vma->vm_ops = &btrfs_file_vm_ops;
1961
1962 return 0;
1963}
1964
1965static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
1966 int slot, u64 start, u64 end)
1967{
1968 struct btrfs_file_extent_item *fi;
1969 struct btrfs_key key;
1970
1971 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1972 return 0;
1973
1974 btrfs_item_key_to_cpu(leaf, &key, slot);
1975 if (key.objectid != btrfs_ino(inode) ||
1976 key.type != BTRFS_EXTENT_DATA_KEY)
1977 return 0;
1978
1979 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1980
1981 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
1982 return 0;
1983
1984 if (btrfs_file_extent_disk_bytenr(leaf, fi))
1985 return 0;
1986
1987 if (key.offset == end)
1988 return 1;
1989 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
1990 return 1;
1991 return 0;
1992}
1993
1994static int fill_holes(struct btrfs_trans_handle *trans,
1995 struct btrfs_inode *inode,
1996 struct btrfs_path *path, u64 offset, u64 end)
1997{
1998 struct btrfs_fs_info *fs_info = trans->fs_info;
1999 struct btrfs_root *root = inode->root;
2000 struct extent_buffer *leaf;
2001 struct btrfs_file_extent_item *fi;
2002 struct extent_map *hole_em;
2003 struct btrfs_key key;
2004 int ret;
2005
2006 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2007 goto out;
2008
2009 key.objectid = btrfs_ino(inode);
2010 key.type = BTRFS_EXTENT_DATA_KEY;
2011 key.offset = offset;
2012
2013 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2014 if (ret <= 0) {
2015 /*
2016 * We should have dropped this offset, so if we find it then
2017 * something has gone horribly wrong.
2018 */
2019 if (ret == 0)
2020 ret = -EINVAL;
2021 return ret;
2022 }
2023
2024 leaf = path->nodes[0];
2025 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2026 u64 num_bytes;
2027
2028 path->slots[0]--;
2029 fi = btrfs_item_ptr(leaf, path->slots[0],
2030 struct btrfs_file_extent_item);
2031 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2032 end - offset;
2033 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2034 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2035 btrfs_set_file_extent_offset(leaf, fi, 0);
2036 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2037 btrfs_mark_buffer_dirty(trans, leaf);
2038 goto out;
2039 }
2040
2041 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2042 u64 num_bytes;
2043
2044 key.offset = offset;
2045 btrfs_set_item_key_safe(trans, path, &key);
2046 fi = btrfs_item_ptr(leaf, path->slots[0],
2047 struct btrfs_file_extent_item);
2048 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2049 offset;
2050 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2051 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2052 btrfs_set_file_extent_offset(leaf, fi, 0);
2053 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2054 btrfs_mark_buffer_dirty(trans, leaf);
2055 goto out;
2056 }
2057 btrfs_release_path(path);
2058
2059 ret = btrfs_insert_hole_extent(trans, root, btrfs_ino(inode), offset,
2060 end - offset);
2061 if (ret)
2062 return ret;
2063
2064out:
2065 btrfs_release_path(path);
2066
2067 hole_em = alloc_extent_map();
2068 if (!hole_em) {
2069 btrfs_drop_extent_map_range(inode, offset, end - 1, false);
2070 btrfs_set_inode_full_sync(inode);
2071 } else {
2072 hole_em->start = offset;
2073 hole_em->len = end - offset;
2074 hole_em->ram_bytes = hole_em->len;
2075
2076 hole_em->disk_bytenr = EXTENT_MAP_HOLE;
2077 hole_em->disk_num_bytes = 0;
2078 hole_em->generation = trans->transid;
2079
2080 ret = btrfs_replace_extent_map_range(inode, hole_em, true);
2081 free_extent_map(hole_em);
2082 if (ret)
2083 btrfs_set_inode_full_sync(inode);
2084 }
2085
2086 return 0;
2087}
2088
2089/*
2090 * Find a hole extent on given inode and change start/len to the end of hole
2091 * extent.(hole/vacuum extent whose em->start <= start &&
2092 * em->start + em->len > start)
2093 * When a hole extent is found, return 1 and modify start/len.
2094 */
2095static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2096{
2097 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2098 struct extent_map *em;
2099 int ret = 0;
2100
2101 em = btrfs_get_extent(inode, NULL,
2102 round_down(*start, fs_info->sectorsize),
2103 round_up(*len, fs_info->sectorsize));
2104 if (IS_ERR(em))
2105 return PTR_ERR(em);
2106
2107 /* Hole or vacuum extent(only exists in no-hole mode) */
2108 if (em->disk_bytenr == EXTENT_MAP_HOLE) {
2109 ret = 1;
2110 *len = em->start + em->len > *start + *len ?
2111 0 : *start + *len - em->start - em->len;
2112 *start = em->start + em->len;
2113 }
2114 free_extent_map(em);
2115 return ret;
2116}
2117
2118static void btrfs_punch_hole_lock_range(struct inode *inode,
2119 const u64 lockstart,
2120 const u64 lockend,
2121 struct extent_state **cached_state)
2122{
2123 /*
2124 * For subpage case, if the range is not at page boundary, we could
2125 * have pages at the leading/tailing part of the range.
2126 * This could lead to dead loop since filemap_range_has_page()
2127 * will always return true.
2128 * So here we need to do extra page alignment for
2129 * filemap_range_has_page().
2130 */
2131 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2132 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2133
2134 while (1) {
2135 truncate_pagecache_range(inode, lockstart, lockend);
2136
2137 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2138 cached_state);
2139 /*
2140 * We can't have ordered extents in the range, nor dirty/writeback
2141 * pages, because we have locked the inode's VFS lock in exclusive
2142 * mode, we have locked the inode's i_mmap_lock in exclusive mode,
2143 * we have flushed all delalloc in the range and we have waited
2144 * for any ordered extents in the range to complete.
2145 * We can race with anyone reading pages from this range, so after
2146 * locking the range check if we have pages in the range, and if
2147 * we do, unlock the range and retry.
2148 */
2149 if (!filemap_range_has_page(inode->i_mapping, page_lockstart,
2150 page_lockend))
2151 break;
2152
2153 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2154 cached_state);
2155 }
2156
2157 btrfs_assert_inode_range_clean(BTRFS_I(inode), lockstart, lockend);
2158}
2159
2160static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2161 struct btrfs_inode *inode,
2162 struct btrfs_path *path,
2163 struct btrfs_replace_extent_info *extent_info,
2164 const u64 replace_len,
2165 const u64 bytes_to_drop)
2166{
2167 struct btrfs_fs_info *fs_info = trans->fs_info;
2168 struct btrfs_root *root = inode->root;
2169 struct btrfs_file_extent_item *extent;
2170 struct extent_buffer *leaf;
2171 struct btrfs_key key;
2172 int slot;
2173 int ret;
2174
2175 if (replace_len == 0)
2176 return 0;
2177
2178 if (extent_info->disk_offset == 0 &&
2179 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2180 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2181 return 0;
2182 }
2183
2184 key.objectid = btrfs_ino(inode);
2185 key.type = BTRFS_EXTENT_DATA_KEY;
2186 key.offset = extent_info->file_offset;
2187 ret = btrfs_insert_empty_item(trans, root, path, &key,
2188 sizeof(struct btrfs_file_extent_item));
2189 if (ret)
2190 return ret;
2191 leaf = path->nodes[0];
2192 slot = path->slots[0];
2193 write_extent_buffer(leaf, extent_info->extent_buf,
2194 btrfs_item_ptr_offset(leaf, slot),
2195 sizeof(struct btrfs_file_extent_item));
2196 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2197 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2198 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2199 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2200 if (extent_info->is_new_extent)
2201 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2202 btrfs_mark_buffer_dirty(trans, leaf);
2203 btrfs_release_path(path);
2204
2205 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2206 replace_len);
2207 if (ret)
2208 return ret;
2209
2210 /* If it's a hole, nothing more needs to be done. */
2211 if (extent_info->disk_offset == 0) {
2212 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2213 return 0;
2214 }
2215
2216 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2217
2218 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2219 key.objectid = extent_info->disk_offset;
2220 key.type = BTRFS_EXTENT_ITEM_KEY;
2221 key.offset = extent_info->disk_len;
2222 ret = btrfs_alloc_reserved_file_extent(trans, root,
2223 btrfs_ino(inode),
2224 extent_info->file_offset,
2225 extent_info->qgroup_reserved,
2226 &key);
2227 } else {
2228 struct btrfs_ref ref = {
2229 .action = BTRFS_ADD_DELAYED_REF,
2230 .bytenr = extent_info->disk_offset,
2231 .num_bytes = extent_info->disk_len,
2232 .owning_root = btrfs_root_id(root),
2233 .ref_root = btrfs_root_id(root),
2234 };
2235 u64 ref_offset;
2236
2237 ref_offset = extent_info->file_offset - extent_info->data_offset;
2238 btrfs_init_data_ref(&ref, btrfs_ino(inode), ref_offset, 0, false);
2239 ret = btrfs_inc_extent_ref(trans, &ref);
2240 }
2241
2242 extent_info->insertions++;
2243
2244 return ret;
2245}
2246
2247/*
2248 * The respective range must have been previously locked, as well as the inode.
2249 * The end offset is inclusive (last byte of the range).
2250 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2251 * the file range with an extent.
2252 * When not punching a hole, we don't want to end up in a state where we dropped
2253 * extents without inserting a new one, so we must abort the transaction to avoid
2254 * a corruption.
2255 */
2256int btrfs_replace_file_extents(struct btrfs_inode *inode,
2257 struct btrfs_path *path, const u64 start,
2258 const u64 end,
2259 struct btrfs_replace_extent_info *extent_info,
2260 struct btrfs_trans_handle **trans_out)
2261{
2262 struct btrfs_drop_extents_args drop_args = { 0 };
2263 struct btrfs_root *root = inode->root;
2264 struct btrfs_fs_info *fs_info = root->fs_info;
2265 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2266 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2267 struct btrfs_trans_handle *trans = NULL;
2268 struct btrfs_block_rsv *rsv;
2269 unsigned int rsv_count;
2270 u64 cur_offset;
2271 u64 len = end - start;
2272 int ret = 0;
2273
2274 if (end <= start)
2275 return -EINVAL;
2276
2277 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2278 if (!rsv) {
2279 ret = -ENOMEM;
2280 goto out;
2281 }
2282 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2283 rsv->failfast = true;
2284
2285 /*
2286 * 1 - update the inode
2287 * 1 - removing the extents in the range
2288 * 1 - adding the hole extent if no_holes isn't set or if we are
2289 * replacing the range with a new extent
2290 */
2291 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2292 rsv_count = 3;
2293 else
2294 rsv_count = 2;
2295
2296 trans = btrfs_start_transaction(root, rsv_count);
2297 if (IS_ERR(trans)) {
2298 ret = PTR_ERR(trans);
2299 trans = NULL;
2300 goto out_free;
2301 }
2302
2303 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2304 min_size, false);
2305 if (WARN_ON(ret))
2306 goto out_trans;
2307 trans->block_rsv = rsv;
2308
2309 cur_offset = start;
2310 drop_args.path = path;
2311 drop_args.end = end + 1;
2312 drop_args.drop_cache = true;
2313 while (cur_offset < end) {
2314 drop_args.start = cur_offset;
2315 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2316 /* If we are punching a hole decrement the inode's byte count */
2317 if (!extent_info)
2318 btrfs_update_inode_bytes(inode, 0,
2319 drop_args.bytes_found);
2320 if (ret != -ENOSPC) {
2321 /*
2322 * The only time we don't want to abort is if we are
2323 * attempting to clone a partial inline extent, in which
2324 * case we'll get EOPNOTSUPP. However if we aren't
2325 * clone we need to abort no matter what, because if we
2326 * got EOPNOTSUPP via prealloc then we messed up and
2327 * need to abort.
2328 */
2329 if (ret &&
2330 (ret != -EOPNOTSUPP ||
2331 (extent_info && extent_info->is_new_extent)))
2332 btrfs_abort_transaction(trans, ret);
2333 break;
2334 }
2335
2336 trans->block_rsv = &fs_info->trans_block_rsv;
2337
2338 if (!extent_info && cur_offset < drop_args.drop_end &&
2339 cur_offset < ino_size) {
2340 ret = fill_holes(trans, inode, path, cur_offset,
2341 drop_args.drop_end);
2342 if (ret) {
2343 /*
2344 * If we failed then we didn't insert our hole
2345 * entries for the area we dropped, so now the
2346 * fs is corrupted, so we must abort the
2347 * transaction.
2348 */
2349 btrfs_abort_transaction(trans, ret);
2350 break;
2351 }
2352 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2353 /*
2354 * We are past the i_size here, but since we didn't
2355 * insert holes we need to clear the mapped area so we
2356 * know to not set disk_i_size in this area until a new
2357 * file extent is inserted here.
2358 */
2359 ret = btrfs_inode_clear_file_extent_range(inode,
2360 cur_offset,
2361 drop_args.drop_end - cur_offset);
2362 if (ret) {
2363 /*
2364 * We couldn't clear our area, so we could
2365 * presumably adjust up and corrupt the fs, so
2366 * we need to abort.
2367 */
2368 btrfs_abort_transaction(trans, ret);
2369 break;
2370 }
2371 }
2372
2373 if (extent_info &&
2374 drop_args.drop_end > extent_info->file_offset) {
2375 u64 replace_len = drop_args.drop_end -
2376 extent_info->file_offset;
2377
2378 ret = btrfs_insert_replace_extent(trans, inode, path,
2379 extent_info, replace_len,
2380 drop_args.bytes_found);
2381 if (ret) {
2382 btrfs_abort_transaction(trans, ret);
2383 break;
2384 }
2385 extent_info->data_len -= replace_len;
2386 extent_info->data_offset += replace_len;
2387 extent_info->file_offset += replace_len;
2388 }
2389
2390 /*
2391 * We are releasing our handle on the transaction, balance the
2392 * dirty pages of the btree inode and flush delayed items, and
2393 * then get a new transaction handle, which may now point to a
2394 * new transaction in case someone else may have committed the
2395 * transaction we used to replace/drop file extent items. So
2396 * bump the inode's iversion and update mtime and ctime except
2397 * if we are called from a dedupe context. This is because a
2398 * power failure/crash may happen after the transaction is
2399 * committed and before we finish replacing/dropping all the
2400 * file extent items we need.
2401 */
2402 inode_inc_iversion(&inode->vfs_inode);
2403
2404 if (!extent_info || extent_info->update_times)
2405 inode_set_mtime_to_ts(&inode->vfs_inode,
2406 inode_set_ctime_current(&inode->vfs_inode));
2407
2408 ret = btrfs_update_inode(trans, inode);
2409 if (ret)
2410 break;
2411
2412 btrfs_end_transaction(trans);
2413 btrfs_btree_balance_dirty(fs_info);
2414
2415 trans = btrfs_start_transaction(root, rsv_count);
2416 if (IS_ERR(trans)) {
2417 ret = PTR_ERR(trans);
2418 trans = NULL;
2419 break;
2420 }
2421
2422 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2423 rsv, min_size, false);
2424 if (WARN_ON(ret))
2425 break;
2426 trans->block_rsv = rsv;
2427
2428 cur_offset = drop_args.drop_end;
2429 len = end - cur_offset;
2430 if (!extent_info && len) {
2431 ret = find_first_non_hole(inode, &cur_offset, &len);
2432 if (unlikely(ret < 0))
2433 break;
2434 if (ret && !len) {
2435 ret = 0;
2436 break;
2437 }
2438 }
2439 }
2440
2441 /*
2442 * If we were cloning, force the next fsync to be a full one since we
2443 * we replaced (or just dropped in the case of cloning holes when
2444 * NO_HOLES is enabled) file extent items and did not setup new extent
2445 * maps for the replacement extents (or holes).
2446 */
2447 if (extent_info && !extent_info->is_new_extent)
2448 btrfs_set_inode_full_sync(inode);
2449
2450 if (ret)
2451 goto out_trans;
2452
2453 trans->block_rsv = &fs_info->trans_block_rsv;
2454 /*
2455 * If we are using the NO_HOLES feature we might have had already an
2456 * hole that overlaps a part of the region [lockstart, lockend] and
2457 * ends at (or beyond) lockend. Since we have no file extent items to
2458 * represent holes, drop_end can be less than lockend and so we must
2459 * make sure we have an extent map representing the existing hole (the
2460 * call to __btrfs_drop_extents() might have dropped the existing extent
2461 * map representing the existing hole), otherwise the fast fsync path
2462 * will not record the existence of the hole region
2463 * [existing_hole_start, lockend].
2464 */
2465 if (drop_args.drop_end <= end)
2466 drop_args.drop_end = end + 1;
2467 /*
2468 * Don't insert file hole extent item if it's for a range beyond eof
2469 * (because it's useless) or if it represents a 0 bytes range (when
2470 * cur_offset == drop_end).
2471 */
2472 if (!extent_info && cur_offset < ino_size &&
2473 cur_offset < drop_args.drop_end) {
2474 ret = fill_holes(trans, inode, path, cur_offset,
2475 drop_args.drop_end);
2476 if (ret) {
2477 /* Same comment as above. */
2478 btrfs_abort_transaction(trans, ret);
2479 goto out_trans;
2480 }
2481 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2482 /* See the comment in the loop above for the reasoning here. */
2483 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2484 drop_args.drop_end - cur_offset);
2485 if (ret) {
2486 btrfs_abort_transaction(trans, ret);
2487 goto out_trans;
2488 }
2489
2490 }
2491 if (extent_info) {
2492 ret = btrfs_insert_replace_extent(trans, inode, path,
2493 extent_info, extent_info->data_len,
2494 drop_args.bytes_found);
2495 if (ret) {
2496 btrfs_abort_transaction(trans, ret);
2497 goto out_trans;
2498 }
2499 }
2500
2501out_trans:
2502 if (!trans)
2503 goto out_free;
2504
2505 trans->block_rsv = &fs_info->trans_block_rsv;
2506 if (ret)
2507 btrfs_end_transaction(trans);
2508 else
2509 *trans_out = trans;
2510out_free:
2511 btrfs_free_block_rsv(fs_info, rsv);
2512out:
2513 return ret;
2514}
2515
2516static int btrfs_punch_hole(struct file *file, loff_t offset, loff_t len)
2517{
2518 struct inode *inode = file_inode(file);
2519 struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
2520 struct btrfs_root *root = BTRFS_I(inode)->root;
2521 struct extent_state *cached_state = NULL;
2522 struct btrfs_path *path;
2523 struct btrfs_trans_handle *trans = NULL;
2524 u64 lockstart;
2525 u64 lockend;
2526 u64 tail_start;
2527 u64 tail_len;
2528 u64 orig_start = offset;
2529 int ret = 0;
2530 bool same_block;
2531 u64 ino_size;
2532 bool truncated_block = false;
2533 bool updated_inode = false;
2534
2535 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2536
2537 ret = btrfs_wait_ordered_range(BTRFS_I(inode), offset, len);
2538 if (ret)
2539 goto out_only_mutex;
2540
2541 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2542 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2543 if (ret < 0)
2544 goto out_only_mutex;
2545 if (ret && !len) {
2546 /* Already in a large hole */
2547 ret = 0;
2548 goto out_only_mutex;
2549 }
2550
2551 ret = file_modified(file);
2552 if (ret)
2553 goto out_only_mutex;
2554
2555 lockstart = round_up(offset, fs_info->sectorsize);
2556 lockend = round_down(offset + len, fs_info->sectorsize) - 1;
2557 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2558 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2559 /*
2560 * We needn't truncate any block which is beyond the end of the file
2561 * because we are sure there is no data there.
2562 */
2563 /*
2564 * Only do this if we are in the same block and we aren't doing the
2565 * entire block.
2566 */
2567 if (same_block && len < fs_info->sectorsize) {
2568 if (offset < ino_size) {
2569 truncated_block = true;
2570 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2571 0);
2572 } else {
2573 ret = 0;
2574 }
2575 goto out_only_mutex;
2576 }
2577
2578 /* zero back part of the first block */
2579 if (offset < ino_size) {
2580 truncated_block = true;
2581 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2582 if (ret) {
2583 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2584 return ret;
2585 }
2586 }
2587
2588 /* Check the aligned pages after the first unaligned page,
2589 * if offset != orig_start, which means the first unaligned page
2590 * including several following pages are already in holes,
2591 * the extra check can be skipped */
2592 if (offset == orig_start) {
2593 /* after truncate page, check hole again */
2594 len = offset + len - lockstart;
2595 offset = lockstart;
2596 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2597 if (ret < 0)
2598 goto out_only_mutex;
2599 if (ret && !len) {
2600 ret = 0;
2601 goto out_only_mutex;
2602 }
2603 lockstart = offset;
2604 }
2605
2606 /* Check the tail unaligned part is in a hole */
2607 tail_start = lockend + 1;
2608 tail_len = offset + len - tail_start;
2609 if (tail_len) {
2610 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2611 if (unlikely(ret < 0))
2612 goto out_only_mutex;
2613 if (!ret) {
2614 /* zero the front end of the last page */
2615 if (tail_start + tail_len < ino_size) {
2616 truncated_block = true;
2617 ret = btrfs_truncate_block(BTRFS_I(inode),
2618 tail_start + tail_len,
2619 0, 1);
2620 if (ret)
2621 goto out_only_mutex;
2622 }
2623 }
2624 }
2625
2626 if (lockend < lockstart) {
2627 ret = 0;
2628 goto out_only_mutex;
2629 }
2630
2631 btrfs_punch_hole_lock_range(inode, lockstart, lockend, &cached_state);
2632
2633 path = btrfs_alloc_path();
2634 if (!path) {
2635 ret = -ENOMEM;
2636 goto out;
2637 }
2638
2639 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
2640 lockend, NULL, &trans);
2641 btrfs_free_path(path);
2642 if (ret)
2643 goto out;
2644
2645 ASSERT(trans != NULL);
2646 inode_inc_iversion(inode);
2647 inode_set_mtime_to_ts(inode, inode_set_ctime_current(inode));
2648 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2649 updated_inode = true;
2650 btrfs_end_transaction(trans);
2651 btrfs_btree_balance_dirty(fs_info);
2652out:
2653 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2654 &cached_state);
2655out_only_mutex:
2656 if (!updated_inode && truncated_block && !ret) {
2657 /*
2658 * If we only end up zeroing part of a page, we still need to
2659 * update the inode item, so that all the time fields are
2660 * updated as well as the necessary btrfs inode in memory fields
2661 * for detecting, at fsync time, if the inode isn't yet in the
2662 * log tree or it's there but not up to date.
2663 */
2664 struct timespec64 now = inode_set_ctime_current(inode);
2665
2666 inode_inc_iversion(inode);
2667 inode_set_mtime_to_ts(inode, now);
2668 trans = btrfs_start_transaction(root, 1);
2669 if (IS_ERR(trans)) {
2670 ret = PTR_ERR(trans);
2671 } else {
2672 int ret2;
2673
2674 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2675 ret2 = btrfs_end_transaction(trans);
2676 if (!ret)
2677 ret = ret2;
2678 }
2679 }
2680 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2681 return ret;
2682}
2683
2684/* Helper structure to record which range is already reserved */
2685struct falloc_range {
2686 struct list_head list;
2687 u64 start;
2688 u64 len;
2689};
2690
2691/*
2692 * Helper function to add falloc range
2693 *
2694 * Caller should have locked the larger range of extent containing
2695 * [start, len)
2696 */
2697static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2698{
2699 struct falloc_range *range = NULL;
2700
2701 if (!list_empty(head)) {
2702 /*
2703 * As fallocate iterates by bytenr order, we only need to check
2704 * the last range.
2705 */
2706 range = list_last_entry(head, struct falloc_range, list);
2707 if (range->start + range->len == start) {
2708 range->len += len;
2709 return 0;
2710 }
2711 }
2712
2713 range = kmalloc(sizeof(*range), GFP_KERNEL);
2714 if (!range)
2715 return -ENOMEM;
2716 range->start = start;
2717 range->len = len;
2718 list_add_tail(&range->list, head);
2719 return 0;
2720}
2721
2722static int btrfs_fallocate_update_isize(struct inode *inode,
2723 const u64 end,
2724 const int mode)
2725{
2726 struct btrfs_trans_handle *trans;
2727 struct btrfs_root *root = BTRFS_I(inode)->root;
2728 int ret;
2729 int ret2;
2730
2731 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2732 return 0;
2733
2734 trans = btrfs_start_transaction(root, 1);
2735 if (IS_ERR(trans))
2736 return PTR_ERR(trans);
2737
2738 inode_set_ctime_current(inode);
2739 i_size_write(inode, end);
2740 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
2741 ret = btrfs_update_inode(trans, BTRFS_I(inode));
2742 ret2 = btrfs_end_transaction(trans);
2743
2744 return ret ? ret : ret2;
2745}
2746
2747enum {
2748 RANGE_BOUNDARY_WRITTEN_EXTENT,
2749 RANGE_BOUNDARY_PREALLOC_EXTENT,
2750 RANGE_BOUNDARY_HOLE,
2751};
2752
2753static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
2754 u64 offset)
2755{
2756 const u64 sectorsize = inode->root->fs_info->sectorsize;
2757 struct extent_map *em;
2758 int ret;
2759
2760 offset = round_down(offset, sectorsize);
2761 em = btrfs_get_extent(inode, NULL, offset, sectorsize);
2762 if (IS_ERR(em))
2763 return PTR_ERR(em);
2764
2765 if (em->disk_bytenr == EXTENT_MAP_HOLE)
2766 ret = RANGE_BOUNDARY_HOLE;
2767 else if (em->flags & EXTENT_FLAG_PREALLOC)
2768 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2769 else
2770 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2771
2772 free_extent_map(em);
2773 return ret;
2774}
2775
2776static int btrfs_zero_range(struct inode *inode,
2777 loff_t offset,
2778 loff_t len,
2779 const int mode)
2780{
2781 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2782 struct extent_map *em;
2783 struct extent_changeset *data_reserved = NULL;
2784 int ret;
2785 u64 alloc_hint = 0;
2786 const u64 sectorsize = fs_info->sectorsize;
2787 u64 alloc_start = round_down(offset, sectorsize);
2788 u64 alloc_end = round_up(offset + len, sectorsize);
2789 u64 bytes_to_reserve = 0;
2790 bool space_reserved = false;
2791
2792 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start,
2793 alloc_end - alloc_start);
2794 if (IS_ERR(em)) {
2795 ret = PTR_ERR(em);
2796 goto out;
2797 }
2798
2799 /*
2800 * Avoid hole punching and extent allocation for some cases. More cases
2801 * could be considered, but these are unlikely common and we keep things
2802 * as simple as possible for now. Also, intentionally, if the target
2803 * range contains one or more prealloc extents together with regular
2804 * extents and holes, we drop all the existing extents and allocate a
2805 * new prealloc extent, so that we get a larger contiguous disk extent.
2806 */
2807 if (em->start <= alloc_start && (em->flags & EXTENT_FLAG_PREALLOC)) {
2808 const u64 em_end = em->start + em->len;
2809
2810 if (em_end >= offset + len) {
2811 /*
2812 * The whole range is already a prealloc extent,
2813 * do nothing except updating the inode's i_size if
2814 * needed.
2815 */
2816 free_extent_map(em);
2817 ret = btrfs_fallocate_update_isize(inode, offset + len,
2818 mode);
2819 goto out;
2820 }
2821 /*
2822 * Part of the range is already a prealloc extent, so operate
2823 * only on the remaining part of the range.
2824 */
2825 alloc_start = em_end;
2826 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2827 len = offset + len - alloc_start;
2828 offset = alloc_start;
2829 alloc_hint = extent_map_block_start(em) + em->len;
2830 }
2831 free_extent_map(em);
2832
2833 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2834 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2835 em = btrfs_get_extent(BTRFS_I(inode), NULL, alloc_start, sectorsize);
2836 if (IS_ERR(em)) {
2837 ret = PTR_ERR(em);
2838 goto out;
2839 }
2840
2841 if (em->flags & EXTENT_FLAG_PREALLOC) {
2842 free_extent_map(em);
2843 ret = btrfs_fallocate_update_isize(inode, offset + len,
2844 mode);
2845 goto out;
2846 }
2847 if (len < sectorsize && em->disk_bytenr != EXTENT_MAP_HOLE) {
2848 free_extent_map(em);
2849 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2850 0);
2851 if (!ret)
2852 ret = btrfs_fallocate_update_isize(inode,
2853 offset + len,
2854 mode);
2855 return ret;
2856 }
2857 free_extent_map(em);
2858 alloc_start = round_down(offset, sectorsize);
2859 alloc_end = alloc_start + sectorsize;
2860 goto reserve_space;
2861 }
2862
2863 alloc_start = round_up(offset, sectorsize);
2864 alloc_end = round_down(offset + len, sectorsize);
2865
2866 /*
2867 * For unaligned ranges, check the pages at the boundaries, they might
2868 * map to an extent, in which case we need to partially zero them, or
2869 * they might map to a hole, in which case we need our allocation range
2870 * to cover them.
2871 */
2872 if (!IS_ALIGNED(offset, sectorsize)) {
2873 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2874 offset);
2875 if (ret < 0)
2876 goto out;
2877 if (ret == RANGE_BOUNDARY_HOLE) {
2878 alloc_start = round_down(offset, sectorsize);
2879 ret = 0;
2880 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2881 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2882 if (ret)
2883 goto out;
2884 } else {
2885 ret = 0;
2886 }
2887 }
2888
2889 if (!IS_ALIGNED(offset + len, sectorsize)) {
2890 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
2891 offset + len);
2892 if (ret < 0)
2893 goto out;
2894 if (ret == RANGE_BOUNDARY_HOLE) {
2895 alloc_end = round_up(offset + len, sectorsize);
2896 ret = 0;
2897 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2898 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
2899 0, 1);
2900 if (ret)
2901 goto out;
2902 } else {
2903 ret = 0;
2904 }
2905 }
2906
2907reserve_space:
2908 if (alloc_start < alloc_end) {
2909 struct extent_state *cached_state = NULL;
2910 const u64 lockstart = alloc_start;
2911 const u64 lockend = alloc_end - 1;
2912
2913 bytes_to_reserve = alloc_end - alloc_start;
2914 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2915 bytes_to_reserve);
2916 if (ret < 0)
2917 goto out;
2918 space_reserved = true;
2919 btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2920 &cached_state);
2921 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
2922 alloc_start, bytes_to_reserve);
2923 if (ret) {
2924 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart,
2925 lockend, &cached_state);
2926 goto out;
2927 }
2928 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
2929 alloc_end - alloc_start,
2930 fs_info->sectorsize,
2931 offset + len, &alloc_hint);
2932 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2933 &cached_state);
2934 /* btrfs_prealloc_file_range releases reserved space on error */
2935 if (ret) {
2936 space_reserved = false;
2937 goto out;
2938 }
2939 }
2940 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
2941 out:
2942 if (ret && space_reserved)
2943 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
2944 alloc_start, bytes_to_reserve);
2945 extent_changeset_free(data_reserved);
2946
2947 return ret;
2948}
2949
2950static long btrfs_fallocate(struct file *file, int mode,
2951 loff_t offset, loff_t len)
2952{
2953 struct inode *inode = file_inode(file);
2954 struct extent_state *cached_state = NULL;
2955 struct extent_changeset *data_reserved = NULL;
2956 struct falloc_range *range;
2957 struct falloc_range *tmp;
2958 LIST_HEAD(reserve_list);
2959 u64 cur_offset;
2960 u64 last_byte;
2961 u64 alloc_start;
2962 u64 alloc_end;
2963 u64 alloc_hint = 0;
2964 u64 locked_end;
2965 u64 actual_end = 0;
2966 u64 data_space_needed = 0;
2967 u64 data_space_reserved = 0;
2968 u64 qgroup_reserved = 0;
2969 struct extent_map *em;
2970 int blocksize = BTRFS_I(inode)->root->fs_info->sectorsize;
2971 int ret;
2972
2973 /* Do not allow fallocate in ZONED mode */
2974 if (btrfs_is_zoned(inode_to_fs_info(inode)))
2975 return -EOPNOTSUPP;
2976
2977 alloc_start = round_down(offset, blocksize);
2978 alloc_end = round_up(offset + len, blocksize);
2979 cur_offset = alloc_start;
2980
2981 /* Make sure we aren't being give some crap mode */
2982 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
2983 FALLOC_FL_ZERO_RANGE))
2984 return -EOPNOTSUPP;
2985
2986 if (mode & FALLOC_FL_PUNCH_HOLE)
2987 return btrfs_punch_hole(file, offset, len);
2988
2989 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
2990
2991 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2992 ret = inode_newsize_ok(inode, offset + len);
2993 if (ret)
2994 goto out;
2995 }
2996
2997 ret = file_modified(file);
2998 if (ret)
2999 goto out;
3000
3001 /*
3002 * TODO: Move these two operations after we have checked
3003 * accurate reserved space, or fallocate can still fail but
3004 * with page truncated or size expanded.
3005 *
3006 * But that's a minor problem and won't do much harm BTW.
3007 */
3008 if (alloc_start > inode->i_size) {
3009 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3010 alloc_start);
3011 if (ret)
3012 goto out;
3013 } else if (offset + len > inode->i_size) {
3014 /*
3015 * If we are fallocating from the end of the file onward we
3016 * need to zero out the end of the block if i_size lands in the
3017 * middle of a block.
3018 */
3019 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3020 if (ret)
3021 goto out;
3022 }
3023
3024 /*
3025 * We have locked the inode at the VFS level (in exclusive mode) and we
3026 * have locked the i_mmap_lock lock (in exclusive mode). Now before
3027 * locking the file range, flush all dealloc in the range and wait for
3028 * all ordered extents in the range to complete. After this we can lock
3029 * the file range and, due to the previous locking we did, we know there
3030 * can't be more delalloc or ordered extents in the range.
3031 */
3032 ret = btrfs_wait_ordered_range(BTRFS_I(inode), alloc_start,
3033 alloc_end - alloc_start);
3034 if (ret)
3035 goto out;
3036
3037 if (mode & FALLOC_FL_ZERO_RANGE) {
3038 ret = btrfs_zero_range(inode, offset, len, mode);
3039 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3040 return ret;
3041 }
3042
3043 locked_end = alloc_end - 1;
3044 lock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3045 &cached_state);
3046
3047 btrfs_assert_inode_range_clean(BTRFS_I(inode), alloc_start, locked_end);
3048
3049 /* First, check if we exceed the qgroup limit */
3050 while (cur_offset < alloc_end) {
3051 em = btrfs_get_extent(BTRFS_I(inode), NULL, cur_offset,
3052 alloc_end - cur_offset);
3053 if (IS_ERR(em)) {
3054 ret = PTR_ERR(em);
3055 break;
3056 }
3057 last_byte = min(extent_map_end(em), alloc_end);
3058 actual_end = min_t(u64, extent_map_end(em), offset + len);
3059 last_byte = ALIGN(last_byte, blocksize);
3060 if (em->disk_bytenr == EXTENT_MAP_HOLE ||
3061 (cur_offset >= inode->i_size &&
3062 !(em->flags & EXTENT_FLAG_PREALLOC))) {
3063 const u64 range_len = last_byte - cur_offset;
3064
3065 ret = add_falloc_range(&reserve_list, cur_offset, range_len);
3066 if (ret < 0) {
3067 free_extent_map(em);
3068 break;
3069 }
3070 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3071 &data_reserved, cur_offset, range_len);
3072 if (ret < 0) {
3073 free_extent_map(em);
3074 break;
3075 }
3076 qgroup_reserved += range_len;
3077 data_space_needed += range_len;
3078 }
3079 free_extent_map(em);
3080 cur_offset = last_byte;
3081 }
3082
3083 if (!ret && data_space_needed > 0) {
3084 /*
3085 * We are safe to reserve space here as we can't have delalloc
3086 * in the range, see above.
3087 */
3088 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3089 data_space_needed);
3090 if (!ret)
3091 data_space_reserved = data_space_needed;
3092 }
3093
3094 /*
3095 * If ret is still 0, means we're OK to fallocate.
3096 * Or just cleanup the list and exit.
3097 */
3098 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3099 if (!ret) {
3100 ret = btrfs_prealloc_file_range(inode, mode,
3101 range->start,
3102 range->len, blocksize,
3103 offset + len, &alloc_hint);
3104 /*
3105 * btrfs_prealloc_file_range() releases space even
3106 * if it returns an error.
3107 */
3108 data_space_reserved -= range->len;
3109 qgroup_reserved -= range->len;
3110 } else if (data_space_reserved > 0) {
3111 btrfs_free_reserved_data_space(BTRFS_I(inode),
3112 data_reserved, range->start,
3113 range->len);
3114 data_space_reserved -= range->len;
3115 qgroup_reserved -= range->len;
3116 } else if (qgroup_reserved > 0) {
3117 btrfs_qgroup_free_data(BTRFS_I(inode), data_reserved,
3118 range->start, range->len, NULL);
3119 qgroup_reserved -= range->len;
3120 }
3121 list_del(&range->list);
3122 kfree(range);
3123 }
3124 if (ret < 0)
3125 goto out_unlock;
3126
3127 /*
3128 * We didn't need to allocate any more space, but we still extended the
3129 * size of the file so we need to update i_size and the inode item.
3130 */
3131 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3132out_unlock:
3133 unlock_extent(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3134 &cached_state);
3135out:
3136 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_MMAP);
3137 extent_changeset_free(data_reserved);
3138 return ret;
3139}
3140
3141/*
3142 * Helper for btrfs_find_delalloc_in_range(). Find a subrange in a given range
3143 * that has unflushed and/or flushing delalloc. There might be other adjacent
3144 * subranges after the one it found, so btrfs_find_delalloc_in_range() keeps
3145 * looping while it gets adjacent subranges, and merging them together.
3146 */
3147static bool find_delalloc_subrange(struct btrfs_inode *inode, u64 start, u64 end,
3148 struct extent_state **cached_state,
3149 bool *search_io_tree,
3150 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3151{
3152 u64 len = end + 1 - start;
3153 u64 delalloc_len = 0;
3154 struct btrfs_ordered_extent *oe;
3155 u64 oe_start;
3156 u64 oe_end;
3157
3158 /*
3159 * Search the io tree first for EXTENT_DELALLOC. If we find any, it
3160 * means we have delalloc (dirty pages) for which writeback has not
3161 * started yet.
3162 */
3163 if (*search_io_tree) {
3164 spin_lock(&inode->lock);
3165 if (inode->delalloc_bytes > 0) {
3166 spin_unlock(&inode->lock);
3167 *delalloc_start_ret = start;
3168 delalloc_len = count_range_bits(&inode->io_tree,
3169 delalloc_start_ret, end,
3170 len, EXTENT_DELALLOC, 1,
3171 cached_state);
3172 } else {
3173 spin_unlock(&inode->lock);
3174 }
3175 }
3176
3177 if (delalloc_len > 0) {
3178 /*
3179 * If delalloc was found then *delalloc_start_ret has a sector size
3180 * aligned value (rounded down).
3181 */
3182 *delalloc_end_ret = *delalloc_start_ret + delalloc_len - 1;
3183
3184 if (*delalloc_start_ret == start) {
3185 /* Delalloc for the whole range, nothing more to do. */
3186 if (*delalloc_end_ret == end)
3187 return true;
3188 /* Else trim our search range for ordered extents. */
3189 start = *delalloc_end_ret + 1;
3190 len = end + 1 - start;
3191 }
3192 } else {
3193 /* No delalloc, future calls don't need to search again. */
3194 *search_io_tree = false;
3195 }
3196
3197 /*
3198 * Now also check if there's any ordered extent in the range.
3199 * We do this because:
3200 *
3201 * 1) When delalloc is flushed, the file range is locked, we clear the
3202 * EXTENT_DELALLOC bit from the io tree and create an extent map and
3203 * an ordered extent for the write. So we might just have been called
3204 * after delalloc is flushed and before the ordered extent completes
3205 * and inserts the new file extent item in the subvolume's btree;
3206 *
3207 * 2) We may have an ordered extent created by flushing delalloc for a
3208 * subrange that starts before the subrange we found marked with
3209 * EXTENT_DELALLOC in the io tree.
3210 *
3211 * We could also use the extent map tree to find such delalloc that is
3212 * being flushed, but using the ordered extents tree is more efficient
3213 * because it's usually much smaller as ordered extents are removed from
3214 * the tree once they complete. With the extent maps, we mau have them
3215 * in the extent map tree for a very long time, and they were either
3216 * created by previous writes or loaded by read operations.
3217 */
3218 oe = btrfs_lookup_first_ordered_range(inode, start, len);
3219 if (!oe)
3220 return (delalloc_len > 0);
3221
3222 /* The ordered extent may span beyond our search range. */
3223 oe_start = max(oe->file_offset, start);
3224 oe_end = min(oe->file_offset + oe->num_bytes - 1, end);
3225
3226 btrfs_put_ordered_extent(oe);
3227
3228 /* Don't have unflushed delalloc, return the ordered extent range. */
3229 if (delalloc_len == 0) {
3230 *delalloc_start_ret = oe_start;
3231 *delalloc_end_ret = oe_end;
3232 return true;
3233 }
3234
3235 /*
3236 * We have both unflushed delalloc (io_tree) and an ordered extent.
3237 * If the ranges are adjacent returned a combined range, otherwise
3238 * return the leftmost range.
3239 */
3240 if (oe_start < *delalloc_start_ret) {
3241 if (oe_end < *delalloc_start_ret)
3242 *delalloc_end_ret = oe_end;
3243 *delalloc_start_ret = oe_start;
3244 } else if (*delalloc_end_ret + 1 == oe_start) {
3245 *delalloc_end_ret = oe_end;
3246 }
3247
3248 return true;
3249}
3250
3251/*
3252 * Check if there's delalloc in a given range.
3253 *
3254 * @inode: The inode.
3255 * @start: The start offset of the range. It does not need to be
3256 * sector size aligned.
3257 * @end: The end offset (inclusive value) of the search range.
3258 * It does not need to be sector size aligned.
3259 * @cached_state: Extent state record used for speeding up delalloc
3260 * searches in the inode's io_tree. Can be NULL.
3261 * @delalloc_start_ret: Output argument, set to the start offset of the
3262 * subrange found with delalloc (may not be sector size
3263 * aligned).
3264 * @delalloc_end_ret: Output argument, set to he end offset (inclusive value)
3265 * of the subrange found with delalloc.
3266 *
3267 * Returns true if a subrange with delalloc is found within the given range, and
3268 * if so it sets @delalloc_start_ret and @delalloc_end_ret with the start and
3269 * end offsets of the subrange.
3270 */
3271bool btrfs_find_delalloc_in_range(struct btrfs_inode *inode, u64 start, u64 end,
3272 struct extent_state **cached_state,
3273 u64 *delalloc_start_ret, u64 *delalloc_end_ret)
3274{
3275 u64 cur_offset = round_down(start, inode->root->fs_info->sectorsize);
3276 u64 prev_delalloc_end = 0;
3277 bool search_io_tree = true;
3278 bool ret = false;
3279
3280 while (cur_offset <= end) {
3281 u64 delalloc_start;
3282 u64 delalloc_end;
3283 bool delalloc;
3284
3285 delalloc = find_delalloc_subrange(inode, cur_offset, end,
3286 cached_state, &search_io_tree,
3287 &delalloc_start,
3288 &delalloc_end);
3289 if (!delalloc)
3290 break;
3291
3292 if (prev_delalloc_end == 0) {
3293 /* First subrange found. */
3294 *delalloc_start_ret = max(delalloc_start, start);
3295 *delalloc_end_ret = delalloc_end;
3296 ret = true;
3297 } else if (delalloc_start == prev_delalloc_end + 1) {
3298 /* Subrange adjacent to the previous one, merge them. */
3299 *delalloc_end_ret = delalloc_end;
3300 } else {
3301 /* Subrange not adjacent to the previous one, exit. */
3302 break;
3303 }
3304
3305 prev_delalloc_end = delalloc_end;
3306 cur_offset = delalloc_end + 1;
3307 cond_resched();
3308 }
3309
3310 return ret;
3311}
3312
3313/*
3314 * Check if there's a hole or delalloc range in a range representing a hole (or
3315 * prealloc extent) found in the inode's subvolume btree.
3316 *
3317 * @inode: The inode.
3318 * @whence: Seek mode (SEEK_DATA or SEEK_HOLE).
3319 * @start: Start offset of the hole region. It does not need to be sector
3320 * size aligned.
3321 * @end: End offset (inclusive value) of the hole region. It does not
3322 * need to be sector size aligned.
3323 * @start_ret: Return parameter, used to set the start of the subrange in the
3324 * hole that matches the search criteria (seek mode), if such
3325 * subrange is found (return value of the function is true).
3326 * The value returned here may not be sector size aligned.
3327 *
3328 * Returns true if a subrange matching the given seek mode is found, and if one
3329 * is found, it updates @start_ret with the start of the subrange.
3330 */
3331static bool find_desired_extent_in_hole(struct btrfs_inode *inode, int whence,
3332 struct extent_state **cached_state,
3333 u64 start, u64 end, u64 *start_ret)
3334{
3335 u64 delalloc_start;
3336 u64 delalloc_end;
3337 bool delalloc;
3338
3339 delalloc = btrfs_find_delalloc_in_range(inode, start, end, cached_state,
3340 &delalloc_start, &delalloc_end);
3341 if (delalloc && whence == SEEK_DATA) {
3342 *start_ret = delalloc_start;
3343 return true;
3344 }
3345
3346 if (delalloc && whence == SEEK_HOLE) {
3347 /*
3348 * We found delalloc but it starts after out start offset. So we
3349 * have a hole between our start offset and the delalloc start.
3350 */
3351 if (start < delalloc_start) {
3352 *start_ret = start;
3353 return true;
3354 }
3355 /*
3356 * Delalloc range starts at our start offset.
3357 * If the delalloc range's length is smaller than our range,
3358 * then it means we have a hole that starts where the delalloc
3359 * subrange ends.
3360 */
3361 if (delalloc_end < end) {
3362 *start_ret = delalloc_end + 1;
3363 return true;
3364 }
3365
3366 /* There's delalloc for the whole range. */
3367 return false;
3368 }
3369
3370 if (!delalloc && whence == SEEK_HOLE) {
3371 *start_ret = start;
3372 return true;
3373 }
3374
3375 /*
3376 * No delalloc in the range and we are seeking for data. The caller has
3377 * to iterate to the next extent item in the subvolume btree.
3378 */
3379 return false;
3380}
3381
3382static loff_t find_desired_extent(struct file *file, loff_t offset, int whence)
3383{
3384 struct btrfs_inode *inode = BTRFS_I(file->f_mapping->host);
3385 struct btrfs_file_private *private;
3386 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3387 struct extent_state *cached_state = NULL;
3388 struct extent_state **delalloc_cached_state;
3389 const loff_t i_size = i_size_read(&inode->vfs_inode);
3390 const u64 ino = btrfs_ino(inode);
3391 struct btrfs_root *root = inode->root;
3392 struct btrfs_path *path;
3393 struct btrfs_key key;
3394 u64 last_extent_end;
3395 u64 lockstart;
3396 u64 lockend;
3397 u64 start;
3398 int ret;
3399 bool found = false;
3400
3401 if (i_size == 0 || offset >= i_size)
3402 return -ENXIO;
3403
3404 /*
3405 * Quick path. If the inode has no prealloc extents and its number of
3406 * bytes used matches its i_size, then it can not have holes.
3407 */
3408 if (whence == SEEK_HOLE &&
3409 !(inode->flags & BTRFS_INODE_PREALLOC) &&
3410 inode_get_bytes(&inode->vfs_inode) == i_size)
3411 return i_size;
3412
3413 spin_lock(&inode->lock);
3414 private = file->private_data;
3415 spin_unlock(&inode->lock);
3416
3417 if (private && private->owner_task != current) {
3418 /*
3419 * Not allocated by us, don't use it as its cached state is used
3420 * by the task that allocated it and we don't want neither to
3421 * mess with it nor get incorrect results because it reflects an
3422 * invalid state for the current task.
3423 */
3424 private = NULL;
3425 } else if (!private) {
3426 private = kzalloc(sizeof(*private), GFP_KERNEL);
3427 /*
3428 * No worries if memory allocation failed.
3429 * The private structure is used only for speeding up multiple
3430 * lseek SEEK_HOLE/DATA calls to a file when there's delalloc,
3431 * so everything will still be correct.
3432 */
3433 if (private) {
3434 bool free = false;
3435
3436 private->owner_task = current;
3437
3438 spin_lock(&inode->lock);
3439 if (file->private_data)
3440 free = true;
3441 else
3442 file->private_data = private;
3443 spin_unlock(&inode->lock);
3444
3445 if (free) {
3446 kfree(private);
3447 private = NULL;
3448 }
3449 }
3450 }
3451
3452 if (private)
3453 delalloc_cached_state = &private->llseek_cached_state;
3454 else
3455 delalloc_cached_state = NULL;
3456
3457 /*
3458 * offset can be negative, in this case we start finding DATA/HOLE from
3459 * the very start of the file.
3460 */
3461 start = max_t(loff_t, 0, offset);
3462
3463 lockstart = round_down(start, fs_info->sectorsize);
3464 lockend = round_up(i_size, fs_info->sectorsize);
3465 if (lockend <= lockstart)
3466 lockend = lockstart + fs_info->sectorsize;
3467 lockend--;
3468
3469 path = btrfs_alloc_path();
3470 if (!path)
3471 return -ENOMEM;
3472 path->reada = READA_FORWARD;
3473
3474 key.objectid = ino;
3475 key.type = BTRFS_EXTENT_DATA_KEY;
3476 key.offset = start;
3477
3478 last_extent_end = lockstart;
3479
3480 lock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3481
3482 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3483 if (ret < 0) {
3484 goto out;
3485 } else if (ret > 0 && path->slots[0] > 0) {
3486 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0] - 1);
3487 if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
3488 path->slots[0]--;
3489 }
3490
3491 while (start < i_size) {
3492 struct extent_buffer *leaf = path->nodes[0];
3493 struct btrfs_file_extent_item *extent;
3494 u64 extent_end;
3495 u8 type;
3496
3497 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3498 ret = btrfs_next_leaf(root, path);
3499 if (ret < 0)
3500 goto out;
3501 else if (ret > 0)
3502 break;
3503
3504 leaf = path->nodes[0];
3505 }
3506
3507 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3508 if (key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY)
3509 break;
3510
3511 extent_end = btrfs_file_extent_end(path);
3512
3513 /*
3514 * In the first iteration we may have a slot that points to an
3515 * extent that ends before our start offset, so skip it.
3516 */
3517 if (extent_end <= start) {
3518 path->slots[0]++;
3519 continue;
3520 }
3521
3522 /* We have an implicit hole, NO_HOLES feature is likely set. */
3523 if (last_extent_end < key.offset) {
3524 u64 search_start = last_extent_end;
3525 u64 found_start;
3526
3527 /*
3528 * First iteration, @start matches @offset and it's
3529 * within the hole.
3530 */
3531 if (start == offset)
3532 search_start = offset;
3533
3534 found = find_desired_extent_in_hole(inode, whence,
3535 delalloc_cached_state,
3536 search_start,
3537 key.offset - 1,
3538 &found_start);
3539 if (found) {
3540 start = found_start;
3541 break;
3542 }
3543 /*
3544 * Didn't find data or a hole (due to delalloc) in the
3545 * implicit hole range, so need to analyze the extent.
3546 */
3547 }
3548
3549 extent = btrfs_item_ptr(leaf, path->slots[0],
3550 struct btrfs_file_extent_item);
3551 type = btrfs_file_extent_type(leaf, extent);
3552
3553 /*
3554 * Can't access the extent's disk_bytenr field if this is an
3555 * inline extent, since at that offset, it's where the extent
3556 * data starts.
3557 */
3558 if (type == BTRFS_FILE_EXTENT_PREALLOC ||
3559 (type == BTRFS_FILE_EXTENT_REG &&
3560 btrfs_file_extent_disk_bytenr(leaf, extent) == 0)) {
3561 /*
3562 * Explicit hole or prealloc extent, search for delalloc.
3563 * A prealloc extent is treated like a hole.
3564 */
3565 u64 search_start = key.offset;
3566 u64 found_start;
3567
3568 /*
3569 * First iteration, @start matches @offset and it's
3570 * within the hole.
3571 */
3572 if (start == offset)
3573 search_start = offset;
3574
3575 found = find_desired_extent_in_hole(inode, whence,
3576 delalloc_cached_state,
3577 search_start,
3578 extent_end - 1,
3579 &found_start);
3580 if (found) {
3581 start = found_start;
3582 break;
3583 }
3584 /*
3585 * Didn't find data or a hole (due to delalloc) in the
3586 * implicit hole range, so need to analyze the next
3587 * extent item.
3588 */
3589 } else {
3590 /*
3591 * Found a regular or inline extent.
3592 * If we are seeking for data, adjust the start offset
3593 * and stop, we're done.
3594 */
3595 if (whence == SEEK_DATA) {
3596 start = max_t(u64, key.offset, offset);
3597 found = true;
3598 break;
3599 }
3600 /*
3601 * Else, we are seeking for a hole, check the next file
3602 * extent item.
3603 */
3604 }
3605
3606 start = extent_end;
3607 last_extent_end = extent_end;
3608 path->slots[0]++;
3609 if (fatal_signal_pending(current)) {
3610 ret = -EINTR;
3611 goto out;
3612 }
3613 cond_resched();
3614 }
3615
3616 /* We have an implicit hole from the last extent found up to i_size. */
3617 if (!found && start < i_size) {
3618 found = find_desired_extent_in_hole(inode, whence,
3619 delalloc_cached_state, start,
3620 i_size - 1, &start);
3621 if (!found)
3622 start = i_size;
3623 }
3624
3625out:
3626 unlock_extent(&inode->io_tree, lockstart, lockend, &cached_state);
3627 btrfs_free_path(path);
3628
3629 if (ret < 0)
3630 return ret;
3631
3632 if (whence == SEEK_DATA && start >= i_size)
3633 return -ENXIO;
3634
3635 return min_t(loff_t, start, i_size);
3636}
3637
3638static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3639{
3640 struct inode *inode = file->f_mapping->host;
3641
3642 switch (whence) {
3643 default:
3644 return generic_file_llseek(file, offset, whence);
3645 case SEEK_DATA:
3646 case SEEK_HOLE:
3647 btrfs_inode_lock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3648 offset = find_desired_extent(file, offset, whence);
3649 btrfs_inode_unlock(BTRFS_I(inode), BTRFS_ILOCK_SHARED);
3650 break;
3651 }
3652
3653 if (offset < 0)
3654 return offset;
3655
3656 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3657}
3658
3659static int btrfs_file_open(struct inode *inode, struct file *filp)
3660{
3661 int ret;
3662
3663 filp->f_mode |= FMODE_NOWAIT | FMODE_CAN_ODIRECT;
3664
3665 ret = fsverity_file_open(inode, filp);
3666 if (ret)
3667 return ret;
3668 return generic_file_open(inode, filp);
3669}
3670
3671static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3672{
3673 ssize_t ret = 0;
3674
3675 if (iocb->ki_flags & IOCB_DIRECT) {
3676 ret = btrfs_direct_read(iocb, to);
3677 if (ret < 0 || !iov_iter_count(to) ||
3678 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3679 return ret;
3680 }
3681
3682 return filemap_read(iocb, to, ret);
3683}
3684
3685const struct file_operations btrfs_file_operations = {
3686 .llseek = btrfs_file_llseek,
3687 .read_iter = btrfs_file_read_iter,
3688 .splice_read = filemap_splice_read,
3689 .write_iter = btrfs_file_write_iter,
3690 .splice_write = iter_file_splice_write,
3691 .mmap = btrfs_file_mmap,
3692 .open = btrfs_file_open,
3693 .release = btrfs_release_file,
3694 .get_unmapped_area = thp_get_unmapped_area,
3695 .fsync = btrfs_sync_file,
3696 .fallocate = btrfs_fallocate,
3697 .unlocked_ioctl = btrfs_ioctl,
3698#ifdef CONFIG_COMPAT
3699 .compat_ioctl = btrfs_compat_ioctl,
3700#endif
3701 .remap_file_range = btrfs_remap_file_range,
3702 .uring_cmd = btrfs_uring_cmd,
3703 .fop_flags = FOP_BUFFER_RASYNC | FOP_BUFFER_WASYNC,
3704};
3705
3706int btrfs_fdatawrite_range(struct btrfs_inode *inode, loff_t start, loff_t end)
3707{
3708 struct address_space *mapping = inode->vfs_inode.i_mapping;
3709 int ret;
3710
3711 /*
3712 * So with compression we will find and lock a dirty page and clear the
3713 * first one as dirty, setup an async extent, and immediately return
3714 * with the entire range locked but with nobody actually marked with
3715 * writeback. So we can't just filemap_write_and_wait_range() and
3716 * expect it to work since it will just kick off a thread to do the
3717 * actual work. So we need to call filemap_fdatawrite_range _again_
3718 * since it will wait on the page lock, which won't be unlocked until
3719 * after the pages have been marked as writeback and so we're good to go
3720 * from there. We have to do this otherwise we'll miss the ordered
3721 * extents and that results in badness. Please Josef, do not think you
3722 * know better and pull this out at some point in the future, it is
3723 * right and you are wrong.
3724 */
3725 ret = filemap_fdatawrite_range(mapping, start, end);
3726 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT, &inode->runtime_flags))
3727 ret = filemap_fdatawrite_range(mapping, start, end);
3728
3729 return ret;
3730}
1/*
2 * Copyright (C) 2007 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/fs.h>
20#include <linux/pagemap.h>
21#include <linux/highmem.h>
22#include <linux/time.h>
23#include <linux/init.h>
24#include <linux/string.h>
25#include <linux/backing-dev.h>
26#include <linux/mpage.h>
27#include <linux/falloc.h>
28#include <linux/swap.h>
29#include <linux/writeback.h>
30#include <linux/statfs.h>
31#include <linux/compat.h>
32#include <linux/slab.h>
33#include <linux/btrfs.h>
34#include <linux/uio.h>
35#include "ctree.h"
36#include "disk-io.h"
37#include "transaction.h"
38#include "btrfs_inode.h"
39#include "print-tree.h"
40#include "tree-log.h"
41#include "locking.h"
42#include "volumes.h"
43#include "qgroup.h"
44#include "compression.h"
45
46static struct kmem_cache *btrfs_inode_defrag_cachep;
47/*
48 * when auto defrag is enabled we
49 * queue up these defrag structs to remember which
50 * inodes need defragging passes
51 */
52struct inode_defrag {
53 struct rb_node rb_node;
54 /* objectid */
55 u64 ino;
56 /*
57 * transid where the defrag was added, we search for
58 * extents newer than this
59 */
60 u64 transid;
61
62 /* root objectid */
63 u64 root;
64
65 /* last offset we were able to defrag */
66 u64 last_offset;
67
68 /* if we've wrapped around back to zero once already */
69 int cycled;
70};
71
72static int __compare_inode_defrag(struct inode_defrag *defrag1,
73 struct inode_defrag *defrag2)
74{
75 if (defrag1->root > defrag2->root)
76 return 1;
77 else if (defrag1->root < defrag2->root)
78 return -1;
79 else if (defrag1->ino > defrag2->ino)
80 return 1;
81 else if (defrag1->ino < defrag2->ino)
82 return -1;
83 else
84 return 0;
85}
86
87/* pop a record for an inode into the defrag tree. The lock
88 * must be held already
89 *
90 * If you're inserting a record for an older transid than an
91 * existing record, the transid already in the tree is lowered
92 *
93 * If an existing record is found the defrag item you
94 * pass in is freed
95 */
96static int __btrfs_add_inode_defrag(struct inode *inode,
97 struct inode_defrag *defrag)
98{
99 struct btrfs_root *root = BTRFS_I(inode)->root;
100 struct inode_defrag *entry;
101 struct rb_node **p;
102 struct rb_node *parent = NULL;
103 int ret;
104
105 p = &root->fs_info->defrag_inodes.rb_node;
106 while (*p) {
107 parent = *p;
108 entry = rb_entry(parent, struct inode_defrag, rb_node);
109
110 ret = __compare_inode_defrag(defrag, entry);
111 if (ret < 0)
112 p = &parent->rb_left;
113 else if (ret > 0)
114 p = &parent->rb_right;
115 else {
116 /* if we're reinserting an entry for
117 * an old defrag run, make sure to
118 * lower the transid of our existing record
119 */
120 if (defrag->transid < entry->transid)
121 entry->transid = defrag->transid;
122 if (defrag->last_offset > entry->last_offset)
123 entry->last_offset = defrag->last_offset;
124 return -EEXIST;
125 }
126 }
127 set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
128 rb_link_node(&defrag->rb_node, parent, p);
129 rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
130 return 0;
131}
132
133static inline int __need_auto_defrag(struct btrfs_root *root)
134{
135 if (!btrfs_test_opt(root, AUTO_DEFRAG))
136 return 0;
137
138 if (btrfs_fs_closing(root->fs_info))
139 return 0;
140
141 return 1;
142}
143
144/*
145 * insert a defrag record for this inode if auto defrag is
146 * enabled
147 */
148int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
149 struct inode *inode)
150{
151 struct btrfs_root *root = BTRFS_I(inode)->root;
152 struct inode_defrag *defrag;
153 u64 transid;
154 int ret;
155
156 if (!__need_auto_defrag(root))
157 return 0;
158
159 if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
160 return 0;
161
162 if (trans)
163 transid = trans->transid;
164 else
165 transid = BTRFS_I(inode)->root->last_trans;
166
167 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
168 if (!defrag)
169 return -ENOMEM;
170
171 defrag->ino = btrfs_ino(inode);
172 defrag->transid = transid;
173 defrag->root = root->root_key.objectid;
174
175 spin_lock(&root->fs_info->defrag_inodes_lock);
176 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
177 /*
178 * If we set IN_DEFRAG flag and evict the inode from memory,
179 * and then re-read this inode, this new inode doesn't have
180 * IN_DEFRAG flag. At the case, we may find the existed defrag.
181 */
182 ret = __btrfs_add_inode_defrag(inode, defrag);
183 if (ret)
184 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
185 } else {
186 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
187 }
188 spin_unlock(&root->fs_info->defrag_inodes_lock);
189 return 0;
190}
191
192/*
193 * Requeue the defrag object. If there is a defrag object that points to
194 * the same inode in the tree, we will merge them together (by
195 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
196 */
197static void btrfs_requeue_inode_defrag(struct inode *inode,
198 struct inode_defrag *defrag)
199{
200 struct btrfs_root *root = BTRFS_I(inode)->root;
201 int ret;
202
203 if (!__need_auto_defrag(root))
204 goto out;
205
206 /*
207 * Here we don't check the IN_DEFRAG flag, because we need merge
208 * them together.
209 */
210 spin_lock(&root->fs_info->defrag_inodes_lock);
211 ret = __btrfs_add_inode_defrag(inode, defrag);
212 spin_unlock(&root->fs_info->defrag_inodes_lock);
213 if (ret)
214 goto out;
215 return;
216out:
217 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
218}
219
220/*
221 * pick the defragable inode that we want, if it doesn't exist, we will get
222 * the next one.
223 */
224static struct inode_defrag *
225btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
226{
227 struct inode_defrag *entry = NULL;
228 struct inode_defrag tmp;
229 struct rb_node *p;
230 struct rb_node *parent = NULL;
231 int ret;
232
233 tmp.ino = ino;
234 tmp.root = root;
235
236 spin_lock(&fs_info->defrag_inodes_lock);
237 p = fs_info->defrag_inodes.rb_node;
238 while (p) {
239 parent = p;
240 entry = rb_entry(parent, struct inode_defrag, rb_node);
241
242 ret = __compare_inode_defrag(&tmp, entry);
243 if (ret < 0)
244 p = parent->rb_left;
245 else if (ret > 0)
246 p = parent->rb_right;
247 else
248 goto out;
249 }
250
251 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
252 parent = rb_next(parent);
253 if (parent)
254 entry = rb_entry(parent, struct inode_defrag, rb_node);
255 else
256 entry = NULL;
257 }
258out:
259 if (entry)
260 rb_erase(parent, &fs_info->defrag_inodes);
261 spin_unlock(&fs_info->defrag_inodes_lock);
262 return entry;
263}
264
265void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
266{
267 struct inode_defrag *defrag;
268 struct rb_node *node;
269
270 spin_lock(&fs_info->defrag_inodes_lock);
271 node = rb_first(&fs_info->defrag_inodes);
272 while (node) {
273 rb_erase(node, &fs_info->defrag_inodes);
274 defrag = rb_entry(node, struct inode_defrag, rb_node);
275 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
276
277 cond_resched_lock(&fs_info->defrag_inodes_lock);
278
279 node = rb_first(&fs_info->defrag_inodes);
280 }
281 spin_unlock(&fs_info->defrag_inodes_lock);
282}
283
284#define BTRFS_DEFRAG_BATCH 1024
285
286static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
287 struct inode_defrag *defrag)
288{
289 struct btrfs_root *inode_root;
290 struct inode *inode;
291 struct btrfs_key key;
292 struct btrfs_ioctl_defrag_range_args range;
293 int num_defrag;
294 int index;
295 int ret;
296
297 /* get the inode */
298 key.objectid = defrag->root;
299 key.type = BTRFS_ROOT_ITEM_KEY;
300 key.offset = (u64)-1;
301
302 index = srcu_read_lock(&fs_info->subvol_srcu);
303
304 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
305 if (IS_ERR(inode_root)) {
306 ret = PTR_ERR(inode_root);
307 goto cleanup;
308 }
309
310 key.objectid = defrag->ino;
311 key.type = BTRFS_INODE_ITEM_KEY;
312 key.offset = 0;
313 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
314 if (IS_ERR(inode)) {
315 ret = PTR_ERR(inode);
316 goto cleanup;
317 }
318 srcu_read_unlock(&fs_info->subvol_srcu, index);
319
320 /* do a chunk of defrag */
321 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
322 memset(&range, 0, sizeof(range));
323 range.len = (u64)-1;
324 range.start = defrag->last_offset;
325
326 sb_start_write(fs_info->sb);
327 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
328 BTRFS_DEFRAG_BATCH);
329 sb_end_write(fs_info->sb);
330 /*
331 * if we filled the whole defrag batch, there
332 * must be more work to do. Queue this defrag
333 * again
334 */
335 if (num_defrag == BTRFS_DEFRAG_BATCH) {
336 defrag->last_offset = range.start;
337 btrfs_requeue_inode_defrag(inode, defrag);
338 } else if (defrag->last_offset && !defrag->cycled) {
339 /*
340 * we didn't fill our defrag batch, but
341 * we didn't start at zero. Make sure we loop
342 * around to the start of the file.
343 */
344 defrag->last_offset = 0;
345 defrag->cycled = 1;
346 btrfs_requeue_inode_defrag(inode, defrag);
347 } else {
348 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
349 }
350
351 iput(inode);
352 return 0;
353cleanup:
354 srcu_read_unlock(&fs_info->subvol_srcu, index);
355 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
356 return ret;
357}
358
359/*
360 * run through the list of inodes in the FS that need
361 * defragging
362 */
363int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
364{
365 struct inode_defrag *defrag;
366 u64 first_ino = 0;
367 u64 root_objectid = 0;
368
369 atomic_inc(&fs_info->defrag_running);
370 while (1) {
371 /* Pause the auto defragger. */
372 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
373 &fs_info->fs_state))
374 break;
375
376 if (!__need_auto_defrag(fs_info->tree_root))
377 break;
378
379 /* find an inode to defrag */
380 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
381 first_ino);
382 if (!defrag) {
383 if (root_objectid || first_ino) {
384 root_objectid = 0;
385 first_ino = 0;
386 continue;
387 } else {
388 break;
389 }
390 }
391
392 first_ino = defrag->ino + 1;
393 root_objectid = defrag->root;
394
395 __btrfs_run_defrag_inode(fs_info, defrag);
396 }
397 atomic_dec(&fs_info->defrag_running);
398
399 /*
400 * during unmount, we use the transaction_wait queue to
401 * wait for the defragger to stop
402 */
403 wake_up(&fs_info->transaction_wait);
404 return 0;
405}
406
407/* simple helper to fault in pages and copy. This should go away
408 * and be replaced with calls into generic code.
409 */
410static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
411 struct page **prepared_pages,
412 struct iov_iter *i)
413{
414 size_t copied = 0;
415 size_t total_copied = 0;
416 int pg = 0;
417 int offset = pos & (PAGE_SIZE - 1);
418
419 while (write_bytes > 0) {
420 size_t count = min_t(size_t,
421 PAGE_SIZE - offset, write_bytes);
422 struct page *page = prepared_pages[pg];
423 /*
424 * Copy data from userspace to the current page
425 */
426 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
427
428 /* Flush processor's dcache for this page */
429 flush_dcache_page(page);
430
431 /*
432 * if we get a partial write, we can end up with
433 * partially up to date pages. These add
434 * a lot of complexity, so make sure they don't
435 * happen by forcing this copy to be retried.
436 *
437 * The rest of the btrfs_file_write code will fall
438 * back to page at a time copies after we return 0.
439 */
440 if (!PageUptodate(page) && copied < count)
441 copied = 0;
442
443 iov_iter_advance(i, copied);
444 write_bytes -= copied;
445 total_copied += copied;
446
447 /* Return to btrfs_file_write_iter to fault page */
448 if (unlikely(copied == 0))
449 break;
450
451 if (copied < PAGE_SIZE - offset) {
452 offset += copied;
453 } else {
454 pg++;
455 offset = 0;
456 }
457 }
458 return total_copied;
459}
460
461/*
462 * unlocks pages after btrfs_file_write is done with them
463 */
464static void btrfs_drop_pages(struct page **pages, size_t num_pages)
465{
466 size_t i;
467 for (i = 0; i < num_pages; i++) {
468 /* page checked is some magic around finding pages that
469 * have been modified without going through btrfs_set_page_dirty
470 * clear it here. There should be no need to mark the pages
471 * accessed as prepare_pages should have marked them accessed
472 * in prepare_pages via find_or_create_page()
473 */
474 ClearPageChecked(pages[i]);
475 unlock_page(pages[i]);
476 put_page(pages[i]);
477 }
478}
479
480/*
481 * after copy_from_user, pages need to be dirtied and we need to make
482 * sure holes are created between the current EOF and the start of
483 * any next extents (if required).
484 *
485 * this also makes the decision about creating an inline extent vs
486 * doing real data extents, marking pages dirty and delalloc as required.
487 */
488int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
489 struct page **pages, size_t num_pages,
490 loff_t pos, size_t write_bytes,
491 struct extent_state **cached)
492{
493 int err = 0;
494 int i;
495 u64 num_bytes;
496 u64 start_pos;
497 u64 end_of_last_block;
498 u64 end_pos = pos + write_bytes;
499 loff_t isize = i_size_read(inode);
500
501 start_pos = pos & ~((u64)root->sectorsize - 1);
502 num_bytes = round_up(write_bytes + pos - start_pos, root->sectorsize);
503
504 end_of_last_block = start_pos + num_bytes - 1;
505 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
506 cached);
507 if (err)
508 return err;
509
510 for (i = 0; i < num_pages; i++) {
511 struct page *p = pages[i];
512 SetPageUptodate(p);
513 ClearPageChecked(p);
514 set_page_dirty(p);
515 }
516
517 /*
518 * we've only changed i_size in ram, and we haven't updated
519 * the disk i_size. There is no need to log the inode
520 * at this time.
521 */
522 if (end_pos > isize)
523 i_size_write(inode, end_pos);
524 return 0;
525}
526
527/*
528 * this drops all the extents in the cache that intersect the range
529 * [start, end]. Existing extents are split as required.
530 */
531void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
532 int skip_pinned)
533{
534 struct extent_map *em;
535 struct extent_map *split = NULL;
536 struct extent_map *split2 = NULL;
537 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
538 u64 len = end - start + 1;
539 u64 gen;
540 int ret;
541 int testend = 1;
542 unsigned long flags;
543 int compressed = 0;
544 bool modified;
545
546 WARN_ON(end < start);
547 if (end == (u64)-1) {
548 len = (u64)-1;
549 testend = 0;
550 }
551 while (1) {
552 int no_splits = 0;
553
554 modified = false;
555 if (!split)
556 split = alloc_extent_map();
557 if (!split2)
558 split2 = alloc_extent_map();
559 if (!split || !split2)
560 no_splits = 1;
561
562 write_lock(&em_tree->lock);
563 em = lookup_extent_mapping(em_tree, start, len);
564 if (!em) {
565 write_unlock(&em_tree->lock);
566 break;
567 }
568 flags = em->flags;
569 gen = em->generation;
570 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
571 if (testend && em->start + em->len >= start + len) {
572 free_extent_map(em);
573 write_unlock(&em_tree->lock);
574 break;
575 }
576 start = em->start + em->len;
577 if (testend)
578 len = start + len - (em->start + em->len);
579 free_extent_map(em);
580 write_unlock(&em_tree->lock);
581 continue;
582 }
583 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
584 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
585 clear_bit(EXTENT_FLAG_LOGGING, &flags);
586 modified = !list_empty(&em->list);
587 if (no_splits)
588 goto next;
589
590 if (em->start < start) {
591 split->start = em->start;
592 split->len = start - em->start;
593
594 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
595 split->orig_start = em->orig_start;
596 split->block_start = em->block_start;
597
598 if (compressed)
599 split->block_len = em->block_len;
600 else
601 split->block_len = split->len;
602 split->orig_block_len = max(split->block_len,
603 em->orig_block_len);
604 split->ram_bytes = em->ram_bytes;
605 } else {
606 split->orig_start = split->start;
607 split->block_len = 0;
608 split->block_start = em->block_start;
609 split->orig_block_len = 0;
610 split->ram_bytes = split->len;
611 }
612
613 split->generation = gen;
614 split->bdev = em->bdev;
615 split->flags = flags;
616 split->compress_type = em->compress_type;
617 replace_extent_mapping(em_tree, em, split, modified);
618 free_extent_map(split);
619 split = split2;
620 split2 = NULL;
621 }
622 if (testend && em->start + em->len > start + len) {
623 u64 diff = start + len - em->start;
624
625 split->start = start + len;
626 split->len = em->start + em->len - (start + len);
627 split->bdev = em->bdev;
628 split->flags = flags;
629 split->compress_type = em->compress_type;
630 split->generation = gen;
631
632 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
633 split->orig_block_len = max(em->block_len,
634 em->orig_block_len);
635
636 split->ram_bytes = em->ram_bytes;
637 if (compressed) {
638 split->block_len = em->block_len;
639 split->block_start = em->block_start;
640 split->orig_start = em->orig_start;
641 } else {
642 split->block_len = split->len;
643 split->block_start = em->block_start
644 + diff;
645 split->orig_start = em->orig_start;
646 }
647 } else {
648 split->ram_bytes = split->len;
649 split->orig_start = split->start;
650 split->block_len = 0;
651 split->block_start = em->block_start;
652 split->orig_block_len = 0;
653 }
654
655 if (extent_map_in_tree(em)) {
656 replace_extent_mapping(em_tree, em, split,
657 modified);
658 } else {
659 ret = add_extent_mapping(em_tree, split,
660 modified);
661 ASSERT(ret == 0); /* Logic error */
662 }
663 free_extent_map(split);
664 split = NULL;
665 }
666next:
667 if (extent_map_in_tree(em))
668 remove_extent_mapping(em_tree, em);
669 write_unlock(&em_tree->lock);
670
671 /* once for us */
672 free_extent_map(em);
673 /* once for the tree*/
674 free_extent_map(em);
675 }
676 if (split)
677 free_extent_map(split);
678 if (split2)
679 free_extent_map(split2);
680}
681
682/*
683 * this is very complex, but the basic idea is to drop all extents
684 * in the range start - end. hint_block is filled in with a block number
685 * that would be a good hint to the block allocator for this file.
686 *
687 * If an extent intersects the range but is not entirely inside the range
688 * it is either truncated or split. Anything entirely inside the range
689 * is deleted from the tree.
690 */
691int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
692 struct btrfs_root *root, struct inode *inode,
693 struct btrfs_path *path, u64 start, u64 end,
694 u64 *drop_end, int drop_cache,
695 int replace_extent,
696 u32 extent_item_size,
697 int *key_inserted)
698{
699 struct extent_buffer *leaf;
700 struct btrfs_file_extent_item *fi;
701 struct btrfs_key key;
702 struct btrfs_key new_key;
703 u64 ino = btrfs_ino(inode);
704 u64 search_start = start;
705 u64 disk_bytenr = 0;
706 u64 num_bytes = 0;
707 u64 extent_offset = 0;
708 u64 extent_end = 0;
709 int del_nr = 0;
710 int del_slot = 0;
711 int extent_type;
712 int recow;
713 int ret;
714 int modify_tree = -1;
715 int update_refs;
716 int found = 0;
717 int leafs_visited = 0;
718
719 if (drop_cache)
720 btrfs_drop_extent_cache(inode, start, end - 1, 0);
721
722 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
723 modify_tree = 0;
724
725 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
726 root == root->fs_info->tree_root);
727 while (1) {
728 recow = 0;
729 ret = btrfs_lookup_file_extent(trans, root, path, ino,
730 search_start, modify_tree);
731 if (ret < 0)
732 break;
733 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
734 leaf = path->nodes[0];
735 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
736 if (key.objectid == ino &&
737 key.type == BTRFS_EXTENT_DATA_KEY)
738 path->slots[0]--;
739 }
740 ret = 0;
741 leafs_visited++;
742next_slot:
743 leaf = path->nodes[0];
744 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
745 BUG_ON(del_nr > 0);
746 ret = btrfs_next_leaf(root, path);
747 if (ret < 0)
748 break;
749 if (ret > 0) {
750 ret = 0;
751 break;
752 }
753 leafs_visited++;
754 leaf = path->nodes[0];
755 recow = 1;
756 }
757
758 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
759
760 if (key.objectid > ino)
761 break;
762 if (WARN_ON_ONCE(key.objectid < ino) ||
763 key.type < BTRFS_EXTENT_DATA_KEY) {
764 ASSERT(del_nr == 0);
765 path->slots[0]++;
766 goto next_slot;
767 }
768 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
769 break;
770
771 fi = btrfs_item_ptr(leaf, path->slots[0],
772 struct btrfs_file_extent_item);
773 extent_type = btrfs_file_extent_type(leaf, fi);
774
775 if (extent_type == BTRFS_FILE_EXTENT_REG ||
776 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
777 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
778 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
779 extent_offset = btrfs_file_extent_offset(leaf, fi);
780 extent_end = key.offset +
781 btrfs_file_extent_num_bytes(leaf, fi);
782 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
783 extent_end = key.offset +
784 btrfs_file_extent_inline_len(leaf,
785 path->slots[0], fi);
786 } else {
787 /* can't happen */
788 BUG();
789 }
790
791 /*
792 * Don't skip extent items representing 0 byte lengths. They
793 * used to be created (bug) if while punching holes we hit
794 * -ENOSPC condition. So if we find one here, just ensure we
795 * delete it, otherwise we would insert a new file extent item
796 * with the same key (offset) as that 0 bytes length file
797 * extent item in the call to setup_items_for_insert() later
798 * in this function.
799 */
800 if (extent_end == key.offset && extent_end >= search_start)
801 goto delete_extent_item;
802
803 if (extent_end <= search_start) {
804 path->slots[0]++;
805 goto next_slot;
806 }
807
808 found = 1;
809 search_start = max(key.offset, start);
810 if (recow || !modify_tree) {
811 modify_tree = -1;
812 btrfs_release_path(path);
813 continue;
814 }
815
816 /*
817 * | - range to drop - |
818 * | -------- extent -------- |
819 */
820 if (start > key.offset && end < extent_end) {
821 BUG_ON(del_nr > 0);
822 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
823 ret = -EOPNOTSUPP;
824 break;
825 }
826
827 memcpy(&new_key, &key, sizeof(new_key));
828 new_key.offset = start;
829 ret = btrfs_duplicate_item(trans, root, path,
830 &new_key);
831 if (ret == -EAGAIN) {
832 btrfs_release_path(path);
833 continue;
834 }
835 if (ret < 0)
836 break;
837
838 leaf = path->nodes[0];
839 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
840 struct btrfs_file_extent_item);
841 btrfs_set_file_extent_num_bytes(leaf, fi,
842 start - key.offset);
843
844 fi = btrfs_item_ptr(leaf, path->slots[0],
845 struct btrfs_file_extent_item);
846
847 extent_offset += start - key.offset;
848 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
849 btrfs_set_file_extent_num_bytes(leaf, fi,
850 extent_end - start);
851 btrfs_mark_buffer_dirty(leaf);
852
853 if (update_refs && disk_bytenr > 0) {
854 ret = btrfs_inc_extent_ref(trans, root,
855 disk_bytenr, num_bytes, 0,
856 root->root_key.objectid,
857 new_key.objectid,
858 start - extent_offset);
859 BUG_ON(ret); /* -ENOMEM */
860 }
861 key.offset = start;
862 }
863 /*
864 * | ---- range to drop ----- |
865 * | -------- extent -------- |
866 */
867 if (start <= key.offset && end < extent_end) {
868 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
869 ret = -EOPNOTSUPP;
870 break;
871 }
872
873 memcpy(&new_key, &key, sizeof(new_key));
874 new_key.offset = end;
875 btrfs_set_item_key_safe(root->fs_info, path, &new_key);
876
877 extent_offset += end - key.offset;
878 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
879 btrfs_set_file_extent_num_bytes(leaf, fi,
880 extent_end - end);
881 btrfs_mark_buffer_dirty(leaf);
882 if (update_refs && disk_bytenr > 0)
883 inode_sub_bytes(inode, end - key.offset);
884 break;
885 }
886
887 search_start = extent_end;
888 /*
889 * | ---- range to drop ----- |
890 * | -------- extent -------- |
891 */
892 if (start > key.offset && end >= extent_end) {
893 BUG_ON(del_nr > 0);
894 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
895 ret = -EOPNOTSUPP;
896 break;
897 }
898
899 btrfs_set_file_extent_num_bytes(leaf, fi,
900 start - key.offset);
901 btrfs_mark_buffer_dirty(leaf);
902 if (update_refs && disk_bytenr > 0)
903 inode_sub_bytes(inode, extent_end - start);
904 if (end == extent_end)
905 break;
906
907 path->slots[0]++;
908 goto next_slot;
909 }
910
911 /*
912 * | ---- range to drop ----- |
913 * | ------ extent ------ |
914 */
915 if (start <= key.offset && end >= extent_end) {
916delete_extent_item:
917 if (del_nr == 0) {
918 del_slot = path->slots[0];
919 del_nr = 1;
920 } else {
921 BUG_ON(del_slot + del_nr != path->slots[0]);
922 del_nr++;
923 }
924
925 if (update_refs &&
926 extent_type == BTRFS_FILE_EXTENT_INLINE) {
927 inode_sub_bytes(inode,
928 extent_end - key.offset);
929 extent_end = ALIGN(extent_end,
930 root->sectorsize);
931 } else if (update_refs && disk_bytenr > 0) {
932 ret = btrfs_free_extent(trans, root,
933 disk_bytenr, num_bytes, 0,
934 root->root_key.objectid,
935 key.objectid, key.offset -
936 extent_offset);
937 BUG_ON(ret); /* -ENOMEM */
938 inode_sub_bytes(inode,
939 extent_end - key.offset);
940 }
941
942 if (end == extent_end)
943 break;
944
945 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
946 path->slots[0]++;
947 goto next_slot;
948 }
949
950 ret = btrfs_del_items(trans, root, path, del_slot,
951 del_nr);
952 if (ret) {
953 btrfs_abort_transaction(trans, root, ret);
954 break;
955 }
956
957 del_nr = 0;
958 del_slot = 0;
959
960 btrfs_release_path(path);
961 continue;
962 }
963
964 BUG_ON(1);
965 }
966
967 if (!ret && del_nr > 0) {
968 /*
969 * Set path->slots[0] to first slot, so that after the delete
970 * if items are move off from our leaf to its immediate left or
971 * right neighbor leafs, we end up with a correct and adjusted
972 * path->slots[0] for our insertion (if replace_extent != 0).
973 */
974 path->slots[0] = del_slot;
975 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
976 if (ret)
977 btrfs_abort_transaction(trans, root, ret);
978 }
979
980 leaf = path->nodes[0];
981 /*
982 * If btrfs_del_items() was called, it might have deleted a leaf, in
983 * which case it unlocked our path, so check path->locks[0] matches a
984 * write lock.
985 */
986 if (!ret && replace_extent && leafs_visited == 1 &&
987 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
988 path->locks[0] == BTRFS_WRITE_LOCK) &&
989 btrfs_leaf_free_space(root, leaf) >=
990 sizeof(struct btrfs_item) + extent_item_size) {
991
992 key.objectid = ino;
993 key.type = BTRFS_EXTENT_DATA_KEY;
994 key.offset = start;
995 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
996 struct btrfs_key slot_key;
997
998 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
999 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1000 path->slots[0]++;
1001 }
1002 setup_items_for_insert(root, path, &key,
1003 &extent_item_size,
1004 extent_item_size,
1005 sizeof(struct btrfs_item) +
1006 extent_item_size, 1);
1007 *key_inserted = 1;
1008 }
1009
1010 if (!replace_extent || !(*key_inserted))
1011 btrfs_release_path(path);
1012 if (drop_end)
1013 *drop_end = found ? min(end, extent_end) : end;
1014 return ret;
1015}
1016
1017int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1018 struct btrfs_root *root, struct inode *inode, u64 start,
1019 u64 end, int drop_cache)
1020{
1021 struct btrfs_path *path;
1022 int ret;
1023
1024 path = btrfs_alloc_path();
1025 if (!path)
1026 return -ENOMEM;
1027 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1028 drop_cache, 0, 0, NULL);
1029 btrfs_free_path(path);
1030 return ret;
1031}
1032
1033static int extent_mergeable(struct extent_buffer *leaf, int slot,
1034 u64 objectid, u64 bytenr, u64 orig_offset,
1035 u64 *start, u64 *end)
1036{
1037 struct btrfs_file_extent_item *fi;
1038 struct btrfs_key key;
1039 u64 extent_end;
1040
1041 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1042 return 0;
1043
1044 btrfs_item_key_to_cpu(leaf, &key, slot);
1045 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1046 return 0;
1047
1048 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1049 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1050 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1051 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1052 btrfs_file_extent_compression(leaf, fi) ||
1053 btrfs_file_extent_encryption(leaf, fi) ||
1054 btrfs_file_extent_other_encoding(leaf, fi))
1055 return 0;
1056
1057 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1058 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1059 return 0;
1060
1061 *start = key.offset;
1062 *end = extent_end;
1063 return 1;
1064}
1065
1066/*
1067 * Mark extent in the range start - end as written.
1068 *
1069 * This changes extent type from 'pre-allocated' to 'regular'. If only
1070 * part of extent is marked as written, the extent will be split into
1071 * two or three.
1072 */
1073int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1074 struct inode *inode, u64 start, u64 end)
1075{
1076 struct btrfs_root *root = BTRFS_I(inode)->root;
1077 struct extent_buffer *leaf;
1078 struct btrfs_path *path;
1079 struct btrfs_file_extent_item *fi;
1080 struct btrfs_key key;
1081 struct btrfs_key new_key;
1082 u64 bytenr;
1083 u64 num_bytes;
1084 u64 extent_end;
1085 u64 orig_offset;
1086 u64 other_start;
1087 u64 other_end;
1088 u64 split;
1089 int del_nr = 0;
1090 int del_slot = 0;
1091 int recow;
1092 int ret;
1093 u64 ino = btrfs_ino(inode);
1094
1095 path = btrfs_alloc_path();
1096 if (!path)
1097 return -ENOMEM;
1098again:
1099 recow = 0;
1100 split = start;
1101 key.objectid = ino;
1102 key.type = BTRFS_EXTENT_DATA_KEY;
1103 key.offset = split;
1104
1105 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1106 if (ret < 0)
1107 goto out;
1108 if (ret > 0 && path->slots[0] > 0)
1109 path->slots[0]--;
1110
1111 leaf = path->nodes[0];
1112 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1113 BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1114 fi = btrfs_item_ptr(leaf, path->slots[0],
1115 struct btrfs_file_extent_item);
1116 BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1117 BTRFS_FILE_EXTENT_PREALLOC);
1118 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1119 BUG_ON(key.offset > start || extent_end < end);
1120
1121 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1122 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1123 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1124 memcpy(&new_key, &key, sizeof(new_key));
1125
1126 if (start == key.offset && end < extent_end) {
1127 other_start = 0;
1128 other_end = start;
1129 if (extent_mergeable(leaf, path->slots[0] - 1,
1130 ino, bytenr, orig_offset,
1131 &other_start, &other_end)) {
1132 new_key.offset = end;
1133 btrfs_set_item_key_safe(root->fs_info, path, &new_key);
1134 fi = btrfs_item_ptr(leaf, path->slots[0],
1135 struct btrfs_file_extent_item);
1136 btrfs_set_file_extent_generation(leaf, fi,
1137 trans->transid);
1138 btrfs_set_file_extent_num_bytes(leaf, fi,
1139 extent_end - end);
1140 btrfs_set_file_extent_offset(leaf, fi,
1141 end - orig_offset);
1142 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1143 struct btrfs_file_extent_item);
1144 btrfs_set_file_extent_generation(leaf, fi,
1145 trans->transid);
1146 btrfs_set_file_extent_num_bytes(leaf, fi,
1147 end - other_start);
1148 btrfs_mark_buffer_dirty(leaf);
1149 goto out;
1150 }
1151 }
1152
1153 if (start > key.offset && end == extent_end) {
1154 other_start = end;
1155 other_end = 0;
1156 if (extent_mergeable(leaf, path->slots[0] + 1,
1157 ino, bytenr, orig_offset,
1158 &other_start, &other_end)) {
1159 fi = btrfs_item_ptr(leaf, path->slots[0],
1160 struct btrfs_file_extent_item);
1161 btrfs_set_file_extent_num_bytes(leaf, fi,
1162 start - key.offset);
1163 btrfs_set_file_extent_generation(leaf, fi,
1164 trans->transid);
1165 path->slots[0]++;
1166 new_key.offset = start;
1167 btrfs_set_item_key_safe(root->fs_info, path, &new_key);
1168
1169 fi = btrfs_item_ptr(leaf, path->slots[0],
1170 struct btrfs_file_extent_item);
1171 btrfs_set_file_extent_generation(leaf, fi,
1172 trans->transid);
1173 btrfs_set_file_extent_num_bytes(leaf, fi,
1174 other_end - start);
1175 btrfs_set_file_extent_offset(leaf, fi,
1176 start - orig_offset);
1177 btrfs_mark_buffer_dirty(leaf);
1178 goto out;
1179 }
1180 }
1181
1182 while (start > key.offset || end < extent_end) {
1183 if (key.offset == start)
1184 split = end;
1185
1186 new_key.offset = split;
1187 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1188 if (ret == -EAGAIN) {
1189 btrfs_release_path(path);
1190 goto again;
1191 }
1192 if (ret < 0) {
1193 btrfs_abort_transaction(trans, root, ret);
1194 goto out;
1195 }
1196
1197 leaf = path->nodes[0];
1198 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1199 struct btrfs_file_extent_item);
1200 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1201 btrfs_set_file_extent_num_bytes(leaf, fi,
1202 split - key.offset);
1203
1204 fi = btrfs_item_ptr(leaf, path->slots[0],
1205 struct btrfs_file_extent_item);
1206
1207 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1208 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1209 btrfs_set_file_extent_num_bytes(leaf, fi,
1210 extent_end - split);
1211 btrfs_mark_buffer_dirty(leaf);
1212
1213 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1214 root->root_key.objectid,
1215 ino, orig_offset);
1216 BUG_ON(ret); /* -ENOMEM */
1217
1218 if (split == start) {
1219 key.offset = start;
1220 } else {
1221 BUG_ON(start != key.offset);
1222 path->slots[0]--;
1223 extent_end = end;
1224 }
1225 recow = 1;
1226 }
1227
1228 other_start = end;
1229 other_end = 0;
1230 if (extent_mergeable(leaf, path->slots[0] + 1,
1231 ino, bytenr, orig_offset,
1232 &other_start, &other_end)) {
1233 if (recow) {
1234 btrfs_release_path(path);
1235 goto again;
1236 }
1237 extent_end = other_end;
1238 del_slot = path->slots[0] + 1;
1239 del_nr++;
1240 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1241 0, root->root_key.objectid,
1242 ino, orig_offset);
1243 BUG_ON(ret); /* -ENOMEM */
1244 }
1245 other_start = 0;
1246 other_end = start;
1247 if (extent_mergeable(leaf, path->slots[0] - 1,
1248 ino, bytenr, orig_offset,
1249 &other_start, &other_end)) {
1250 if (recow) {
1251 btrfs_release_path(path);
1252 goto again;
1253 }
1254 key.offset = other_start;
1255 del_slot = path->slots[0];
1256 del_nr++;
1257 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1258 0, root->root_key.objectid,
1259 ino, orig_offset);
1260 BUG_ON(ret); /* -ENOMEM */
1261 }
1262 if (del_nr == 0) {
1263 fi = btrfs_item_ptr(leaf, path->slots[0],
1264 struct btrfs_file_extent_item);
1265 btrfs_set_file_extent_type(leaf, fi,
1266 BTRFS_FILE_EXTENT_REG);
1267 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1268 btrfs_mark_buffer_dirty(leaf);
1269 } else {
1270 fi = btrfs_item_ptr(leaf, del_slot - 1,
1271 struct btrfs_file_extent_item);
1272 btrfs_set_file_extent_type(leaf, fi,
1273 BTRFS_FILE_EXTENT_REG);
1274 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1275 btrfs_set_file_extent_num_bytes(leaf, fi,
1276 extent_end - key.offset);
1277 btrfs_mark_buffer_dirty(leaf);
1278
1279 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1280 if (ret < 0) {
1281 btrfs_abort_transaction(trans, root, ret);
1282 goto out;
1283 }
1284 }
1285out:
1286 btrfs_free_path(path);
1287 return 0;
1288}
1289
1290/*
1291 * on error we return an unlocked page and the error value
1292 * on success we return a locked page and 0
1293 */
1294static int prepare_uptodate_page(struct inode *inode,
1295 struct page *page, u64 pos,
1296 bool force_uptodate)
1297{
1298 int ret = 0;
1299
1300 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1301 !PageUptodate(page)) {
1302 ret = btrfs_readpage(NULL, page);
1303 if (ret)
1304 return ret;
1305 lock_page(page);
1306 if (!PageUptodate(page)) {
1307 unlock_page(page);
1308 return -EIO;
1309 }
1310 if (page->mapping != inode->i_mapping) {
1311 unlock_page(page);
1312 return -EAGAIN;
1313 }
1314 }
1315 return 0;
1316}
1317
1318/*
1319 * this just gets pages into the page cache and locks them down.
1320 */
1321static noinline int prepare_pages(struct inode *inode, struct page **pages,
1322 size_t num_pages, loff_t pos,
1323 size_t write_bytes, bool force_uptodate)
1324{
1325 int i;
1326 unsigned long index = pos >> PAGE_SHIFT;
1327 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1328 int err = 0;
1329 int faili;
1330
1331 for (i = 0; i < num_pages; i++) {
1332again:
1333 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1334 mask | __GFP_WRITE);
1335 if (!pages[i]) {
1336 faili = i - 1;
1337 err = -ENOMEM;
1338 goto fail;
1339 }
1340
1341 if (i == 0)
1342 err = prepare_uptodate_page(inode, pages[i], pos,
1343 force_uptodate);
1344 if (!err && i == num_pages - 1)
1345 err = prepare_uptodate_page(inode, pages[i],
1346 pos + write_bytes, false);
1347 if (err) {
1348 put_page(pages[i]);
1349 if (err == -EAGAIN) {
1350 err = 0;
1351 goto again;
1352 }
1353 faili = i - 1;
1354 goto fail;
1355 }
1356 wait_on_page_writeback(pages[i]);
1357 }
1358
1359 return 0;
1360fail:
1361 while (faili >= 0) {
1362 unlock_page(pages[faili]);
1363 put_page(pages[faili]);
1364 faili--;
1365 }
1366 return err;
1367
1368}
1369
1370/*
1371 * This function locks the extent and properly waits for data=ordered extents
1372 * to finish before allowing the pages to be modified if need.
1373 *
1374 * The return value:
1375 * 1 - the extent is locked
1376 * 0 - the extent is not locked, and everything is OK
1377 * -EAGAIN - need re-prepare the pages
1378 * the other < 0 number - Something wrong happens
1379 */
1380static noinline int
1381lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1382 size_t num_pages, loff_t pos,
1383 size_t write_bytes,
1384 u64 *lockstart, u64 *lockend,
1385 struct extent_state **cached_state)
1386{
1387 struct btrfs_root *root = BTRFS_I(inode)->root;
1388 u64 start_pos;
1389 u64 last_pos;
1390 int i;
1391 int ret = 0;
1392
1393 start_pos = round_down(pos, root->sectorsize);
1394 last_pos = start_pos
1395 + round_up(pos + write_bytes - start_pos, root->sectorsize) - 1;
1396
1397 if (start_pos < inode->i_size) {
1398 struct btrfs_ordered_extent *ordered;
1399 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1400 start_pos, last_pos, cached_state);
1401 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1402 last_pos - start_pos + 1);
1403 if (ordered &&
1404 ordered->file_offset + ordered->len > start_pos &&
1405 ordered->file_offset <= last_pos) {
1406 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1407 start_pos, last_pos,
1408 cached_state, GFP_NOFS);
1409 for (i = 0; i < num_pages; i++) {
1410 unlock_page(pages[i]);
1411 put_page(pages[i]);
1412 }
1413 btrfs_start_ordered_extent(inode, ordered, 1);
1414 btrfs_put_ordered_extent(ordered);
1415 return -EAGAIN;
1416 }
1417 if (ordered)
1418 btrfs_put_ordered_extent(ordered);
1419
1420 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1421 last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1422 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1423 0, 0, cached_state, GFP_NOFS);
1424 *lockstart = start_pos;
1425 *lockend = last_pos;
1426 ret = 1;
1427 }
1428
1429 for (i = 0; i < num_pages; i++) {
1430 if (clear_page_dirty_for_io(pages[i]))
1431 account_page_redirty(pages[i]);
1432 set_page_extent_mapped(pages[i]);
1433 WARN_ON(!PageLocked(pages[i]));
1434 }
1435
1436 return ret;
1437}
1438
1439static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1440 size_t *write_bytes)
1441{
1442 struct btrfs_root *root = BTRFS_I(inode)->root;
1443 struct btrfs_ordered_extent *ordered;
1444 u64 lockstart, lockend;
1445 u64 num_bytes;
1446 int ret;
1447
1448 ret = btrfs_start_write_no_snapshoting(root);
1449 if (!ret)
1450 return -ENOSPC;
1451
1452 lockstart = round_down(pos, root->sectorsize);
1453 lockend = round_up(pos + *write_bytes, root->sectorsize) - 1;
1454
1455 while (1) {
1456 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1457 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1458 lockend - lockstart + 1);
1459 if (!ordered) {
1460 break;
1461 }
1462 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1463 btrfs_start_ordered_extent(inode, ordered, 1);
1464 btrfs_put_ordered_extent(ordered);
1465 }
1466
1467 num_bytes = lockend - lockstart + 1;
1468 ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1469 if (ret <= 0) {
1470 ret = 0;
1471 btrfs_end_write_no_snapshoting(root);
1472 } else {
1473 *write_bytes = min_t(size_t, *write_bytes ,
1474 num_bytes - pos + lockstart);
1475 }
1476
1477 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1478
1479 return ret;
1480}
1481
1482static noinline ssize_t __btrfs_buffered_write(struct file *file,
1483 struct iov_iter *i,
1484 loff_t pos)
1485{
1486 struct inode *inode = file_inode(file);
1487 struct btrfs_root *root = BTRFS_I(inode)->root;
1488 struct page **pages = NULL;
1489 struct extent_state *cached_state = NULL;
1490 u64 release_bytes = 0;
1491 u64 lockstart;
1492 u64 lockend;
1493 size_t num_written = 0;
1494 int nrptrs;
1495 int ret = 0;
1496 bool only_release_metadata = false;
1497 bool force_page_uptodate = false;
1498 bool need_unlock;
1499
1500 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1501 PAGE_SIZE / (sizeof(struct page *)));
1502 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1503 nrptrs = max(nrptrs, 8);
1504 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1505 if (!pages)
1506 return -ENOMEM;
1507
1508 while (iov_iter_count(i) > 0) {
1509 size_t offset = pos & (PAGE_SIZE - 1);
1510 size_t sector_offset;
1511 size_t write_bytes = min(iov_iter_count(i),
1512 nrptrs * (size_t)PAGE_SIZE -
1513 offset);
1514 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1515 PAGE_SIZE);
1516 size_t reserve_bytes;
1517 size_t dirty_pages;
1518 size_t copied;
1519 size_t dirty_sectors;
1520 size_t num_sectors;
1521
1522 WARN_ON(num_pages > nrptrs);
1523
1524 /*
1525 * Fault pages before locking them in prepare_pages
1526 * to avoid recursive lock
1527 */
1528 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1529 ret = -EFAULT;
1530 break;
1531 }
1532
1533 sector_offset = pos & (root->sectorsize - 1);
1534 reserve_bytes = round_up(write_bytes + sector_offset,
1535 root->sectorsize);
1536
1537 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1538 BTRFS_INODE_PREALLOC)) &&
1539 check_can_nocow(inode, pos, &write_bytes) > 0) {
1540 /*
1541 * For nodata cow case, no need to reserve
1542 * data space.
1543 */
1544 only_release_metadata = true;
1545 /*
1546 * our prealloc extent may be smaller than
1547 * write_bytes, so scale down.
1548 */
1549 num_pages = DIV_ROUND_UP(write_bytes + offset,
1550 PAGE_SIZE);
1551 reserve_bytes = round_up(write_bytes + sector_offset,
1552 root->sectorsize);
1553 goto reserve_metadata;
1554 }
1555
1556 ret = btrfs_check_data_free_space(inode, pos, write_bytes);
1557 if (ret < 0)
1558 break;
1559
1560reserve_metadata:
1561 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1562 if (ret) {
1563 if (!only_release_metadata)
1564 btrfs_free_reserved_data_space(inode, pos,
1565 write_bytes);
1566 else
1567 btrfs_end_write_no_snapshoting(root);
1568 break;
1569 }
1570
1571 release_bytes = reserve_bytes;
1572 need_unlock = false;
1573again:
1574 /*
1575 * This is going to setup the pages array with the number of
1576 * pages we want, so we don't really need to worry about the
1577 * contents of pages from loop to loop
1578 */
1579 ret = prepare_pages(inode, pages, num_pages,
1580 pos, write_bytes,
1581 force_page_uptodate);
1582 if (ret)
1583 break;
1584
1585 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1586 pos, write_bytes, &lockstart,
1587 &lockend, &cached_state);
1588 if (ret < 0) {
1589 if (ret == -EAGAIN)
1590 goto again;
1591 break;
1592 } else if (ret > 0) {
1593 need_unlock = true;
1594 ret = 0;
1595 }
1596
1597 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1598
1599 /*
1600 * if we have trouble faulting in the pages, fall
1601 * back to one page at a time
1602 */
1603 if (copied < write_bytes)
1604 nrptrs = 1;
1605
1606 if (copied == 0) {
1607 force_page_uptodate = true;
1608 dirty_pages = 0;
1609 } else {
1610 force_page_uptodate = false;
1611 dirty_pages = DIV_ROUND_UP(copied + offset,
1612 PAGE_SIZE);
1613 }
1614
1615 /*
1616 * If we had a short copy we need to release the excess delaloc
1617 * bytes we reserved. We need to increment outstanding_extents
1618 * because btrfs_delalloc_release_space will decrement it, but
1619 * we still have an outstanding extent for the chunk we actually
1620 * managed to copy.
1621 */
1622 num_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
1623 reserve_bytes);
1624 dirty_sectors = round_up(copied + sector_offset,
1625 root->sectorsize);
1626 dirty_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
1627 dirty_sectors);
1628
1629 if (num_sectors > dirty_sectors) {
1630 release_bytes = (write_bytes - copied)
1631 & ~((u64)root->sectorsize - 1);
1632 if (copied > 0) {
1633 spin_lock(&BTRFS_I(inode)->lock);
1634 BTRFS_I(inode)->outstanding_extents++;
1635 spin_unlock(&BTRFS_I(inode)->lock);
1636 }
1637 if (only_release_metadata) {
1638 btrfs_delalloc_release_metadata(inode,
1639 release_bytes);
1640 } else {
1641 u64 __pos;
1642
1643 __pos = round_down(pos, root->sectorsize) +
1644 (dirty_pages << PAGE_SHIFT);
1645 btrfs_delalloc_release_space(inode, __pos,
1646 release_bytes);
1647 }
1648 }
1649
1650 release_bytes = round_up(copied + sector_offset,
1651 root->sectorsize);
1652
1653 if (copied > 0)
1654 ret = btrfs_dirty_pages(root, inode, pages,
1655 dirty_pages, pos, copied,
1656 NULL);
1657 if (need_unlock)
1658 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1659 lockstart, lockend, &cached_state,
1660 GFP_NOFS);
1661 if (ret) {
1662 btrfs_drop_pages(pages, num_pages);
1663 break;
1664 }
1665
1666 release_bytes = 0;
1667 if (only_release_metadata)
1668 btrfs_end_write_no_snapshoting(root);
1669
1670 if (only_release_metadata && copied > 0) {
1671 lockstart = round_down(pos, root->sectorsize);
1672 lockend = round_up(pos + copied, root->sectorsize) - 1;
1673
1674 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1675 lockend, EXTENT_NORESERVE, NULL,
1676 NULL, GFP_NOFS);
1677 only_release_metadata = false;
1678 }
1679
1680 btrfs_drop_pages(pages, num_pages);
1681
1682 cond_resched();
1683
1684 balance_dirty_pages_ratelimited(inode->i_mapping);
1685 if (dirty_pages < (root->nodesize >> PAGE_SHIFT) + 1)
1686 btrfs_btree_balance_dirty(root);
1687
1688 pos += copied;
1689 num_written += copied;
1690 }
1691
1692 kfree(pages);
1693
1694 if (release_bytes) {
1695 if (only_release_metadata) {
1696 btrfs_end_write_no_snapshoting(root);
1697 btrfs_delalloc_release_metadata(inode, release_bytes);
1698 } else {
1699 btrfs_delalloc_release_space(inode, pos, release_bytes);
1700 }
1701 }
1702
1703 return num_written ? num_written : ret;
1704}
1705
1706static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1707 struct iov_iter *from,
1708 loff_t pos)
1709{
1710 struct file *file = iocb->ki_filp;
1711 struct inode *inode = file_inode(file);
1712 ssize_t written;
1713 ssize_t written_buffered;
1714 loff_t endbyte;
1715 int err;
1716
1717 written = generic_file_direct_write(iocb, from, pos);
1718
1719 if (written < 0 || !iov_iter_count(from))
1720 return written;
1721
1722 pos += written;
1723 written_buffered = __btrfs_buffered_write(file, from, pos);
1724 if (written_buffered < 0) {
1725 err = written_buffered;
1726 goto out;
1727 }
1728 /*
1729 * Ensure all data is persisted. We want the next direct IO read to be
1730 * able to read what was just written.
1731 */
1732 endbyte = pos + written_buffered - 1;
1733 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1734 if (err)
1735 goto out;
1736 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1737 if (err)
1738 goto out;
1739 written += written_buffered;
1740 iocb->ki_pos = pos + written_buffered;
1741 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1742 endbyte >> PAGE_SHIFT);
1743out:
1744 return written ? written : err;
1745}
1746
1747static void update_time_for_write(struct inode *inode)
1748{
1749 struct timespec now;
1750
1751 if (IS_NOCMTIME(inode))
1752 return;
1753
1754 now = current_fs_time(inode->i_sb);
1755 if (!timespec_equal(&inode->i_mtime, &now))
1756 inode->i_mtime = now;
1757
1758 if (!timespec_equal(&inode->i_ctime, &now))
1759 inode->i_ctime = now;
1760
1761 if (IS_I_VERSION(inode))
1762 inode_inc_iversion(inode);
1763}
1764
1765static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1766 struct iov_iter *from)
1767{
1768 struct file *file = iocb->ki_filp;
1769 struct inode *inode = file_inode(file);
1770 struct btrfs_root *root = BTRFS_I(inode)->root;
1771 u64 start_pos;
1772 u64 end_pos;
1773 ssize_t num_written = 0;
1774 bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1775 ssize_t err;
1776 loff_t pos;
1777 size_t count;
1778 loff_t oldsize;
1779 int clean_page = 0;
1780
1781 inode_lock(inode);
1782 err = generic_write_checks(iocb, from);
1783 if (err <= 0) {
1784 inode_unlock(inode);
1785 return err;
1786 }
1787
1788 current->backing_dev_info = inode_to_bdi(inode);
1789 err = file_remove_privs(file);
1790 if (err) {
1791 inode_unlock(inode);
1792 goto out;
1793 }
1794
1795 /*
1796 * If BTRFS flips readonly due to some impossible error
1797 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1798 * although we have opened a file as writable, we have
1799 * to stop this write operation to ensure FS consistency.
1800 */
1801 if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1802 inode_unlock(inode);
1803 err = -EROFS;
1804 goto out;
1805 }
1806
1807 /*
1808 * We reserve space for updating the inode when we reserve space for the
1809 * extent we are going to write, so we will enospc out there. We don't
1810 * need to start yet another transaction to update the inode as we will
1811 * update the inode when we finish writing whatever data we write.
1812 */
1813 update_time_for_write(inode);
1814
1815 pos = iocb->ki_pos;
1816 count = iov_iter_count(from);
1817 start_pos = round_down(pos, root->sectorsize);
1818 oldsize = i_size_read(inode);
1819 if (start_pos > oldsize) {
1820 /* Expand hole size to cover write data, preventing empty gap */
1821 end_pos = round_up(pos + count, root->sectorsize);
1822 err = btrfs_cont_expand(inode, oldsize, end_pos);
1823 if (err) {
1824 inode_unlock(inode);
1825 goto out;
1826 }
1827 if (start_pos > round_up(oldsize, root->sectorsize))
1828 clean_page = 1;
1829 }
1830
1831 if (sync)
1832 atomic_inc(&BTRFS_I(inode)->sync_writers);
1833
1834 if (iocb->ki_flags & IOCB_DIRECT) {
1835 num_written = __btrfs_direct_write(iocb, from, pos);
1836 } else {
1837 num_written = __btrfs_buffered_write(file, from, pos);
1838 if (num_written > 0)
1839 iocb->ki_pos = pos + num_written;
1840 if (clean_page)
1841 pagecache_isize_extended(inode, oldsize,
1842 i_size_read(inode));
1843 }
1844
1845 inode_unlock(inode);
1846
1847 /*
1848 * We also have to set last_sub_trans to the current log transid,
1849 * otherwise subsequent syncs to a file that's been synced in this
1850 * transaction will appear to have already occurred.
1851 */
1852 spin_lock(&BTRFS_I(inode)->lock);
1853 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1854 spin_unlock(&BTRFS_I(inode)->lock);
1855 if (num_written > 0) {
1856 err = generic_write_sync(file, pos, num_written);
1857 if (err < 0)
1858 num_written = err;
1859 }
1860
1861 if (sync)
1862 atomic_dec(&BTRFS_I(inode)->sync_writers);
1863out:
1864 current->backing_dev_info = NULL;
1865 return num_written ? num_written : err;
1866}
1867
1868int btrfs_release_file(struct inode *inode, struct file *filp)
1869{
1870 if (filp->private_data)
1871 btrfs_ioctl_trans_end(filp);
1872 /*
1873 * ordered_data_close is set by settattr when we are about to truncate
1874 * a file from a non-zero size to a zero size. This tries to
1875 * flush down new bytes that may have been written if the
1876 * application were using truncate to replace a file in place.
1877 */
1878 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1879 &BTRFS_I(inode)->runtime_flags))
1880 filemap_flush(inode->i_mapping);
1881 return 0;
1882}
1883
1884static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1885{
1886 int ret;
1887
1888 atomic_inc(&BTRFS_I(inode)->sync_writers);
1889 ret = btrfs_fdatawrite_range(inode, start, end);
1890 atomic_dec(&BTRFS_I(inode)->sync_writers);
1891
1892 return ret;
1893}
1894
1895/*
1896 * fsync call for both files and directories. This logs the inode into
1897 * the tree log instead of forcing full commits whenever possible.
1898 *
1899 * It needs to call filemap_fdatawait so that all ordered extent updates are
1900 * in the metadata btree are up to date for copying to the log.
1901 *
1902 * It drops the inode mutex before doing the tree log commit. This is an
1903 * important optimization for directories because holding the mutex prevents
1904 * new operations on the dir while we write to disk.
1905 */
1906int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1907{
1908 struct dentry *dentry = file_dentry(file);
1909 struct inode *inode = d_inode(dentry);
1910 struct btrfs_root *root = BTRFS_I(inode)->root;
1911 struct btrfs_trans_handle *trans;
1912 struct btrfs_log_ctx ctx;
1913 int ret = 0;
1914 bool full_sync = 0;
1915 u64 len;
1916
1917 /*
1918 * The range length can be represented by u64, we have to do the typecasts
1919 * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
1920 */
1921 len = (u64)end - (u64)start + 1;
1922 trace_btrfs_sync_file(file, datasync);
1923
1924 /*
1925 * We write the dirty pages in the range and wait until they complete
1926 * out of the ->i_mutex. If so, we can flush the dirty pages by
1927 * multi-task, and make the performance up. See
1928 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1929 */
1930 ret = start_ordered_ops(inode, start, end);
1931 if (ret)
1932 return ret;
1933
1934 inode_lock(inode);
1935 atomic_inc(&root->log_batch);
1936 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1937 &BTRFS_I(inode)->runtime_flags);
1938 /*
1939 * We might have have had more pages made dirty after calling
1940 * start_ordered_ops and before acquiring the inode's i_mutex.
1941 */
1942 if (full_sync) {
1943 /*
1944 * For a full sync, we need to make sure any ordered operations
1945 * start and finish before we start logging the inode, so that
1946 * all extents are persisted and the respective file extent
1947 * items are in the fs/subvol btree.
1948 */
1949 ret = btrfs_wait_ordered_range(inode, start, len);
1950 } else {
1951 /*
1952 * Start any new ordered operations before starting to log the
1953 * inode. We will wait for them to finish in btrfs_sync_log().
1954 *
1955 * Right before acquiring the inode's mutex, we might have new
1956 * writes dirtying pages, which won't immediately start the
1957 * respective ordered operations - that is done through the
1958 * fill_delalloc callbacks invoked from the writepage and
1959 * writepages address space operations. So make sure we start
1960 * all ordered operations before starting to log our inode. Not
1961 * doing this means that while logging the inode, writeback
1962 * could start and invoke writepage/writepages, which would call
1963 * the fill_delalloc callbacks (cow_file_range,
1964 * submit_compressed_extents). These callbacks add first an
1965 * extent map to the modified list of extents and then create
1966 * the respective ordered operation, which means in
1967 * tree-log.c:btrfs_log_inode() we might capture all existing
1968 * ordered operations (with btrfs_get_logged_extents()) before
1969 * the fill_delalloc callback adds its ordered operation, and by
1970 * the time we visit the modified list of extent maps (with
1971 * btrfs_log_changed_extents()), we see and process the extent
1972 * map they created. We then use the extent map to construct a
1973 * file extent item for logging without waiting for the
1974 * respective ordered operation to finish - this file extent
1975 * item points to a disk location that might not have yet been
1976 * written to, containing random data - so after a crash a log
1977 * replay will make our inode have file extent items that point
1978 * to disk locations containing invalid data, as we returned
1979 * success to userspace without waiting for the respective
1980 * ordered operation to finish, because it wasn't captured by
1981 * btrfs_get_logged_extents().
1982 */
1983 ret = start_ordered_ops(inode, start, end);
1984 }
1985 if (ret) {
1986 inode_unlock(inode);
1987 goto out;
1988 }
1989 atomic_inc(&root->log_batch);
1990
1991 /*
1992 * If the last transaction that changed this file was before the current
1993 * transaction and we have the full sync flag set in our inode, we can
1994 * bail out now without any syncing.
1995 *
1996 * Note that we can't bail out if the full sync flag isn't set. This is
1997 * because when the full sync flag is set we start all ordered extents
1998 * and wait for them to fully complete - when they complete they update
1999 * the inode's last_trans field through:
2000 *
2001 * btrfs_finish_ordered_io() ->
2002 * btrfs_update_inode_fallback() ->
2003 * btrfs_update_inode() ->
2004 * btrfs_set_inode_last_trans()
2005 *
2006 * So we are sure that last_trans is up to date and can do this check to
2007 * bail out safely. For the fast path, when the full sync flag is not
2008 * set in our inode, we can not do it because we start only our ordered
2009 * extents and don't wait for them to complete (that is when
2010 * btrfs_finish_ordered_io runs), so here at this point their last_trans
2011 * value might be less than or equals to fs_info->last_trans_committed,
2012 * and setting a speculative last_trans for an inode when a buffered
2013 * write is made (such as fs_info->generation + 1 for example) would not
2014 * be reliable since after setting the value and before fsync is called
2015 * any number of transactions can start and commit (transaction kthread
2016 * commits the current transaction periodically), and a transaction
2017 * commit does not start nor waits for ordered extents to complete.
2018 */
2019 smp_mb();
2020 if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
2021 (full_sync && BTRFS_I(inode)->last_trans <=
2022 root->fs_info->last_trans_committed) ||
2023 (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2024 BTRFS_I(inode)->last_trans
2025 <= root->fs_info->last_trans_committed)) {
2026 /*
2027 * We'v had everything committed since the last time we were
2028 * modified so clear this flag in case it was set for whatever
2029 * reason, it's no longer relevant.
2030 */
2031 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2032 &BTRFS_I(inode)->runtime_flags);
2033 inode_unlock(inode);
2034 goto out;
2035 }
2036
2037 /*
2038 * ok we haven't committed the transaction yet, lets do a commit
2039 */
2040 if (file->private_data)
2041 btrfs_ioctl_trans_end(file);
2042
2043 /*
2044 * We use start here because we will need to wait on the IO to complete
2045 * in btrfs_sync_log, which could require joining a transaction (for
2046 * example checking cross references in the nocow path). If we use join
2047 * here we could get into a situation where we're waiting on IO to
2048 * happen that is blocked on a transaction trying to commit. With start
2049 * we inc the extwriter counter, so we wait for all extwriters to exit
2050 * before we start blocking join'ers. This comment is to keep somebody
2051 * from thinking they are super smart and changing this to
2052 * btrfs_join_transaction *cough*Josef*cough*.
2053 */
2054 trans = btrfs_start_transaction(root, 0);
2055 if (IS_ERR(trans)) {
2056 ret = PTR_ERR(trans);
2057 inode_unlock(inode);
2058 goto out;
2059 }
2060 trans->sync = true;
2061
2062 btrfs_init_log_ctx(&ctx);
2063
2064 ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2065 if (ret < 0) {
2066 /* Fallthrough and commit/free transaction. */
2067 ret = 1;
2068 }
2069
2070 /* we've logged all the items and now have a consistent
2071 * version of the file in the log. It is possible that
2072 * someone will come in and modify the file, but that's
2073 * fine because the log is consistent on disk, and we
2074 * have references to all of the file's extents
2075 *
2076 * It is possible that someone will come in and log the
2077 * file again, but that will end up using the synchronization
2078 * inside btrfs_sync_log to keep things safe.
2079 */
2080 inode_unlock(inode);
2081
2082 /*
2083 * If any of the ordered extents had an error, just return it to user
2084 * space, so that the application knows some writes didn't succeed and
2085 * can take proper action (retry for e.g.). Blindly committing the
2086 * transaction in this case, would fool userspace that everything was
2087 * successful. And we also want to make sure our log doesn't contain
2088 * file extent items pointing to extents that weren't fully written to -
2089 * just like in the non fast fsync path, where we check for the ordered
2090 * operation's error flag before writing to the log tree and return -EIO
2091 * if any of them had this flag set (btrfs_wait_ordered_range) -
2092 * therefore we need to check for errors in the ordered operations,
2093 * which are indicated by ctx.io_err.
2094 */
2095 if (ctx.io_err) {
2096 btrfs_end_transaction(trans, root);
2097 ret = ctx.io_err;
2098 goto out;
2099 }
2100
2101 if (ret != BTRFS_NO_LOG_SYNC) {
2102 if (!ret) {
2103 ret = btrfs_sync_log(trans, root, &ctx);
2104 if (!ret) {
2105 ret = btrfs_end_transaction(trans, root);
2106 goto out;
2107 }
2108 }
2109 if (!full_sync) {
2110 ret = btrfs_wait_ordered_range(inode, start, len);
2111 if (ret) {
2112 btrfs_end_transaction(trans, root);
2113 goto out;
2114 }
2115 }
2116 ret = btrfs_commit_transaction(trans, root);
2117 } else {
2118 ret = btrfs_end_transaction(trans, root);
2119 }
2120out:
2121 return ret > 0 ? -EIO : ret;
2122}
2123
2124static const struct vm_operations_struct btrfs_file_vm_ops = {
2125 .fault = filemap_fault,
2126 .map_pages = filemap_map_pages,
2127 .page_mkwrite = btrfs_page_mkwrite,
2128};
2129
2130static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2131{
2132 struct address_space *mapping = filp->f_mapping;
2133
2134 if (!mapping->a_ops->readpage)
2135 return -ENOEXEC;
2136
2137 file_accessed(filp);
2138 vma->vm_ops = &btrfs_file_vm_ops;
2139
2140 return 0;
2141}
2142
2143static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2144 int slot, u64 start, u64 end)
2145{
2146 struct btrfs_file_extent_item *fi;
2147 struct btrfs_key key;
2148
2149 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2150 return 0;
2151
2152 btrfs_item_key_to_cpu(leaf, &key, slot);
2153 if (key.objectid != btrfs_ino(inode) ||
2154 key.type != BTRFS_EXTENT_DATA_KEY)
2155 return 0;
2156
2157 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2158
2159 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2160 return 0;
2161
2162 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2163 return 0;
2164
2165 if (key.offset == end)
2166 return 1;
2167 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2168 return 1;
2169 return 0;
2170}
2171
2172static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2173 struct btrfs_path *path, u64 offset, u64 end)
2174{
2175 struct btrfs_root *root = BTRFS_I(inode)->root;
2176 struct extent_buffer *leaf;
2177 struct btrfs_file_extent_item *fi;
2178 struct extent_map *hole_em;
2179 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2180 struct btrfs_key key;
2181 int ret;
2182
2183 if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
2184 goto out;
2185
2186 key.objectid = btrfs_ino(inode);
2187 key.type = BTRFS_EXTENT_DATA_KEY;
2188 key.offset = offset;
2189
2190 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2191 if (ret < 0)
2192 return ret;
2193 BUG_ON(!ret);
2194
2195 leaf = path->nodes[0];
2196 if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2197 u64 num_bytes;
2198
2199 path->slots[0]--;
2200 fi = btrfs_item_ptr(leaf, path->slots[0],
2201 struct btrfs_file_extent_item);
2202 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2203 end - offset;
2204 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2205 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2206 btrfs_set_file_extent_offset(leaf, fi, 0);
2207 btrfs_mark_buffer_dirty(leaf);
2208 goto out;
2209 }
2210
2211 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2212 u64 num_bytes;
2213
2214 key.offset = offset;
2215 btrfs_set_item_key_safe(root->fs_info, path, &key);
2216 fi = btrfs_item_ptr(leaf, path->slots[0],
2217 struct btrfs_file_extent_item);
2218 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2219 offset;
2220 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2221 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2222 btrfs_set_file_extent_offset(leaf, fi, 0);
2223 btrfs_mark_buffer_dirty(leaf);
2224 goto out;
2225 }
2226 btrfs_release_path(path);
2227
2228 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2229 0, 0, end - offset, 0, end - offset,
2230 0, 0, 0);
2231 if (ret)
2232 return ret;
2233
2234out:
2235 btrfs_release_path(path);
2236
2237 hole_em = alloc_extent_map();
2238 if (!hole_em) {
2239 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2240 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2241 &BTRFS_I(inode)->runtime_flags);
2242 } else {
2243 hole_em->start = offset;
2244 hole_em->len = end - offset;
2245 hole_em->ram_bytes = hole_em->len;
2246 hole_em->orig_start = offset;
2247
2248 hole_em->block_start = EXTENT_MAP_HOLE;
2249 hole_em->block_len = 0;
2250 hole_em->orig_block_len = 0;
2251 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2252 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2253 hole_em->generation = trans->transid;
2254
2255 do {
2256 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2257 write_lock(&em_tree->lock);
2258 ret = add_extent_mapping(em_tree, hole_em, 1);
2259 write_unlock(&em_tree->lock);
2260 } while (ret == -EEXIST);
2261 free_extent_map(hole_em);
2262 if (ret)
2263 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2264 &BTRFS_I(inode)->runtime_flags);
2265 }
2266
2267 return 0;
2268}
2269
2270/*
2271 * Find a hole extent on given inode and change start/len to the end of hole
2272 * extent.(hole/vacuum extent whose em->start <= start &&
2273 * em->start + em->len > start)
2274 * When a hole extent is found, return 1 and modify start/len.
2275 */
2276static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2277{
2278 struct extent_map *em;
2279 int ret = 0;
2280
2281 em = btrfs_get_extent(inode, NULL, 0, *start, *len, 0);
2282 if (IS_ERR_OR_NULL(em)) {
2283 if (!em)
2284 ret = -ENOMEM;
2285 else
2286 ret = PTR_ERR(em);
2287 return ret;
2288 }
2289
2290 /* Hole or vacuum extent(only exists in no-hole mode) */
2291 if (em->block_start == EXTENT_MAP_HOLE) {
2292 ret = 1;
2293 *len = em->start + em->len > *start + *len ?
2294 0 : *start + *len - em->start - em->len;
2295 *start = em->start + em->len;
2296 }
2297 free_extent_map(em);
2298 return ret;
2299}
2300
2301static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2302{
2303 struct btrfs_root *root = BTRFS_I(inode)->root;
2304 struct extent_state *cached_state = NULL;
2305 struct btrfs_path *path;
2306 struct btrfs_block_rsv *rsv;
2307 struct btrfs_trans_handle *trans;
2308 u64 lockstart;
2309 u64 lockend;
2310 u64 tail_start;
2311 u64 tail_len;
2312 u64 orig_start = offset;
2313 u64 cur_offset;
2314 u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2315 u64 drop_end;
2316 int ret = 0;
2317 int err = 0;
2318 unsigned int rsv_count;
2319 bool same_block;
2320 bool no_holes = btrfs_fs_incompat(root->fs_info, NO_HOLES);
2321 u64 ino_size;
2322 bool truncated_block = false;
2323 bool updated_inode = false;
2324
2325 ret = btrfs_wait_ordered_range(inode, offset, len);
2326 if (ret)
2327 return ret;
2328
2329 inode_lock(inode);
2330 ino_size = round_up(inode->i_size, root->sectorsize);
2331 ret = find_first_non_hole(inode, &offset, &len);
2332 if (ret < 0)
2333 goto out_only_mutex;
2334 if (ret && !len) {
2335 /* Already in a large hole */
2336 ret = 0;
2337 goto out_only_mutex;
2338 }
2339
2340 lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
2341 lockend = round_down(offset + len,
2342 BTRFS_I(inode)->root->sectorsize) - 1;
2343 same_block = (BTRFS_BYTES_TO_BLKS(root->fs_info, offset))
2344 == (BTRFS_BYTES_TO_BLKS(root->fs_info, offset + len - 1));
2345 /*
2346 * We needn't truncate any block which is beyond the end of the file
2347 * because we are sure there is no data there.
2348 */
2349 /*
2350 * Only do this if we are in the same block and we aren't doing the
2351 * entire block.
2352 */
2353 if (same_block && len < root->sectorsize) {
2354 if (offset < ino_size) {
2355 truncated_block = true;
2356 ret = btrfs_truncate_block(inode, offset, len, 0);
2357 } else {
2358 ret = 0;
2359 }
2360 goto out_only_mutex;
2361 }
2362
2363 /* zero back part of the first block */
2364 if (offset < ino_size) {
2365 truncated_block = true;
2366 ret = btrfs_truncate_block(inode, offset, 0, 0);
2367 if (ret) {
2368 inode_unlock(inode);
2369 return ret;
2370 }
2371 }
2372
2373 /* Check the aligned pages after the first unaligned page,
2374 * if offset != orig_start, which means the first unaligned page
2375 * including serveral following pages are already in holes,
2376 * the extra check can be skipped */
2377 if (offset == orig_start) {
2378 /* after truncate page, check hole again */
2379 len = offset + len - lockstart;
2380 offset = lockstart;
2381 ret = find_first_non_hole(inode, &offset, &len);
2382 if (ret < 0)
2383 goto out_only_mutex;
2384 if (ret && !len) {
2385 ret = 0;
2386 goto out_only_mutex;
2387 }
2388 lockstart = offset;
2389 }
2390
2391 /* Check the tail unaligned part is in a hole */
2392 tail_start = lockend + 1;
2393 tail_len = offset + len - tail_start;
2394 if (tail_len) {
2395 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2396 if (unlikely(ret < 0))
2397 goto out_only_mutex;
2398 if (!ret) {
2399 /* zero the front end of the last page */
2400 if (tail_start + tail_len < ino_size) {
2401 truncated_block = true;
2402 ret = btrfs_truncate_block(inode,
2403 tail_start + tail_len,
2404 0, 1);
2405 if (ret)
2406 goto out_only_mutex;
2407 }
2408 }
2409 }
2410
2411 if (lockend < lockstart) {
2412 ret = 0;
2413 goto out_only_mutex;
2414 }
2415
2416 while (1) {
2417 struct btrfs_ordered_extent *ordered;
2418
2419 truncate_pagecache_range(inode, lockstart, lockend);
2420
2421 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2422 &cached_state);
2423 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2424
2425 /*
2426 * We need to make sure we have no ordered extents in this range
2427 * and nobody raced in and read a page in this range, if we did
2428 * we need to try again.
2429 */
2430 if ((!ordered ||
2431 (ordered->file_offset + ordered->len <= lockstart ||
2432 ordered->file_offset > lockend)) &&
2433 !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2434 if (ordered)
2435 btrfs_put_ordered_extent(ordered);
2436 break;
2437 }
2438 if (ordered)
2439 btrfs_put_ordered_extent(ordered);
2440 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2441 lockend, &cached_state, GFP_NOFS);
2442 ret = btrfs_wait_ordered_range(inode, lockstart,
2443 lockend - lockstart + 1);
2444 if (ret) {
2445 inode_unlock(inode);
2446 return ret;
2447 }
2448 }
2449
2450 path = btrfs_alloc_path();
2451 if (!path) {
2452 ret = -ENOMEM;
2453 goto out;
2454 }
2455
2456 rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2457 if (!rsv) {
2458 ret = -ENOMEM;
2459 goto out_free;
2460 }
2461 rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2462 rsv->failfast = 1;
2463
2464 /*
2465 * 1 - update the inode
2466 * 1 - removing the extents in the range
2467 * 1 - adding the hole extent if no_holes isn't set
2468 */
2469 rsv_count = no_holes ? 2 : 3;
2470 trans = btrfs_start_transaction(root, rsv_count);
2471 if (IS_ERR(trans)) {
2472 err = PTR_ERR(trans);
2473 goto out_free;
2474 }
2475
2476 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2477 min_size);
2478 BUG_ON(ret);
2479 trans->block_rsv = rsv;
2480
2481 cur_offset = lockstart;
2482 len = lockend - cur_offset;
2483 while (cur_offset < lockend) {
2484 ret = __btrfs_drop_extents(trans, root, inode, path,
2485 cur_offset, lockend + 1,
2486 &drop_end, 1, 0, 0, NULL);
2487 if (ret != -ENOSPC)
2488 break;
2489
2490 trans->block_rsv = &root->fs_info->trans_block_rsv;
2491
2492 if (cur_offset < ino_size) {
2493 ret = fill_holes(trans, inode, path, cur_offset,
2494 drop_end);
2495 if (ret) {
2496 err = ret;
2497 break;
2498 }
2499 }
2500
2501 cur_offset = drop_end;
2502
2503 ret = btrfs_update_inode(trans, root, inode);
2504 if (ret) {
2505 err = ret;
2506 break;
2507 }
2508
2509 btrfs_end_transaction(trans, root);
2510 btrfs_btree_balance_dirty(root);
2511
2512 trans = btrfs_start_transaction(root, rsv_count);
2513 if (IS_ERR(trans)) {
2514 ret = PTR_ERR(trans);
2515 trans = NULL;
2516 break;
2517 }
2518
2519 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2520 rsv, min_size);
2521 BUG_ON(ret); /* shouldn't happen */
2522 trans->block_rsv = rsv;
2523
2524 ret = find_first_non_hole(inode, &cur_offset, &len);
2525 if (unlikely(ret < 0))
2526 break;
2527 if (ret && !len) {
2528 ret = 0;
2529 break;
2530 }
2531 }
2532
2533 if (ret) {
2534 err = ret;
2535 goto out_trans;
2536 }
2537
2538 trans->block_rsv = &root->fs_info->trans_block_rsv;
2539 /*
2540 * If we are using the NO_HOLES feature we might have had already an
2541 * hole that overlaps a part of the region [lockstart, lockend] and
2542 * ends at (or beyond) lockend. Since we have no file extent items to
2543 * represent holes, drop_end can be less than lockend and so we must
2544 * make sure we have an extent map representing the existing hole (the
2545 * call to __btrfs_drop_extents() might have dropped the existing extent
2546 * map representing the existing hole), otherwise the fast fsync path
2547 * will not record the existence of the hole region
2548 * [existing_hole_start, lockend].
2549 */
2550 if (drop_end <= lockend)
2551 drop_end = lockend + 1;
2552 /*
2553 * Don't insert file hole extent item if it's for a range beyond eof
2554 * (because it's useless) or if it represents a 0 bytes range (when
2555 * cur_offset == drop_end).
2556 */
2557 if (cur_offset < ino_size && cur_offset < drop_end) {
2558 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2559 if (ret) {
2560 err = ret;
2561 goto out_trans;
2562 }
2563 }
2564
2565out_trans:
2566 if (!trans)
2567 goto out_free;
2568
2569 inode_inc_iversion(inode);
2570 inode->i_mtime = inode->i_ctime = current_fs_time(inode->i_sb);
2571
2572 trans->block_rsv = &root->fs_info->trans_block_rsv;
2573 ret = btrfs_update_inode(trans, root, inode);
2574 updated_inode = true;
2575 btrfs_end_transaction(trans, root);
2576 btrfs_btree_balance_dirty(root);
2577out_free:
2578 btrfs_free_path(path);
2579 btrfs_free_block_rsv(root, rsv);
2580out:
2581 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2582 &cached_state, GFP_NOFS);
2583out_only_mutex:
2584 if (!updated_inode && truncated_block && !ret && !err) {
2585 /*
2586 * If we only end up zeroing part of a page, we still need to
2587 * update the inode item, so that all the time fields are
2588 * updated as well as the necessary btrfs inode in memory fields
2589 * for detecting, at fsync time, if the inode isn't yet in the
2590 * log tree or it's there but not up to date.
2591 */
2592 trans = btrfs_start_transaction(root, 1);
2593 if (IS_ERR(trans)) {
2594 err = PTR_ERR(trans);
2595 } else {
2596 err = btrfs_update_inode(trans, root, inode);
2597 ret = btrfs_end_transaction(trans, root);
2598 }
2599 }
2600 inode_unlock(inode);
2601 if (ret && !err)
2602 err = ret;
2603 return err;
2604}
2605
2606/* Helper structure to record which range is already reserved */
2607struct falloc_range {
2608 struct list_head list;
2609 u64 start;
2610 u64 len;
2611};
2612
2613/*
2614 * Helper function to add falloc range
2615 *
2616 * Caller should have locked the larger range of extent containing
2617 * [start, len)
2618 */
2619static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2620{
2621 struct falloc_range *prev = NULL;
2622 struct falloc_range *range = NULL;
2623
2624 if (list_empty(head))
2625 goto insert;
2626
2627 /*
2628 * As fallocate iterate by bytenr order, we only need to check
2629 * the last range.
2630 */
2631 prev = list_entry(head->prev, struct falloc_range, list);
2632 if (prev->start + prev->len == start) {
2633 prev->len += len;
2634 return 0;
2635 }
2636insert:
2637 range = kmalloc(sizeof(*range), GFP_KERNEL);
2638 if (!range)
2639 return -ENOMEM;
2640 range->start = start;
2641 range->len = len;
2642 list_add_tail(&range->list, head);
2643 return 0;
2644}
2645
2646static long btrfs_fallocate(struct file *file, int mode,
2647 loff_t offset, loff_t len)
2648{
2649 struct inode *inode = file_inode(file);
2650 struct extent_state *cached_state = NULL;
2651 struct falloc_range *range;
2652 struct falloc_range *tmp;
2653 struct list_head reserve_list;
2654 u64 cur_offset;
2655 u64 last_byte;
2656 u64 alloc_start;
2657 u64 alloc_end;
2658 u64 alloc_hint = 0;
2659 u64 locked_end;
2660 u64 actual_end = 0;
2661 struct extent_map *em;
2662 int blocksize = BTRFS_I(inode)->root->sectorsize;
2663 int ret;
2664
2665 alloc_start = round_down(offset, blocksize);
2666 alloc_end = round_up(offset + len, blocksize);
2667
2668 /* Make sure we aren't being give some crap mode */
2669 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2670 return -EOPNOTSUPP;
2671
2672 if (mode & FALLOC_FL_PUNCH_HOLE)
2673 return btrfs_punch_hole(inode, offset, len);
2674
2675 /*
2676 * Only trigger disk allocation, don't trigger qgroup reserve
2677 *
2678 * For qgroup space, it will be checked later.
2679 */
2680 ret = btrfs_alloc_data_chunk_ondemand(inode, alloc_end - alloc_start);
2681 if (ret < 0)
2682 return ret;
2683
2684 inode_lock(inode);
2685
2686 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2687 ret = inode_newsize_ok(inode, offset + len);
2688 if (ret)
2689 goto out;
2690 }
2691
2692 /*
2693 * TODO: Move these two operations after we have checked
2694 * accurate reserved space, or fallocate can still fail but
2695 * with page truncated or size expanded.
2696 *
2697 * But that's a minor problem and won't do much harm BTW.
2698 */
2699 if (alloc_start > inode->i_size) {
2700 ret = btrfs_cont_expand(inode, i_size_read(inode),
2701 alloc_start);
2702 if (ret)
2703 goto out;
2704 } else if (offset + len > inode->i_size) {
2705 /*
2706 * If we are fallocating from the end of the file onward we
2707 * need to zero out the end of the block if i_size lands in the
2708 * middle of a block.
2709 */
2710 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2711 if (ret)
2712 goto out;
2713 }
2714
2715 /*
2716 * wait for ordered IO before we have any locks. We'll loop again
2717 * below with the locks held.
2718 */
2719 ret = btrfs_wait_ordered_range(inode, alloc_start,
2720 alloc_end - alloc_start);
2721 if (ret)
2722 goto out;
2723
2724 locked_end = alloc_end - 1;
2725 while (1) {
2726 struct btrfs_ordered_extent *ordered;
2727
2728 /* the extent lock is ordered inside the running
2729 * transaction
2730 */
2731 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2732 locked_end, &cached_state);
2733 ordered = btrfs_lookup_first_ordered_extent(inode,
2734 alloc_end - 1);
2735 if (ordered &&
2736 ordered->file_offset + ordered->len > alloc_start &&
2737 ordered->file_offset < alloc_end) {
2738 btrfs_put_ordered_extent(ordered);
2739 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2740 alloc_start, locked_end,
2741 &cached_state, GFP_KERNEL);
2742 /*
2743 * we can't wait on the range with the transaction
2744 * running or with the extent lock held
2745 */
2746 ret = btrfs_wait_ordered_range(inode, alloc_start,
2747 alloc_end - alloc_start);
2748 if (ret)
2749 goto out;
2750 } else {
2751 if (ordered)
2752 btrfs_put_ordered_extent(ordered);
2753 break;
2754 }
2755 }
2756
2757 /* First, check if we exceed the qgroup limit */
2758 INIT_LIST_HEAD(&reserve_list);
2759 cur_offset = alloc_start;
2760 while (1) {
2761 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2762 alloc_end - cur_offset, 0);
2763 if (IS_ERR_OR_NULL(em)) {
2764 if (!em)
2765 ret = -ENOMEM;
2766 else
2767 ret = PTR_ERR(em);
2768 break;
2769 }
2770 last_byte = min(extent_map_end(em), alloc_end);
2771 actual_end = min_t(u64, extent_map_end(em), offset + len);
2772 last_byte = ALIGN(last_byte, blocksize);
2773 if (em->block_start == EXTENT_MAP_HOLE ||
2774 (cur_offset >= inode->i_size &&
2775 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2776 ret = add_falloc_range(&reserve_list, cur_offset,
2777 last_byte - cur_offset);
2778 if (ret < 0) {
2779 free_extent_map(em);
2780 break;
2781 }
2782 ret = btrfs_qgroup_reserve_data(inode, cur_offset,
2783 last_byte - cur_offset);
2784 if (ret < 0)
2785 break;
2786 }
2787 free_extent_map(em);
2788 cur_offset = last_byte;
2789 if (cur_offset >= alloc_end)
2790 break;
2791 }
2792
2793 /*
2794 * If ret is still 0, means we're OK to fallocate.
2795 * Or just cleanup the list and exit.
2796 */
2797 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2798 if (!ret)
2799 ret = btrfs_prealloc_file_range(inode, mode,
2800 range->start,
2801 range->len, 1 << inode->i_blkbits,
2802 offset + len, &alloc_hint);
2803 list_del(&range->list);
2804 kfree(range);
2805 }
2806 if (ret < 0)
2807 goto out_unlock;
2808
2809 if (actual_end > inode->i_size &&
2810 !(mode & FALLOC_FL_KEEP_SIZE)) {
2811 struct btrfs_trans_handle *trans;
2812 struct btrfs_root *root = BTRFS_I(inode)->root;
2813
2814 /*
2815 * We didn't need to allocate any more space, but we
2816 * still extended the size of the file so we need to
2817 * update i_size and the inode item.
2818 */
2819 trans = btrfs_start_transaction(root, 1);
2820 if (IS_ERR(trans)) {
2821 ret = PTR_ERR(trans);
2822 } else {
2823 inode->i_ctime = current_fs_time(inode->i_sb);
2824 i_size_write(inode, actual_end);
2825 btrfs_ordered_update_i_size(inode, actual_end, NULL);
2826 ret = btrfs_update_inode(trans, root, inode);
2827 if (ret)
2828 btrfs_end_transaction(trans, root);
2829 else
2830 ret = btrfs_end_transaction(trans, root);
2831 }
2832 }
2833out_unlock:
2834 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2835 &cached_state, GFP_KERNEL);
2836out:
2837 /*
2838 * As we waited the extent range, the data_rsv_map must be empty
2839 * in the range, as written data range will be released from it.
2840 * And for prealloacted extent, it will also be released when
2841 * its metadata is written.
2842 * So this is completely used as cleanup.
2843 */
2844 btrfs_qgroup_free_data(inode, alloc_start, alloc_end - alloc_start);
2845 inode_unlock(inode);
2846 /* Let go of our reservation. */
2847 btrfs_free_reserved_data_space(inode, alloc_start,
2848 alloc_end - alloc_start);
2849 return ret;
2850}
2851
2852static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2853{
2854 struct btrfs_root *root = BTRFS_I(inode)->root;
2855 struct extent_map *em = NULL;
2856 struct extent_state *cached_state = NULL;
2857 u64 lockstart;
2858 u64 lockend;
2859 u64 start;
2860 u64 len;
2861 int ret = 0;
2862
2863 if (inode->i_size == 0)
2864 return -ENXIO;
2865
2866 /*
2867 * *offset can be negative, in this case we start finding DATA/HOLE from
2868 * the very start of the file.
2869 */
2870 start = max_t(loff_t, 0, *offset);
2871
2872 lockstart = round_down(start, root->sectorsize);
2873 lockend = round_up(i_size_read(inode), root->sectorsize);
2874 if (lockend <= lockstart)
2875 lockend = lockstart + root->sectorsize;
2876 lockend--;
2877 len = lockend - lockstart + 1;
2878
2879 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2880 &cached_state);
2881
2882 while (start < inode->i_size) {
2883 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2884 if (IS_ERR(em)) {
2885 ret = PTR_ERR(em);
2886 em = NULL;
2887 break;
2888 }
2889
2890 if (whence == SEEK_HOLE &&
2891 (em->block_start == EXTENT_MAP_HOLE ||
2892 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2893 break;
2894 else if (whence == SEEK_DATA &&
2895 (em->block_start != EXTENT_MAP_HOLE &&
2896 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2897 break;
2898
2899 start = em->start + em->len;
2900 free_extent_map(em);
2901 em = NULL;
2902 cond_resched();
2903 }
2904 free_extent_map(em);
2905 if (!ret) {
2906 if (whence == SEEK_DATA && start >= inode->i_size)
2907 ret = -ENXIO;
2908 else
2909 *offset = min_t(loff_t, start, inode->i_size);
2910 }
2911 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2912 &cached_state, GFP_NOFS);
2913 return ret;
2914}
2915
2916static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2917{
2918 struct inode *inode = file->f_mapping->host;
2919 int ret;
2920
2921 inode_lock(inode);
2922 switch (whence) {
2923 case SEEK_END:
2924 case SEEK_CUR:
2925 offset = generic_file_llseek(file, offset, whence);
2926 goto out;
2927 case SEEK_DATA:
2928 case SEEK_HOLE:
2929 if (offset >= i_size_read(inode)) {
2930 inode_unlock(inode);
2931 return -ENXIO;
2932 }
2933
2934 ret = find_desired_extent(inode, &offset, whence);
2935 if (ret) {
2936 inode_unlock(inode);
2937 return ret;
2938 }
2939 }
2940
2941 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2942out:
2943 inode_unlock(inode);
2944 return offset;
2945}
2946
2947const struct file_operations btrfs_file_operations = {
2948 .llseek = btrfs_file_llseek,
2949 .read_iter = generic_file_read_iter,
2950 .splice_read = generic_file_splice_read,
2951 .write_iter = btrfs_file_write_iter,
2952 .mmap = btrfs_file_mmap,
2953 .open = generic_file_open,
2954 .release = btrfs_release_file,
2955 .fsync = btrfs_sync_file,
2956 .fallocate = btrfs_fallocate,
2957 .unlocked_ioctl = btrfs_ioctl,
2958#ifdef CONFIG_COMPAT
2959 .compat_ioctl = btrfs_ioctl,
2960#endif
2961 .copy_file_range = btrfs_copy_file_range,
2962 .clone_file_range = btrfs_clone_file_range,
2963 .dedupe_file_range = btrfs_dedupe_file_range,
2964};
2965
2966void btrfs_auto_defrag_exit(void)
2967{
2968 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2969}
2970
2971int btrfs_auto_defrag_init(void)
2972{
2973 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2974 sizeof(struct inode_defrag), 0,
2975 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2976 NULL);
2977 if (!btrfs_inode_defrag_cachep)
2978 return -ENOMEM;
2979
2980 return 0;
2981}
2982
2983int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
2984{
2985 int ret;
2986
2987 /*
2988 * So with compression we will find and lock a dirty page and clear the
2989 * first one as dirty, setup an async extent, and immediately return
2990 * with the entire range locked but with nobody actually marked with
2991 * writeback. So we can't just filemap_write_and_wait_range() and
2992 * expect it to work since it will just kick off a thread to do the
2993 * actual work. So we need to call filemap_fdatawrite_range _again_
2994 * since it will wait on the page lock, which won't be unlocked until
2995 * after the pages have been marked as writeback and so we're good to go
2996 * from there. We have to do this otherwise we'll miss the ordered
2997 * extents and that results in badness. Please Josef, do not think you
2998 * know better and pull this out at some point in the future, it is
2999 * right and you are wrong.
3000 */
3001 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3002 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3003 &BTRFS_I(inode)->runtime_flags))
3004 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3005
3006 return ret;
3007}