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