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