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 "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}
1/*
2 * Copyright (C) 2007 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/fs.h>
20#include <linux/pagemap.h>
21#include <linux/highmem.h>
22#include <linux/time.h>
23#include <linux/init.h>
24#include <linux/string.h>
25#include <linux/backing-dev.h>
26#include <linux/mpage.h>
27#include <linux/falloc.h>
28#include <linux/swap.h>
29#include <linux/writeback.h>
30#include <linux/compat.h>
31#include <linux/slab.h>
32#include <linux/btrfs.h>
33#include <linux/uio.h>
34#include "ctree.h"
35#include "disk-io.h"
36#include "transaction.h"
37#include "btrfs_inode.h"
38#include "print-tree.h"
39#include "tree-log.h"
40#include "locking.h"
41#include "volumes.h"
42#include "qgroup.h"
43#include "compression.h"
44
45static struct kmem_cache *btrfs_inode_defrag_cachep;
46/*
47 * when auto defrag is enabled we
48 * queue up these defrag structs to remember which
49 * inodes need defragging passes
50 */
51struct inode_defrag {
52 struct rb_node rb_node;
53 /* objectid */
54 u64 ino;
55 /*
56 * transid where the defrag was added, we search for
57 * extents newer than this
58 */
59 u64 transid;
60
61 /* root objectid */
62 u64 root;
63
64 /* last offset we were able to defrag */
65 u64 last_offset;
66
67 /* if we've wrapped around back to zero once already */
68 int cycled;
69};
70
71static int __compare_inode_defrag(struct inode_defrag *defrag1,
72 struct inode_defrag *defrag2)
73{
74 if (defrag1->root > defrag2->root)
75 return 1;
76 else if (defrag1->root < defrag2->root)
77 return -1;
78 else if (defrag1->ino > defrag2->ino)
79 return 1;
80 else if (defrag1->ino < defrag2->ino)
81 return -1;
82 else
83 return 0;
84}
85
86/* pop a record for an inode into the defrag tree. The lock
87 * must be held already
88 *
89 * If you're inserting a record for an older transid than an
90 * existing record, the transid already in the tree is lowered
91 *
92 * If an existing record is found the defrag item you
93 * pass in is freed
94 */
95static int __btrfs_add_inode_defrag(struct inode *inode,
96 struct inode_defrag *defrag)
97{
98 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
99 struct inode_defrag *entry;
100 struct rb_node **p;
101 struct rb_node *parent = NULL;
102 int ret;
103
104 p = &fs_info->defrag_inodes.rb_node;
105 while (*p) {
106 parent = *p;
107 entry = rb_entry(parent, struct inode_defrag, rb_node);
108
109 ret = __compare_inode_defrag(defrag, entry);
110 if (ret < 0)
111 p = &parent->rb_left;
112 else if (ret > 0)
113 p = &parent->rb_right;
114 else {
115 /* if we're reinserting an entry for
116 * an old defrag run, make sure to
117 * lower the transid of our existing record
118 */
119 if (defrag->transid < entry->transid)
120 entry->transid = defrag->transid;
121 if (defrag->last_offset > entry->last_offset)
122 entry->last_offset = defrag->last_offset;
123 return -EEXIST;
124 }
125 }
126 set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
127 rb_link_node(&defrag->rb_node, parent, p);
128 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
129 return 0;
130}
131
132static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
133{
134 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
135 return 0;
136
137 if (btrfs_fs_closing(fs_info))
138 return 0;
139
140 return 1;
141}
142
143/*
144 * insert a defrag record for this inode if auto defrag is
145 * enabled
146 */
147int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
148 struct inode *inode)
149{
150 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
151 struct btrfs_root *root = BTRFS_I(inode)->root;
152 struct inode_defrag *defrag;
153 u64 transid;
154 int ret;
155
156 if (!__need_auto_defrag(fs_info))
157 return 0;
158
159 if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
160 return 0;
161
162 if (trans)
163 transid = trans->transid;
164 else
165 transid = BTRFS_I(inode)->root->last_trans;
166
167 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
168 if (!defrag)
169 return -ENOMEM;
170
171 defrag->ino = btrfs_ino(inode);
172 defrag->transid = transid;
173 defrag->root = root->root_key.objectid;
174
175 spin_lock(&fs_info->defrag_inodes_lock);
176 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
177 /*
178 * If we set IN_DEFRAG flag and evict the inode from memory,
179 * and then re-read this inode, this new inode doesn't have
180 * IN_DEFRAG flag. At the case, we may find the existed defrag.
181 */
182 ret = __btrfs_add_inode_defrag(inode, defrag);
183 if (ret)
184 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
185 } else {
186 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
187 }
188 spin_unlock(&fs_info->defrag_inodes_lock);
189 return 0;
190}
191
192/*
193 * Requeue the defrag object. If there is a defrag object that points to
194 * the same inode in the tree, we will merge them together (by
195 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
196 */
197static void btrfs_requeue_inode_defrag(struct inode *inode,
198 struct inode_defrag *defrag)
199{
200 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
201 int ret;
202
203 if (!__need_auto_defrag(fs_info))
204 goto out;
205
206 /*
207 * Here we don't check the IN_DEFRAG flag, because we need merge
208 * them together.
209 */
210 spin_lock(&fs_info->defrag_inodes_lock);
211 ret = __btrfs_add_inode_defrag(inode, defrag);
212 spin_unlock(&fs_info->defrag_inodes_lock);
213 if (ret)
214 goto out;
215 return;
216out:
217 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
218}
219
220/*
221 * pick the defragable inode that we want, if it doesn't exist, we will get
222 * the next one.
223 */
224static struct inode_defrag *
225btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
226{
227 struct inode_defrag *entry = NULL;
228 struct inode_defrag tmp;
229 struct rb_node *p;
230 struct rb_node *parent = NULL;
231 int ret;
232
233 tmp.ino = ino;
234 tmp.root = root;
235
236 spin_lock(&fs_info->defrag_inodes_lock);
237 p = fs_info->defrag_inodes.rb_node;
238 while (p) {
239 parent = p;
240 entry = rb_entry(parent, struct inode_defrag, rb_node);
241
242 ret = __compare_inode_defrag(&tmp, entry);
243 if (ret < 0)
244 p = parent->rb_left;
245 else if (ret > 0)
246 p = parent->rb_right;
247 else
248 goto out;
249 }
250
251 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
252 parent = rb_next(parent);
253 if (parent)
254 entry = rb_entry(parent, struct inode_defrag, rb_node);
255 else
256 entry = NULL;
257 }
258out:
259 if (entry)
260 rb_erase(parent, &fs_info->defrag_inodes);
261 spin_unlock(&fs_info->defrag_inodes_lock);
262 return entry;
263}
264
265void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
266{
267 struct inode_defrag *defrag;
268 struct rb_node *node;
269
270 spin_lock(&fs_info->defrag_inodes_lock);
271 node = rb_first(&fs_info->defrag_inodes);
272 while (node) {
273 rb_erase(node, &fs_info->defrag_inodes);
274 defrag = rb_entry(node, struct inode_defrag, rb_node);
275 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
276
277 cond_resched_lock(&fs_info->defrag_inodes_lock);
278
279 node = rb_first(&fs_info->defrag_inodes);
280 }
281 spin_unlock(&fs_info->defrag_inodes_lock);
282}
283
284#define BTRFS_DEFRAG_BATCH 1024
285
286static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
287 struct inode_defrag *defrag)
288{
289 struct btrfs_root *inode_root;
290 struct inode *inode;
291 struct btrfs_key key;
292 struct btrfs_ioctl_defrag_range_args range;
293 int num_defrag;
294 int index;
295 int ret;
296
297 /* get the inode */
298 key.objectid = defrag->root;
299 key.type = BTRFS_ROOT_ITEM_KEY;
300 key.offset = (u64)-1;
301
302 index = srcu_read_lock(&fs_info->subvol_srcu);
303
304 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
305 if (IS_ERR(inode_root)) {
306 ret = PTR_ERR(inode_root);
307 goto cleanup;
308 }
309
310 key.objectid = defrag->ino;
311 key.type = BTRFS_INODE_ITEM_KEY;
312 key.offset = 0;
313 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
314 if (IS_ERR(inode)) {
315 ret = PTR_ERR(inode);
316 goto cleanup;
317 }
318 srcu_read_unlock(&fs_info->subvol_srcu, index);
319
320 /* do a chunk of defrag */
321 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
322 memset(&range, 0, sizeof(range));
323 range.len = (u64)-1;
324 range.start = defrag->last_offset;
325
326 sb_start_write(fs_info->sb);
327 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
328 BTRFS_DEFRAG_BATCH);
329 sb_end_write(fs_info->sb);
330 /*
331 * if we filled the whole defrag batch, there
332 * must be more work to do. Queue this defrag
333 * again
334 */
335 if (num_defrag == BTRFS_DEFRAG_BATCH) {
336 defrag->last_offset = range.start;
337 btrfs_requeue_inode_defrag(inode, defrag);
338 } else if (defrag->last_offset && !defrag->cycled) {
339 /*
340 * we didn't fill our defrag batch, but
341 * we didn't start at zero. Make sure we loop
342 * around to the start of the file.
343 */
344 defrag->last_offset = 0;
345 defrag->cycled = 1;
346 btrfs_requeue_inode_defrag(inode, defrag);
347 } else {
348 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
349 }
350
351 iput(inode);
352 return 0;
353cleanup:
354 srcu_read_unlock(&fs_info->subvol_srcu, index);
355 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
356 return ret;
357}
358
359/*
360 * run through the list of inodes in the FS that need
361 * defragging
362 */
363int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
364{
365 struct inode_defrag *defrag;
366 u64 first_ino = 0;
367 u64 root_objectid = 0;
368
369 atomic_inc(&fs_info->defrag_running);
370 while (1) {
371 /* Pause the auto defragger. */
372 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
373 &fs_info->fs_state))
374 break;
375
376 if (!__need_auto_defrag(fs_info))
377 break;
378
379 /* find an inode to defrag */
380 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
381 first_ino);
382 if (!defrag) {
383 if (root_objectid || first_ino) {
384 root_objectid = 0;
385 first_ino = 0;
386 continue;
387 } else {
388 break;
389 }
390 }
391
392 first_ino = defrag->ino + 1;
393 root_objectid = defrag->root;
394
395 __btrfs_run_defrag_inode(fs_info, defrag);
396 }
397 atomic_dec(&fs_info->defrag_running);
398
399 /*
400 * during unmount, we use the transaction_wait queue to
401 * wait for the defragger to stop
402 */
403 wake_up(&fs_info->transaction_wait);
404 return 0;
405}
406
407/* simple helper to fault in pages and copy. This should go away
408 * and be replaced with calls into generic code.
409 */
410static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
411 struct page **prepared_pages,
412 struct iov_iter *i)
413{
414 size_t copied = 0;
415 size_t total_copied = 0;
416 int pg = 0;
417 int offset = pos & (PAGE_SIZE - 1);
418
419 while (write_bytes > 0) {
420 size_t count = min_t(size_t,
421 PAGE_SIZE - offset, write_bytes);
422 struct page *page = prepared_pages[pg];
423 /*
424 * Copy data from userspace to the current page
425 */
426 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
427
428 /* Flush processor's dcache for this page */
429 flush_dcache_page(page);
430
431 /*
432 * if we get a partial write, we can end up with
433 * partially up to date pages. These add
434 * a lot of complexity, so make sure they don't
435 * happen by forcing this copy to be retried.
436 *
437 * The rest of the btrfs_file_write code will fall
438 * back to page at a time copies after we return 0.
439 */
440 if (!PageUptodate(page) && copied < count)
441 copied = 0;
442
443 iov_iter_advance(i, copied);
444 write_bytes -= copied;
445 total_copied += copied;
446
447 /* Return to btrfs_file_write_iter to fault page */
448 if (unlikely(copied == 0))
449 break;
450
451 if (copied < PAGE_SIZE - offset) {
452 offset += copied;
453 } else {
454 pg++;
455 offset = 0;
456 }
457 }
458 return total_copied;
459}
460
461/*
462 * unlocks pages after btrfs_file_write is done with them
463 */
464static void btrfs_drop_pages(struct page **pages, size_t num_pages)
465{
466 size_t i;
467 for (i = 0; i < num_pages; i++) {
468 /* page checked is some magic around finding pages that
469 * have been modified without going through btrfs_set_page_dirty
470 * clear it here. There should be no need to mark the pages
471 * accessed as prepare_pages should have marked them accessed
472 * in prepare_pages via find_or_create_page()
473 */
474 ClearPageChecked(pages[i]);
475 unlock_page(pages[i]);
476 put_page(pages[i]);
477 }
478}
479
480/*
481 * after copy_from_user, pages need to be dirtied and we need to make
482 * sure holes are created between the current EOF and the start of
483 * any next extents (if required).
484 *
485 * this also makes the decision about creating an inline extent vs
486 * doing real data extents, marking pages dirty and delalloc as required.
487 */
488int btrfs_dirty_pages(struct inode *inode, struct page **pages,
489 size_t num_pages, loff_t pos, size_t write_bytes,
490 struct extent_state **cached)
491{
492 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
493 int err = 0;
494 int i;
495 u64 num_bytes;
496 u64 start_pos;
497 u64 end_of_last_block;
498 u64 end_pos = pos + write_bytes;
499 loff_t isize = i_size_read(inode);
500
501 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
502 num_bytes = round_up(write_bytes + pos - start_pos,
503 fs_info->sectorsize);
504
505 end_of_last_block = start_pos + num_bytes - 1;
506 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
507 cached, 0);
508 if (err)
509 return err;
510
511 for (i = 0; i < num_pages; i++) {
512 struct page *p = pages[i];
513 SetPageUptodate(p);
514 ClearPageChecked(p);
515 set_page_dirty(p);
516 }
517
518 /*
519 * we've only changed i_size in ram, and we haven't updated
520 * the disk i_size. There is no need to log the inode
521 * at this time.
522 */
523 if (end_pos > isize)
524 i_size_write(inode, end_pos);
525 return 0;
526}
527
528/*
529 * this drops all the extents in the cache that intersect the range
530 * [start, end]. Existing extents are split as required.
531 */
532void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
533 int skip_pinned)
534{
535 struct extent_map *em;
536 struct extent_map *split = NULL;
537 struct extent_map *split2 = NULL;
538 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
539 u64 len = end - start + 1;
540 u64 gen;
541 int ret;
542 int testend = 1;
543 unsigned long flags;
544 int compressed = 0;
545 bool modified;
546
547 WARN_ON(end < start);
548 if (end == (u64)-1) {
549 len = (u64)-1;
550 testend = 0;
551 }
552 while (1) {
553 int no_splits = 0;
554
555 modified = false;
556 if (!split)
557 split = alloc_extent_map();
558 if (!split2)
559 split2 = alloc_extent_map();
560 if (!split || !split2)
561 no_splits = 1;
562
563 write_lock(&em_tree->lock);
564 em = lookup_extent_mapping(em_tree, start, len);
565 if (!em) {
566 write_unlock(&em_tree->lock);
567 break;
568 }
569 flags = em->flags;
570 gen = em->generation;
571 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
572 if (testend && em->start + em->len >= start + len) {
573 free_extent_map(em);
574 write_unlock(&em_tree->lock);
575 break;
576 }
577 start = em->start + em->len;
578 if (testend)
579 len = start + len - (em->start + em->len);
580 free_extent_map(em);
581 write_unlock(&em_tree->lock);
582 continue;
583 }
584 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
585 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
586 clear_bit(EXTENT_FLAG_LOGGING, &flags);
587 modified = !list_empty(&em->list);
588 if (no_splits)
589 goto next;
590
591 if (em->start < start) {
592 split->start = em->start;
593 split->len = start - em->start;
594
595 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
596 split->orig_start = em->orig_start;
597 split->block_start = em->block_start;
598
599 if (compressed)
600 split->block_len = em->block_len;
601 else
602 split->block_len = split->len;
603 split->orig_block_len = max(split->block_len,
604 em->orig_block_len);
605 split->ram_bytes = em->ram_bytes;
606 } else {
607 split->orig_start = split->start;
608 split->block_len = 0;
609 split->block_start = em->block_start;
610 split->orig_block_len = 0;
611 split->ram_bytes = split->len;
612 }
613
614 split->generation = gen;
615 split->bdev = em->bdev;
616 split->flags = flags;
617 split->compress_type = em->compress_type;
618 replace_extent_mapping(em_tree, em, split, modified);
619 free_extent_map(split);
620 split = split2;
621 split2 = NULL;
622 }
623 if (testend && em->start + em->len > start + len) {
624 u64 diff = start + len - em->start;
625
626 split->start = start + len;
627 split->len = em->start + em->len - (start + len);
628 split->bdev = em->bdev;
629 split->flags = flags;
630 split->compress_type = em->compress_type;
631 split->generation = gen;
632
633 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
634 split->orig_block_len = max(em->block_len,
635 em->orig_block_len);
636
637 split->ram_bytes = em->ram_bytes;
638 if (compressed) {
639 split->block_len = em->block_len;
640 split->block_start = em->block_start;
641 split->orig_start = em->orig_start;
642 } else {
643 split->block_len = split->len;
644 split->block_start = em->block_start
645 + diff;
646 split->orig_start = em->orig_start;
647 }
648 } else {
649 split->ram_bytes = split->len;
650 split->orig_start = split->start;
651 split->block_len = 0;
652 split->block_start = em->block_start;
653 split->orig_block_len = 0;
654 }
655
656 if (extent_map_in_tree(em)) {
657 replace_extent_mapping(em_tree, em, split,
658 modified);
659 } else {
660 ret = add_extent_mapping(em_tree, split,
661 modified);
662 ASSERT(ret == 0); /* Logic error */
663 }
664 free_extent_map(split);
665 split = NULL;
666 }
667next:
668 if (extent_map_in_tree(em))
669 remove_extent_mapping(em_tree, em);
670 write_unlock(&em_tree->lock);
671
672 /* once for us */
673 free_extent_map(em);
674 /* once for the tree*/
675 free_extent_map(em);
676 }
677 if (split)
678 free_extent_map(split);
679 if (split2)
680 free_extent_map(split2);
681}
682
683/*
684 * this is very complex, but the basic idea is to drop all extents
685 * in the range start - end. hint_block is filled in with a block number
686 * that would be a good hint to the block allocator for this file.
687 *
688 * If an extent intersects the range but is not entirely inside the range
689 * it is either truncated or split. Anything entirely inside the range
690 * is deleted from the tree.
691 */
692int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
693 struct btrfs_root *root, struct inode *inode,
694 struct btrfs_path *path, u64 start, u64 end,
695 u64 *drop_end, int drop_cache,
696 int replace_extent,
697 u32 extent_item_size,
698 int *key_inserted)
699{
700 struct btrfs_fs_info *fs_info = root->fs_info;
701 struct extent_buffer *leaf;
702 struct btrfs_file_extent_item *fi;
703 struct btrfs_key key;
704 struct btrfs_key new_key;
705 u64 ino = btrfs_ino(inode);
706 u64 search_start = start;
707 u64 disk_bytenr = 0;
708 u64 num_bytes = 0;
709 u64 extent_offset = 0;
710 u64 extent_end = 0;
711 u64 last_end = start;
712 int del_nr = 0;
713 int del_slot = 0;
714 int extent_type;
715 int recow;
716 int ret;
717 int modify_tree = -1;
718 int update_refs;
719 int found = 0;
720 int leafs_visited = 0;
721
722 if (drop_cache)
723 btrfs_drop_extent_cache(inode, start, end - 1, 0);
724
725 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
726 modify_tree = 0;
727
728 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
729 root == fs_info->tree_root);
730 while (1) {
731 recow = 0;
732 ret = btrfs_lookup_file_extent(trans, root, path, ino,
733 search_start, modify_tree);
734 if (ret < 0)
735 break;
736 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
737 leaf = path->nodes[0];
738 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
739 if (key.objectid == ino &&
740 key.type == BTRFS_EXTENT_DATA_KEY)
741 path->slots[0]--;
742 }
743 ret = 0;
744 leafs_visited++;
745next_slot:
746 leaf = path->nodes[0];
747 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
748 BUG_ON(del_nr > 0);
749 ret = btrfs_next_leaf(root, path);
750 if (ret < 0)
751 break;
752 if (ret > 0) {
753 ret = 0;
754 break;
755 }
756 leafs_visited++;
757 leaf = path->nodes[0];
758 recow = 1;
759 }
760
761 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
762
763 if (key.objectid > ino)
764 break;
765 if (WARN_ON_ONCE(key.objectid < ino) ||
766 key.type < BTRFS_EXTENT_DATA_KEY) {
767 ASSERT(del_nr == 0);
768 path->slots[0]++;
769 goto next_slot;
770 }
771 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
772 break;
773
774 fi = btrfs_item_ptr(leaf, path->slots[0],
775 struct btrfs_file_extent_item);
776 extent_type = btrfs_file_extent_type(leaf, fi);
777
778 if (extent_type == BTRFS_FILE_EXTENT_REG ||
779 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
780 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
781 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
782 extent_offset = btrfs_file_extent_offset(leaf, fi);
783 extent_end = key.offset +
784 btrfs_file_extent_num_bytes(leaf, fi);
785 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
786 extent_end = key.offset +
787 btrfs_file_extent_inline_len(leaf,
788 path->slots[0], fi);
789 } else {
790 /* can't happen */
791 BUG();
792 }
793
794 /*
795 * Don't skip extent items representing 0 byte lengths. They
796 * used to be created (bug) if while punching holes we hit
797 * -ENOSPC condition. So if we find one here, just ensure we
798 * delete it, otherwise we would insert a new file extent item
799 * with the same key (offset) as that 0 bytes length file
800 * extent item in the call to setup_items_for_insert() later
801 * in this function.
802 */
803 if (extent_end == key.offset && extent_end >= search_start) {
804 last_end = extent_end;
805 goto delete_extent_item;
806 }
807
808 if (extent_end <= search_start) {
809 path->slots[0]++;
810 goto next_slot;
811 }
812
813 found = 1;
814 search_start = max(key.offset, start);
815 if (recow || !modify_tree) {
816 modify_tree = -1;
817 btrfs_release_path(path);
818 continue;
819 }
820
821 /*
822 * | - range to drop - |
823 * | -------- extent -------- |
824 */
825 if (start > key.offset && end < extent_end) {
826 BUG_ON(del_nr > 0);
827 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
828 ret = -EOPNOTSUPP;
829 break;
830 }
831
832 memcpy(&new_key, &key, sizeof(new_key));
833 new_key.offset = start;
834 ret = btrfs_duplicate_item(trans, root, path,
835 &new_key);
836 if (ret == -EAGAIN) {
837 btrfs_release_path(path);
838 continue;
839 }
840 if (ret < 0)
841 break;
842
843 leaf = path->nodes[0];
844 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
845 struct btrfs_file_extent_item);
846 btrfs_set_file_extent_num_bytes(leaf, fi,
847 start - key.offset);
848
849 fi = btrfs_item_ptr(leaf, path->slots[0],
850 struct btrfs_file_extent_item);
851
852 extent_offset += start - key.offset;
853 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
854 btrfs_set_file_extent_num_bytes(leaf, fi,
855 extent_end - start);
856 btrfs_mark_buffer_dirty(leaf);
857
858 if (update_refs && disk_bytenr > 0) {
859 ret = btrfs_inc_extent_ref(trans, fs_info,
860 disk_bytenr, num_bytes, 0,
861 root->root_key.objectid,
862 new_key.objectid,
863 start - extent_offset);
864 BUG_ON(ret); /* -ENOMEM */
865 }
866 key.offset = start;
867 }
868 /*
869 * From here on out we will have actually dropped something, so
870 * last_end can be updated.
871 */
872 last_end = extent_end;
873
874 /*
875 * | ---- range to drop ----- |
876 * | -------- extent -------- |
877 */
878 if (start <= key.offset && end < extent_end) {
879 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
880 ret = -EOPNOTSUPP;
881 break;
882 }
883
884 memcpy(&new_key, &key, sizeof(new_key));
885 new_key.offset = end;
886 btrfs_set_item_key_safe(fs_info, path, &new_key);
887
888 extent_offset += end - key.offset;
889 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
890 btrfs_set_file_extent_num_bytes(leaf, fi,
891 extent_end - end);
892 btrfs_mark_buffer_dirty(leaf);
893 if (update_refs && disk_bytenr > 0)
894 inode_sub_bytes(inode, end - key.offset);
895 break;
896 }
897
898 search_start = extent_end;
899 /*
900 * | ---- range to drop ----- |
901 * | -------- extent -------- |
902 */
903 if (start > key.offset && end >= extent_end) {
904 BUG_ON(del_nr > 0);
905 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
906 ret = -EOPNOTSUPP;
907 break;
908 }
909
910 btrfs_set_file_extent_num_bytes(leaf, fi,
911 start - key.offset);
912 btrfs_mark_buffer_dirty(leaf);
913 if (update_refs && disk_bytenr > 0)
914 inode_sub_bytes(inode, extent_end - start);
915 if (end == extent_end)
916 break;
917
918 path->slots[0]++;
919 goto next_slot;
920 }
921
922 /*
923 * | ---- range to drop ----- |
924 * | ------ extent ------ |
925 */
926 if (start <= key.offset && end >= extent_end) {
927delete_extent_item:
928 if (del_nr == 0) {
929 del_slot = path->slots[0];
930 del_nr = 1;
931 } else {
932 BUG_ON(del_slot + del_nr != path->slots[0]);
933 del_nr++;
934 }
935
936 if (update_refs &&
937 extent_type == BTRFS_FILE_EXTENT_INLINE) {
938 inode_sub_bytes(inode,
939 extent_end - key.offset);
940 extent_end = ALIGN(extent_end,
941 fs_info->sectorsize);
942 } else if (update_refs && disk_bytenr > 0) {
943 ret = btrfs_free_extent(trans, fs_info,
944 disk_bytenr, num_bytes, 0,
945 root->root_key.objectid,
946 key.objectid, key.offset -
947 extent_offset);
948 BUG_ON(ret); /* -ENOMEM */
949 inode_sub_bytes(inode,
950 extent_end - key.offset);
951 }
952
953 if (end == extent_end)
954 break;
955
956 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
957 path->slots[0]++;
958 goto next_slot;
959 }
960
961 ret = btrfs_del_items(trans, root, path, del_slot,
962 del_nr);
963 if (ret) {
964 btrfs_abort_transaction(trans, ret);
965 break;
966 }
967
968 del_nr = 0;
969 del_slot = 0;
970
971 btrfs_release_path(path);
972 continue;
973 }
974
975 BUG_ON(1);
976 }
977
978 if (!ret && del_nr > 0) {
979 /*
980 * Set path->slots[0] to first slot, so that after the delete
981 * if items are move off from our leaf to its immediate left or
982 * right neighbor leafs, we end up with a correct and adjusted
983 * path->slots[0] for our insertion (if replace_extent != 0).
984 */
985 path->slots[0] = del_slot;
986 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
987 if (ret)
988 btrfs_abort_transaction(trans, ret);
989 }
990
991 leaf = path->nodes[0];
992 /*
993 * If btrfs_del_items() was called, it might have deleted a leaf, in
994 * which case it unlocked our path, so check path->locks[0] matches a
995 * write lock.
996 */
997 if (!ret && replace_extent && leafs_visited == 1 &&
998 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
999 path->locks[0] == BTRFS_WRITE_LOCK) &&
1000 btrfs_leaf_free_space(fs_info, leaf) >=
1001 sizeof(struct btrfs_item) + extent_item_size) {
1002
1003 key.objectid = ino;
1004 key.type = BTRFS_EXTENT_DATA_KEY;
1005 key.offset = start;
1006 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1007 struct btrfs_key slot_key;
1008
1009 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1010 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1011 path->slots[0]++;
1012 }
1013 setup_items_for_insert(root, path, &key,
1014 &extent_item_size,
1015 extent_item_size,
1016 sizeof(struct btrfs_item) +
1017 extent_item_size, 1);
1018 *key_inserted = 1;
1019 }
1020
1021 if (!replace_extent || !(*key_inserted))
1022 btrfs_release_path(path);
1023 if (drop_end)
1024 *drop_end = found ? min(end, last_end) : end;
1025 return ret;
1026}
1027
1028int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1029 struct btrfs_root *root, struct inode *inode, u64 start,
1030 u64 end, int drop_cache)
1031{
1032 struct btrfs_path *path;
1033 int ret;
1034
1035 path = btrfs_alloc_path();
1036 if (!path)
1037 return -ENOMEM;
1038 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1039 drop_cache, 0, 0, NULL);
1040 btrfs_free_path(path);
1041 return ret;
1042}
1043
1044static int extent_mergeable(struct extent_buffer *leaf, int slot,
1045 u64 objectid, u64 bytenr, u64 orig_offset,
1046 u64 *start, u64 *end)
1047{
1048 struct btrfs_file_extent_item *fi;
1049 struct btrfs_key key;
1050 u64 extent_end;
1051
1052 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1053 return 0;
1054
1055 btrfs_item_key_to_cpu(leaf, &key, slot);
1056 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1057 return 0;
1058
1059 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1060 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1061 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1062 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1063 btrfs_file_extent_compression(leaf, fi) ||
1064 btrfs_file_extent_encryption(leaf, fi) ||
1065 btrfs_file_extent_other_encoding(leaf, fi))
1066 return 0;
1067
1068 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1069 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1070 return 0;
1071
1072 *start = key.offset;
1073 *end = extent_end;
1074 return 1;
1075}
1076
1077/*
1078 * Mark extent in the range start - end as written.
1079 *
1080 * This changes extent type from 'pre-allocated' to 'regular'. If only
1081 * part of extent is marked as written, the extent will be split into
1082 * two or three.
1083 */
1084int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1085 struct inode *inode, u64 start, u64 end)
1086{
1087 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1088 struct btrfs_root *root = BTRFS_I(inode)->root;
1089 struct extent_buffer *leaf;
1090 struct btrfs_path *path;
1091 struct btrfs_file_extent_item *fi;
1092 struct btrfs_key key;
1093 struct btrfs_key new_key;
1094 u64 bytenr;
1095 u64 num_bytes;
1096 u64 extent_end;
1097 u64 orig_offset;
1098 u64 other_start;
1099 u64 other_end;
1100 u64 split;
1101 int del_nr = 0;
1102 int del_slot = 0;
1103 int recow;
1104 int ret;
1105 u64 ino = btrfs_ino(inode);
1106
1107 path = btrfs_alloc_path();
1108 if (!path)
1109 return -ENOMEM;
1110again:
1111 recow = 0;
1112 split = start;
1113 key.objectid = ino;
1114 key.type = BTRFS_EXTENT_DATA_KEY;
1115 key.offset = split;
1116
1117 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1118 if (ret < 0)
1119 goto out;
1120 if (ret > 0 && path->slots[0] > 0)
1121 path->slots[0]--;
1122
1123 leaf = path->nodes[0];
1124 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1125 if (key.objectid != ino ||
1126 key.type != BTRFS_EXTENT_DATA_KEY) {
1127 ret = -EINVAL;
1128 btrfs_abort_transaction(trans, ret);
1129 goto out;
1130 }
1131 fi = btrfs_item_ptr(leaf, path->slots[0],
1132 struct btrfs_file_extent_item);
1133 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1134 ret = -EINVAL;
1135 btrfs_abort_transaction(trans, ret);
1136 goto out;
1137 }
1138 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1139 if (key.offset > start || extent_end < end) {
1140 ret = -EINVAL;
1141 btrfs_abort_transaction(trans, ret);
1142 goto out;
1143 }
1144
1145 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1146 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1147 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1148 memcpy(&new_key, &key, sizeof(new_key));
1149
1150 if (start == key.offset && end < extent_end) {
1151 other_start = 0;
1152 other_end = start;
1153 if (extent_mergeable(leaf, path->slots[0] - 1,
1154 ino, bytenr, orig_offset,
1155 &other_start, &other_end)) {
1156 new_key.offset = end;
1157 btrfs_set_item_key_safe(fs_info, path, &new_key);
1158 fi = btrfs_item_ptr(leaf, path->slots[0],
1159 struct btrfs_file_extent_item);
1160 btrfs_set_file_extent_generation(leaf, fi,
1161 trans->transid);
1162 btrfs_set_file_extent_num_bytes(leaf, fi,
1163 extent_end - end);
1164 btrfs_set_file_extent_offset(leaf, fi,
1165 end - orig_offset);
1166 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1167 struct btrfs_file_extent_item);
1168 btrfs_set_file_extent_generation(leaf, fi,
1169 trans->transid);
1170 btrfs_set_file_extent_num_bytes(leaf, fi,
1171 end - other_start);
1172 btrfs_mark_buffer_dirty(leaf);
1173 goto out;
1174 }
1175 }
1176
1177 if (start > key.offset && end == extent_end) {
1178 other_start = end;
1179 other_end = 0;
1180 if (extent_mergeable(leaf, path->slots[0] + 1,
1181 ino, bytenr, orig_offset,
1182 &other_start, &other_end)) {
1183 fi = btrfs_item_ptr(leaf, path->slots[0],
1184 struct btrfs_file_extent_item);
1185 btrfs_set_file_extent_num_bytes(leaf, fi,
1186 start - key.offset);
1187 btrfs_set_file_extent_generation(leaf, fi,
1188 trans->transid);
1189 path->slots[0]++;
1190 new_key.offset = start;
1191 btrfs_set_item_key_safe(fs_info, path, &new_key);
1192
1193 fi = btrfs_item_ptr(leaf, path->slots[0],
1194 struct btrfs_file_extent_item);
1195 btrfs_set_file_extent_generation(leaf, fi,
1196 trans->transid);
1197 btrfs_set_file_extent_num_bytes(leaf, fi,
1198 other_end - start);
1199 btrfs_set_file_extent_offset(leaf, fi,
1200 start - orig_offset);
1201 btrfs_mark_buffer_dirty(leaf);
1202 goto out;
1203 }
1204 }
1205
1206 while (start > key.offset || end < extent_end) {
1207 if (key.offset == start)
1208 split = end;
1209
1210 new_key.offset = split;
1211 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1212 if (ret == -EAGAIN) {
1213 btrfs_release_path(path);
1214 goto again;
1215 }
1216 if (ret < 0) {
1217 btrfs_abort_transaction(trans, ret);
1218 goto out;
1219 }
1220
1221 leaf = path->nodes[0];
1222 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1223 struct btrfs_file_extent_item);
1224 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1225 btrfs_set_file_extent_num_bytes(leaf, fi,
1226 split - key.offset);
1227
1228 fi = btrfs_item_ptr(leaf, path->slots[0],
1229 struct btrfs_file_extent_item);
1230
1231 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1232 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1233 btrfs_set_file_extent_num_bytes(leaf, fi,
1234 extent_end - split);
1235 btrfs_mark_buffer_dirty(leaf);
1236
1237 ret = btrfs_inc_extent_ref(trans, fs_info, bytenr, num_bytes,
1238 0, root->root_key.objectid,
1239 ino, orig_offset);
1240 if (ret) {
1241 btrfs_abort_transaction(trans, ret);
1242 goto out;
1243 }
1244
1245 if (split == start) {
1246 key.offset = start;
1247 } else {
1248 if (start != key.offset) {
1249 ret = -EINVAL;
1250 btrfs_abort_transaction(trans, ret);
1251 goto out;
1252 }
1253 path->slots[0]--;
1254 extent_end = end;
1255 }
1256 recow = 1;
1257 }
1258
1259 other_start = end;
1260 other_end = 0;
1261 if (extent_mergeable(leaf, path->slots[0] + 1,
1262 ino, bytenr, orig_offset,
1263 &other_start, &other_end)) {
1264 if (recow) {
1265 btrfs_release_path(path);
1266 goto again;
1267 }
1268 extent_end = other_end;
1269 del_slot = path->slots[0] + 1;
1270 del_nr++;
1271 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1272 0, root->root_key.objectid,
1273 ino, orig_offset);
1274 if (ret) {
1275 btrfs_abort_transaction(trans, ret);
1276 goto out;
1277 }
1278 }
1279 other_start = 0;
1280 other_end = start;
1281 if (extent_mergeable(leaf, path->slots[0] - 1,
1282 ino, bytenr, orig_offset,
1283 &other_start, &other_end)) {
1284 if (recow) {
1285 btrfs_release_path(path);
1286 goto again;
1287 }
1288 key.offset = other_start;
1289 del_slot = path->slots[0];
1290 del_nr++;
1291 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1292 0, root->root_key.objectid,
1293 ino, orig_offset);
1294 if (ret) {
1295 btrfs_abort_transaction(trans, ret);
1296 goto out;
1297 }
1298 }
1299 if (del_nr == 0) {
1300 fi = btrfs_item_ptr(leaf, path->slots[0],
1301 struct btrfs_file_extent_item);
1302 btrfs_set_file_extent_type(leaf, fi,
1303 BTRFS_FILE_EXTENT_REG);
1304 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1305 btrfs_mark_buffer_dirty(leaf);
1306 } else {
1307 fi = btrfs_item_ptr(leaf, del_slot - 1,
1308 struct btrfs_file_extent_item);
1309 btrfs_set_file_extent_type(leaf, fi,
1310 BTRFS_FILE_EXTENT_REG);
1311 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1312 btrfs_set_file_extent_num_bytes(leaf, fi,
1313 extent_end - key.offset);
1314 btrfs_mark_buffer_dirty(leaf);
1315
1316 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1317 if (ret < 0) {
1318 btrfs_abort_transaction(trans, ret);
1319 goto out;
1320 }
1321 }
1322out:
1323 btrfs_free_path(path);
1324 return 0;
1325}
1326
1327/*
1328 * on error we return an unlocked page and the error value
1329 * on success we return a locked page and 0
1330 */
1331static int prepare_uptodate_page(struct inode *inode,
1332 struct page *page, u64 pos,
1333 bool force_uptodate)
1334{
1335 int ret = 0;
1336
1337 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1338 !PageUptodate(page)) {
1339 ret = btrfs_readpage(NULL, page);
1340 if (ret)
1341 return ret;
1342 lock_page(page);
1343 if (!PageUptodate(page)) {
1344 unlock_page(page);
1345 return -EIO;
1346 }
1347 if (page->mapping != inode->i_mapping) {
1348 unlock_page(page);
1349 return -EAGAIN;
1350 }
1351 }
1352 return 0;
1353}
1354
1355/*
1356 * this just gets pages into the page cache and locks them down.
1357 */
1358static noinline int prepare_pages(struct inode *inode, struct page **pages,
1359 size_t num_pages, loff_t pos,
1360 size_t write_bytes, bool force_uptodate)
1361{
1362 int i;
1363 unsigned long index = pos >> PAGE_SHIFT;
1364 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1365 int err = 0;
1366 int faili;
1367
1368 for (i = 0; i < num_pages; i++) {
1369again:
1370 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1371 mask | __GFP_WRITE);
1372 if (!pages[i]) {
1373 faili = i - 1;
1374 err = -ENOMEM;
1375 goto fail;
1376 }
1377
1378 if (i == 0)
1379 err = prepare_uptodate_page(inode, pages[i], pos,
1380 force_uptodate);
1381 if (!err && i == num_pages - 1)
1382 err = prepare_uptodate_page(inode, pages[i],
1383 pos + write_bytes, false);
1384 if (err) {
1385 put_page(pages[i]);
1386 if (err == -EAGAIN) {
1387 err = 0;
1388 goto again;
1389 }
1390 faili = i - 1;
1391 goto fail;
1392 }
1393 wait_on_page_writeback(pages[i]);
1394 }
1395
1396 return 0;
1397fail:
1398 while (faili >= 0) {
1399 unlock_page(pages[faili]);
1400 put_page(pages[faili]);
1401 faili--;
1402 }
1403 return err;
1404
1405}
1406
1407/*
1408 * This function locks the extent and properly waits for data=ordered extents
1409 * to finish before allowing the pages to be modified if need.
1410 *
1411 * The return value:
1412 * 1 - the extent is locked
1413 * 0 - the extent is not locked, and everything is OK
1414 * -EAGAIN - need re-prepare the pages
1415 * the other < 0 number - Something wrong happens
1416 */
1417static noinline int
1418lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1419 size_t num_pages, loff_t pos,
1420 size_t write_bytes,
1421 u64 *lockstart, u64 *lockend,
1422 struct extent_state **cached_state)
1423{
1424 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1425 u64 start_pos;
1426 u64 last_pos;
1427 int i;
1428 int ret = 0;
1429
1430 start_pos = round_down(pos, fs_info->sectorsize);
1431 last_pos = start_pos
1432 + round_up(pos + write_bytes - start_pos,
1433 fs_info->sectorsize) - 1;
1434
1435 if (start_pos < inode->i_size) {
1436 struct btrfs_ordered_extent *ordered;
1437 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1438 start_pos, last_pos, cached_state);
1439 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1440 last_pos - start_pos + 1);
1441 if (ordered &&
1442 ordered->file_offset + ordered->len > start_pos &&
1443 ordered->file_offset <= last_pos) {
1444 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1445 start_pos, last_pos,
1446 cached_state, GFP_NOFS);
1447 for (i = 0; i < num_pages; i++) {
1448 unlock_page(pages[i]);
1449 put_page(pages[i]);
1450 }
1451 btrfs_start_ordered_extent(inode, ordered, 1);
1452 btrfs_put_ordered_extent(ordered);
1453 return -EAGAIN;
1454 }
1455 if (ordered)
1456 btrfs_put_ordered_extent(ordered);
1457
1458 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1459 last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1460 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1461 0, 0, cached_state, GFP_NOFS);
1462 *lockstart = start_pos;
1463 *lockend = last_pos;
1464 ret = 1;
1465 }
1466
1467 for (i = 0; i < num_pages; i++) {
1468 if (clear_page_dirty_for_io(pages[i]))
1469 account_page_redirty(pages[i]);
1470 set_page_extent_mapped(pages[i]);
1471 WARN_ON(!PageLocked(pages[i]));
1472 }
1473
1474 return ret;
1475}
1476
1477static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1478 size_t *write_bytes)
1479{
1480 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1481 struct btrfs_root *root = BTRFS_I(inode)->root;
1482 struct btrfs_ordered_extent *ordered;
1483 u64 lockstart, lockend;
1484 u64 num_bytes;
1485 int ret;
1486
1487 ret = btrfs_start_write_no_snapshoting(root);
1488 if (!ret)
1489 return -ENOSPC;
1490
1491 lockstart = round_down(pos, fs_info->sectorsize);
1492 lockend = round_up(pos + *write_bytes,
1493 fs_info->sectorsize) - 1;
1494
1495 while (1) {
1496 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1497 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1498 lockend - lockstart + 1);
1499 if (!ordered) {
1500 break;
1501 }
1502 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1503 btrfs_start_ordered_extent(inode, ordered, 1);
1504 btrfs_put_ordered_extent(ordered);
1505 }
1506
1507 num_bytes = lockend - lockstart + 1;
1508 ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1509 if (ret <= 0) {
1510 ret = 0;
1511 btrfs_end_write_no_snapshoting(root);
1512 } else {
1513 *write_bytes = min_t(size_t, *write_bytes ,
1514 num_bytes - pos + lockstart);
1515 }
1516
1517 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1518
1519 return ret;
1520}
1521
1522static noinline ssize_t __btrfs_buffered_write(struct file *file,
1523 struct iov_iter *i,
1524 loff_t pos)
1525{
1526 struct inode *inode = file_inode(file);
1527 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1528 struct btrfs_root *root = BTRFS_I(inode)->root;
1529 struct page **pages = NULL;
1530 struct extent_state *cached_state = NULL;
1531 u64 release_bytes = 0;
1532 u64 lockstart;
1533 u64 lockend;
1534 size_t num_written = 0;
1535 int nrptrs;
1536 int ret = 0;
1537 bool only_release_metadata = false;
1538 bool force_page_uptodate = false;
1539 bool need_unlock;
1540
1541 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1542 PAGE_SIZE / (sizeof(struct page *)));
1543 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1544 nrptrs = max(nrptrs, 8);
1545 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1546 if (!pages)
1547 return -ENOMEM;
1548
1549 while (iov_iter_count(i) > 0) {
1550 size_t offset = pos & (PAGE_SIZE - 1);
1551 size_t sector_offset;
1552 size_t write_bytes = min(iov_iter_count(i),
1553 nrptrs * (size_t)PAGE_SIZE -
1554 offset);
1555 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1556 PAGE_SIZE);
1557 size_t reserve_bytes;
1558 size_t dirty_pages;
1559 size_t copied;
1560 size_t dirty_sectors;
1561 size_t num_sectors;
1562
1563 WARN_ON(num_pages > nrptrs);
1564
1565 /*
1566 * Fault pages before locking them in prepare_pages
1567 * to avoid recursive lock
1568 */
1569 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1570 ret = -EFAULT;
1571 break;
1572 }
1573
1574 sector_offset = pos & (fs_info->sectorsize - 1);
1575 reserve_bytes = round_up(write_bytes + sector_offset,
1576 fs_info->sectorsize);
1577
1578 ret = btrfs_check_data_free_space(inode, pos, write_bytes);
1579 if (ret < 0) {
1580 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1581 BTRFS_INODE_PREALLOC)) &&
1582 check_can_nocow(inode, pos, &write_bytes) > 0) {
1583 /*
1584 * For nodata cow case, no need to reserve
1585 * data space.
1586 */
1587 only_release_metadata = true;
1588 /*
1589 * our prealloc extent may be smaller than
1590 * write_bytes, so scale down.
1591 */
1592 num_pages = DIV_ROUND_UP(write_bytes + offset,
1593 PAGE_SIZE);
1594 reserve_bytes = round_up(write_bytes +
1595 sector_offset,
1596 fs_info->sectorsize);
1597 } else {
1598 break;
1599 }
1600 }
1601
1602 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1603 if (ret) {
1604 if (!only_release_metadata)
1605 btrfs_free_reserved_data_space(inode, pos,
1606 write_bytes);
1607 else
1608 btrfs_end_write_no_snapshoting(root);
1609 break;
1610 }
1611
1612 release_bytes = reserve_bytes;
1613 need_unlock = false;
1614again:
1615 /*
1616 * This is going to setup the pages array with the number of
1617 * pages we want, so we don't really need to worry about the
1618 * contents of pages from loop to loop
1619 */
1620 ret = prepare_pages(inode, pages, num_pages,
1621 pos, write_bytes,
1622 force_page_uptodate);
1623 if (ret)
1624 break;
1625
1626 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1627 pos, write_bytes, &lockstart,
1628 &lockend, &cached_state);
1629 if (ret < 0) {
1630 if (ret == -EAGAIN)
1631 goto again;
1632 break;
1633 } else if (ret > 0) {
1634 need_unlock = true;
1635 ret = 0;
1636 }
1637
1638 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1639
1640 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1641 dirty_sectors = round_up(copied + sector_offset,
1642 fs_info->sectorsize);
1643 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1644
1645 /*
1646 * if we have trouble faulting in the pages, fall
1647 * back to one page at a time
1648 */
1649 if (copied < write_bytes)
1650 nrptrs = 1;
1651
1652 if (copied == 0) {
1653 force_page_uptodate = true;
1654 dirty_sectors = 0;
1655 dirty_pages = 0;
1656 } else {
1657 force_page_uptodate = false;
1658 dirty_pages = DIV_ROUND_UP(copied + offset,
1659 PAGE_SIZE);
1660 }
1661
1662 /*
1663 * If we had a short copy we need to release the excess delaloc
1664 * bytes we reserved. We need to increment outstanding_extents
1665 * because btrfs_delalloc_release_space and
1666 * btrfs_delalloc_release_metadata will decrement it, but
1667 * we still have an outstanding extent for the chunk we actually
1668 * managed to copy.
1669 */
1670 if (num_sectors > dirty_sectors) {
1671 /* release everything except the sectors we dirtied */
1672 release_bytes -= dirty_sectors <<
1673 fs_info->sb->s_blocksize_bits;
1674 if (copied > 0) {
1675 spin_lock(&BTRFS_I(inode)->lock);
1676 BTRFS_I(inode)->outstanding_extents++;
1677 spin_unlock(&BTRFS_I(inode)->lock);
1678 }
1679 if (only_release_metadata) {
1680 btrfs_delalloc_release_metadata(inode,
1681 release_bytes);
1682 } else {
1683 u64 __pos;
1684
1685 __pos = round_down(pos,
1686 fs_info->sectorsize) +
1687 (dirty_pages << PAGE_SHIFT);
1688 btrfs_delalloc_release_space(inode, __pos,
1689 release_bytes);
1690 }
1691 }
1692
1693 release_bytes = round_up(copied + sector_offset,
1694 fs_info->sectorsize);
1695
1696 if (copied > 0)
1697 ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1698 pos, copied, NULL);
1699 if (need_unlock)
1700 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1701 lockstart, lockend, &cached_state,
1702 GFP_NOFS);
1703 if (ret) {
1704 btrfs_drop_pages(pages, num_pages);
1705 break;
1706 }
1707
1708 release_bytes = 0;
1709 if (only_release_metadata)
1710 btrfs_end_write_no_snapshoting(root);
1711
1712 if (only_release_metadata && copied > 0) {
1713 lockstart = round_down(pos,
1714 fs_info->sectorsize);
1715 lockend = round_up(pos + copied,
1716 fs_info->sectorsize) - 1;
1717
1718 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1719 lockend, EXTENT_NORESERVE, NULL,
1720 NULL, GFP_NOFS);
1721 only_release_metadata = false;
1722 }
1723
1724 btrfs_drop_pages(pages, num_pages);
1725
1726 cond_resched();
1727
1728 balance_dirty_pages_ratelimited(inode->i_mapping);
1729 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1730 btrfs_btree_balance_dirty(fs_info);
1731
1732 pos += copied;
1733 num_written += copied;
1734 }
1735
1736 kfree(pages);
1737
1738 if (release_bytes) {
1739 if (only_release_metadata) {
1740 btrfs_end_write_no_snapshoting(root);
1741 btrfs_delalloc_release_metadata(inode, release_bytes);
1742 } else {
1743 btrfs_delalloc_release_space(inode,
1744 round_down(pos, fs_info->sectorsize),
1745 release_bytes);
1746 }
1747 }
1748
1749 return num_written ? num_written : ret;
1750}
1751
1752static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1753{
1754 struct file *file = iocb->ki_filp;
1755 struct inode *inode = file_inode(file);
1756 loff_t pos = iocb->ki_pos;
1757 ssize_t written;
1758 ssize_t written_buffered;
1759 loff_t endbyte;
1760 int err;
1761
1762 written = generic_file_direct_write(iocb, from);
1763
1764 if (written < 0 || !iov_iter_count(from))
1765 return written;
1766
1767 pos += written;
1768 written_buffered = __btrfs_buffered_write(file, from, pos);
1769 if (written_buffered < 0) {
1770 err = written_buffered;
1771 goto out;
1772 }
1773 /*
1774 * Ensure all data is persisted. We want the next direct IO read to be
1775 * able to read what was just written.
1776 */
1777 endbyte = pos + written_buffered - 1;
1778 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1779 if (err)
1780 goto out;
1781 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1782 if (err)
1783 goto out;
1784 written += written_buffered;
1785 iocb->ki_pos = pos + written_buffered;
1786 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1787 endbyte >> PAGE_SHIFT);
1788out:
1789 return written ? written : err;
1790}
1791
1792static void update_time_for_write(struct inode *inode)
1793{
1794 struct timespec now;
1795
1796 if (IS_NOCMTIME(inode))
1797 return;
1798
1799 now = current_time(inode);
1800 if (!timespec_equal(&inode->i_mtime, &now))
1801 inode->i_mtime = now;
1802
1803 if (!timespec_equal(&inode->i_ctime, &now))
1804 inode->i_ctime = now;
1805
1806 if (IS_I_VERSION(inode))
1807 inode_inc_iversion(inode);
1808}
1809
1810static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1811 struct iov_iter *from)
1812{
1813 struct file *file = iocb->ki_filp;
1814 struct inode *inode = file_inode(file);
1815 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1816 struct btrfs_root *root = BTRFS_I(inode)->root;
1817 u64 start_pos;
1818 u64 end_pos;
1819 ssize_t num_written = 0;
1820 bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1821 ssize_t err;
1822 loff_t pos;
1823 size_t count;
1824 loff_t oldsize;
1825 int clean_page = 0;
1826
1827 inode_lock(inode);
1828 err = generic_write_checks(iocb, from);
1829 if (err <= 0) {
1830 inode_unlock(inode);
1831 return err;
1832 }
1833
1834 current->backing_dev_info = inode_to_bdi(inode);
1835 err = file_remove_privs(file);
1836 if (err) {
1837 inode_unlock(inode);
1838 goto out;
1839 }
1840
1841 /*
1842 * If BTRFS flips readonly due to some impossible error
1843 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1844 * although we have opened a file as writable, we have
1845 * to stop this write operation to ensure FS consistency.
1846 */
1847 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1848 inode_unlock(inode);
1849 err = -EROFS;
1850 goto out;
1851 }
1852
1853 /*
1854 * We reserve space for updating the inode when we reserve space for the
1855 * extent we are going to write, so we will enospc out there. We don't
1856 * need to start yet another transaction to update the inode as we will
1857 * update the inode when we finish writing whatever data we write.
1858 */
1859 update_time_for_write(inode);
1860
1861 pos = iocb->ki_pos;
1862 count = iov_iter_count(from);
1863 start_pos = round_down(pos, fs_info->sectorsize);
1864 oldsize = i_size_read(inode);
1865 if (start_pos > oldsize) {
1866 /* Expand hole size to cover write data, preventing empty gap */
1867 end_pos = round_up(pos + count,
1868 fs_info->sectorsize);
1869 err = btrfs_cont_expand(inode, oldsize, end_pos);
1870 if (err) {
1871 inode_unlock(inode);
1872 goto out;
1873 }
1874 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1875 clean_page = 1;
1876 }
1877
1878 if (sync)
1879 atomic_inc(&BTRFS_I(inode)->sync_writers);
1880
1881 if (iocb->ki_flags & IOCB_DIRECT) {
1882 num_written = __btrfs_direct_write(iocb, from);
1883 } else {
1884 num_written = __btrfs_buffered_write(file, from, pos);
1885 if (num_written > 0)
1886 iocb->ki_pos = pos + num_written;
1887 if (clean_page)
1888 pagecache_isize_extended(inode, oldsize,
1889 i_size_read(inode));
1890 }
1891
1892 inode_unlock(inode);
1893
1894 /*
1895 * We also have to set last_sub_trans to the current log transid,
1896 * otherwise subsequent syncs to a file that's been synced in this
1897 * transaction will appear to have already occurred.
1898 */
1899 spin_lock(&BTRFS_I(inode)->lock);
1900 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1901 spin_unlock(&BTRFS_I(inode)->lock);
1902 if (num_written > 0)
1903 num_written = generic_write_sync(iocb, num_written);
1904
1905 if (sync)
1906 atomic_dec(&BTRFS_I(inode)->sync_writers);
1907out:
1908 current->backing_dev_info = NULL;
1909 return num_written ? num_written : err;
1910}
1911
1912int btrfs_release_file(struct inode *inode, struct file *filp)
1913{
1914 if (filp->private_data)
1915 btrfs_ioctl_trans_end(filp);
1916 /*
1917 * ordered_data_close is set by settattr when we are about to truncate
1918 * a file from a non-zero size to a zero size. This tries to
1919 * flush down new bytes that may have been written if the
1920 * application were using truncate to replace a file in place.
1921 */
1922 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1923 &BTRFS_I(inode)->runtime_flags))
1924 filemap_flush(inode->i_mapping);
1925 return 0;
1926}
1927
1928static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1929{
1930 int ret;
1931
1932 atomic_inc(&BTRFS_I(inode)->sync_writers);
1933 ret = btrfs_fdatawrite_range(inode, start, end);
1934 atomic_dec(&BTRFS_I(inode)->sync_writers);
1935
1936 return ret;
1937}
1938
1939/*
1940 * fsync call for both files and directories. This logs the inode into
1941 * the tree log instead of forcing full commits whenever possible.
1942 *
1943 * It needs to call filemap_fdatawait so that all ordered extent updates are
1944 * in the metadata btree are up to date for copying to the log.
1945 *
1946 * It drops the inode mutex before doing the tree log commit. This is an
1947 * important optimization for directories because holding the mutex prevents
1948 * new operations on the dir while we write to disk.
1949 */
1950int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1951{
1952 struct dentry *dentry = file_dentry(file);
1953 struct inode *inode = d_inode(dentry);
1954 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1955 struct btrfs_root *root = BTRFS_I(inode)->root;
1956 struct btrfs_trans_handle *trans;
1957 struct btrfs_log_ctx ctx;
1958 int ret = 0;
1959 bool full_sync = 0;
1960 u64 len;
1961
1962 /*
1963 * The range length can be represented by u64, we have to do the typecasts
1964 * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
1965 */
1966 len = (u64)end - (u64)start + 1;
1967 trace_btrfs_sync_file(file, datasync);
1968
1969 /*
1970 * We write the dirty pages in the range and wait until they complete
1971 * out of the ->i_mutex. If so, we can flush the dirty pages by
1972 * multi-task, and make the performance up. See
1973 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1974 */
1975 ret = start_ordered_ops(inode, start, end);
1976 if (ret)
1977 return ret;
1978
1979 inode_lock(inode);
1980 atomic_inc(&root->log_batch);
1981 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1982 &BTRFS_I(inode)->runtime_flags);
1983 /*
1984 * We might have have had more pages made dirty after calling
1985 * start_ordered_ops and before acquiring the inode's i_mutex.
1986 */
1987 if (full_sync) {
1988 /*
1989 * For a full sync, we need to make sure any ordered operations
1990 * start and finish before we start logging the inode, so that
1991 * all extents are persisted and the respective file extent
1992 * items are in the fs/subvol btree.
1993 */
1994 ret = btrfs_wait_ordered_range(inode, start, len);
1995 } else {
1996 /*
1997 * Start any new ordered operations before starting to log the
1998 * inode. We will wait for them to finish in btrfs_sync_log().
1999 *
2000 * Right before acquiring the inode's mutex, we might have new
2001 * writes dirtying pages, which won't immediately start the
2002 * respective ordered operations - that is done through the
2003 * fill_delalloc callbacks invoked from the writepage and
2004 * writepages address space operations. So make sure we start
2005 * all ordered operations before starting to log our inode. Not
2006 * doing this means that while logging the inode, writeback
2007 * could start and invoke writepage/writepages, which would call
2008 * the fill_delalloc callbacks (cow_file_range,
2009 * submit_compressed_extents). These callbacks add first an
2010 * extent map to the modified list of extents and then create
2011 * the respective ordered operation, which means in
2012 * tree-log.c:btrfs_log_inode() we might capture all existing
2013 * ordered operations (with btrfs_get_logged_extents()) before
2014 * the fill_delalloc callback adds its ordered operation, and by
2015 * the time we visit the modified list of extent maps (with
2016 * btrfs_log_changed_extents()), we see and process the extent
2017 * map they created. We then use the extent map to construct a
2018 * file extent item for logging without waiting for the
2019 * respective ordered operation to finish - this file extent
2020 * item points to a disk location that might not have yet been
2021 * written to, containing random data - so after a crash a log
2022 * replay will make our inode have file extent items that point
2023 * to disk locations containing invalid data, as we returned
2024 * success to userspace without waiting for the respective
2025 * ordered operation to finish, because it wasn't captured by
2026 * btrfs_get_logged_extents().
2027 */
2028 ret = start_ordered_ops(inode, start, end);
2029 }
2030 if (ret) {
2031 inode_unlock(inode);
2032 goto out;
2033 }
2034 atomic_inc(&root->log_batch);
2035
2036 /*
2037 * If the last transaction that changed this file was before the current
2038 * transaction and we have the full sync flag set in our inode, we can
2039 * bail out now without any syncing.
2040 *
2041 * Note that we can't bail out if the full sync flag isn't set. This is
2042 * because when the full sync flag is set we start all ordered extents
2043 * and wait for them to fully complete - when they complete they update
2044 * the inode's last_trans field through:
2045 *
2046 * btrfs_finish_ordered_io() ->
2047 * btrfs_update_inode_fallback() ->
2048 * btrfs_update_inode() ->
2049 * btrfs_set_inode_last_trans()
2050 *
2051 * So we are sure that last_trans is up to date and can do this check to
2052 * bail out safely. For the fast path, when the full sync flag is not
2053 * set in our inode, we can not do it because we start only our ordered
2054 * extents and don't wait for them to complete (that is when
2055 * btrfs_finish_ordered_io runs), so here at this point their last_trans
2056 * value might be less than or equals to fs_info->last_trans_committed,
2057 * and setting a speculative last_trans for an inode when a buffered
2058 * write is made (such as fs_info->generation + 1 for example) would not
2059 * be reliable since after setting the value and before fsync is called
2060 * any number of transactions can start and commit (transaction kthread
2061 * commits the current transaction periodically), and a transaction
2062 * commit does not start nor waits for ordered extents to complete.
2063 */
2064 smp_mb();
2065 if (btrfs_inode_in_log(inode, fs_info->generation) ||
2066 (full_sync && BTRFS_I(inode)->last_trans <=
2067 fs_info->last_trans_committed) ||
2068 (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2069 BTRFS_I(inode)->last_trans
2070 <= fs_info->last_trans_committed)) {
2071 /*
2072 * We've had everything committed since the last time we were
2073 * modified so clear this flag in case it was set for whatever
2074 * reason, it's no longer relevant.
2075 */
2076 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2077 &BTRFS_I(inode)->runtime_flags);
2078 /*
2079 * An ordered extent might have started before and completed
2080 * already with io errors, in which case the inode was not
2081 * updated and we end up here. So check the inode's mapping
2082 * flags for any errors that might have happened while doing
2083 * writeback of file data.
2084 */
2085 ret = filemap_check_errors(inode->i_mapping);
2086 inode_unlock(inode);
2087 goto out;
2088 }
2089
2090 /*
2091 * ok we haven't committed the transaction yet, lets do a commit
2092 */
2093 if (file->private_data)
2094 btrfs_ioctl_trans_end(file);
2095
2096 /*
2097 * We use start here because we will need to wait on the IO to complete
2098 * in btrfs_sync_log, which could require joining a transaction (for
2099 * example checking cross references in the nocow path). If we use join
2100 * here we could get into a situation where we're waiting on IO to
2101 * happen that is blocked on a transaction trying to commit. With start
2102 * we inc the extwriter counter, so we wait for all extwriters to exit
2103 * before we start blocking join'ers. This comment is to keep somebody
2104 * from thinking they are super smart and changing this to
2105 * btrfs_join_transaction *cough*Josef*cough*.
2106 */
2107 trans = btrfs_start_transaction(root, 0);
2108 if (IS_ERR(trans)) {
2109 ret = PTR_ERR(trans);
2110 inode_unlock(inode);
2111 goto out;
2112 }
2113 trans->sync = true;
2114
2115 btrfs_init_log_ctx(&ctx, inode);
2116
2117 ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2118 if (ret < 0) {
2119 /* Fallthrough and commit/free transaction. */
2120 ret = 1;
2121 }
2122
2123 /* we've logged all the items and now have a consistent
2124 * version of the file in the log. It is possible that
2125 * someone will come in and modify the file, but that's
2126 * fine because the log is consistent on disk, and we
2127 * have references to all of the file's extents
2128 *
2129 * It is possible that someone will come in and log the
2130 * file again, but that will end up using the synchronization
2131 * inside btrfs_sync_log to keep things safe.
2132 */
2133 inode_unlock(inode);
2134
2135 /*
2136 * If any of the ordered extents had an error, just return it to user
2137 * space, so that the application knows some writes didn't succeed and
2138 * can take proper action (retry for e.g.). Blindly committing the
2139 * transaction in this case, would fool userspace that everything was
2140 * successful. And we also want to make sure our log doesn't contain
2141 * file extent items pointing to extents that weren't fully written to -
2142 * just like in the non fast fsync path, where we check for the ordered
2143 * operation's error flag before writing to the log tree and return -EIO
2144 * if any of them had this flag set (btrfs_wait_ordered_range) -
2145 * therefore we need to check for errors in the ordered operations,
2146 * which are indicated by ctx.io_err.
2147 */
2148 if (ctx.io_err) {
2149 btrfs_end_transaction(trans);
2150 ret = ctx.io_err;
2151 goto out;
2152 }
2153
2154 if (ret != BTRFS_NO_LOG_SYNC) {
2155 if (!ret) {
2156 ret = btrfs_sync_log(trans, root, &ctx);
2157 if (!ret) {
2158 ret = btrfs_end_transaction(trans);
2159 goto out;
2160 }
2161 }
2162 if (!full_sync) {
2163 ret = btrfs_wait_ordered_range(inode, start, len);
2164 if (ret) {
2165 btrfs_end_transaction(trans);
2166 goto out;
2167 }
2168 }
2169 ret = btrfs_commit_transaction(trans);
2170 } else {
2171 ret = btrfs_end_transaction(trans);
2172 }
2173out:
2174 return ret > 0 ? -EIO : ret;
2175}
2176
2177static const struct vm_operations_struct btrfs_file_vm_ops = {
2178 .fault = filemap_fault,
2179 .map_pages = filemap_map_pages,
2180 .page_mkwrite = btrfs_page_mkwrite,
2181};
2182
2183static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2184{
2185 struct address_space *mapping = filp->f_mapping;
2186
2187 if (!mapping->a_ops->readpage)
2188 return -ENOEXEC;
2189
2190 file_accessed(filp);
2191 vma->vm_ops = &btrfs_file_vm_ops;
2192
2193 return 0;
2194}
2195
2196static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2197 int slot, u64 start, u64 end)
2198{
2199 struct btrfs_file_extent_item *fi;
2200 struct btrfs_key key;
2201
2202 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2203 return 0;
2204
2205 btrfs_item_key_to_cpu(leaf, &key, slot);
2206 if (key.objectid != btrfs_ino(inode) ||
2207 key.type != BTRFS_EXTENT_DATA_KEY)
2208 return 0;
2209
2210 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2211
2212 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2213 return 0;
2214
2215 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2216 return 0;
2217
2218 if (key.offset == end)
2219 return 1;
2220 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2221 return 1;
2222 return 0;
2223}
2224
2225static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2226 struct btrfs_path *path, u64 offset, u64 end)
2227{
2228 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2229 struct btrfs_root *root = BTRFS_I(inode)->root;
2230 struct extent_buffer *leaf;
2231 struct btrfs_file_extent_item *fi;
2232 struct extent_map *hole_em;
2233 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2234 struct btrfs_key key;
2235 int ret;
2236
2237 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2238 goto out;
2239
2240 key.objectid = btrfs_ino(inode);
2241 key.type = BTRFS_EXTENT_DATA_KEY;
2242 key.offset = offset;
2243
2244 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2245 if (ret <= 0) {
2246 /*
2247 * We should have dropped this offset, so if we find it then
2248 * something has gone horribly wrong.
2249 */
2250 if (ret == 0)
2251 ret = -EINVAL;
2252 return ret;
2253 }
2254
2255 leaf = path->nodes[0];
2256 if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2257 u64 num_bytes;
2258
2259 path->slots[0]--;
2260 fi = btrfs_item_ptr(leaf, path->slots[0],
2261 struct btrfs_file_extent_item);
2262 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2263 end - offset;
2264 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2265 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2266 btrfs_set_file_extent_offset(leaf, fi, 0);
2267 btrfs_mark_buffer_dirty(leaf);
2268 goto out;
2269 }
2270
2271 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2272 u64 num_bytes;
2273
2274 key.offset = offset;
2275 btrfs_set_item_key_safe(fs_info, path, &key);
2276 fi = btrfs_item_ptr(leaf, path->slots[0],
2277 struct btrfs_file_extent_item);
2278 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2279 offset;
2280 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2281 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2282 btrfs_set_file_extent_offset(leaf, fi, 0);
2283 btrfs_mark_buffer_dirty(leaf);
2284 goto out;
2285 }
2286 btrfs_release_path(path);
2287
2288 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2289 0, 0, end - offset, 0, end - offset,
2290 0, 0, 0);
2291 if (ret)
2292 return ret;
2293
2294out:
2295 btrfs_release_path(path);
2296
2297 hole_em = alloc_extent_map();
2298 if (!hole_em) {
2299 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2300 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2301 &BTRFS_I(inode)->runtime_flags);
2302 } else {
2303 hole_em->start = offset;
2304 hole_em->len = end - offset;
2305 hole_em->ram_bytes = hole_em->len;
2306 hole_em->orig_start = offset;
2307
2308 hole_em->block_start = EXTENT_MAP_HOLE;
2309 hole_em->block_len = 0;
2310 hole_em->orig_block_len = 0;
2311 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2312 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2313 hole_em->generation = trans->transid;
2314
2315 do {
2316 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2317 write_lock(&em_tree->lock);
2318 ret = add_extent_mapping(em_tree, hole_em, 1);
2319 write_unlock(&em_tree->lock);
2320 } while (ret == -EEXIST);
2321 free_extent_map(hole_em);
2322 if (ret)
2323 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2324 &BTRFS_I(inode)->runtime_flags);
2325 }
2326
2327 return 0;
2328}
2329
2330/*
2331 * Find a hole extent on given inode and change start/len to the end of hole
2332 * extent.(hole/vacuum extent whose em->start <= start &&
2333 * em->start + em->len > start)
2334 * When a hole extent is found, return 1 and modify start/len.
2335 */
2336static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2337{
2338 struct extent_map *em;
2339 int ret = 0;
2340
2341 em = btrfs_get_extent(inode, NULL, 0, *start, *len, 0);
2342 if (IS_ERR_OR_NULL(em)) {
2343 if (!em)
2344 ret = -ENOMEM;
2345 else
2346 ret = PTR_ERR(em);
2347 return ret;
2348 }
2349
2350 /* Hole or vacuum extent(only exists in no-hole mode) */
2351 if (em->block_start == EXTENT_MAP_HOLE) {
2352 ret = 1;
2353 *len = em->start + em->len > *start + *len ?
2354 0 : *start + *len - em->start - em->len;
2355 *start = em->start + em->len;
2356 }
2357 free_extent_map(em);
2358 return ret;
2359}
2360
2361static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2362{
2363 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2364 struct btrfs_root *root = BTRFS_I(inode)->root;
2365 struct extent_state *cached_state = NULL;
2366 struct btrfs_path *path;
2367 struct btrfs_block_rsv *rsv;
2368 struct btrfs_trans_handle *trans;
2369 u64 lockstart;
2370 u64 lockend;
2371 u64 tail_start;
2372 u64 tail_len;
2373 u64 orig_start = offset;
2374 u64 cur_offset;
2375 u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2376 u64 drop_end;
2377 int ret = 0;
2378 int err = 0;
2379 unsigned int rsv_count;
2380 bool same_block;
2381 bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2382 u64 ino_size;
2383 bool truncated_block = false;
2384 bool updated_inode = false;
2385
2386 ret = btrfs_wait_ordered_range(inode, offset, len);
2387 if (ret)
2388 return ret;
2389
2390 inode_lock(inode);
2391 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2392 ret = find_first_non_hole(inode, &offset, &len);
2393 if (ret < 0)
2394 goto out_only_mutex;
2395 if (ret && !len) {
2396 /* Already in a large hole */
2397 ret = 0;
2398 goto out_only_mutex;
2399 }
2400
2401 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2402 lockend = round_down(offset + len,
2403 btrfs_inode_sectorsize(inode)) - 1;
2404 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2405 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2406 /*
2407 * We needn't truncate any block which is beyond the end of the file
2408 * because we are sure there is no data there.
2409 */
2410 /*
2411 * Only do this if we are in the same block and we aren't doing the
2412 * entire block.
2413 */
2414 if (same_block && len < fs_info->sectorsize) {
2415 if (offset < ino_size) {
2416 truncated_block = true;
2417 ret = btrfs_truncate_block(inode, offset, len, 0);
2418 } else {
2419 ret = 0;
2420 }
2421 goto out_only_mutex;
2422 }
2423
2424 /* zero back part of the first block */
2425 if (offset < ino_size) {
2426 truncated_block = true;
2427 ret = btrfs_truncate_block(inode, offset, 0, 0);
2428 if (ret) {
2429 inode_unlock(inode);
2430 return ret;
2431 }
2432 }
2433
2434 /* Check the aligned pages after the first unaligned page,
2435 * if offset != orig_start, which means the first unaligned page
2436 * including several following pages are already in holes,
2437 * the extra check can be skipped */
2438 if (offset == orig_start) {
2439 /* after truncate page, check hole again */
2440 len = offset + len - lockstart;
2441 offset = lockstart;
2442 ret = find_first_non_hole(inode, &offset, &len);
2443 if (ret < 0)
2444 goto out_only_mutex;
2445 if (ret && !len) {
2446 ret = 0;
2447 goto out_only_mutex;
2448 }
2449 lockstart = offset;
2450 }
2451
2452 /* Check the tail unaligned part is in a hole */
2453 tail_start = lockend + 1;
2454 tail_len = offset + len - tail_start;
2455 if (tail_len) {
2456 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2457 if (unlikely(ret < 0))
2458 goto out_only_mutex;
2459 if (!ret) {
2460 /* zero the front end of the last page */
2461 if (tail_start + tail_len < ino_size) {
2462 truncated_block = true;
2463 ret = btrfs_truncate_block(inode,
2464 tail_start + tail_len,
2465 0, 1);
2466 if (ret)
2467 goto out_only_mutex;
2468 }
2469 }
2470 }
2471
2472 if (lockend < lockstart) {
2473 ret = 0;
2474 goto out_only_mutex;
2475 }
2476
2477 while (1) {
2478 struct btrfs_ordered_extent *ordered;
2479
2480 truncate_pagecache_range(inode, lockstart, lockend);
2481
2482 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2483 &cached_state);
2484 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2485
2486 /*
2487 * We need to make sure we have no ordered extents in this range
2488 * and nobody raced in and read a page in this range, if we did
2489 * we need to try again.
2490 */
2491 if ((!ordered ||
2492 (ordered->file_offset + ordered->len <= lockstart ||
2493 ordered->file_offset > lockend)) &&
2494 !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2495 if (ordered)
2496 btrfs_put_ordered_extent(ordered);
2497 break;
2498 }
2499 if (ordered)
2500 btrfs_put_ordered_extent(ordered);
2501 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2502 lockend, &cached_state, GFP_NOFS);
2503 ret = btrfs_wait_ordered_range(inode, lockstart,
2504 lockend - lockstart + 1);
2505 if (ret) {
2506 inode_unlock(inode);
2507 return ret;
2508 }
2509 }
2510
2511 path = btrfs_alloc_path();
2512 if (!path) {
2513 ret = -ENOMEM;
2514 goto out;
2515 }
2516
2517 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2518 if (!rsv) {
2519 ret = -ENOMEM;
2520 goto out_free;
2521 }
2522 rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2523 rsv->failfast = 1;
2524
2525 /*
2526 * 1 - update the inode
2527 * 1 - removing the extents in the range
2528 * 1 - adding the hole extent if no_holes isn't set
2529 */
2530 rsv_count = no_holes ? 2 : 3;
2531 trans = btrfs_start_transaction(root, rsv_count);
2532 if (IS_ERR(trans)) {
2533 err = PTR_ERR(trans);
2534 goto out_free;
2535 }
2536
2537 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2538 min_size, 0);
2539 BUG_ON(ret);
2540 trans->block_rsv = rsv;
2541
2542 cur_offset = lockstart;
2543 len = lockend - cur_offset;
2544 while (cur_offset < lockend) {
2545 ret = __btrfs_drop_extents(trans, root, inode, path,
2546 cur_offset, lockend + 1,
2547 &drop_end, 1, 0, 0, NULL);
2548 if (ret != -ENOSPC)
2549 break;
2550
2551 trans->block_rsv = &fs_info->trans_block_rsv;
2552
2553 if (cur_offset < drop_end && cur_offset < ino_size) {
2554 ret = fill_holes(trans, inode, path, cur_offset,
2555 drop_end);
2556 if (ret) {
2557 /*
2558 * If we failed then we didn't insert our hole
2559 * entries for the area we dropped, so now the
2560 * fs is corrupted, so we must abort the
2561 * transaction.
2562 */
2563 btrfs_abort_transaction(trans, ret);
2564 err = ret;
2565 break;
2566 }
2567 }
2568
2569 cur_offset = drop_end;
2570
2571 ret = btrfs_update_inode(trans, root, inode);
2572 if (ret) {
2573 err = ret;
2574 break;
2575 }
2576
2577 btrfs_end_transaction(trans);
2578 btrfs_btree_balance_dirty(fs_info);
2579
2580 trans = btrfs_start_transaction(root, rsv_count);
2581 if (IS_ERR(trans)) {
2582 ret = PTR_ERR(trans);
2583 trans = NULL;
2584 break;
2585 }
2586
2587 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2588 rsv, min_size, 0);
2589 BUG_ON(ret); /* shouldn't happen */
2590 trans->block_rsv = rsv;
2591
2592 ret = find_first_non_hole(inode, &cur_offset, &len);
2593 if (unlikely(ret < 0))
2594 break;
2595 if (ret && !len) {
2596 ret = 0;
2597 break;
2598 }
2599 }
2600
2601 if (ret) {
2602 err = ret;
2603 goto out_trans;
2604 }
2605
2606 trans->block_rsv = &fs_info->trans_block_rsv;
2607 /*
2608 * If we are using the NO_HOLES feature we might have had already an
2609 * hole that overlaps a part of the region [lockstart, lockend] and
2610 * ends at (or beyond) lockend. Since we have no file extent items to
2611 * represent holes, drop_end can be less than lockend and so we must
2612 * make sure we have an extent map representing the existing hole (the
2613 * call to __btrfs_drop_extents() might have dropped the existing extent
2614 * map representing the existing hole), otherwise the fast fsync path
2615 * will not record the existence of the hole region
2616 * [existing_hole_start, lockend].
2617 */
2618 if (drop_end <= lockend)
2619 drop_end = lockend + 1;
2620 /*
2621 * Don't insert file hole extent item if it's for a range beyond eof
2622 * (because it's useless) or if it represents a 0 bytes range (when
2623 * cur_offset == drop_end).
2624 */
2625 if (cur_offset < ino_size && cur_offset < drop_end) {
2626 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2627 if (ret) {
2628 /* Same comment as above. */
2629 btrfs_abort_transaction(trans, ret);
2630 err = ret;
2631 goto out_trans;
2632 }
2633 }
2634
2635out_trans:
2636 if (!trans)
2637 goto out_free;
2638
2639 inode_inc_iversion(inode);
2640 inode->i_mtime = inode->i_ctime = current_time(inode);
2641
2642 trans->block_rsv = &fs_info->trans_block_rsv;
2643 ret = btrfs_update_inode(trans, root, inode);
2644 updated_inode = true;
2645 btrfs_end_transaction(trans);
2646 btrfs_btree_balance_dirty(fs_info);
2647out_free:
2648 btrfs_free_path(path);
2649 btrfs_free_block_rsv(fs_info, rsv);
2650out:
2651 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2652 &cached_state, GFP_NOFS);
2653out_only_mutex:
2654 if (!updated_inode && truncated_block && !ret && !err) {
2655 /*
2656 * If we only end up zeroing part of a page, we still need to
2657 * update the inode item, so that all the time fields are
2658 * updated as well as the necessary btrfs inode in memory fields
2659 * for detecting, at fsync time, if the inode isn't yet in the
2660 * log tree or it's there but not up to date.
2661 */
2662 trans = btrfs_start_transaction(root, 1);
2663 if (IS_ERR(trans)) {
2664 err = PTR_ERR(trans);
2665 } else {
2666 err = btrfs_update_inode(trans, root, inode);
2667 ret = btrfs_end_transaction(trans);
2668 }
2669 }
2670 inode_unlock(inode);
2671 if (ret && !err)
2672 err = ret;
2673 return err;
2674}
2675
2676/* Helper structure to record which range is already reserved */
2677struct falloc_range {
2678 struct list_head list;
2679 u64 start;
2680 u64 len;
2681};
2682
2683/*
2684 * Helper function to add falloc range
2685 *
2686 * Caller should have locked the larger range of extent containing
2687 * [start, len)
2688 */
2689static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2690{
2691 struct falloc_range *prev = NULL;
2692 struct falloc_range *range = NULL;
2693
2694 if (list_empty(head))
2695 goto insert;
2696
2697 /*
2698 * As fallocate iterate by bytenr order, we only need to check
2699 * the last range.
2700 */
2701 prev = list_entry(head->prev, struct falloc_range, list);
2702 if (prev->start + prev->len == start) {
2703 prev->len += len;
2704 return 0;
2705 }
2706insert:
2707 range = kmalloc(sizeof(*range), GFP_KERNEL);
2708 if (!range)
2709 return -ENOMEM;
2710 range->start = start;
2711 range->len = len;
2712 list_add_tail(&range->list, head);
2713 return 0;
2714}
2715
2716static long btrfs_fallocate(struct file *file, int mode,
2717 loff_t offset, loff_t len)
2718{
2719 struct inode *inode = file_inode(file);
2720 struct extent_state *cached_state = NULL;
2721 struct falloc_range *range;
2722 struct falloc_range *tmp;
2723 struct list_head reserve_list;
2724 u64 cur_offset;
2725 u64 last_byte;
2726 u64 alloc_start;
2727 u64 alloc_end;
2728 u64 alloc_hint = 0;
2729 u64 locked_end;
2730 u64 actual_end = 0;
2731 struct extent_map *em;
2732 int blocksize = btrfs_inode_sectorsize(inode);
2733 int ret;
2734
2735 alloc_start = round_down(offset, blocksize);
2736 alloc_end = round_up(offset + len, blocksize);
2737 cur_offset = alloc_start;
2738
2739 /* Make sure we aren't being give some crap mode */
2740 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2741 return -EOPNOTSUPP;
2742
2743 if (mode & FALLOC_FL_PUNCH_HOLE)
2744 return btrfs_punch_hole(inode, offset, len);
2745
2746 /*
2747 * Only trigger disk allocation, don't trigger qgroup reserve
2748 *
2749 * For qgroup space, it will be checked later.
2750 */
2751 ret = btrfs_alloc_data_chunk_ondemand(inode, alloc_end - alloc_start);
2752 if (ret < 0)
2753 return ret;
2754
2755 inode_lock(inode);
2756
2757 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2758 ret = inode_newsize_ok(inode, offset + len);
2759 if (ret)
2760 goto out;
2761 }
2762
2763 /*
2764 * TODO: Move these two operations after we have checked
2765 * accurate reserved space, or fallocate can still fail but
2766 * with page truncated or size expanded.
2767 *
2768 * But that's a minor problem and won't do much harm BTW.
2769 */
2770 if (alloc_start > inode->i_size) {
2771 ret = btrfs_cont_expand(inode, i_size_read(inode),
2772 alloc_start);
2773 if (ret)
2774 goto out;
2775 } else if (offset + len > inode->i_size) {
2776 /*
2777 * If we are fallocating from the end of the file onward we
2778 * need to zero out the end of the block if i_size lands in the
2779 * middle of a block.
2780 */
2781 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2782 if (ret)
2783 goto out;
2784 }
2785
2786 /*
2787 * wait for ordered IO before we have any locks. We'll loop again
2788 * below with the locks held.
2789 */
2790 ret = btrfs_wait_ordered_range(inode, alloc_start,
2791 alloc_end - alloc_start);
2792 if (ret)
2793 goto out;
2794
2795 locked_end = alloc_end - 1;
2796 while (1) {
2797 struct btrfs_ordered_extent *ordered;
2798
2799 /* the extent lock is ordered inside the running
2800 * transaction
2801 */
2802 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2803 locked_end, &cached_state);
2804 ordered = btrfs_lookup_first_ordered_extent(inode,
2805 alloc_end - 1);
2806 if (ordered &&
2807 ordered->file_offset + ordered->len > alloc_start &&
2808 ordered->file_offset < alloc_end) {
2809 btrfs_put_ordered_extent(ordered);
2810 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2811 alloc_start, locked_end,
2812 &cached_state, GFP_KERNEL);
2813 /*
2814 * we can't wait on the range with the transaction
2815 * running or with the extent lock held
2816 */
2817 ret = btrfs_wait_ordered_range(inode, alloc_start,
2818 alloc_end - alloc_start);
2819 if (ret)
2820 goto out;
2821 } else {
2822 if (ordered)
2823 btrfs_put_ordered_extent(ordered);
2824 break;
2825 }
2826 }
2827
2828 /* First, check if we exceed the qgroup limit */
2829 INIT_LIST_HEAD(&reserve_list);
2830 while (1) {
2831 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2832 alloc_end - cur_offset, 0);
2833 if (IS_ERR_OR_NULL(em)) {
2834 if (!em)
2835 ret = -ENOMEM;
2836 else
2837 ret = PTR_ERR(em);
2838 break;
2839 }
2840 last_byte = min(extent_map_end(em), alloc_end);
2841 actual_end = min_t(u64, extent_map_end(em), offset + len);
2842 last_byte = ALIGN(last_byte, blocksize);
2843 if (em->block_start == EXTENT_MAP_HOLE ||
2844 (cur_offset >= inode->i_size &&
2845 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2846 ret = add_falloc_range(&reserve_list, cur_offset,
2847 last_byte - cur_offset);
2848 if (ret < 0) {
2849 free_extent_map(em);
2850 break;
2851 }
2852 ret = btrfs_qgroup_reserve_data(inode, cur_offset,
2853 last_byte - cur_offset);
2854 if (ret < 0)
2855 break;
2856 } else {
2857 /*
2858 * Do not need to reserve unwritten extent for this
2859 * range, free reserved data space first, otherwise
2860 * it'll result in false ENOSPC error.
2861 */
2862 btrfs_free_reserved_data_space(inode, cur_offset,
2863 last_byte - cur_offset);
2864 }
2865 free_extent_map(em);
2866 cur_offset = last_byte;
2867 if (cur_offset >= alloc_end)
2868 break;
2869 }
2870
2871 /*
2872 * If ret is still 0, means we're OK to fallocate.
2873 * Or just cleanup the list and exit.
2874 */
2875 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2876 if (!ret)
2877 ret = btrfs_prealloc_file_range(inode, mode,
2878 range->start,
2879 range->len, 1 << inode->i_blkbits,
2880 offset + len, &alloc_hint);
2881 else
2882 btrfs_free_reserved_data_space(inode, range->start,
2883 range->len);
2884 list_del(&range->list);
2885 kfree(range);
2886 }
2887 if (ret < 0)
2888 goto out_unlock;
2889
2890 if (actual_end > inode->i_size &&
2891 !(mode & FALLOC_FL_KEEP_SIZE)) {
2892 struct btrfs_trans_handle *trans;
2893 struct btrfs_root *root = BTRFS_I(inode)->root;
2894
2895 /*
2896 * We didn't need to allocate any more space, but we
2897 * still extended the size of the file so we need to
2898 * update i_size and the inode item.
2899 */
2900 trans = btrfs_start_transaction(root, 1);
2901 if (IS_ERR(trans)) {
2902 ret = PTR_ERR(trans);
2903 } else {
2904 inode->i_ctime = current_time(inode);
2905 i_size_write(inode, actual_end);
2906 btrfs_ordered_update_i_size(inode, actual_end, NULL);
2907 ret = btrfs_update_inode(trans, root, inode);
2908 if (ret)
2909 btrfs_end_transaction(trans);
2910 else
2911 ret = btrfs_end_transaction(trans);
2912 }
2913 }
2914out_unlock:
2915 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2916 &cached_state, GFP_KERNEL);
2917out:
2918 inode_unlock(inode);
2919 /* Let go of our reservation. */
2920 if (ret != 0)
2921 btrfs_free_reserved_data_space(inode, alloc_start,
2922 alloc_end - cur_offset);
2923 return ret;
2924}
2925
2926static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2927{
2928 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2929 struct extent_map *em = NULL;
2930 struct extent_state *cached_state = NULL;
2931 u64 lockstart;
2932 u64 lockend;
2933 u64 start;
2934 u64 len;
2935 int ret = 0;
2936
2937 if (inode->i_size == 0)
2938 return -ENXIO;
2939
2940 /*
2941 * *offset can be negative, in this case we start finding DATA/HOLE from
2942 * the very start of the file.
2943 */
2944 start = max_t(loff_t, 0, *offset);
2945
2946 lockstart = round_down(start, fs_info->sectorsize);
2947 lockend = round_up(i_size_read(inode),
2948 fs_info->sectorsize);
2949 if (lockend <= lockstart)
2950 lockend = lockstart + fs_info->sectorsize;
2951 lockend--;
2952 len = lockend - lockstart + 1;
2953
2954 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2955 &cached_state);
2956
2957 while (start < inode->i_size) {
2958 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2959 if (IS_ERR(em)) {
2960 ret = PTR_ERR(em);
2961 em = NULL;
2962 break;
2963 }
2964
2965 if (whence == SEEK_HOLE &&
2966 (em->block_start == EXTENT_MAP_HOLE ||
2967 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2968 break;
2969 else if (whence == SEEK_DATA &&
2970 (em->block_start != EXTENT_MAP_HOLE &&
2971 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2972 break;
2973
2974 start = em->start + em->len;
2975 free_extent_map(em);
2976 em = NULL;
2977 cond_resched();
2978 }
2979 free_extent_map(em);
2980 if (!ret) {
2981 if (whence == SEEK_DATA && start >= inode->i_size)
2982 ret = -ENXIO;
2983 else
2984 *offset = min_t(loff_t, start, inode->i_size);
2985 }
2986 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2987 &cached_state, GFP_NOFS);
2988 return ret;
2989}
2990
2991static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2992{
2993 struct inode *inode = file->f_mapping->host;
2994 int ret;
2995
2996 inode_lock(inode);
2997 switch (whence) {
2998 case SEEK_END:
2999 case SEEK_CUR:
3000 offset = generic_file_llseek(file, offset, whence);
3001 goto out;
3002 case SEEK_DATA:
3003 case SEEK_HOLE:
3004 if (offset >= i_size_read(inode)) {
3005 inode_unlock(inode);
3006 return -ENXIO;
3007 }
3008
3009 ret = find_desired_extent(inode, &offset, whence);
3010 if (ret) {
3011 inode_unlock(inode);
3012 return ret;
3013 }
3014 }
3015
3016 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3017out:
3018 inode_unlock(inode);
3019 return offset;
3020}
3021
3022const struct file_operations btrfs_file_operations = {
3023 .llseek = btrfs_file_llseek,
3024 .read_iter = generic_file_read_iter,
3025 .splice_read = generic_file_splice_read,
3026 .write_iter = btrfs_file_write_iter,
3027 .mmap = btrfs_file_mmap,
3028 .open = generic_file_open,
3029 .release = btrfs_release_file,
3030 .fsync = btrfs_sync_file,
3031 .fallocate = btrfs_fallocate,
3032 .unlocked_ioctl = btrfs_ioctl,
3033#ifdef CONFIG_COMPAT
3034 .compat_ioctl = btrfs_compat_ioctl,
3035#endif
3036 .clone_file_range = btrfs_clone_file_range,
3037 .dedupe_file_range = btrfs_dedupe_file_range,
3038};
3039
3040void btrfs_auto_defrag_exit(void)
3041{
3042 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3043}
3044
3045int btrfs_auto_defrag_init(void)
3046{
3047 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3048 sizeof(struct inode_defrag), 0,
3049 SLAB_MEM_SPREAD,
3050 NULL);
3051 if (!btrfs_inode_defrag_cachep)
3052 return -ENOMEM;
3053
3054 return 0;
3055}
3056
3057int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3058{
3059 int ret;
3060
3061 /*
3062 * So with compression we will find and lock a dirty page and clear the
3063 * first one as dirty, setup an async extent, and immediately return
3064 * with the entire range locked but with nobody actually marked with
3065 * writeback. So we can't just filemap_write_and_wait_range() and
3066 * expect it to work since it will just kick off a thread to do the
3067 * actual work. So we need to call filemap_fdatawrite_range _again_
3068 * since it will wait on the page lock, which won't be unlocked until
3069 * after the pages have been marked as writeback and so we're good to go
3070 * from there. We have to do this otherwise we'll miss the ordered
3071 * extents and that results in badness. Please Josef, do not think you
3072 * know better and pull this out at some point in the future, it is
3073 * right and you are wrong.
3074 */
3075 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3076 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3077 &BTRFS_I(inode)->runtime_flags))
3078 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3079
3080 return ret;
3081}