Loading...
1/*
2 * Copyright (C) 2007 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/fs.h>
20#include <linux/pagemap.h>
21#include <linux/highmem.h>
22#include <linux/time.h>
23#include <linux/init.h>
24#include <linux/string.h>
25#include <linux/backing-dev.h>
26#include <linux/mpage.h>
27#include <linux/falloc.h>
28#include <linux/swap.h>
29#include <linux/writeback.h>
30#include <linux/statfs.h>
31#include <linux/compat.h>
32#include <linux/slab.h>
33#include <linux/btrfs.h>
34#include <linux/uio.h>
35#include "ctree.h"
36#include "disk-io.h"
37#include "transaction.h"
38#include "btrfs_inode.h"
39#include "print-tree.h"
40#include "tree-log.h"
41#include "locking.h"
42#include "volumes.h"
43#include "qgroup.h"
44#include "compression.h"
45
46static struct kmem_cache *btrfs_inode_defrag_cachep;
47/*
48 * when auto defrag is enabled we
49 * queue up these defrag structs to remember which
50 * inodes need defragging passes
51 */
52struct inode_defrag {
53 struct rb_node rb_node;
54 /* objectid */
55 u64 ino;
56 /*
57 * transid where the defrag was added, we search for
58 * extents newer than this
59 */
60 u64 transid;
61
62 /* root objectid */
63 u64 root;
64
65 /* last offset we were able to defrag */
66 u64 last_offset;
67
68 /* if we've wrapped around back to zero once already */
69 int cycled;
70};
71
72static int __compare_inode_defrag(struct inode_defrag *defrag1,
73 struct inode_defrag *defrag2)
74{
75 if (defrag1->root > defrag2->root)
76 return 1;
77 else if (defrag1->root < defrag2->root)
78 return -1;
79 else if (defrag1->ino > defrag2->ino)
80 return 1;
81 else if (defrag1->ino < defrag2->ino)
82 return -1;
83 else
84 return 0;
85}
86
87/* pop a record for an inode into the defrag tree. The lock
88 * must be held already
89 *
90 * If you're inserting a record for an older transid than an
91 * existing record, the transid already in the tree is lowered
92 *
93 * If an existing record is found the defrag item you
94 * pass in is freed
95 */
96static int __btrfs_add_inode_defrag(struct inode *inode,
97 struct inode_defrag *defrag)
98{
99 struct btrfs_root *root = BTRFS_I(inode)->root;
100 struct inode_defrag *entry;
101 struct rb_node **p;
102 struct rb_node *parent = NULL;
103 int ret;
104
105 p = &root->fs_info->defrag_inodes.rb_node;
106 while (*p) {
107 parent = *p;
108 entry = rb_entry(parent, struct inode_defrag, rb_node);
109
110 ret = __compare_inode_defrag(defrag, entry);
111 if (ret < 0)
112 p = &parent->rb_left;
113 else if (ret > 0)
114 p = &parent->rb_right;
115 else {
116 /* if we're reinserting an entry for
117 * an old defrag run, make sure to
118 * lower the transid of our existing record
119 */
120 if (defrag->transid < entry->transid)
121 entry->transid = defrag->transid;
122 if (defrag->last_offset > entry->last_offset)
123 entry->last_offset = defrag->last_offset;
124 return -EEXIST;
125 }
126 }
127 set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
128 rb_link_node(&defrag->rb_node, parent, p);
129 rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
130 return 0;
131}
132
133static inline int __need_auto_defrag(struct btrfs_root *root)
134{
135 if (!btrfs_test_opt(root, AUTO_DEFRAG))
136 return 0;
137
138 if (btrfs_fs_closing(root->fs_info))
139 return 0;
140
141 return 1;
142}
143
144/*
145 * insert a defrag record for this inode if auto defrag is
146 * enabled
147 */
148int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
149 struct inode *inode)
150{
151 struct btrfs_root *root = BTRFS_I(inode)->root;
152 struct inode_defrag *defrag;
153 u64 transid;
154 int ret;
155
156 if (!__need_auto_defrag(root))
157 return 0;
158
159 if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
160 return 0;
161
162 if (trans)
163 transid = trans->transid;
164 else
165 transid = BTRFS_I(inode)->root->last_trans;
166
167 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
168 if (!defrag)
169 return -ENOMEM;
170
171 defrag->ino = btrfs_ino(inode);
172 defrag->transid = transid;
173 defrag->root = root->root_key.objectid;
174
175 spin_lock(&root->fs_info->defrag_inodes_lock);
176 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
177 /*
178 * If we set IN_DEFRAG flag and evict the inode from memory,
179 * and then re-read this inode, this new inode doesn't have
180 * IN_DEFRAG flag. At the case, we may find the existed defrag.
181 */
182 ret = __btrfs_add_inode_defrag(inode, defrag);
183 if (ret)
184 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
185 } else {
186 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
187 }
188 spin_unlock(&root->fs_info->defrag_inodes_lock);
189 return 0;
190}
191
192/*
193 * Requeue the defrag object. If there is a defrag object that points to
194 * the same inode in the tree, we will merge them together (by
195 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
196 */
197static void btrfs_requeue_inode_defrag(struct inode *inode,
198 struct inode_defrag *defrag)
199{
200 struct btrfs_root *root = BTRFS_I(inode)->root;
201 int ret;
202
203 if (!__need_auto_defrag(root))
204 goto out;
205
206 /*
207 * Here we don't check the IN_DEFRAG flag, because we need merge
208 * them together.
209 */
210 spin_lock(&root->fs_info->defrag_inodes_lock);
211 ret = __btrfs_add_inode_defrag(inode, defrag);
212 spin_unlock(&root->fs_info->defrag_inodes_lock);
213 if (ret)
214 goto out;
215 return;
216out:
217 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
218}
219
220/*
221 * pick the defragable inode that we want, if it doesn't exist, we will get
222 * the next one.
223 */
224static struct inode_defrag *
225btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
226{
227 struct inode_defrag *entry = NULL;
228 struct inode_defrag tmp;
229 struct rb_node *p;
230 struct rb_node *parent = NULL;
231 int ret;
232
233 tmp.ino = ino;
234 tmp.root = root;
235
236 spin_lock(&fs_info->defrag_inodes_lock);
237 p = fs_info->defrag_inodes.rb_node;
238 while (p) {
239 parent = p;
240 entry = rb_entry(parent, struct inode_defrag, rb_node);
241
242 ret = __compare_inode_defrag(&tmp, entry);
243 if (ret < 0)
244 p = parent->rb_left;
245 else if (ret > 0)
246 p = parent->rb_right;
247 else
248 goto out;
249 }
250
251 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
252 parent = rb_next(parent);
253 if (parent)
254 entry = rb_entry(parent, struct inode_defrag, rb_node);
255 else
256 entry = NULL;
257 }
258out:
259 if (entry)
260 rb_erase(parent, &fs_info->defrag_inodes);
261 spin_unlock(&fs_info->defrag_inodes_lock);
262 return entry;
263}
264
265void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
266{
267 struct inode_defrag *defrag;
268 struct rb_node *node;
269
270 spin_lock(&fs_info->defrag_inodes_lock);
271 node = rb_first(&fs_info->defrag_inodes);
272 while (node) {
273 rb_erase(node, &fs_info->defrag_inodes);
274 defrag = rb_entry(node, struct inode_defrag, rb_node);
275 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
276
277 cond_resched_lock(&fs_info->defrag_inodes_lock);
278
279 node = rb_first(&fs_info->defrag_inodes);
280 }
281 spin_unlock(&fs_info->defrag_inodes_lock);
282}
283
284#define BTRFS_DEFRAG_BATCH 1024
285
286static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
287 struct inode_defrag *defrag)
288{
289 struct btrfs_root *inode_root;
290 struct inode *inode;
291 struct btrfs_key key;
292 struct btrfs_ioctl_defrag_range_args range;
293 int num_defrag;
294 int index;
295 int ret;
296
297 /* get the inode */
298 key.objectid = defrag->root;
299 key.type = BTRFS_ROOT_ITEM_KEY;
300 key.offset = (u64)-1;
301
302 index = srcu_read_lock(&fs_info->subvol_srcu);
303
304 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
305 if (IS_ERR(inode_root)) {
306 ret = PTR_ERR(inode_root);
307 goto cleanup;
308 }
309
310 key.objectid = defrag->ino;
311 key.type = BTRFS_INODE_ITEM_KEY;
312 key.offset = 0;
313 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
314 if (IS_ERR(inode)) {
315 ret = PTR_ERR(inode);
316 goto cleanup;
317 }
318 srcu_read_unlock(&fs_info->subvol_srcu, index);
319
320 /* do a chunk of defrag */
321 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
322 memset(&range, 0, sizeof(range));
323 range.len = (u64)-1;
324 range.start = defrag->last_offset;
325
326 sb_start_write(fs_info->sb);
327 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
328 BTRFS_DEFRAG_BATCH);
329 sb_end_write(fs_info->sb);
330 /*
331 * if we filled the whole defrag batch, there
332 * must be more work to do. Queue this defrag
333 * again
334 */
335 if (num_defrag == BTRFS_DEFRAG_BATCH) {
336 defrag->last_offset = range.start;
337 btrfs_requeue_inode_defrag(inode, defrag);
338 } else if (defrag->last_offset && !defrag->cycled) {
339 /*
340 * we didn't fill our defrag batch, but
341 * we didn't start at zero. Make sure we loop
342 * around to the start of the file.
343 */
344 defrag->last_offset = 0;
345 defrag->cycled = 1;
346 btrfs_requeue_inode_defrag(inode, defrag);
347 } else {
348 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
349 }
350
351 iput(inode);
352 return 0;
353cleanup:
354 srcu_read_unlock(&fs_info->subvol_srcu, index);
355 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
356 return ret;
357}
358
359/*
360 * run through the list of inodes in the FS that need
361 * defragging
362 */
363int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
364{
365 struct inode_defrag *defrag;
366 u64 first_ino = 0;
367 u64 root_objectid = 0;
368
369 atomic_inc(&fs_info->defrag_running);
370 while (1) {
371 /* Pause the auto defragger. */
372 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
373 &fs_info->fs_state))
374 break;
375
376 if (!__need_auto_defrag(fs_info->tree_root))
377 break;
378
379 /* find an inode to defrag */
380 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
381 first_ino);
382 if (!defrag) {
383 if (root_objectid || first_ino) {
384 root_objectid = 0;
385 first_ino = 0;
386 continue;
387 } else {
388 break;
389 }
390 }
391
392 first_ino = defrag->ino + 1;
393 root_objectid = defrag->root;
394
395 __btrfs_run_defrag_inode(fs_info, defrag);
396 }
397 atomic_dec(&fs_info->defrag_running);
398
399 /*
400 * during unmount, we use the transaction_wait queue to
401 * wait for the defragger to stop
402 */
403 wake_up(&fs_info->transaction_wait);
404 return 0;
405}
406
407/* simple helper to fault in pages and copy. This should go away
408 * and be replaced with calls into generic code.
409 */
410static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
411 struct page **prepared_pages,
412 struct iov_iter *i)
413{
414 size_t copied = 0;
415 size_t total_copied = 0;
416 int pg = 0;
417 int offset = pos & (PAGE_SIZE - 1);
418
419 while (write_bytes > 0) {
420 size_t count = min_t(size_t,
421 PAGE_SIZE - offset, write_bytes);
422 struct page *page = prepared_pages[pg];
423 /*
424 * Copy data from userspace to the current page
425 */
426 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
427
428 /* Flush processor's dcache for this page */
429 flush_dcache_page(page);
430
431 /*
432 * if we get a partial write, we can end up with
433 * partially up to date pages. These add
434 * a lot of complexity, so make sure they don't
435 * happen by forcing this copy to be retried.
436 *
437 * The rest of the btrfs_file_write code will fall
438 * back to page at a time copies after we return 0.
439 */
440 if (!PageUptodate(page) && copied < count)
441 copied = 0;
442
443 iov_iter_advance(i, copied);
444 write_bytes -= copied;
445 total_copied += copied;
446
447 /* Return to btrfs_file_write_iter to fault page */
448 if (unlikely(copied == 0))
449 break;
450
451 if (copied < PAGE_SIZE - offset) {
452 offset += copied;
453 } else {
454 pg++;
455 offset = 0;
456 }
457 }
458 return total_copied;
459}
460
461/*
462 * unlocks pages after btrfs_file_write is done with them
463 */
464static void btrfs_drop_pages(struct page **pages, size_t num_pages)
465{
466 size_t i;
467 for (i = 0; i < num_pages; i++) {
468 /* page checked is some magic around finding pages that
469 * have been modified without going through btrfs_set_page_dirty
470 * clear it here. There should be no need to mark the pages
471 * accessed as prepare_pages should have marked them accessed
472 * in prepare_pages via find_or_create_page()
473 */
474 ClearPageChecked(pages[i]);
475 unlock_page(pages[i]);
476 put_page(pages[i]);
477 }
478}
479
480/*
481 * after copy_from_user, pages need to be dirtied and we need to make
482 * sure holes are created between the current EOF and the start of
483 * any next extents (if required).
484 *
485 * this also makes the decision about creating an inline extent vs
486 * doing real data extents, marking pages dirty and delalloc as required.
487 */
488int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
489 struct page **pages, size_t num_pages,
490 loff_t pos, size_t write_bytes,
491 struct extent_state **cached)
492{
493 int err = 0;
494 int i;
495 u64 num_bytes;
496 u64 start_pos;
497 u64 end_of_last_block;
498 u64 end_pos = pos + write_bytes;
499 loff_t isize = i_size_read(inode);
500
501 start_pos = pos & ~((u64)root->sectorsize - 1);
502 num_bytes = round_up(write_bytes + pos - start_pos, root->sectorsize);
503
504 end_of_last_block = start_pos + num_bytes - 1;
505 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
506 cached);
507 if (err)
508 return err;
509
510 for (i = 0; i < num_pages; i++) {
511 struct page *p = pages[i];
512 SetPageUptodate(p);
513 ClearPageChecked(p);
514 set_page_dirty(p);
515 }
516
517 /*
518 * we've only changed i_size in ram, and we haven't updated
519 * the disk i_size. There is no need to log the inode
520 * at this time.
521 */
522 if (end_pos > isize)
523 i_size_write(inode, end_pos);
524 return 0;
525}
526
527/*
528 * this drops all the extents in the cache that intersect the range
529 * [start, end]. Existing extents are split as required.
530 */
531void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
532 int skip_pinned)
533{
534 struct extent_map *em;
535 struct extent_map *split = NULL;
536 struct extent_map *split2 = NULL;
537 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
538 u64 len = end - start + 1;
539 u64 gen;
540 int ret;
541 int testend = 1;
542 unsigned long flags;
543 int compressed = 0;
544 bool modified;
545
546 WARN_ON(end < start);
547 if (end == (u64)-1) {
548 len = (u64)-1;
549 testend = 0;
550 }
551 while (1) {
552 int no_splits = 0;
553
554 modified = false;
555 if (!split)
556 split = alloc_extent_map();
557 if (!split2)
558 split2 = alloc_extent_map();
559 if (!split || !split2)
560 no_splits = 1;
561
562 write_lock(&em_tree->lock);
563 em = lookup_extent_mapping(em_tree, start, len);
564 if (!em) {
565 write_unlock(&em_tree->lock);
566 break;
567 }
568 flags = em->flags;
569 gen = em->generation;
570 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
571 if (testend && em->start + em->len >= start + len) {
572 free_extent_map(em);
573 write_unlock(&em_tree->lock);
574 break;
575 }
576 start = em->start + em->len;
577 if (testend)
578 len = start + len - (em->start + em->len);
579 free_extent_map(em);
580 write_unlock(&em_tree->lock);
581 continue;
582 }
583 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
584 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
585 clear_bit(EXTENT_FLAG_LOGGING, &flags);
586 modified = !list_empty(&em->list);
587 if (no_splits)
588 goto next;
589
590 if (em->start < start) {
591 split->start = em->start;
592 split->len = start - em->start;
593
594 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
595 split->orig_start = em->orig_start;
596 split->block_start = em->block_start;
597
598 if (compressed)
599 split->block_len = em->block_len;
600 else
601 split->block_len = split->len;
602 split->orig_block_len = max(split->block_len,
603 em->orig_block_len);
604 split->ram_bytes = em->ram_bytes;
605 } else {
606 split->orig_start = split->start;
607 split->block_len = 0;
608 split->block_start = em->block_start;
609 split->orig_block_len = 0;
610 split->ram_bytes = split->len;
611 }
612
613 split->generation = gen;
614 split->bdev = em->bdev;
615 split->flags = flags;
616 split->compress_type = em->compress_type;
617 replace_extent_mapping(em_tree, em, split, modified);
618 free_extent_map(split);
619 split = split2;
620 split2 = NULL;
621 }
622 if (testend && em->start + em->len > start + len) {
623 u64 diff = start + len - em->start;
624
625 split->start = start + len;
626 split->len = em->start + em->len - (start + len);
627 split->bdev = em->bdev;
628 split->flags = flags;
629 split->compress_type = em->compress_type;
630 split->generation = gen;
631
632 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
633 split->orig_block_len = max(em->block_len,
634 em->orig_block_len);
635
636 split->ram_bytes = em->ram_bytes;
637 if (compressed) {
638 split->block_len = em->block_len;
639 split->block_start = em->block_start;
640 split->orig_start = em->orig_start;
641 } else {
642 split->block_len = split->len;
643 split->block_start = em->block_start
644 + diff;
645 split->orig_start = em->orig_start;
646 }
647 } else {
648 split->ram_bytes = split->len;
649 split->orig_start = split->start;
650 split->block_len = 0;
651 split->block_start = em->block_start;
652 split->orig_block_len = 0;
653 }
654
655 if (extent_map_in_tree(em)) {
656 replace_extent_mapping(em_tree, em, split,
657 modified);
658 } else {
659 ret = add_extent_mapping(em_tree, split,
660 modified);
661 ASSERT(ret == 0); /* Logic error */
662 }
663 free_extent_map(split);
664 split = NULL;
665 }
666next:
667 if (extent_map_in_tree(em))
668 remove_extent_mapping(em_tree, em);
669 write_unlock(&em_tree->lock);
670
671 /* once for us */
672 free_extent_map(em);
673 /* once for the tree*/
674 free_extent_map(em);
675 }
676 if (split)
677 free_extent_map(split);
678 if (split2)
679 free_extent_map(split2);
680}
681
682/*
683 * this is very complex, but the basic idea is to drop all extents
684 * in the range start - end. hint_block is filled in with a block number
685 * that would be a good hint to the block allocator for this file.
686 *
687 * If an extent intersects the range but is not entirely inside the range
688 * it is either truncated or split. Anything entirely inside the range
689 * is deleted from the tree.
690 */
691int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
692 struct btrfs_root *root, struct inode *inode,
693 struct btrfs_path *path, u64 start, u64 end,
694 u64 *drop_end, int drop_cache,
695 int replace_extent,
696 u32 extent_item_size,
697 int *key_inserted)
698{
699 struct extent_buffer *leaf;
700 struct btrfs_file_extent_item *fi;
701 struct btrfs_key key;
702 struct btrfs_key new_key;
703 u64 ino = btrfs_ino(inode);
704 u64 search_start = start;
705 u64 disk_bytenr = 0;
706 u64 num_bytes = 0;
707 u64 extent_offset = 0;
708 u64 extent_end = 0;
709 int del_nr = 0;
710 int del_slot = 0;
711 int extent_type;
712 int recow;
713 int ret;
714 int modify_tree = -1;
715 int update_refs;
716 int found = 0;
717 int leafs_visited = 0;
718
719 if (drop_cache)
720 btrfs_drop_extent_cache(inode, start, end - 1, 0);
721
722 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
723 modify_tree = 0;
724
725 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
726 root == root->fs_info->tree_root);
727 while (1) {
728 recow = 0;
729 ret = btrfs_lookup_file_extent(trans, root, path, ino,
730 search_start, modify_tree);
731 if (ret < 0)
732 break;
733 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
734 leaf = path->nodes[0];
735 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
736 if (key.objectid == ino &&
737 key.type == BTRFS_EXTENT_DATA_KEY)
738 path->slots[0]--;
739 }
740 ret = 0;
741 leafs_visited++;
742next_slot:
743 leaf = path->nodes[0];
744 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
745 BUG_ON(del_nr > 0);
746 ret = btrfs_next_leaf(root, path);
747 if (ret < 0)
748 break;
749 if (ret > 0) {
750 ret = 0;
751 break;
752 }
753 leafs_visited++;
754 leaf = path->nodes[0];
755 recow = 1;
756 }
757
758 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
759
760 if (key.objectid > ino)
761 break;
762 if (WARN_ON_ONCE(key.objectid < ino) ||
763 key.type < BTRFS_EXTENT_DATA_KEY) {
764 ASSERT(del_nr == 0);
765 path->slots[0]++;
766 goto next_slot;
767 }
768 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
769 break;
770
771 fi = btrfs_item_ptr(leaf, path->slots[0],
772 struct btrfs_file_extent_item);
773 extent_type = btrfs_file_extent_type(leaf, fi);
774
775 if (extent_type == BTRFS_FILE_EXTENT_REG ||
776 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
777 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
778 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
779 extent_offset = btrfs_file_extent_offset(leaf, fi);
780 extent_end = key.offset +
781 btrfs_file_extent_num_bytes(leaf, fi);
782 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
783 extent_end = key.offset +
784 btrfs_file_extent_inline_len(leaf,
785 path->slots[0], fi);
786 } else {
787 /* can't happen */
788 BUG();
789 }
790
791 /*
792 * Don't skip extent items representing 0 byte lengths. They
793 * used to be created (bug) if while punching holes we hit
794 * -ENOSPC condition. So if we find one here, just ensure we
795 * delete it, otherwise we would insert a new file extent item
796 * with the same key (offset) as that 0 bytes length file
797 * extent item in the call to setup_items_for_insert() later
798 * in this function.
799 */
800 if (extent_end == key.offset && extent_end >= search_start)
801 goto delete_extent_item;
802
803 if (extent_end <= search_start) {
804 path->slots[0]++;
805 goto next_slot;
806 }
807
808 found = 1;
809 search_start = max(key.offset, start);
810 if (recow || !modify_tree) {
811 modify_tree = -1;
812 btrfs_release_path(path);
813 continue;
814 }
815
816 /*
817 * | - range to drop - |
818 * | -------- extent -------- |
819 */
820 if (start > key.offset && end < extent_end) {
821 BUG_ON(del_nr > 0);
822 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
823 ret = -EOPNOTSUPP;
824 break;
825 }
826
827 memcpy(&new_key, &key, sizeof(new_key));
828 new_key.offset = start;
829 ret = btrfs_duplicate_item(trans, root, path,
830 &new_key);
831 if (ret == -EAGAIN) {
832 btrfs_release_path(path);
833 continue;
834 }
835 if (ret < 0)
836 break;
837
838 leaf = path->nodes[0];
839 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
840 struct btrfs_file_extent_item);
841 btrfs_set_file_extent_num_bytes(leaf, fi,
842 start - key.offset);
843
844 fi = btrfs_item_ptr(leaf, path->slots[0],
845 struct btrfs_file_extent_item);
846
847 extent_offset += start - key.offset;
848 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
849 btrfs_set_file_extent_num_bytes(leaf, fi,
850 extent_end - start);
851 btrfs_mark_buffer_dirty(leaf);
852
853 if (update_refs && disk_bytenr > 0) {
854 ret = btrfs_inc_extent_ref(trans, root,
855 disk_bytenr, num_bytes, 0,
856 root->root_key.objectid,
857 new_key.objectid,
858 start - extent_offset);
859 BUG_ON(ret); /* -ENOMEM */
860 }
861 key.offset = start;
862 }
863 /*
864 * | ---- range to drop ----- |
865 * | -------- extent -------- |
866 */
867 if (start <= key.offset && end < extent_end) {
868 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
869 ret = -EOPNOTSUPP;
870 break;
871 }
872
873 memcpy(&new_key, &key, sizeof(new_key));
874 new_key.offset = end;
875 btrfs_set_item_key_safe(root->fs_info, path, &new_key);
876
877 extent_offset += end - key.offset;
878 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
879 btrfs_set_file_extent_num_bytes(leaf, fi,
880 extent_end - end);
881 btrfs_mark_buffer_dirty(leaf);
882 if (update_refs && disk_bytenr > 0)
883 inode_sub_bytes(inode, end - key.offset);
884 break;
885 }
886
887 search_start = extent_end;
888 /*
889 * | ---- range to drop ----- |
890 * | -------- extent -------- |
891 */
892 if (start > key.offset && end >= extent_end) {
893 BUG_ON(del_nr > 0);
894 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
895 ret = -EOPNOTSUPP;
896 break;
897 }
898
899 btrfs_set_file_extent_num_bytes(leaf, fi,
900 start - key.offset);
901 btrfs_mark_buffer_dirty(leaf);
902 if (update_refs && disk_bytenr > 0)
903 inode_sub_bytes(inode, extent_end - start);
904 if (end == extent_end)
905 break;
906
907 path->slots[0]++;
908 goto next_slot;
909 }
910
911 /*
912 * | ---- range to drop ----- |
913 * | ------ extent ------ |
914 */
915 if (start <= key.offset && end >= extent_end) {
916delete_extent_item:
917 if (del_nr == 0) {
918 del_slot = path->slots[0];
919 del_nr = 1;
920 } else {
921 BUG_ON(del_slot + del_nr != path->slots[0]);
922 del_nr++;
923 }
924
925 if (update_refs &&
926 extent_type == BTRFS_FILE_EXTENT_INLINE) {
927 inode_sub_bytes(inode,
928 extent_end - key.offset);
929 extent_end = ALIGN(extent_end,
930 root->sectorsize);
931 } else if (update_refs && disk_bytenr > 0) {
932 ret = btrfs_free_extent(trans, root,
933 disk_bytenr, num_bytes, 0,
934 root->root_key.objectid,
935 key.objectid, key.offset -
936 extent_offset);
937 BUG_ON(ret); /* -ENOMEM */
938 inode_sub_bytes(inode,
939 extent_end - key.offset);
940 }
941
942 if (end == extent_end)
943 break;
944
945 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
946 path->slots[0]++;
947 goto next_slot;
948 }
949
950 ret = btrfs_del_items(trans, root, path, del_slot,
951 del_nr);
952 if (ret) {
953 btrfs_abort_transaction(trans, root, ret);
954 break;
955 }
956
957 del_nr = 0;
958 del_slot = 0;
959
960 btrfs_release_path(path);
961 continue;
962 }
963
964 BUG_ON(1);
965 }
966
967 if (!ret && del_nr > 0) {
968 /*
969 * Set path->slots[0] to first slot, so that after the delete
970 * if items are move off from our leaf to its immediate left or
971 * right neighbor leafs, we end up with a correct and adjusted
972 * path->slots[0] for our insertion (if replace_extent != 0).
973 */
974 path->slots[0] = del_slot;
975 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
976 if (ret)
977 btrfs_abort_transaction(trans, root, ret);
978 }
979
980 leaf = path->nodes[0];
981 /*
982 * If btrfs_del_items() was called, it might have deleted a leaf, in
983 * which case it unlocked our path, so check path->locks[0] matches a
984 * write lock.
985 */
986 if (!ret && replace_extent && leafs_visited == 1 &&
987 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
988 path->locks[0] == BTRFS_WRITE_LOCK) &&
989 btrfs_leaf_free_space(root, leaf) >=
990 sizeof(struct btrfs_item) + extent_item_size) {
991
992 key.objectid = ino;
993 key.type = BTRFS_EXTENT_DATA_KEY;
994 key.offset = start;
995 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
996 struct btrfs_key slot_key;
997
998 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
999 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1000 path->slots[0]++;
1001 }
1002 setup_items_for_insert(root, path, &key,
1003 &extent_item_size,
1004 extent_item_size,
1005 sizeof(struct btrfs_item) +
1006 extent_item_size, 1);
1007 *key_inserted = 1;
1008 }
1009
1010 if (!replace_extent || !(*key_inserted))
1011 btrfs_release_path(path);
1012 if (drop_end)
1013 *drop_end = found ? min(end, extent_end) : end;
1014 return ret;
1015}
1016
1017int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1018 struct btrfs_root *root, struct inode *inode, u64 start,
1019 u64 end, int drop_cache)
1020{
1021 struct btrfs_path *path;
1022 int ret;
1023
1024 path = btrfs_alloc_path();
1025 if (!path)
1026 return -ENOMEM;
1027 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1028 drop_cache, 0, 0, NULL);
1029 btrfs_free_path(path);
1030 return ret;
1031}
1032
1033static int extent_mergeable(struct extent_buffer *leaf, int slot,
1034 u64 objectid, u64 bytenr, u64 orig_offset,
1035 u64 *start, u64 *end)
1036{
1037 struct btrfs_file_extent_item *fi;
1038 struct btrfs_key key;
1039 u64 extent_end;
1040
1041 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1042 return 0;
1043
1044 btrfs_item_key_to_cpu(leaf, &key, slot);
1045 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1046 return 0;
1047
1048 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1049 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1050 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1051 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1052 btrfs_file_extent_compression(leaf, fi) ||
1053 btrfs_file_extent_encryption(leaf, fi) ||
1054 btrfs_file_extent_other_encoding(leaf, fi))
1055 return 0;
1056
1057 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1058 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1059 return 0;
1060
1061 *start = key.offset;
1062 *end = extent_end;
1063 return 1;
1064}
1065
1066/*
1067 * Mark extent in the range start - end as written.
1068 *
1069 * This changes extent type from 'pre-allocated' to 'regular'. If only
1070 * part of extent is marked as written, the extent will be split into
1071 * two or three.
1072 */
1073int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1074 struct inode *inode, u64 start, u64 end)
1075{
1076 struct btrfs_root *root = BTRFS_I(inode)->root;
1077 struct extent_buffer *leaf;
1078 struct btrfs_path *path;
1079 struct btrfs_file_extent_item *fi;
1080 struct btrfs_key key;
1081 struct btrfs_key new_key;
1082 u64 bytenr;
1083 u64 num_bytes;
1084 u64 extent_end;
1085 u64 orig_offset;
1086 u64 other_start;
1087 u64 other_end;
1088 u64 split;
1089 int del_nr = 0;
1090 int del_slot = 0;
1091 int recow;
1092 int ret;
1093 u64 ino = btrfs_ino(inode);
1094
1095 path = btrfs_alloc_path();
1096 if (!path)
1097 return -ENOMEM;
1098again:
1099 recow = 0;
1100 split = start;
1101 key.objectid = ino;
1102 key.type = BTRFS_EXTENT_DATA_KEY;
1103 key.offset = split;
1104
1105 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1106 if (ret < 0)
1107 goto out;
1108 if (ret > 0 && path->slots[0] > 0)
1109 path->slots[0]--;
1110
1111 leaf = path->nodes[0];
1112 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1113 BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1114 fi = btrfs_item_ptr(leaf, path->slots[0],
1115 struct btrfs_file_extent_item);
1116 BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1117 BTRFS_FILE_EXTENT_PREALLOC);
1118 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1119 BUG_ON(key.offset > start || extent_end < end);
1120
1121 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1122 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1123 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1124 memcpy(&new_key, &key, sizeof(new_key));
1125
1126 if (start == key.offset && end < extent_end) {
1127 other_start = 0;
1128 other_end = start;
1129 if (extent_mergeable(leaf, path->slots[0] - 1,
1130 ino, bytenr, orig_offset,
1131 &other_start, &other_end)) {
1132 new_key.offset = end;
1133 btrfs_set_item_key_safe(root->fs_info, path, &new_key);
1134 fi = btrfs_item_ptr(leaf, path->slots[0],
1135 struct btrfs_file_extent_item);
1136 btrfs_set_file_extent_generation(leaf, fi,
1137 trans->transid);
1138 btrfs_set_file_extent_num_bytes(leaf, fi,
1139 extent_end - end);
1140 btrfs_set_file_extent_offset(leaf, fi,
1141 end - orig_offset);
1142 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1143 struct btrfs_file_extent_item);
1144 btrfs_set_file_extent_generation(leaf, fi,
1145 trans->transid);
1146 btrfs_set_file_extent_num_bytes(leaf, fi,
1147 end - other_start);
1148 btrfs_mark_buffer_dirty(leaf);
1149 goto out;
1150 }
1151 }
1152
1153 if (start > key.offset && end == extent_end) {
1154 other_start = end;
1155 other_end = 0;
1156 if (extent_mergeable(leaf, path->slots[0] + 1,
1157 ino, bytenr, orig_offset,
1158 &other_start, &other_end)) {
1159 fi = btrfs_item_ptr(leaf, path->slots[0],
1160 struct btrfs_file_extent_item);
1161 btrfs_set_file_extent_num_bytes(leaf, fi,
1162 start - key.offset);
1163 btrfs_set_file_extent_generation(leaf, fi,
1164 trans->transid);
1165 path->slots[0]++;
1166 new_key.offset = start;
1167 btrfs_set_item_key_safe(root->fs_info, path, &new_key);
1168
1169 fi = btrfs_item_ptr(leaf, path->slots[0],
1170 struct btrfs_file_extent_item);
1171 btrfs_set_file_extent_generation(leaf, fi,
1172 trans->transid);
1173 btrfs_set_file_extent_num_bytes(leaf, fi,
1174 other_end - start);
1175 btrfs_set_file_extent_offset(leaf, fi,
1176 start - orig_offset);
1177 btrfs_mark_buffer_dirty(leaf);
1178 goto out;
1179 }
1180 }
1181
1182 while (start > key.offset || end < extent_end) {
1183 if (key.offset == start)
1184 split = end;
1185
1186 new_key.offset = split;
1187 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1188 if (ret == -EAGAIN) {
1189 btrfs_release_path(path);
1190 goto again;
1191 }
1192 if (ret < 0) {
1193 btrfs_abort_transaction(trans, root, ret);
1194 goto out;
1195 }
1196
1197 leaf = path->nodes[0];
1198 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1199 struct btrfs_file_extent_item);
1200 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1201 btrfs_set_file_extent_num_bytes(leaf, fi,
1202 split - key.offset);
1203
1204 fi = btrfs_item_ptr(leaf, path->slots[0],
1205 struct btrfs_file_extent_item);
1206
1207 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1208 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1209 btrfs_set_file_extent_num_bytes(leaf, fi,
1210 extent_end - split);
1211 btrfs_mark_buffer_dirty(leaf);
1212
1213 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1214 root->root_key.objectid,
1215 ino, orig_offset);
1216 BUG_ON(ret); /* -ENOMEM */
1217
1218 if (split == start) {
1219 key.offset = start;
1220 } else {
1221 BUG_ON(start != key.offset);
1222 path->slots[0]--;
1223 extent_end = end;
1224 }
1225 recow = 1;
1226 }
1227
1228 other_start = end;
1229 other_end = 0;
1230 if (extent_mergeable(leaf, path->slots[0] + 1,
1231 ino, bytenr, orig_offset,
1232 &other_start, &other_end)) {
1233 if (recow) {
1234 btrfs_release_path(path);
1235 goto again;
1236 }
1237 extent_end = other_end;
1238 del_slot = path->slots[0] + 1;
1239 del_nr++;
1240 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1241 0, root->root_key.objectid,
1242 ino, orig_offset);
1243 BUG_ON(ret); /* -ENOMEM */
1244 }
1245 other_start = 0;
1246 other_end = start;
1247 if (extent_mergeable(leaf, path->slots[0] - 1,
1248 ino, bytenr, orig_offset,
1249 &other_start, &other_end)) {
1250 if (recow) {
1251 btrfs_release_path(path);
1252 goto again;
1253 }
1254 key.offset = other_start;
1255 del_slot = path->slots[0];
1256 del_nr++;
1257 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1258 0, root->root_key.objectid,
1259 ino, orig_offset);
1260 BUG_ON(ret); /* -ENOMEM */
1261 }
1262 if (del_nr == 0) {
1263 fi = btrfs_item_ptr(leaf, path->slots[0],
1264 struct btrfs_file_extent_item);
1265 btrfs_set_file_extent_type(leaf, fi,
1266 BTRFS_FILE_EXTENT_REG);
1267 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1268 btrfs_mark_buffer_dirty(leaf);
1269 } else {
1270 fi = btrfs_item_ptr(leaf, del_slot - 1,
1271 struct btrfs_file_extent_item);
1272 btrfs_set_file_extent_type(leaf, fi,
1273 BTRFS_FILE_EXTENT_REG);
1274 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1275 btrfs_set_file_extent_num_bytes(leaf, fi,
1276 extent_end - key.offset);
1277 btrfs_mark_buffer_dirty(leaf);
1278
1279 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1280 if (ret < 0) {
1281 btrfs_abort_transaction(trans, root, ret);
1282 goto out;
1283 }
1284 }
1285out:
1286 btrfs_free_path(path);
1287 return 0;
1288}
1289
1290/*
1291 * on error we return an unlocked page and the error value
1292 * on success we return a locked page and 0
1293 */
1294static int prepare_uptodate_page(struct inode *inode,
1295 struct page *page, u64 pos,
1296 bool force_uptodate)
1297{
1298 int ret = 0;
1299
1300 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1301 !PageUptodate(page)) {
1302 ret = btrfs_readpage(NULL, page);
1303 if (ret)
1304 return ret;
1305 lock_page(page);
1306 if (!PageUptodate(page)) {
1307 unlock_page(page);
1308 return -EIO;
1309 }
1310 if (page->mapping != inode->i_mapping) {
1311 unlock_page(page);
1312 return -EAGAIN;
1313 }
1314 }
1315 return 0;
1316}
1317
1318/*
1319 * this just gets pages into the page cache and locks them down.
1320 */
1321static noinline int prepare_pages(struct inode *inode, struct page **pages,
1322 size_t num_pages, loff_t pos,
1323 size_t write_bytes, bool force_uptodate)
1324{
1325 int i;
1326 unsigned long index = pos >> PAGE_SHIFT;
1327 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1328 int err = 0;
1329 int faili;
1330
1331 for (i = 0; i < num_pages; i++) {
1332again:
1333 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1334 mask | __GFP_WRITE);
1335 if (!pages[i]) {
1336 faili = i - 1;
1337 err = -ENOMEM;
1338 goto fail;
1339 }
1340
1341 if (i == 0)
1342 err = prepare_uptodate_page(inode, pages[i], pos,
1343 force_uptodate);
1344 if (!err && i == num_pages - 1)
1345 err = prepare_uptodate_page(inode, pages[i],
1346 pos + write_bytes, false);
1347 if (err) {
1348 put_page(pages[i]);
1349 if (err == -EAGAIN) {
1350 err = 0;
1351 goto again;
1352 }
1353 faili = i - 1;
1354 goto fail;
1355 }
1356 wait_on_page_writeback(pages[i]);
1357 }
1358
1359 return 0;
1360fail:
1361 while (faili >= 0) {
1362 unlock_page(pages[faili]);
1363 put_page(pages[faili]);
1364 faili--;
1365 }
1366 return err;
1367
1368}
1369
1370/*
1371 * This function locks the extent and properly waits for data=ordered extents
1372 * to finish before allowing the pages to be modified if need.
1373 *
1374 * The return value:
1375 * 1 - the extent is locked
1376 * 0 - the extent is not locked, and everything is OK
1377 * -EAGAIN - need re-prepare the pages
1378 * the other < 0 number - Something wrong happens
1379 */
1380static noinline int
1381lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1382 size_t num_pages, loff_t pos,
1383 size_t write_bytes,
1384 u64 *lockstart, u64 *lockend,
1385 struct extent_state **cached_state)
1386{
1387 struct btrfs_root *root = BTRFS_I(inode)->root;
1388 u64 start_pos;
1389 u64 last_pos;
1390 int i;
1391 int ret = 0;
1392
1393 start_pos = round_down(pos, root->sectorsize);
1394 last_pos = start_pos
1395 + round_up(pos + write_bytes - start_pos, root->sectorsize) - 1;
1396
1397 if (start_pos < inode->i_size) {
1398 struct btrfs_ordered_extent *ordered;
1399 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1400 start_pos, last_pos, cached_state);
1401 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1402 last_pos - start_pos + 1);
1403 if (ordered &&
1404 ordered->file_offset + ordered->len > start_pos &&
1405 ordered->file_offset <= last_pos) {
1406 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1407 start_pos, last_pos,
1408 cached_state, GFP_NOFS);
1409 for (i = 0; i < num_pages; i++) {
1410 unlock_page(pages[i]);
1411 put_page(pages[i]);
1412 }
1413 btrfs_start_ordered_extent(inode, ordered, 1);
1414 btrfs_put_ordered_extent(ordered);
1415 return -EAGAIN;
1416 }
1417 if (ordered)
1418 btrfs_put_ordered_extent(ordered);
1419
1420 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1421 last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1422 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1423 0, 0, cached_state, GFP_NOFS);
1424 *lockstart = start_pos;
1425 *lockend = last_pos;
1426 ret = 1;
1427 }
1428
1429 for (i = 0; i < num_pages; i++) {
1430 if (clear_page_dirty_for_io(pages[i]))
1431 account_page_redirty(pages[i]);
1432 set_page_extent_mapped(pages[i]);
1433 WARN_ON(!PageLocked(pages[i]));
1434 }
1435
1436 return ret;
1437}
1438
1439static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1440 size_t *write_bytes)
1441{
1442 struct btrfs_root *root = BTRFS_I(inode)->root;
1443 struct btrfs_ordered_extent *ordered;
1444 u64 lockstart, lockend;
1445 u64 num_bytes;
1446 int ret;
1447
1448 ret = btrfs_start_write_no_snapshoting(root);
1449 if (!ret)
1450 return -ENOSPC;
1451
1452 lockstart = round_down(pos, root->sectorsize);
1453 lockend = round_up(pos + *write_bytes, root->sectorsize) - 1;
1454
1455 while (1) {
1456 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1457 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1458 lockend - lockstart + 1);
1459 if (!ordered) {
1460 break;
1461 }
1462 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1463 btrfs_start_ordered_extent(inode, ordered, 1);
1464 btrfs_put_ordered_extent(ordered);
1465 }
1466
1467 num_bytes = lockend - lockstart + 1;
1468 ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1469 if (ret <= 0) {
1470 ret = 0;
1471 btrfs_end_write_no_snapshoting(root);
1472 } else {
1473 *write_bytes = min_t(size_t, *write_bytes ,
1474 num_bytes - pos + lockstart);
1475 }
1476
1477 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1478
1479 return ret;
1480}
1481
1482static noinline ssize_t __btrfs_buffered_write(struct file *file,
1483 struct iov_iter *i,
1484 loff_t pos)
1485{
1486 struct inode *inode = file_inode(file);
1487 struct btrfs_root *root = BTRFS_I(inode)->root;
1488 struct page **pages = NULL;
1489 struct extent_state *cached_state = NULL;
1490 u64 release_bytes = 0;
1491 u64 lockstart;
1492 u64 lockend;
1493 size_t num_written = 0;
1494 int nrptrs;
1495 int ret = 0;
1496 bool only_release_metadata = false;
1497 bool force_page_uptodate = false;
1498 bool need_unlock;
1499
1500 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1501 PAGE_SIZE / (sizeof(struct page *)));
1502 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1503 nrptrs = max(nrptrs, 8);
1504 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1505 if (!pages)
1506 return -ENOMEM;
1507
1508 while (iov_iter_count(i) > 0) {
1509 size_t offset = pos & (PAGE_SIZE - 1);
1510 size_t sector_offset;
1511 size_t write_bytes = min(iov_iter_count(i),
1512 nrptrs * (size_t)PAGE_SIZE -
1513 offset);
1514 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1515 PAGE_SIZE);
1516 size_t reserve_bytes;
1517 size_t dirty_pages;
1518 size_t copied;
1519 size_t dirty_sectors;
1520 size_t num_sectors;
1521
1522 WARN_ON(num_pages > nrptrs);
1523
1524 /*
1525 * Fault pages before locking them in prepare_pages
1526 * to avoid recursive lock
1527 */
1528 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1529 ret = -EFAULT;
1530 break;
1531 }
1532
1533 sector_offset = pos & (root->sectorsize - 1);
1534 reserve_bytes = round_up(write_bytes + sector_offset,
1535 root->sectorsize);
1536
1537 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1538 BTRFS_INODE_PREALLOC)) &&
1539 check_can_nocow(inode, pos, &write_bytes) > 0) {
1540 /*
1541 * For nodata cow case, no need to reserve
1542 * data space.
1543 */
1544 only_release_metadata = true;
1545 /*
1546 * our prealloc extent may be smaller than
1547 * write_bytes, so scale down.
1548 */
1549 num_pages = DIV_ROUND_UP(write_bytes + offset,
1550 PAGE_SIZE);
1551 reserve_bytes = round_up(write_bytes + sector_offset,
1552 root->sectorsize);
1553 goto reserve_metadata;
1554 }
1555
1556 ret = btrfs_check_data_free_space(inode, pos, write_bytes);
1557 if (ret < 0)
1558 break;
1559
1560reserve_metadata:
1561 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1562 if (ret) {
1563 if (!only_release_metadata)
1564 btrfs_free_reserved_data_space(inode, pos,
1565 write_bytes);
1566 else
1567 btrfs_end_write_no_snapshoting(root);
1568 break;
1569 }
1570
1571 release_bytes = reserve_bytes;
1572 need_unlock = false;
1573again:
1574 /*
1575 * This is going to setup the pages array with the number of
1576 * pages we want, so we don't really need to worry about the
1577 * contents of pages from loop to loop
1578 */
1579 ret = prepare_pages(inode, pages, num_pages,
1580 pos, write_bytes,
1581 force_page_uptodate);
1582 if (ret)
1583 break;
1584
1585 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1586 pos, write_bytes, &lockstart,
1587 &lockend, &cached_state);
1588 if (ret < 0) {
1589 if (ret == -EAGAIN)
1590 goto again;
1591 break;
1592 } else if (ret > 0) {
1593 need_unlock = true;
1594 ret = 0;
1595 }
1596
1597 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1598
1599 /*
1600 * if we have trouble faulting in the pages, fall
1601 * back to one page at a time
1602 */
1603 if (copied < write_bytes)
1604 nrptrs = 1;
1605
1606 if (copied == 0) {
1607 force_page_uptodate = true;
1608 dirty_pages = 0;
1609 } else {
1610 force_page_uptodate = false;
1611 dirty_pages = DIV_ROUND_UP(copied + offset,
1612 PAGE_SIZE);
1613 }
1614
1615 /*
1616 * If we had a short copy we need to release the excess delaloc
1617 * bytes we reserved. We need to increment outstanding_extents
1618 * because btrfs_delalloc_release_space will decrement it, but
1619 * we still have an outstanding extent for the chunk we actually
1620 * managed to copy.
1621 */
1622 num_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
1623 reserve_bytes);
1624 dirty_sectors = round_up(copied + sector_offset,
1625 root->sectorsize);
1626 dirty_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
1627 dirty_sectors);
1628
1629 if (num_sectors > dirty_sectors) {
1630 release_bytes = (write_bytes - copied)
1631 & ~((u64)root->sectorsize - 1);
1632 if (copied > 0) {
1633 spin_lock(&BTRFS_I(inode)->lock);
1634 BTRFS_I(inode)->outstanding_extents++;
1635 spin_unlock(&BTRFS_I(inode)->lock);
1636 }
1637 if (only_release_metadata) {
1638 btrfs_delalloc_release_metadata(inode,
1639 release_bytes);
1640 } else {
1641 u64 __pos;
1642
1643 __pos = round_down(pos, root->sectorsize) +
1644 (dirty_pages << PAGE_SHIFT);
1645 btrfs_delalloc_release_space(inode, __pos,
1646 release_bytes);
1647 }
1648 }
1649
1650 release_bytes = round_up(copied + sector_offset,
1651 root->sectorsize);
1652
1653 if (copied > 0)
1654 ret = btrfs_dirty_pages(root, inode, pages,
1655 dirty_pages, pos, copied,
1656 NULL);
1657 if (need_unlock)
1658 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1659 lockstart, lockend, &cached_state,
1660 GFP_NOFS);
1661 if (ret) {
1662 btrfs_drop_pages(pages, num_pages);
1663 break;
1664 }
1665
1666 release_bytes = 0;
1667 if (only_release_metadata)
1668 btrfs_end_write_no_snapshoting(root);
1669
1670 if (only_release_metadata && copied > 0) {
1671 lockstart = round_down(pos, root->sectorsize);
1672 lockend = round_up(pos + copied, root->sectorsize) - 1;
1673
1674 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1675 lockend, EXTENT_NORESERVE, NULL,
1676 NULL, GFP_NOFS);
1677 only_release_metadata = false;
1678 }
1679
1680 btrfs_drop_pages(pages, num_pages);
1681
1682 cond_resched();
1683
1684 balance_dirty_pages_ratelimited(inode->i_mapping);
1685 if (dirty_pages < (root->nodesize >> PAGE_SHIFT) + 1)
1686 btrfs_btree_balance_dirty(root);
1687
1688 pos += copied;
1689 num_written += copied;
1690 }
1691
1692 kfree(pages);
1693
1694 if (release_bytes) {
1695 if (only_release_metadata) {
1696 btrfs_end_write_no_snapshoting(root);
1697 btrfs_delalloc_release_metadata(inode, release_bytes);
1698 } else {
1699 btrfs_delalloc_release_space(inode, pos, release_bytes);
1700 }
1701 }
1702
1703 return num_written ? num_written : ret;
1704}
1705
1706static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1707 struct iov_iter *from,
1708 loff_t pos)
1709{
1710 struct file *file = iocb->ki_filp;
1711 struct inode *inode = file_inode(file);
1712 ssize_t written;
1713 ssize_t written_buffered;
1714 loff_t endbyte;
1715 int err;
1716
1717 written = generic_file_direct_write(iocb, from, pos);
1718
1719 if (written < 0 || !iov_iter_count(from))
1720 return written;
1721
1722 pos += written;
1723 written_buffered = __btrfs_buffered_write(file, from, pos);
1724 if (written_buffered < 0) {
1725 err = written_buffered;
1726 goto out;
1727 }
1728 /*
1729 * Ensure all data is persisted. We want the next direct IO read to be
1730 * able to read what was just written.
1731 */
1732 endbyte = pos + written_buffered - 1;
1733 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1734 if (err)
1735 goto out;
1736 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1737 if (err)
1738 goto out;
1739 written += written_buffered;
1740 iocb->ki_pos = pos + written_buffered;
1741 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1742 endbyte >> PAGE_SHIFT);
1743out:
1744 return written ? written : err;
1745}
1746
1747static void update_time_for_write(struct inode *inode)
1748{
1749 struct timespec now;
1750
1751 if (IS_NOCMTIME(inode))
1752 return;
1753
1754 now = current_fs_time(inode->i_sb);
1755 if (!timespec_equal(&inode->i_mtime, &now))
1756 inode->i_mtime = now;
1757
1758 if (!timespec_equal(&inode->i_ctime, &now))
1759 inode->i_ctime = now;
1760
1761 if (IS_I_VERSION(inode))
1762 inode_inc_iversion(inode);
1763}
1764
1765static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1766 struct iov_iter *from)
1767{
1768 struct file *file = iocb->ki_filp;
1769 struct inode *inode = file_inode(file);
1770 struct btrfs_root *root = BTRFS_I(inode)->root;
1771 u64 start_pos;
1772 u64 end_pos;
1773 ssize_t num_written = 0;
1774 bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1775 ssize_t err;
1776 loff_t pos;
1777 size_t count;
1778 loff_t oldsize;
1779 int clean_page = 0;
1780
1781 inode_lock(inode);
1782 err = generic_write_checks(iocb, from);
1783 if (err <= 0) {
1784 inode_unlock(inode);
1785 return err;
1786 }
1787
1788 current->backing_dev_info = inode_to_bdi(inode);
1789 err = file_remove_privs(file);
1790 if (err) {
1791 inode_unlock(inode);
1792 goto out;
1793 }
1794
1795 /*
1796 * If BTRFS flips readonly due to some impossible error
1797 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1798 * although we have opened a file as writable, we have
1799 * to stop this write operation to ensure FS consistency.
1800 */
1801 if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1802 inode_unlock(inode);
1803 err = -EROFS;
1804 goto out;
1805 }
1806
1807 /*
1808 * We reserve space for updating the inode when we reserve space for the
1809 * extent we are going to write, so we will enospc out there. We don't
1810 * need to start yet another transaction to update the inode as we will
1811 * update the inode when we finish writing whatever data we write.
1812 */
1813 update_time_for_write(inode);
1814
1815 pos = iocb->ki_pos;
1816 count = iov_iter_count(from);
1817 start_pos = round_down(pos, root->sectorsize);
1818 oldsize = i_size_read(inode);
1819 if (start_pos > oldsize) {
1820 /* Expand hole size to cover write data, preventing empty gap */
1821 end_pos = round_up(pos + count, root->sectorsize);
1822 err = btrfs_cont_expand(inode, oldsize, end_pos);
1823 if (err) {
1824 inode_unlock(inode);
1825 goto out;
1826 }
1827 if (start_pos > round_up(oldsize, root->sectorsize))
1828 clean_page = 1;
1829 }
1830
1831 if (sync)
1832 atomic_inc(&BTRFS_I(inode)->sync_writers);
1833
1834 if (iocb->ki_flags & IOCB_DIRECT) {
1835 num_written = __btrfs_direct_write(iocb, from, pos);
1836 } else {
1837 num_written = __btrfs_buffered_write(file, from, pos);
1838 if (num_written > 0)
1839 iocb->ki_pos = pos + num_written;
1840 if (clean_page)
1841 pagecache_isize_extended(inode, oldsize,
1842 i_size_read(inode));
1843 }
1844
1845 inode_unlock(inode);
1846
1847 /*
1848 * We also have to set last_sub_trans to the current log transid,
1849 * otherwise subsequent syncs to a file that's been synced in this
1850 * transaction will appear to have already occurred.
1851 */
1852 spin_lock(&BTRFS_I(inode)->lock);
1853 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1854 spin_unlock(&BTRFS_I(inode)->lock);
1855 if (num_written > 0) {
1856 err = generic_write_sync(file, pos, num_written);
1857 if (err < 0)
1858 num_written = err;
1859 }
1860
1861 if (sync)
1862 atomic_dec(&BTRFS_I(inode)->sync_writers);
1863out:
1864 current->backing_dev_info = NULL;
1865 return num_written ? num_written : err;
1866}
1867
1868int btrfs_release_file(struct inode *inode, struct file *filp)
1869{
1870 if (filp->private_data)
1871 btrfs_ioctl_trans_end(filp);
1872 /*
1873 * ordered_data_close is set by settattr when we are about to truncate
1874 * a file from a non-zero size to a zero size. This tries to
1875 * flush down new bytes that may have been written if the
1876 * application were using truncate to replace a file in place.
1877 */
1878 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1879 &BTRFS_I(inode)->runtime_flags))
1880 filemap_flush(inode->i_mapping);
1881 return 0;
1882}
1883
1884static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1885{
1886 int ret;
1887
1888 atomic_inc(&BTRFS_I(inode)->sync_writers);
1889 ret = btrfs_fdatawrite_range(inode, start, end);
1890 atomic_dec(&BTRFS_I(inode)->sync_writers);
1891
1892 return ret;
1893}
1894
1895/*
1896 * fsync call for both files and directories. This logs the inode into
1897 * the tree log instead of forcing full commits whenever possible.
1898 *
1899 * It needs to call filemap_fdatawait so that all ordered extent updates are
1900 * in the metadata btree are up to date for copying to the log.
1901 *
1902 * It drops the inode mutex before doing the tree log commit. This is an
1903 * important optimization for directories because holding the mutex prevents
1904 * new operations on the dir while we write to disk.
1905 */
1906int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1907{
1908 struct dentry *dentry = file_dentry(file);
1909 struct inode *inode = d_inode(dentry);
1910 struct btrfs_root *root = BTRFS_I(inode)->root;
1911 struct btrfs_trans_handle *trans;
1912 struct btrfs_log_ctx ctx;
1913 int ret = 0;
1914 bool full_sync = 0;
1915 u64 len;
1916
1917 /*
1918 * The range length can be represented by u64, we have to do the typecasts
1919 * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
1920 */
1921 len = (u64)end - (u64)start + 1;
1922 trace_btrfs_sync_file(file, datasync);
1923
1924 /*
1925 * We write the dirty pages in the range and wait until they complete
1926 * out of the ->i_mutex. If so, we can flush the dirty pages by
1927 * multi-task, and make the performance up. See
1928 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1929 */
1930 ret = start_ordered_ops(inode, start, end);
1931 if (ret)
1932 return ret;
1933
1934 inode_lock(inode);
1935 atomic_inc(&root->log_batch);
1936 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1937 &BTRFS_I(inode)->runtime_flags);
1938 /*
1939 * We might have have had more pages made dirty after calling
1940 * start_ordered_ops and before acquiring the inode's i_mutex.
1941 */
1942 if (full_sync) {
1943 /*
1944 * For a full sync, we need to make sure any ordered operations
1945 * start and finish before we start logging the inode, so that
1946 * all extents are persisted and the respective file extent
1947 * items are in the fs/subvol btree.
1948 */
1949 ret = btrfs_wait_ordered_range(inode, start, len);
1950 } else {
1951 /*
1952 * Start any new ordered operations before starting to log the
1953 * inode. We will wait for them to finish in btrfs_sync_log().
1954 *
1955 * Right before acquiring the inode's mutex, we might have new
1956 * writes dirtying pages, which won't immediately start the
1957 * respective ordered operations - that is done through the
1958 * fill_delalloc callbacks invoked from the writepage and
1959 * writepages address space operations. So make sure we start
1960 * all ordered operations before starting to log our inode. Not
1961 * doing this means that while logging the inode, writeback
1962 * could start and invoke writepage/writepages, which would call
1963 * the fill_delalloc callbacks (cow_file_range,
1964 * submit_compressed_extents). These callbacks add first an
1965 * extent map to the modified list of extents and then create
1966 * the respective ordered operation, which means in
1967 * tree-log.c:btrfs_log_inode() we might capture all existing
1968 * ordered operations (with btrfs_get_logged_extents()) before
1969 * the fill_delalloc callback adds its ordered operation, and by
1970 * the time we visit the modified list of extent maps (with
1971 * btrfs_log_changed_extents()), we see and process the extent
1972 * map they created. We then use the extent map to construct a
1973 * file extent item for logging without waiting for the
1974 * respective ordered operation to finish - this file extent
1975 * item points to a disk location that might not have yet been
1976 * written to, containing random data - so after a crash a log
1977 * replay will make our inode have file extent items that point
1978 * to disk locations containing invalid data, as we returned
1979 * success to userspace without waiting for the respective
1980 * ordered operation to finish, because it wasn't captured by
1981 * btrfs_get_logged_extents().
1982 */
1983 ret = start_ordered_ops(inode, start, end);
1984 }
1985 if (ret) {
1986 inode_unlock(inode);
1987 goto out;
1988 }
1989 atomic_inc(&root->log_batch);
1990
1991 /*
1992 * If the last transaction that changed this file was before the current
1993 * transaction and we have the full sync flag set in our inode, we can
1994 * bail out now without any syncing.
1995 *
1996 * Note that we can't bail out if the full sync flag isn't set. This is
1997 * because when the full sync flag is set we start all ordered extents
1998 * and wait for them to fully complete - when they complete they update
1999 * the inode's last_trans field through:
2000 *
2001 * btrfs_finish_ordered_io() ->
2002 * btrfs_update_inode_fallback() ->
2003 * btrfs_update_inode() ->
2004 * btrfs_set_inode_last_trans()
2005 *
2006 * So we are sure that last_trans is up to date and can do this check to
2007 * bail out safely. For the fast path, when the full sync flag is not
2008 * set in our inode, we can not do it because we start only our ordered
2009 * extents and don't wait for them to complete (that is when
2010 * btrfs_finish_ordered_io runs), so here at this point their last_trans
2011 * value might be less than or equals to fs_info->last_trans_committed,
2012 * and setting a speculative last_trans for an inode when a buffered
2013 * write is made (such as fs_info->generation + 1 for example) would not
2014 * be reliable since after setting the value and before fsync is called
2015 * any number of transactions can start and commit (transaction kthread
2016 * commits the current transaction periodically), and a transaction
2017 * commit does not start nor waits for ordered extents to complete.
2018 */
2019 smp_mb();
2020 if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
2021 (full_sync && BTRFS_I(inode)->last_trans <=
2022 root->fs_info->last_trans_committed) ||
2023 (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2024 BTRFS_I(inode)->last_trans
2025 <= root->fs_info->last_trans_committed)) {
2026 /*
2027 * We'v had everything committed since the last time we were
2028 * modified so clear this flag in case it was set for whatever
2029 * reason, it's no longer relevant.
2030 */
2031 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2032 &BTRFS_I(inode)->runtime_flags);
2033 inode_unlock(inode);
2034 goto out;
2035 }
2036
2037 /*
2038 * ok we haven't committed the transaction yet, lets do a commit
2039 */
2040 if (file->private_data)
2041 btrfs_ioctl_trans_end(file);
2042
2043 /*
2044 * We use start here because we will need to wait on the IO to complete
2045 * in btrfs_sync_log, which could require joining a transaction (for
2046 * example checking cross references in the nocow path). If we use join
2047 * here we could get into a situation where we're waiting on IO to
2048 * happen that is blocked on a transaction trying to commit. With start
2049 * we inc the extwriter counter, so we wait for all extwriters to exit
2050 * before we start blocking join'ers. This comment is to keep somebody
2051 * from thinking they are super smart and changing this to
2052 * btrfs_join_transaction *cough*Josef*cough*.
2053 */
2054 trans = btrfs_start_transaction(root, 0);
2055 if (IS_ERR(trans)) {
2056 ret = PTR_ERR(trans);
2057 inode_unlock(inode);
2058 goto out;
2059 }
2060 trans->sync = true;
2061
2062 btrfs_init_log_ctx(&ctx);
2063
2064 ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2065 if (ret < 0) {
2066 /* Fallthrough and commit/free transaction. */
2067 ret = 1;
2068 }
2069
2070 /* we've logged all the items and now have a consistent
2071 * version of the file in the log. It is possible that
2072 * someone will come in and modify the file, but that's
2073 * fine because the log is consistent on disk, and we
2074 * have references to all of the file's extents
2075 *
2076 * It is possible that someone will come in and log the
2077 * file again, but that will end up using the synchronization
2078 * inside btrfs_sync_log to keep things safe.
2079 */
2080 inode_unlock(inode);
2081
2082 /*
2083 * If any of the ordered extents had an error, just return it to user
2084 * space, so that the application knows some writes didn't succeed and
2085 * can take proper action (retry for e.g.). Blindly committing the
2086 * transaction in this case, would fool userspace that everything was
2087 * successful. And we also want to make sure our log doesn't contain
2088 * file extent items pointing to extents that weren't fully written to -
2089 * just like in the non fast fsync path, where we check for the ordered
2090 * operation's error flag before writing to the log tree and return -EIO
2091 * if any of them had this flag set (btrfs_wait_ordered_range) -
2092 * therefore we need to check for errors in the ordered operations,
2093 * which are indicated by ctx.io_err.
2094 */
2095 if (ctx.io_err) {
2096 btrfs_end_transaction(trans, root);
2097 ret = ctx.io_err;
2098 goto out;
2099 }
2100
2101 if (ret != BTRFS_NO_LOG_SYNC) {
2102 if (!ret) {
2103 ret = btrfs_sync_log(trans, root, &ctx);
2104 if (!ret) {
2105 ret = btrfs_end_transaction(trans, root);
2106 goto out;
2107 }
2108 }
2109 if (!full_sync) {
2110 ret = btrfs_wait_ordered_range(inode, start, len);
2111 if (ret) {
2112 btrfs_end_transaction(trans, root);
2113 goto out;
2114 }
2115 }
2116 ret = btrfs_commit_transaction(trans, root);
2117 } else {
2118 ret = btrfs_end_transaction(trans, root);
2119 }
2120out:
2121 return ret > 0 ? -EIO : ret;
2122}
2123
2124static const struct vm_operations_struct btrfs_file_vm_ops = {
2125 .fault = filemap_fault,
2126 .map_pages = filemap_map_pages,
2127 .page_mkwrite = btrfs_page_mkwrite,
2128};
2129
2130static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2131{
2132 struct address_space *mapping = filp->f_mapping;
2133
2134 if (!mapping->a_ops->readpage)
2135 return -ENOEXEC;
2136
2137 file_accessed(filp);
2138 vma->vm_ops = &btrfs_file_vm_ops;
2139
2140 return 0;
2141}
2142
2143static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2144 int slot, u64 start, u64 end)
2145{
2146 struct btrfs_file_extent_item *fi;
2147 struct btrfs_key key;
2148
2149 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2150 return 0;
2151
2152 btrfs_item_key_to_cpu(leaf, &key, slot);
2153 if (key.objectid != btrfs_ino(inode) ||
2154 key.type != BTRFS_EXTENT_DATA_KEY)
2155 return 0;
2156
2157 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2158
2159 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2160 return 0;
2161
2162 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2163 return 0;
2164
2165 if (key.offset == end)
2166 return 1;
2167 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2168 return 1;
2169 return 0;
2170}
2171
2172static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2173 struct btrfs_path *path, u64 offset, u64 end)
2174{
2175 struct btrfs_root *root = BTRFS_I(inode)->root;
2176 struct extent_buffer *leaf;
2177 struct btrfs_file_extent_item *fi;
2178 struct extent_map *hole_em;
2179 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2180 struct btrfs_key key;
2181 int ret;
2182
2183 if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
2184 goto out;
2185
2186 key.objectid = btrfs_ino(inode);
2187 key.type = BTRFS_EXTENT_DATA_KEY;
2188 key.offset = offset;
2189
2190 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2191 if (ret < 0)
2192 return ret;
2193 BUG_ON(!ret);
2194
2195 leaf = path->nodes[0];
2196 if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2197 u64 num_bytes;
2198
2199 path->slots[0]--;
2200 fi = btrfs_item_ptr(leaf, path->slots[0],
2201 struct btrfs_file_extent_item);
2202 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2203 end - offset;
2204 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2205 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2206 btrfs_set_file_extent_offset(leaf, fi, 0);
2207 btrfs_mark_buffer_dirty(leaf);
2208 goto out;
2209 }
2210
2211 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2212 u64 num_bytes;
2213
2214 key.offset = offset;
2215 btrfs_set_item_key_safe(root->fs_info, path, &key);
2216 fi = btrfs_item_ptr(leaf, path->slots[0],
2217 struct btrfs_file_extent_item);
2218 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2219 offset;
2220 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2221 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2222 btrfs_set_file_extent_offset(leaf, fi, 0);
2223 btrfs_mark_buffer_dirty(leaf);
2224 goto out;
2225 }
2226 btrfs_release_path(path);
2227
2228 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2229 0, 0, end - offset, 0, end - offset,
2230 0, 0, 0);
2231 if (ret)
2232 return ret;
2233
2234out:
2235 btrfs_release_path(path);
2236
2237 hole_em = alloc_extent_map();
2238 if (!hole_em) {
2239 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2240 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2241 &BTRFS_I(inode)->runtime_flags);
2242 } else {
2243 hole_em->start = offset;
2244 hole_em->len = end - offset;
2245 hole_em->ram_bytes = hole_em->len;
2246 hole_em->orig_start = offset;
2247
2248 hole_em->block_start = EXTENT_MAP_HOLE;
2249 hole_em->block_len = 0;
2250 hole_em->orig_block_len = 0;
2251 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2252 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2253 hole_em->generation = trans->transid;
2254
2255 do {
2256 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2257 write_lock(&em_tree->lock);
2258 ret = add_extent_mapping(em_tree, hole_em, 1);
2259 write_unlock(&em_tree->lock);
2260 } while (ret == -EEXIST);
2261 free_extent_map(hole_em);
2262 if (ret)
2263 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2264 &BTRFS_I(inode)->runtime_flags);
2265 }
2266
2267 return 0;
2268}
2269
2270/*
2271 * Find a hole extent on given inode and change start/len to the end of hole
2272 * extent.(hole/vacuum extent whose em->start <= start &&
2273 * em->start + em->len > start)
2274 * When a hole extent is found, return 1 and modify start/len.
2275 */
2276static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2277{
2278 struct extent_map *em;
2279 int ret = 0;
2280
2281 em = btrfs_get_extent(inode, NULL, 0, *start, *len, 0);
2282 if (IS_ERR_OR_NULL(em)) {
2283 if (!em)
2284 ret = -ENOMEM;
2285 else
2286 ret = PTR_ERR(em);
2287 return ret;
2288 }
2289
2290 /* Hole or vacuum extent(only exists in no-hole mode) */
2291 if (em->block_start == EXTENT_MAP_HOLE) {
2292 ret = 1;
2293 *len = em->start + em->len > *start + *len ?
2294 0 : *start + *len - em->start - em->len;
2295 *start = em->start + em->len;
2296 }
2297 free_extent_map(em);
2298 return ret;
2299}
2300
2301static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2302{
2303 struct btrfs_root *root = BTRFS_I(inode)->root;
2304 struct extent_state *cached_state = NULL;
2305 struct btrfs_path *path;
2306 struct btrfs_block_rsv *rsv;
2307 struct btrfs_trans_handle *trans;
2308 u64 lockstart;
2309 u64 lockend;
2310 u64 tail_start;
2311 u64 tail_len;
2312 u64 orig_start = offset;
2313 u64 cur_offset;
2314 u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2315 u64 drop_end;
2316 int ret = 0;
2317 int err = 0;
2318 unsigned int rsv_count;
2319 bool same_block;
2320 bool no_holes = btrfs_fs_incompat(root->fs_info, NO_HOLES);
2321 u64 ino_size;
2322 bool truncated_block = false;
2323 bool updated_inode = false;
2324
2325 ret = btrfs_wait_ordered_range(inode, offset, len);
2326 if (ret)
2327 return ret;
2328
2329 inode_lock(inode);
2330 ino_size = round_up(inode->i_size, root->sectorsize);
2331 ret = find_first_non_hole(inode, &offset, &len);
2332 if (ret < 0)
2333 goto out_only_mutex;
2334 if (ret && !len) {
2335 /* Already in a large hole */
2336 ret = 0;
2337 goto out_only_mutex;
2338 }
2339
2340 lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
2341 lockend = round_down(offset + len,
2342 BTRFS_I(inode)->root->sectorsize) - 1;
2343 same_block = (BTRFS_BYTES_TO_BLKS(root->fs_info, offset))
2344 == (BTRFS_BYTES_TO_BLKS(root->fs_info, offset + len - 1));
2345 /*
2346 * We needn't truncate any block which is beyond the end of the file
2347 * because we are sure there is no data there.
2348 */
2349 /*
2350 * Only do this if we are in the same block and we aren't doing the
2351 * entire block.
2352 */
2353 if (same_block && len < root->sectorsize) {
2354 if (offset < ino_size) {
2355 truncated_block = true;
2356 ret = btrfs_truncate_block(inode, offset, len, 0);
2357 } else {
2358 ret = 0;
2359 }
2360 goto out_only_mutex;
2361 }
2362
2363 /* zero back part of the first block */
2364 if (offset < ino_size) {
2365 truncated_block = true;
2366 ret = btrfs_truncate_block(inode, offset, 0, 0);
2367 if (ret) {
2368 inode_unlock(inode);
2369 return ret;
2370 }
2371 }
2372
2373 /* Check the aligned pages after the first unaligned page,
2374 * if offset != orig_start, which means the first unaligned page
2375 * including serveral following pages are already in holes,
2376 * the extra check can be skipped */
2377 if (offset == orig_start) {
2378 /* after truncate page, check hole again */
2379 len = offset + len - lockstart;
2380 offset = lockstart;
2381 ret = find_first_non_hole(inode, &offset, &len);
2382 if (ret < 0)
2383 goto out_only_mutex;
2384 if (ret && !len) {
2385 ret = 0;
2386 goto out_only_mutex;
2387 }
2388 lockstart = offset;
2389 }
2390
2391 /* Check the tail unaligned part is in a hole */
2392 tail_start = lockend + 1;
2393 tail_len = offset + len - tail_start;
2394 if (tail_len) {
2395 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2396 if (unlikely(ret < 0))
2397 goto out_only_mutex;
2398 if (!ret) {
2399 /* zero the front end of the last page */
2400 if (tail_start + tail_len < ino_size) {
2401 truncated_block = true;
2402 ret = btrfs_truncate_block(inode,
2403 tail_start + tail_len,
2404 0, 1);
2405 if (ret)
2406 goto out_only_mutex;
2407 }
2408 }
2409 }
2410
2411 if (lockend < lockstart) {
2412 ret = 0;
2413 goto out_only_mutex;
2414 }
2415
2416 while (1) {
2417 struct btrfs_ordered_extent *ordered;
2418
2419 truncate_pagecache_range(inode, lockstart, lockend);
2420
2421 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2422 &cached_state);
2423 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2424
2425 /*
2426 * We need to make sure we have no ordered extents in this range
2427 * and nobody raced in and read a page in this range, if we did
2428 * we need to try again.
2429 */
2430 if ((!ordered ||
2431 (ordered->file_offset + ordered->len <= lockstart ||
2432 ordered->file_offset > lockend)) &&
2433 !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2434 if (ordered)
2435 btrfs_put_ordered_extent(ordered);
2436 break;
2437 }
2438 if (ordered)
2439 btrfs_put_ordered_extent(ordered);
2440 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2441 lockend, &cached_state, GFP_NOFS);
2442 ret = btrfs_wait_ordered_range(inode, lockstart,
2443 lockend - lockstart + 1);
2444 if (ret) {
2445 inode_unlock(inode);
2446 return ret;
2447 }
2448 }
2449
2450 path = btrfs_alloc_path();
2451 if (!path) {
2452 ret = -ENOMEM;
2453 goto out;
2454 }
2455
2456 rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2457 if (!rsv) {
2458 ret = -ENOMEM;
2459 goto out_free;
2460 }
2461 rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2462 rsv->failfast = 1;
2463
2464 /*
2465 * 1 - update the inode
2466 * 1 - removing the extents in the range
2467 * 1 - adding the hole extent if no_holes isn't set
2468 */
2469 rsv_count = no_holes ? 2 : 3;
2470 trans = btrfs_start_transaction(root, rsv_count);
2471 if (IS_ERR(trans)) {
2472 err = PTR_ERR(trans);
2473 goto out_free;
2474 }
2475
2476 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2477 min_size);
2478 BUG_ON(ret);
2479 trans->block_rsv = rsv;
2480
2481 cur_offset = lockstart;
2482 len = lockend - cur_offset;
2483 while (cur_offset < lockend) {
2484 ret = __btrfs_drop_extents(trans, root, inode, path,
2485 cur_offset, lockend + 1,
2486 &drop_end, 1, 0, 0, NULL);
2487 if (ret != -ENOSPC)
2488 break;
2489
2490 trans->block_rsv = &root->fs_info->trans_block_rsv;
2491
2492 if (cur_offset < ino_size) {
2493 ret = fill_holes(trans, inode, path, cur_offset,
2494 drop_end);
2495 if (ret) {
2496 err = ret;
2497 break;
2498 }
2499 }
2500
2501 cur_offset = drop_end;
2502
2503 ret = btrfs_update_inode(trans, root, inode);
2504 if (ret) {
2505 err = ret;
2506 break;
2507 }
2508
2509 btrfs_end_transaction(trans, root);
2510 btrfs_btree_balance_dirty(root);
2511
2512 trans = btrfs_start_transaction(root, rsv_count);
2513 if (IS_ERR(trans)) {
2514 ret = PTR_ERR(trans);
2515 trans = NULL;
2516 break;
2517 }
2518
2519 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2520 rsv, min_size);
2521 BUG_ON(ret); /* shouldn't happen */
2522 trans->block_rsv = rsv;
2523
2524 ret = find_first_non_hole(inode, &cur_offset, &len);
2525 if (unlikely(ret < 0))
2526 break;
2527 if (ret && !len) {
2528 ret = 0;
2529 break;
2530 }
2531 }
2532
2533 if (ret) {
2534 err = ret;
2535 goto out_trans;
2536 }
2537
2538 trans->block_rsv = &root->fs_info->trans_block_rsv;
2539 /*
2540 * If we are using the NO_HOLES feature we might have had already an
2541 * hole that overlaps a part of the region [lockstart, lockend] and
2542 * ends at (or beyond) lockend. Since we have no file extent items to
2543 * represent holes, drop_end can be less than lockend and so we must
2544 * make sure we have an extent map representing the existing hole (the
2545 * call to __btrfs_drop_extents() might have dropped the existing extent
2546 * map representing the existing hole), otherwise the fast fsync path
2547 * will not record the existence of the hole region
2548 * [existing_hole_start, lockend].
2549 */
2550 if (drop_end <= lockend)
2551 drop_end = lockend + 1;
2552 /*
2553 * Don't insert file hole extent item if it's for a range beyond eof
2554 * (because it's useless) or if it represents a 0 bytes range (when
2555 * cur_offset == drop_end).
2556 */
2557 if (cur_offset < ino_size && cur_offset < drop_end) {
2558 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2559 if (ret) {
2560 err = ret;
2561 goto out_trans;
2562 }
2563 }
2564
2565out_trans:
2566 if (!trans)
2567 goto out_free;
2568
2569 inode_inc_iversion(inode);
2570 inode->i_mtime = inode->i_ctime = current_fs_time(inode->i_sb);
2571
2572 trans->block_rsv = &root->fs_info->trans_block_rsv;
2573 ret = btrfs_update_inode(trans, root, inode);
2574 updated_inode = true;
2575 btrfs_end_transaction(trans, root);
2576 btrfs_btree_balance_dirty(root);
2577out_free:
2578 btrfs_free_path(path);
2579 btrfs_free_block_rsv(root, rsv);
2580out:
2581 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2582 &cached_state, GFP_NOFS);
2583out_only_mutex:
2584 if (!updated_inode && truncated_block && !ret && !err) {
2585 /*
2586 * If we only end up zeroing part of a page, we still need to
2587 * update the inode item, so that all the time fields are
2588 * updated as well as the necessary btrfs inode in memory fields
2589 * for detecting, at fsync time, if the inode isn't yet in the
2590 * log tree or it's there but not up to date.
2591 */
2592 trans = btrfs_start_transaction(root, 1);
2593 if (IS_ERR(trans)) {
2594 err = PTR_ERR(trans);
2595 } else {
2596 err = btrfs_update_inode(trans, root, inode);
2597 ret = btrfs_end_transaction(trans, root);
2598 }
2599 }
2600 inode_unlock(inode);
2601 if (ret && !err)
2602 err = ret;
2603 return err;
2604}
2605
2606/* Helper structure to record which range is already reserved */
2607struct falloc_range {
2608 struct list_head list;
2609 u64 start;
2610 u64 len;
2611};
2612
2613/*
2614 * Helper function to add falloc range
2615 *
2616 * Caller should have locked the larger range of extent containing
2617 * [start, len)
2618 */
2619static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2620{
2621 struct falloc_range *prev = NULL;
2622 struct falloc_range *range = NULL;
2623
2624 if (list_empty(head))
2625 goto insert;
2626
2627 /*
2628 * As fallocate iterate by bytenr order, we only need to check
2629 * the last range.
2630 */
2631 prev = list_entry(head->prev, struct falloc_range, list);
2632 if (prev->start + prev->len == start) {
2633 prev->len += len;
2634 return 0;
2635 }
2636insert:
2637 range = kmalloc(sizeof(*range), GFP_KERNEL);
2638 if (!range)
2639 return -ENOMEM;
2640 range->start = start;
2641 range->len = len;
2642 list_add_tail(&range->list, head);
2643 return 0;
2644}
2645
2646static long btrfs_fallocate(struct file *file, int mode,
2647 loff_t offset, loff_t len)
2648{
2649 struct inode *inode = file_inode(file);
2650 struct extent_state *cached_state = NULL;
2651 struct falloc_range *range;
2652 struct falloc_range *tmp;
2653 struct list_head reserve_list;
2654 u64 cur_offset;
2655 u64 last_byte;
2656 u64 alloc_start;
2657 u64 alloc_end;
2658 u64 alloc_hint = 0;
2659 u64 locked_end;
2660 u64 actual_end = 0;
2661 struct extent_map *em;
2662 int blocksize = BTRFS_I(inode)->root->sectorsize;
2663 int ret;
2664
2665 alloc_start = round_down(offset, blocksize);
2666 alloc_end = round_up(offset + len, blocksize);
2667
2668 /* Make sure we aren't being give some crap mode */
2669 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2670 return -EOPNOTSUPP;
2671
2672 if (mode & FALLOC_FL_PUNCH_HOLE)
2673 return btrfs_punch_hole(inode, offset, len);
2674
2675 /*
2676 * Only trigger disk allocation, don't trigger qgroup reserve
2677 *
2678 * For qgroup space, it will be checked later.
2679 */
2680 ret = btrfs_alloc_data_chunk_ondemand(inode, alloc_end - alloc_start);
2681 if (ret < 0)
2682 return ret;
2683
2684 inode_lock(inode);
2685
2686 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2687 ret = inode_newsize_ok(inode, offset + len);
2688 if (ret)
2689 goto out;
2690 }
2691
2692 /*
2693 * TODO: Move these two operations after we have checked
2694 * accurate reserved space, or fallocate can still fail but
2695 * with page truncated or size expanded.
2696 *
2697 * But that's a minor problem and won't do much harm BTW.
2698 */
2699 if (alloc_start > inode->i_size) {
2700 ret = btrfs_cont_expand(inode, i_size_read(inode),
2701 alloc_start);
2702 if (ret)
2703 goto out;
2704 } else if (offset + len > inode->i_size) {
2705 /*
2706 * If we are fallocating from the end of the file onward we
2707 * need to zero out the end of the block if i_size lands in the
2708 * middle of a block.
2709 */
2710 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2711 if (ret)
2712 goto out;
2713 }
2714
2715 /*
2716 * wait for ordered IO before we have any locks. We'll loop again
2717 * below with the locks held.
2718 */
2719 ret = btrfs_wait_ordered_range(inode, alloc_start,
2720 alloc_end - alloc_start);
2721 if (ret)
2722 goto out;
2723
2724 locked_end = alloc_end - 1;
2725 while (1) {
2726 struct btrfs_ordered_extent *ordered;
2727
2728 /* the extent lock is ordered inside the running
2729 * transaction
2730 */
2731 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2732 locked_end, &cached_state);
2733 ordered = btrfs_lookup_first_ordered_extent(inode,
2734 alloc_end - 1);
2735 if (ordered &&
2736 ordered->file_offset + ordered->len > alloc_start &&
2737 ordered->file_offset < alloc_end) {
2738 btrfs_put_ordered_extent(ordered);
2739 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2740 alloc_start, locked_end,
2741 &cached_state, GFP_KERNEL);
2742 /*
2743 * we can't wait on the range with the transaction
2744 * running or with the extent lock held
2745 */
2746 ret = btrfs_wait_ordered_range(inode, alloc_start,
2747 alloc_end - alloc_start);
2748 if (ret)
2749 goto out;
2750 } else {
2751 if (ordered)
2752 btrfs_put_ordered_extent(ordered);
2753 break;
2754 }
2755 }
2756
2757 /* First, check if we exceed the qgroup limit */
2758 INIT_LIST_HEAD(&reserve_list);
2759 cur_offset = alloc_start;
2760 while (1) {
2761 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2762 alloc_end - cur_offset, 0);
2763 if (IS_ERR_OR_NULL(em)) {
2764 if (!em)
2765 ret = -ENOMEM;
2766 else
2767 ret = PTR_ERR(em);
2768 break;
2769 }
2770 last_byte = min(extent_map_end(em), alloc_end);
2771 actual_end = min_t(u64, extent_map_end(em), offset + len);
2772 last_byte = ALIGN(last_byte, blocksize);
2773 if (em->block_start == EXTENT_MAP_HOLE ||
2774 (cur_offset >= inode->i_size &&
2775 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2776 ret = add_falloc_range(&reserve_list, cur_offset,
2777 last_byte - cur_offset);
2778 if (ret < 0) {
2779 free_extent_map(em);
2780 break;
2781 }
2782 ret = btrfs_qgroup_reserve_data(inode, cur_offset,
2783 last_byte - cur_offset);
2784 if (ret < 0)
2785 break;
2786 }
2787 free_extent_map(em);
2788 cur_offset = last_byte;
2789 if (cur_offset >= alloc_end)
2790 break;
2791 }
2792
2793 /*
2794 * If ret is still 0, means we're OK to fallocate.
2795 * Or just cleanup the list and exit.
2796 */
2797 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2798 if (!ret)
2799 ret = btrfs_prealloc_file_range(inode, mode,
2800 range->start,
2801 range->len, 1 << inode->i_blkbits,
2802 offset + len, &alloc_hint);
2803 list_del(&range->list);
2804 kfree(range);
2805 }
2806 if (ret < 0)
2807 goto out_unlock;
2808
2809 if (actual_end > inode->i_size &&
2810 !(mode & FALLOC_FL_KEEP_SIZE)) {
2811 struct btrfs_trans_handle *trans;
2812 struct btrfs_root *root = BTRFS_I(inode)->root;
2813
2814 /*
2815 * We didn't need to allocate any more space, but we
2816 * still extended the size of the file so we need to
2817 * update i_size and the inode item.
2818 */
2819 trans = btrfs_start_transaction(root, 1);
2820 if (IS_ERR(trans)) {
2821 ret = PTR_ERR(trans);
2822 } else {
2823 inode->i_ctime = current_fs_time(inode->i_sb);
2824 i_size_write(inode, actual_end);
2825 btrfs_ordered_update_i_size(inode, actual_end, NULL);
2826 ret = btrfs_update_inode(trans, root, inode);
2827 if (ret)
2828 btrfs_end_transaction(trans, root);
2829 else
2830 ret = btrfs_end_transaction(trans, root);
2831 }
2832 }
2833out_unlock:
2834 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2835 &cached_state, GFP_KERNEL);
2836out:
2837 /*
2838 * As we waited the extent range, the data_rsv_map must be empty
2839 * in the range, as written data range will be released from it.
2840 * And for prealloacted extent, it will also be released when
2841 * its metadata is written.
2842 * So this is completely used as cleanup.
2843 */
2844 btrfs_qgroup_free_data(inode, alloc_start, alloc_end - alloc_start);
2845 inode_unlock(inode);
2846 /* Let go of our reservation. */
2847 btrfs_free_reserved_data_space(inode, alloc_start,
2848 alloc_end - alloc_start);
2849 return ret;
2850}
2851
2852static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2853{
2854 struct btrfs_root *root = BTRFS_I(inode)->root;
2855 struct extent_map *em = NULL;
2856 struct extent_state *cached_state = NULL;
2857 u64 lockstart;
2858 u64 lockend;
2859 u64 start;
2860 u64 len;
2861 int ret = 0;
2862
2863 if (inode->i_size == 0)
2864 return -ENXIO;
2865
2866 /*
2867 * *offset can be negative, in this case we start finding DATA/HOLE from
2868 * the very start of the file.
2869 */
2870 start = max_t(loff_t, 0, *offset);
2871
2872 lockstart = round_down(start, root->sectorsize);
2873 lockend = round_up(i_size_read(inode), root->sectorsize);
2874 if (lockend <= lockstart)
2875 lockend = lockstart + root->sectorsize;
2876 lockend--;
2877 len = lockend - lockstart + 1;
2878
2879 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2880 &cached_state);
2881
2882 while (start < inode->i_size) {
2883 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2884 if (IS_ERR(em)) {
2885 ret = PTR_ERR(em);
2886 em = NULL;
2887 break;
2888 }
2889
2890 if (whence == SEEK_HOLE &&
2891 (em->block_start == EXTENT_MAP_HOLE ||
2892 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2893 break;
2894 else if (whence == SEEK_DATA &&
2895 (em->block_start != EXTENT_MAP_HOLE &&
2896 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2897 break;
2898
2899 start = em->start + em->len;
2900 free_extent_map(em);
2901 em = NULL;
2902 cond_resched();
2903 }
2904 free_extent_map(em);
2905 if (!ret) {
2906 if (whence == SEEK_DATA && start >= inode->i_size)
2907 ret = -ENXIO;
2908 else
2909 *offset = min_t(loff_t, start, inode->i_size);
2910 }
2911 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2912 &cached_state, GFP_NOFS);
2913 return ret;
2914}
2915
2916static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2917{
2918 struct inode *inode = file->f_mapping->host;
2919 int ret;
2920
2921 inode_lock(inode);
2922 switch (whence) {
2923 case SEEK_END:
2924 case SEEK_CUR:
2925 offset = generic_file_llseek(file, offset, whence);
2926 goto out;
2927 case SEEK_DATA:
2928 case SEEK_HOLE:
2929 if (offset >= i_size_read(inode)) {
2930 inode_unlock(inode);
2931 return -ENXIO;
2932 }
2933
2934 ret = find_desired_extent(inode, &offset, whence);
2935 if (ret) {
2936 inode_unlock(inode);
2937 return ret;
2938 }
2939 }
2940
2941 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2942out:
2943 inode_unlock(inode);
2944 return offset;
2945}
2946
2947const struct file_operations btrfs_file_operations = {
2948 .llseek = btrfs_file_llseek,
2949 .read_iter = generic_file_read_iter,
2950 .splice_read = generic_file_splice_read,
2951 .write_iter = btrfs_file_write_iter,
2952 .mmap = btrfs_file_mmap,
2953 .open = generic_file_open,
2954 .release = btrfs_release_file,
2955 .fsync = btrfs_sync_file,
2956 .fallocate = btrfs_fallocate,
2957 .unlocked_ioctl = btrfs_ioctl,
2958#ifdef CONFIG_COMPAT
2959 .compat_ioctl = btrfs_ioctl,
2960#endif
2961 .copy_file_range = btrfs_copy_file_range,
2962 .clone_file_range = btrfs_clone_file_range,
2963 .dedupe_file_range = btrfs_dedupe_file_range,
2964};
2965
2966void btrfs_auto_defrag_exit(void)
2967{
2968 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2969}
2970
2971int btrfs_auto_defrag_init(void)
2972{
2973 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2974 sizeof(struct inode_defrag), 0,
2975 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2976 NULL);
2977 if (!btrfs_inode_defrag_cachep)
2978 return -ENOMEM;
2979
2980 return 0;
2981}
2982
2983int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
2984{
2985 int ret;
2986
2987 /*
2988 * So with compression we will find and lock a dirty page and clear the
2989 * first one as dirty, setup an async extent, and immediately return
2990 * with the entire range locked but with nobody actually marked with
2991 * writeback. So we can't just filemap_write_and_wait_range() and
2992 * expect it to work since it will just kick off a thread to do the
2993 * actual work. So we need to call filemap_fdatawrite_range _again_
2994 * since it will wait on the page lock, which won't be unlocked until
2995 * after the pages have been marked as writeback and so we're good to go
2996 * from there. We have to do this otherwise we'll miss the ordered
2997 * extents and that results in badness. Please Josef, do not think you
2998 * know better and pull this out at some point in the future, it is
2999 * right and you are wrong.
3000 */
3001 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3002 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3003 &BTRFS_I(inode)->runtime_flags))
3004 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3005
3006 return ret;
3007}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6#include <linux/fs.h>
7#include <linux/pagemap.h>
8#include <linux/time.h>
9#include <linux/init.h>
10#include <linux/string.h>
11#include <linux/backing-dev.h>
12#include <linux/falloc.h>
13#include <linux/writeback.h>
14#include <linux/compat.h>
15#include <linux/slab.h>
16#include <linux/btrfs.h>
17#include <linux/uio.h>
18#include <linux/iversion.h>
19#include "ctree.h"
20#include "disk-io.h"
21#include "transaction.h"
22#include "btrfs_inode.h"
23#include "print-tree.h"
24#include "tree-log.h"
25#include "locking.h"
26#include "volumes.h"
27#include "qgroup.h"
28#include "compression.h"
29#include "delalloc-space.h"
30#include "reflink.h"
31
32static struct kmem_cache *btrfs_inode_defrag_cachep;
33/*
34 * when auto defrag is enabled we
35 * queue up these defrag structs to remember which
36 * inodes need defragging passes
37 */
38struct inode_defrag {
39 struct rb_node rb_node;
40 /* objectid */
41 u64 ino;
42 /*
43 * transid where the defrag was added, we search for
44 * extents newer than this
45 */
46 u64 transid;
47
48 /* root objectid */
49 u64 root;
50
51 /* last offset we were able to defrag */
52 u64 last_offset;
53
54 /* if we've wrapped around back to zero once already */
55 int cycled;
56};
57
58static int __compare_inode_defrag(struct inode_defrag *defrag1,
59 struct inode_defrag *defrag2)
60{
61 if (defrag1->root > defrag2->root)
62 return 1;
63 else if (defrag1->root < defrag2->root)
64 return -1;
65 else if (defrag1->ino > defrag2->ino)
66 return 1;
67 else if (defrag1->ino < defrag2->ino)
68 return -1;
69 else
70 return 0;
71}
72
73/* pop a record for an inode into the defrag tree. The lock
74 * must be held already
75 *
76 * If you're inserting a record for an older transid than an
77 * existing record, the transid already in the tree is lowered
78 *
79 * If an existing record is found the defrag item you
80 * pass in is freed
81 */
82static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
83 struct inode_defrag *defrag)
84{
85 struct btrfs_fs_info *fs_info = inode->root->fs_info;
86 struct inode_defrag *entry;
87 struct rb_node **p;
88 struct rb_node *parent = NULL;
89 int ret;
90
91 p = &fs_info->defrag_inodes.rb_node;
92 while (*p) {
93 parent = *p;
94 entry = rb_entry(parent, struct inode_defrag, rb_node);
95
96 ret = __compare_inode_defrag(defrag, entry);
97 if (ret < 0)
98 p = &parent->rb_left;
99 else if (ret > 0)
100 p = &parent->rb_right;
101 else {
102 /* if we're reinserting an entry for
103 * an old defrag run, make sure to
104 * lower the transid of our existing record
105 */
106 if (defrag->transid < entry->transid)
107 entry->transid = defrag->transid;
108 if (defrag->last_offset > entry->last_offset)
109 entry->last_offset = defrag->last_offset;
110 return -EEXIST;
111 }
112 }
113 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
114 rb_link_node(&defrag->rb_node, parent, p);
115 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
116 return 0;
117}
118
119static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
120{
121 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
122 return 0;
123
124 if (btrfs_fs_closing(fs_info))
125 return 0;
126
127 return 1;
128}
129
130/*
131 * insert a defrag record for this inode if auto defrag is
132 * enabled
133 */
134int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
135 struct btrfs_inode *inode)
136{
137 struct btrfs_root *root = inode->root;
138 struct btrfs_fs_info *fs_info = root->fs_info;
139 struct inode_defrag *defrag;
140 u64 transid;
141 int ret;
142
143 if (!__need_auto_defrag(fs_info))
144 return 0;
145
146 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
147 return 0;
148
149 if (trans)
150 transid = trans->transid;
151 else
152 transid = inode->root->last_trans;
153
154 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
155 if (!defrag)
156 return -ENOMEM;
157
158 defrag->ino = btrfs_ino(inode);
159 defrag->transid = transid;
160 defrag->root = root->root_key.objectid;
161
162 spin_lock(&fs_info->defrag_inodes_lock);
163 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
164 /*
165 * If we set IN_DEFRAG flag and evict the inode from memory,
166 * and then re-read this inode, this new inode doesn't have
167 * IN_DEFRAG flag. At the case, we may find the existed defrag.
168 */
169 ret = __btrfs_add_inode_defrag(inode, defrag);
170 if (ret)
171 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
172 } else {
173 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
174 }
175 spin_unlock(&fs_info->defrag_inodes_lock);
176 return 0;
177}
178
179/*
180 * Requeue the defrag object. If there is a defrag object that points to
181 * the same inode in the tree, we will merge them together (by
182 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
183 */
184static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
185 struct inode_defrag *defrag)
186{
187 struct btrfs_fs_info *fs_info = inode->root->fs_info;
188 int ret;
189
190 if (!__need_auto_defrag(fs_info))
191 goto out;
192
193 /*
194 * Here we don't check the IN_DEFRAG flag, because we need merge
195 * them together.
196 */
197 spin_lock(&fs_info->defrag_inodes_lock);
198 ret = __btrfs_add_inode_defrag(inode, defrag);
199 spin_unlock(&fs_info->defrag_inodes_lock);
200 if (ret)
201 goto out;
202 return;
203out:
204 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
205}
206
207/*
208 * pick the defragable inode that we want, if it doesn't exist, we will get
209 * the next one.
210 */
211static struct inode_defrag *
212btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
213{
214 struct inode_defrag *entry = NULL;
215 struct inode_defrag tmp;
216 struct rb_node *p;
217 struct rb_node *parent = NULL;
218 int ret;
219
220 tmp.ino = ino;
221 tmp.root = root;
222
223 spin_lock(&fs_info->defrag_inodes_lock);
224 p = fs_info->defrag_inodes.rb_node;
225 while (p) {
226 parent = p;
227 entry = rb_entry(parent, struct inode_defrag, rb_node);
228
229 ret = __compare_inode_defrag(&tmp, entry);
230 if (ret < 0)
231 p = parent->rb_left;
232 else if (ret > 0)
233 p = parent->rb_right;
234 else
235 goto out;
236 }
237
238 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
239 parent = rb_next(parent);
240 if (parent)
241 entry = rb_entry(parent, struct inode_defrag, rb_node);
242 else
243 entry = NULL;
244 }
245out:
246 if (entry)
247 rb_erase(parent, &fs_info->defrag_inodes);
248 spin_unlock(&fs_info->defrag_inodes_lock);
249 return entry;
250}
251
252void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
253{
254 struct inode_defrag *defrag;
255 struct rb_node *node;
256
257 spin_lock(&fs_info->defrag_inodes_lock);
258 node = rb_first(&fs_info->defrag_inodes);
259 while (node) {
260 rb_erase(node, &fs_info->defrag_inodes);
261 defrag = rb_entry(node, struct inode_defrag, rb_node);
262 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
263
264 cond_resched_lock(&fs_info->defrag_inodes_lock);
265
266 node = rb_first(&fs_info->defrag_inodes);
267 }
268 spin_unlock(&fs_info->defrag_inodes_lock);
269}
270
271#define BTRFS_DEFRAG_BATCH 1024
272
273static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
274 struct inode_defrag *defrag)
275{
276 struct btrfs_root *inode_root;
277 struct inode *inode;
278 struct btrfs_ioctl_defrag_range_args range;
279 int num_defrag;
280 int ret;
281
282 /* get the inode */
283 inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
284 if (IS_ERR(inode_root)) {
285 ret = PTR_ERR(inode_root);
286 goto cleanup;
287 }
288
289 inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
290 btrfs_put_root(inode_root);
291 if (IS_ERR(inode)) {
292 ret = PTR_ERR(inode);
293 goto cleanup;
294 }
295
296 /* do a chunk of defrag */
297 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
298 memset(&range, 0, sizeof(range));
299 range.len = (u64)-1;
300 range.start = defrag->last_offset;
301
302 sb_start_write(fs_info->sb);
303 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
304 BTRFS_DEFRAG_BATCH);
305 sb_end_write(fs_info->sb);
306 /*
307 * if we filled the whole defrag batch, there
308 * must be more work to do. Queue this defrag
309 * again
310 */
311 if (num_defrag == BTRFS_DEFRAG_BATCH) {
312 defrag->last_offset = range.start;
313 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
314 } else if (defrag->last_offset && !defrag->cycled) {
315 /*
316 * we didn't fill our defrag batch, but
317 * we didn't start at zero. Make sure we loop
318 * around to the start of the file.
319 */
320 defrag->last_offset = 0;
321 defrag->cycled = 1;
322 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
323 } else {
324 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
325 }
326
327 iput(inode);
328 return 0;
329cleanup:
330 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
331 return ret;
332}
333
334/*
335 * run through the list of inodes in the FS that need
336 * defragging
337 */
338int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
339{
340 struct inode_defrag *defrag;
341 u64 first_ino = 0;
342 u64 root_objectid = 0;
343
344 atomic_inc(&fs_info->defrag_running);
345 while (1) {
346 /* Pause the auto defragger. */
347 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
348 &fs_info->fs_state))
349 break;
350
351 if (!__need_auto_defrag(fs_info))
352 break;
353
354 /* find an inode to defrag */
355 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
356 first_ino);
357 if (!defrag) {
358 if (root_objectid || first_ino) {
359 root_objectid = 0;
360 first_ino = 0;
361 continue;
362 } else {
363 break;
364 }
365 }
366
367 first_ino = defrag->ino + 1;
368 root_objectid = defrag->root;
369
370 __btrfs_run_defrag_inode(fs_info, defrag);
371 }
372 atomic_dec(&fs_info->defrag_running);
373
374 /*
375 * during unmount, we use the transaction_wait queue to
376 * wait for the defragger to stop
377 */
378 wake_up(&fs_info->transaction_wait);
379 return 0;
380}
381
382/* simple helper to fault in pages and copy. This should go away
383 * and be replaced with calls into generic code.
384 */
385static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
386 struct page **prepared_pages,
387 struct iov_iter *i)
388{
389 size_t copied = 0;
390 size_t total_copied = 0;
391 int pg = 0;
392 int offset = offset_in_page(pos);
393
394 while (write_bytes > 0) {
395 size_t count = min_t(size_t,
396 PAGE_SIZE - offset, write_bytes);
397 struct page *page = prepared_pages[pg];
398 /*
399 * Copy data from userspace to the current page
400 */
401 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
402
403 /* Flush processor's dcache for this page */
404 flush_dcache_page(page);
405
406 /*
407 * if we get a partial write, we can end up with
408 * partially up to date pages. These add
409 * a lot of complexity, so make sure they don't
410 * happen by forcing this copy to be retried.
411 *
412 * The rest of the btrfs_file_write code will fall
413 * back to page at a time copies after we return 0.
414 */
415 if (!PageUptodate(page) && copied < count)
416 copied = 0;
417
418 iov_iter_advance(i, copied);
419 write_bytes -= copied;
420 total_copied += copied;
421
422 /* Return to btrfs_file_write_iter to fault page */
423 if (unlikely(copied == 0))
424 break;
425
426 if (copied < PAGE_SIZE - offset) {
427 offset += copied;
428 } else {
429 pg++;
430 offset = 0;
431 }
432 }
433 return total_copied;
434}
435
436/*
437 * unlocks pages after btrfs_file_write is done with them
438 */
439static void btrfs_drop_pages(struct page **pages, size_t num_pages)
440{
441 size_t i;
442 for (i = 0; i < num_pages; i++) {
443 /* page checked is some magic around finding pages that
444 * have been modified without going through btrfs_set_page_dirty
445 * clear it here. There should be no need to mark the pages
446 * accessed as prepare_pages should have marked them accessed
447 * in prepare_pages via find_or_create_page()
448 */
449 ClearPageChecked(pages[i]);
450 unlock_page(pages[i]);
451 put_page(pages[i]);
452 }
453}
454
455static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
456 const u64 start,
457 const u64 len,
458 struct extent_state **cached_state)
459{
460 u64 search_start = start;
461 const u64 end = start + len - 1;
462
463 while (search_start < end) {
464 const u64 search_len = end - search_start + 1;
465 struct extent_map *em;
466 u64 em_len;
467 int ret = 0;
468
469 em = btrfs_get_extent(inode, NULL, 0, search_start, search_len);
470 if (IS_ERR(em))
471 return PTR_ERR(em);
472
473 if (em->block_start != EXTENT_MAP_HOLE)
474 goto next;
475
476 em_len = em->len;
477 if (em->start < search_start)
478 em_len -= search_start - em->start;
479 if (em_len > search_len)
480 em_len = search_len;
481
482 ret = set_extent_bit(&inode->io_tree, search_start,
483 search_start + em_len - 1,
484 EXTENT_DELALLOC_NEW,
485 NULL, cached_state, GFP_NOFS);
486next:
487 search_start = extent_map_end(em);
488 free_extent_map(em);
489 if (ret)
490 return ret;
491 }
492 return 0;
493}
494
495/*
496 * after copy_from_user, pages need to be dirtied and we need to make
497 * sure holes are created between the current EOF and the start of
498 * any next extents (if required).
499 *
500 * this also makes the decision about creating an inline extent vs
501 * doing real data extents, marking pages dirty and delalloc as required.
502 */
503int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
504 size_t num_pages, loff_t pos, size_t write_bytes,
505 struct extent_state **cached)
506{
507 struct btrfs_fs_info *fs_info = inode->root->fs_info;
508 int err = 0;
509 int i;
510 u64 num_bytes;
511 u64 start_pos;
512 u64 end_of_last_block;
513 u64 end_pos = pos + write_bytes;
514 loff_t isize = i_size_read(&inode->vfs_inode);
515 unsigned int extra_bits = 0;
516
517 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
518 num_bytes = round_up(write_bytes + pos - start_pos,
519 fs_info->sectorsize);
520
521 end_of_last_block = start_pos + num_bytes - 1;
522
523 /*
524 * The pages may have already been dirty, clear out old accounting so
525 * we can set things up properly
526 */
527 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
528 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
529 0, 0, cached);
530
531 if (!btrfs_is_free_space_inode(inode)) {
532 if (start_pos >= isize &&
533 !(inode->flags & BTRFS_INODE_PREALLOC)) {
534 /*
535 * There can't be any extents following eof in this case
536 * so just set the delalloc new bit for the range
537 * directly.
538 */
539 extra_bits |= EXTENT_DELALLOC_NEW;
540 } else {
541 err = btrfs_find_new_delalloc_bytes(inode, start_pos,
542 num_bytes, cached);
543 if (err)
544 return err;
545 }
546 }
547
548 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
549 extra_bits, cached);
550 if (err)
551 return err;
552
553 for (i = 0; i < num_pages; i++) {
554 struct page *p = pages[i];
555 SetPageUptodate(p);
556 ClearPageChecked(p);
557 set_page_dirty(p);
558 }
559
560 /*
561 * we've only changed i_size in ram, and we haven't updated
562 * the disk i_size. There is no need to log the inode
563 * at this time.
564 */
565 if (end_pos > isize)
566 i_size_write(&inode->vfs_inode, end_pos);
567 return 0;
568}
569
570/*
571 * this drops all the extents in the cache that intersect the range
572 * [start, end]. Existing extents are split as required.
573 */
574void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
575 int skip_pinned)
576{
577 struct extent_map *em;
578 struct extent_map *split = NULL;
579 struct extent_map *split2 = NULL;
580 struct extent_map_tree *em_tree = &inode->extent_tree;
581 u64 len = end - start + 1;
582 u64 gen;
583 int ret;
584 int testend = 1;
585 unsigned long flags;
586 int compressed = 0;
587 bool modified;
588
589 WARN_ON(end < start);
590 if (end == (u64)-1) {
591 len = (u64)-1;
592 testend = 0;
593 }
594 while (1) {
595 int no_splits = 0;
596
597 modified = false;
598 if (!split)
599 split = alloc_extent_map();
600 if (!split2)
601 split2 = alloc_extent_map();
602 if (!split || !split2)
603 no_splits = 1;
604
605 write_lock(&em_tree->lock);
606 em = lookup_extent_mapping(em_tree, start, len);
607 if (!em) {
608 write_unlock(&em_tree->lock);
609 break;
610 }
611 flags = em->flags;
612 gen = em->generation;
613 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
614 if (testend && em->start + em->len >= start + len) {
615 free_extent_map(em);
616 write_unlock(&em_tree->lock);
617 break;
618 }
619 start = em->start + em->len;
620 if (testend)
621 len = start + len - (em->start + em->len);
622 free_extent_map(em);
623 write_unlock(&em_tree->lock);
624 continue;
625 }
626 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
627 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
628 clear_bit(EXTENT_FLAG_LOGGING, &flags);
629 modified = !list_empty(&em->list);
630 if (no_splits)
631 goto next;
632
633 if (em->start < start) {
634 split->start = em->start;
635 split->len = start - em->start;
636
637 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
638 split->orig_start = em->orig_start;
639 split->block_start = em->block_start;
640
641 if (compressed)
642 split->block_len = em->block_len;
643 else
644 split->block_len = split->len;
645 split->orig_block_len = max(split->block_len,
646 em->orig_block_len);
647 split->ram_bytes = em->ram_bytes;
648 } else {
649 split->orig_start = split->start;
650 split->block_len = 0;
651 split->block_start = em->block_start;
652 split->orig_block_len = 0;
653 split->ram_bytes = split->len;
654 }
655
656 split->generation = gen;
657 split->flags = flags;
658 split->compress_type = em->compress_type;
659 replace_extent_mapping(em_tree, em, split, modified);
660 free_extent_map(split);
661 split = split2;
662 split2 = NULL;
663 }
664 if (testend && em->start + em->len > start + len) {
665 u64 diff = start + len - em->start;
666
667 split->start = start + len;
668 split->len = em->start + em->len - (start + len);
669 split->flags = flags;
670 split->compress_type = em->compress_type;
671 split->generation = gen;
672
673 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
674 split->orig_block_len = max(em->block_len,
675 em->orig_block_len);
676
677 split->ram_bytes = em->ram_bytes;
678 if (compressed) {
679 split->block_len = em->block_len;
680 split->block_start = em->block_start;
681 split->orig_start = em->orig_start;
682 } else {
683 split->block_len = split->len;
684 split->block_start = em->block_start
685 + diff;
686 split->orig_start = em->orig_start;
687 }
688 } else {
689 split->ram_bytes = split->len;
690 split->orig_start = split->start;
691 split->block_len = 0;
692 split->block_start = em->block_start;
693 split->orig_block_len = 0;
694 }
695
696 if (extent_map_in_tree(em)) {
697 replace_extent_mapping(em_tree, em, split,
698 modified);
699 } else {
700 ret = add_extent_mapping(em_tree, split,
701 modified);
702 ASSERT(ret == 0); /* Logic error */
703 }
704 free_extent_map(split);
705 split = NULL;
706 }
707next:
708 if (extent_map_in_tree(em))
709 remove_extent_mapping(em_tree, em);
710 write_unlock(&em_tree->lock);
711
712 /* once for us */
713 free_extent_map(em);
714 /* once for the tree*/
715 free_extent_map(em);
716 }
717 if (split)
718 free_extent_map(split);
719 if (split2)
720 free_extent_map(split2);
721}
722
723/*
724 * this is very complex, but the basic idea is to drop all extents
725 * in the range start - end. hint_block is filled in with a block number
726 * that would be a good hint to the block allocator for this file.
727 *
728 * If an extent intersects the range but is not entirely inside the range
729 * it is either truncated or split. Anything entirely inside the range
730 * is deleted from the tree.
731 */
732int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
733 struct btrfs_root *root, struct btrfs_inode *inode,
734 struct btrfs_path *path, u64 start, u64 end,
735 u64 *drop_end, int drop_cache,
736 int replace_extent,
737 u32 extent_item_size,
738 int *key_inserted)
739{
740 struct btrfs_fs_info *fs_info = root->fs_info;
741 struct extent_buffer *leaf;
742 struct btrfs_file_extent_item *fi;
743 struct btrfs_ref ref = { 0 };
744 struct btrfs_key key;
745 struct btrfs_key new_key;
746 struct inode *vfs_inode = &inode->vfs_inode;
747 u64 ino = btrfs_ino(inode);
748 u64 search_start = start;
749 u64 disk_bytenr = 0;
750 u64 num_bytes = 0;
751 u64 extent_offset = 0;
752 u64 extent_end = 0;
753 u64 last_end = start;
754 int del_nr = 0;
755 int del_slot = 0;
756 int extent_type;
757 int recow;
758 int ret;
759 int modify_tree = -1;
760 int update_refs;
761 int found = 0;
762 int leafs_visited = 0;
763
764 if (drop_cache)
765 btrfs_drop_extent_cache(inode, start, end - 1, 0);
766
767 if (start >= inode->disk_i_size && !replace_extent)
768 modify_tree = 0;
769
770 update_refs = (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
771 root == fs_info->tree_root);
772 while (1) {
773 recow = 0;
774 ret = btrfs_lookup_file_extent(trans, root, path, ino,
775 search_start, modify_tree);
776 if (ret < 0)
777 break;
778 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
779 leaf = path->nodes[0];
780 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
781 if (key.objectid == ino &&
782 key.type == BTRFS_EXTENT_DATA_KEY)
783 path->slots[0]--;
784 }
785 ret = 0;
786 leafs_visited++;
787next_slot:
788 leaf = path->nodes[0];
789 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
790 BUG_ON(del_nr > 0);
791 ret = btrfs_next_leaf(root, path);
792 if (ret < 0)
793 break;
794 if (ret > 0) {
795 ret = 0;
796 break;
797 }
798 leafs_visited++;
799 leaf = path->nodes[0];
800 recow = 1;
801 }
802
803 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
804
805 if (key.objectid > ino)
806 break;
807 if (WARN_ON_ONCE(key.objectid < ino) ||
808 key.type < BTRFS_EXTENT_DATA_KEY) {
809 ASSERT(del_nr == 0);
810 path->slots[0]++;
811 goto next_slot;
812 }
813 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
814 break;
815
816 fi = btrfs_item_ptr(leaf, path->slots[0],
817 struct btrfs_file_extent_item);
818 extent_type = btrfs_file_extent_type(leaf, fi);
819
820 if (extent_type == BTRFS_FILE_EXTENT_REG ||
821 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
822 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
823 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
824 extent_offset = btrfs_file_extent_offset(leaf, fi);
825 extent_end = key.offset +
826 btrfs_file_extent_num_bytes(leaf, fi);
827 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
828 extent_end = key.offset +
829 btrfs_file_extent_ram_bytes(leaf, fi);
830 } else {
831 /* can't happen */
832 BUG();
833 }
834
835 /*
836 * Don't skip extent items representing 0 byte lengths. They
837 * used to be created (bug) if while punching holes we hit
838 * -ENOSPC condition. So if we find one here, just ensure we
839 * delete it, otherwise we would insert a new file extent item
840 * with the same key (offset) as that 0 bytes length file
841 * extent item in the call to setup_items_for_insert() later
842 * in this function.
843 */
844 if (extent_end == key.offset && extent_end >= search_start) {
845 last_end = extent_end;
846 goto delete_extent_item;
847 }
848
849 if (extent_end <= search_start) {
850 path->slots[0]++;
851 goto next_slot;
852 }
853
854 found = 1;
855 search_start = max(key.offset, start);
856 if (recow || !modify_tree) {
857 modify_tree = -1;
858 btrfs_release_path(path);
859 continue;
860 }
861
862 /*
863 * | - range to drop - |
864 * | -------- extent -------- |
865 */
866 if (start > key.offset && end < extent_end) {
867 BUG_ON(del_nr > 0);
868 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
869 ret = -EOPNOTSUPP;
870 break;
871 }
872
873 memcpy(&new_key, &key, sizeof(new_key));
874 new_key.offset = start;
875 ret = btrfs_duplicate_item(trans, root, path,
876 &new_key);
877 if (ret == -EAGAIN) {
878 btrfs_release_path(path);
879 continue;
880 }
881 if (ret < 0)
882 break;
883
884 leaf = path->nodes[0];
885 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
886 struct btrfs_file_extent_item);
887 btrfs_set_file_extent_num_bytes(leaf, fi,
888 start - key.offset);
889
890 fi = btrfs_item_ptr(leaf, path->slots[0],
891 struct btrfs_file_extent_item);
892
893 extent_offset += start - key.offset;
894 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
895 btrfs_set_file_extent_num_bytes(leaf, fi,
896 extent_end - start);
897 btrfs_mark_buffer_dirty(leaf);
898
899 if (update_refs && disk_bytenr > 0) {
900 btrfs_init_generic_ref(&ref,
901 BTRFS_ADD_DELAYED_REF,
902 disk_bytenr, num_bytes, 0);
903 btrfs_init_data_ref(&ref,
904 root->root_key.objectid,
905 new_key.objectid,
906 start - extent_offset);
907 ret = btrfs_inc_extent_ref(trans, &ref);
908 BUG_ON(ret); /* -ENOMEM */
909 }
910 key.offset = start;
911 }
912 /*
913 * From here on out we will have actually dropped something, so
914 * last_end can be updated.
915 */
916 last_end = extent_end;
917
918 /*
919 * | ---- range to drop ----- |
920 * | -------- extent -------- |
921 */
922 if (start <= key.offset && end < extent_end) {
923 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
924 ret = -EOPNOTSUPP;
925 break;
926 }
927
928 memcpy(&new_key, &key, sizeof(new_key));
929 new_key.offset = end;
930 btrfs_set_item_key_safe(fs_info, path, &new_key);
931
932 extent_offset += end - key.offset;
933 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
934 btrfs_set_file_extent_num_bytes(leaf, fi,
935 extent_end - end);
936 btrfs_mark_buffer_dirty(leaf);
937 if (update_refs && disk_bytenr > 0)
938 inode_sub_bytes(vfs_inode, end - key.offset);
939 break;
940 }
941
942 search_start = extent_end;
943 /*
944 * | ---- range to drop ----- |
945 * | -------- extent -------- |
946 */
947 if (start > key.offset && end >= extent_end) {
948 BUG_ON(del_nr > 0);
949 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
950 ret = -EOPNOTSUPP;
951 break;
952 }
953
954 btrfs_set_file_extent_num_bytes(leaf, fi,
955 start - key.offset);
956 btrfs_mark_buffer_dirty(leaf);
957 if (update_refs && disk_bytenr > 0)
958 inode_sub_bytes(vfs_inode, extent_end - start);
959 if (end == extent_end)
960 break;
961
962 path->slots[0]++;
963 goto next_slot;
964 }
965
966 /*
967 * | ---- range to drop ----- |
968 * | ------ extent ------ |
969 */
970 if (start <= key.offset && end >= extent_end) {
971delete_extent_item:
972 if (del_nr == 0) {
973 del_slot = path->slots[0];
974 del_nr = 1;
975 } else {
976 BUG_ON(del_slot + del_nr != path->slots[0]);
977 del_nr++;
978 }
979
980 if (update_refs &&
981 extent_type == BTRFS_FILE_EXTENT_INLINE) {
982 inode_sub_bytes(vfs_inode,
983 extent_end - key.offset);
984 extent_end = ALIGN(extent_end,
985 fs_info->sectorsize);
986 } else if (update_refs && disk_bytenr > 0) {
987 btrfs_init_generic_ref(&ref,
988 BTRFS_DROP_DELAYED_REF,
989 disk_bytenr, num_bytes, 0);
990 btrfs_init_data_ref(&ref,
991 root->root_key.objectid,
992 key.objectid,
993 key.offset - extent_offset);
994 ret = btrfs_free_extent(trans, &ref);
995 BUG_ON(ret); /* -ENOMEM */
996 inode_sub_bytes(vfs_inode,
997 extent_end - key.offset);
998 }
999
1000 if (end == extent_end)
1001 break;
1002
1003 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1004 path->slots[0]++;
1005 goto next_slot;
1006 }
1007
1008 ret = btrfs_del_items(trans, root, path, del_slot,
1009 del_nr);
1010 if (ret) {
1011 btrfs_abort_transaction(trans, ret);
1012 break;
1013 }
1014
1015 del_nr = 0;
1016 del_slot = 0;
1017
1018 btrfs_release_path(path);
1019 continue;
1020 }
1021
1022 BUG();
1023 }
1024
1025 if (!ret && del_nr > 0) {
1026 /*
1027 * Set path->slots[0] to first slot, so that after the delete
1028 * if items are move off from our leaf to its immediate left or
1029 * right neighbor leafs, we end up with a correct and adjusted
1030 * path->slots[0] for our insertion (if replace_extent != 0).
1031 */
1032 path->slots[0] = del_slot;
1033 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1034 if (ret)
1035 btrfs_abort_transaction(trans, ret);
1036 }
1037
1038 leaf = path->nodes[0];
1039 /*
1040 * If btrfs_del_items() was called, it might have deleted a leaf, in
1041 * which case it unlocked our path, so check path->locks[0] matches a
1042 * write lock.
1043 */
1044 if (!ret && replace_extent && leafs_visited == 1 &&
1045 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1046 path->locks[0] == BTRFS_WRITE_LOCK) &&
1047 btrfs_leaf_free_space(leaf) >=
1048 sizeof(struct btrfs_item) + extent_item_size) {
1049
1050 key.objectid = ino;
1051 key.type = BTRFS_EXTENT_DATA_KEY;
1052 key.offset = start;
1053 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1054 struct btrfs_key slot_key;
1055
1056 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1057 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1058 path->slots[0]++;
1059 }
1060 setup_items_for_insert(root, path, &key,
1061 &extent_item_size,
1062 extent_item_size,
1063 sizeof(struct btrfs_item) +
1064 extent_item_size, 1);
1065 *key_inserted = 1;
1066 }
1067
1068 if (!replace_extent || !(*key_inserted))
1069 btrfs_release_path(path);
1070 if (drop_end)
1071 *drop_end = found ? min(end, last_end) : end;
1072 return ret;
1073}
1074
1075int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1076 struct btrfs_root *root, struct inode *inode, u64 start,
1077 u64 end, int drop_cache)
1078{
1079 struct btrfs_path *path;
1080 int ret;
1081
1082 path = btrfs_alloc_path();
1083 if (!path)
1084 return -ENOMEM;
1085 ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path, start,
1086 end, NULL, drop_cache, 0, 0, NULL);
1087 btrfs_free_path(path);
1088 return ret;
1089}
1090
1091static int extent_mergeable(struct extent_buffer *leaf, int slot,
1092 u64 objectid, u64 bytenr, u64 orig_offset,
1093 u64 *start, u64 *end)
1094{
1095 struct btrfs_file_extent_item *fi;
1096 struct btrfs_key key;
1097 u64 extent_end;
1098
1099 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1100 return 0;
1101
1102 btrfs_item_key_to_cpu(leaf, &key, slot);
1103 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1104 return 0;
1105
1106 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1107 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1108 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1109 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1110 btrfs_file_extent_compression(leaf, fi) ||
1111 btrfs_file_extent_encryption(leaf, fi) ||
1112 btrfs_file_extent_other_encoding(leaf, fi))
1113 return 0;
1114
1115 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1116 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1117 return 0;
1118
1119 *start = key.offset;
1120 *end = extent_end;
1121 return 1;
1122}
1123
1124/*
1125 * Mark extent in the range start - end as written.
1126 *
1127 * This changes extent type from 'pre-allocated' to 'regular'. If only
1128 * part of extent is marked as written, the extent will be split into
1129 * two or three.
1130 */
1131int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1132 struct btrfs_inode *inode, u64 start, u64 end)
1133{
1134 struct btrfs_fs_info *fs_info = trans->fs_info;
1135 struct btrfs_root *root = inode->root;
1136 struct extent_buffer *leaf;
1137 struct btrfs_path *path;
1138 struct btrfs_file_extent_item *fi;
1139 struct btrfs_ref ref = { 0 };
1140 struct btrfs_key key;
1141 struct btrfs_key new_key;
1142 u64 bytenr;
1143 u64 num_bytes;
1144 u64 extent_end;
1145 u64 orig_offset;
1146 u64 other_start;
1147 u64 other_end;
1148 u64 split;
1149 int del_nr = 0;
1150 int del_slot = 0;
1151 int recow;
1152 int ret;
1153 u64 ino = btrfs_ino(inode);
1154
1155 path = btrfs_alloc_path();
1156 if (!path)
1157 return -ENOMEM;
1158again:
1159 recow = 0;
1160 split = start;
1161 key.objectid = ino;
1162 key.type = BTRFS_EXTENT_DATA_KEY;
1163 key.offset = split;
1164
1165 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1166 if (ret < 0)
1167 goto out;
1168 if (ret > 0 && path->slots[0] > 0)
1169 path->slots[0]--;
1170
1171 leaf = path->nodes[0];
1172 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1173 if (key.objectid != ino ||
1174 key.type != BTRFS_EXTENT_DATA_KEY) {
1175 ret = -EINVAL;
1176 btrfs_abort_transaction(trans, ret);
1177 goto out;
1178 }
1179 fi = btrfs_item_ptr(leaf, path->slots[0],
1180 struct btrfs_file_extent_item);
1181 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1182 ret = -EINVAL;
1183 btrfs_abort_transaction(trans, ret);
1184 goto out;
1185 }
1186 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1187 if (key.offset > start || extent_end < end) {
1188 ret = -EINVAL;
1189 btrfs_abort_transaction(trans, ret);
1190 goto out;
1191 }
1192
1193 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1194 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1195 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1196 memcpy(&new_key, &key, sizeof(new_key));
1197
1198 if (start == key.offset && end < extent_end) {
1199 other_start = 0;
1200 other_end = start;
1201 if (extent_mergeable(leaf, path->slots[0] - 1,
1202 ino, bytenr, orig_offset,
1203 &other_start, &other_end)) {
1204 new_key.offset = end;
1205 btrfs_set_item_key_safe(fs_info, path, &new_key);
1206 fi = btrfs_item_ptr(leaf, path->slots[0],
1207 struct btrfs_file_extent_item);
1208 btrfs_set_file_extent_generation(leaf, fi,
1209 trans->transid);
1210 btrfs_set_file_extent_num_bytes(leaf, fi,
1211 extent_end - end);
1212 btrfs_set_file_extent_offset(leaf, fi,
1213 end - orig_offset);
1214 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1215 struct btrfs_file_extent_item);
1216 btrfs_set_file_extent_generation(leaf, fi,
1217 trans->transid);
1218 btrfs_set_file_extent_num_bytes(leaf, fi,
1219 end - other_start);
1220 btrfs_mark_buffer_dirty(leaf);
1221 goto out;
1222 }
1223 }
1224
1225 if (start > key.offset && end == extent_end) {
1226 other_start = end;
1227 other_end = 0;
1228 if (extent_mergeable(leaf, path->slots[0] + 1,
1229 ino, bytenr, orig_offset,
1230 &other_start, &other_end)) {
1231 fi = btrfs_item_ptr(leaf, path->slots[0],
1232 struct btrfs_file_extent_item);
1233 btrfs_set_file_extent_num_bytes(leaf, fi,
1234 start - key.offset);
1235 btrfs_set_file_extent_generation(leaf, fi,
1236 trans->transid);
1237 path->slots[0]++;
1238 new_key.offset = start;
1239 btrfs_set_item_key_safe(fs_info, path, &new_key);
1240
1241 fi = btrfs_item_ptr(leaf, path->slots[0],
1242 struct btrfs_file_extent_item);
1243 btrfs_set_file_extent_generation(leaf, fi,
1244 trans->transid);
1245 btrfs_set_file_extent_num_bytes(leaf, fi,
1246 other_end - start);
1247 btrfs_set_file_extent_offset(leaf, fi,
1248 start - orig_offset);
1249 btrfs_mark_buffer_dirty(leaf);
1250 goto out;
1251 }
1252 }
1253
1254 while (start > key.offset || end < extent_end) {
1255 if (key.offset == start)
1256 split = end;
1257
1258 new_key.offset = split;
1259 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1260 if (ret == -EAGAIN) {
1261 btrfs_release_path(path);
1262 goto again;
1263 }
1264 if (ret < 0) {
1265 btrfs_abort_transaction(trans, ret);
1266 goto out;
1267 }
1268
1269 leaf = path->nodes[0];
1270 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1271 struct btrfs_file_extent_item);
1272 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1273 btrfs_set_file_extent_num_bytes(leaf, fi,
1274 split - key.offset);
1275
1276 fi = btrfs_item_ptr(leaf, path->slots[0],
1277 struct btrfs_file_extent_item);
1278
1279 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1280 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1281 btrfs_set_file_extent_num_bytes(leaf, fi,
1282 extent_end - split);
1283 btrfs_mark_buffer_dirty(leaf);
1284
1285 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1286 num_bytes, 0);
1287 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1288 orig_offset);
1289 ret = btrfs_inc_extent_ref(trans, &ref);
1290 if (ret) {
1291 btrfs_abort_transaction(trans, ret);
1292 goto out;
1293 }
1294
1295 if (split == start) {
1296 key.offset = start;
1297 } else {
1298 if (start != key.offset) {
1299 ret = -EINVAL;
1300 btrfs_abort_transaction(trans, ret);
1301 goto out;
1302 }
1303 path->slots[0]--;
1304 extent_end = end;
1305 }
1306 recow = 1;
1307 }
1308
1309 other_start = end;
1310 other_end = 0;
1311 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1312 num_bytes, 0);
1313 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1314 if (extent_mergeable(leaf, path->slots[0] + 1,
1315 ino, bytenr, orig_offset,
1316 &other_start, &other_end)) {
1317 if (recow) {
1318 btrfs_release_path(path);
1319 goto again;
1320 }
1321 extent_end = other_end;
1322 del_slot = path->slots[0] + 1;
1323 del_nr++;
1324 ret = btrfs_free_extent(trans, &ref);
1325 if (ret) {
1326 btrfs_abort_transaction(trans, ret);
1327 goto out;
1328 }
1329 }
1330 other_start = 0;
1331 other_end = start;
1332 if (extent_mergeable(leaf, path->slots[0] - 1,
1333 ino, bytenr, orig_offset,
1334 &other_start, &other_end)) {
1335 if (recow) {
1336 btrfs_release_path(path);
1337 goto again;
1338 }
1339 key.offset = other_start;
1340 del_slot = path->slots[0];
1341 del_nr++;
1342 ret = btrfs_free_extent(trans, &ref);
1343 if (ret) {
1344 btrfs_abort_transaction(trans, ret);
1345 goto out;
1346 }
1347 }
1348 if (del_nr == 0) {
1349 fi = btrfs_item_ptr(leaf, path->slots[0],
1350 struct btrfs_file_extent_item);
1351 btrfs_set_file_extent_type(leaf, fi,
1352 BTRFS_FILE_EXTENT_REG);
1353 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1354 btrfs_mark_buffer_dirty(leaf);
1355 } else {
1356 fi = btrfs_item_ptr(leaf, del_slot - 1,
1357 struct btrfs_file_extent_item);
1358 btrfs_set_file_extent_type(leaf, fi,
1359 BTRFS_FILE_EXTENT_REG);
1360 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1361 btrfs_set_file_extent_num_bytes(leaf, fi,
1362 extent_end - key.offset);
1363 btrfs_mark_buffer_dirty(leaf);
1364
1365 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1366 if (ret < 0) {
1367 btrfs_abort_transaction(trans, ret);
1368 goto out;
1369 }
1370 }
1371out:
1372 btrfs_free_path(path);
1373 return 0;
1374}
1375
1376/*
1377 * on error we return an unlocked page and the error value
1378 * on success we return a locked page and 0
1379 */
1380static int prepare_uptodate_page(struct inode *inode,
1381 struct page *page, u64 pos,
1382 bool force_uptodate)
1383{
1384 int ret = 0;
1385
1386 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1387 !PageUptodate(page)) {
1388 ret = btrfs_readpage(NULL, page);
1389 if (ret)
1390 return ret;
1391 lock_page(page);
1392 if (!PageUptodate(page)) {
1393 unlock_page(page);
1394 return -EIO;
1395 }
1396 if (page->mapping != inode->i_mapping) {
1397 unlock_page(page);
1398 return -EAGAIN;
1399 }
1400 }
1401 return 0;
1402}
1403
1404/*
1405 * this just gets pages into the page cache and locks them down.
1406 */
1407static noinline int prepare_pages(struct inode *inode, struct page **pages,
1408 size_t num_pages, loff_t pos,
1409 size_t write_bytes, bool force_uptodate)
1410{
1411 int i;
1412 unsigned long index = pos >> PAGE_SHIFT;
1413 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1414 int err = 0;
1415 int faili;
1416
1417 for (i = 0; i < num_pages; i++) {
1418again:
1419 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1420 mask | __GFP_WRITE);
1421 if (!pages[i]) {
1422 faili = i - 1;
1423 err = -ENOMEM;
1424 goto fail;
1425 }
1426
1427 if (i == 0)
1428 err = prepare_uptodate_page(inode, pages[i], pos,
1429 force_uptodate);
1430 if (!err && i == num_pages - 1)
1431 err = prepare_uptodate_page(inode, pages[i],
1432 pos + write_bytes, false);
1433 if (err) {
1434 put_page(pages[i]);
1435 if (err == -EAGAIN) {
1436 err = 0;
1437 goto again;
1438 }
1439 faili = i - 1;
1440 goto fail;
1441 }
1442 wait_on_page_writeback(pages[i]);
1443 }
1444
1445 return 0;
1446fail:
1447 while (faili >= 0) {
1448 unlock_page(pages[faili]);
1449 put_page(pages[faili]);
1450 faili--;
1451 }
1452 return err;
1453
1454}
1455
1456/*
1457 * This function locks the extent and properly waits for data=ordered extents
1458 * to finish before allowing the pages to be modified if need.
1459 *
1460 * The return value:
1461 * 1 - the extent is locked
1462 * 0 - the extent is not locked, and everything is OK
1463 * -EAGAIN - need re-prepare the pages
1464 * the other < 0 number - Something wrong happens
1465 */
1466static noinline int
1467lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1468 size_t num_pages, loff_t pos,
1469 size_t write_bytes,
1470 u64 *lockstart, u64 *lockend,
1471 struct extent_state **cached_state)
1472{
1473 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1474 u64 start_pos;
1475 u64 last_pos;
1476 int i;
1477 int ret = 0;
1478
1479 start_pos = round_down(pos, fs_info->sectorsize);
1480 last_pos = start_pos
1481 + round_up(pos + write_bytes - start_pos,
1482 fs_info->sectorsize) - 1;
1483
1484 if (start_pos < inode->vfs_inode.i_size) {
1485 struct btrfs_ordered_extent *ordered;
1486
1487 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1488 cached_state);
1489 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1490 last_pos - start_pos + 1);
1491 if (ordered &&
1492 ordered->file_offset + ordered->num_bytes > start_pos &&
1493 ordered->file_offset <= last_pos) {
1494 unlock_extent_cached(&inode->io_tree, start_pos,
1495 last_pos, cached_state);
1496 for (i = 0; i < num_pages; i++) {
1497 unlock_page(pages[i]);
1498 put_page(pages[i]);
1499 }
1500 btrfs_start_ordered_extent(&inode->vfs_inode,
1501 ordered, 1);
1502 btrfs_put_ordered_extent(ordered);
1503 return -EAGAIN;
1504 }
1505 if (ordered)
1506 btrfs_put_ordered_extent(ordered);
1507
1508 *lockstart = start_pos;
1509 *lockend = last_pos;
1510 ret = 1;
1511 }
1512
1513 /*
1514 * It's possible the pages are dirty right now, but we don't want
1515 * to clean them yet because copy_from_user may catch a page fault
1516 * and we might have to fall back to one page at a time. If that
1517 * happens, we'll unlock these pages and we'd have a window where
1518 * reclaim could sneak in and drop the once-dirty page on the floor
1519 * without writing it.
1520 *
1521 * We have the pages locked and the extent range locked, so there's
1522 * no way someone can start IO on any dirty pages in this range.
1523 *
1524 * We'll call btrfs_dirty_pages() later on, and that will flip around
1525 * delalloc bits and dirty the pages as required.
1526 */
1527 for (i = 0; i < num_pages; i++) {
1528 set_page_extent_mapped(pages[i]);
1529 WARN_ON(!PageLocked(pages[i]));
1530 }
1531
1532 return ret;
1533}
1534
1535static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1536 size_t *write_bytes, bool nowait)
1537{
1538 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1539 struct btrfs_root *root = inode->root;
1540 u64 lockstart, lockend;
1541 u64 num_bytes;
1542 int ret;
1543
1544 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1545 return 0;
1546
1547 if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1548 return -EAGAIN;
1549
1550 lockstart = round_down(pos, fs_info->sectorsize);
1551 lockend = round_up(pos + *write_bytes,
1552 fs_info->sectorsize) - 1;
1553 num_bytes = lockend - lockstart + 1;
1554
1555 if (nowait) {
1556 struct btrfs_ordered_extent *ordered;
1557
1558 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1559 return -EAGAIN;
1560
1561 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1562 num_bytes);
1563 if (ordered) {
1564 btrfs_put_ordered_extent(ordered);
1565 ret = -EAGAIN;
1566 goto out_unlock;
1567 }
1568 } else {
1569 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1570 lockend, NULL);
1571 }
1572
1573 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1574 NULL, NULL, NULL, false);
1575 if (ret <= 0) {
1576 ret = 0;
1577 if (!nowait)
1578 btrfs_drew_write_unlock(&root->snapshot_lock);
1579 } else {
1580 *write_bytes = min_t(size_t, *write_bytes ,
1581 num_bytes - pos + lockstart);
1582 }
1583out_unlock:
1584 unlock_extent(&inode->io_tree, lockstart, lockend);
1585
1586 return ret;
1587}
1588
1589static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1590 size_t *write_bytes)
1591{
1592 return check_can_nocow(inode, pos, write_bytes, true);
1593}
1594
1595/*
1596 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1597 *
1598 * @pos: File offset
1599 * @write_bytes: The length to write, will be updated to the nocow writeable
1600 * range
1601 *
1602 * This function will flush ordered extents in the range to ensure proper
1603 * nocow checks.
1604 *
1605 * Return:
1606 * >0 and update @write_bytes if we can do nocow write
1607 * 0 if we can't do nocow write
1608 * -EAGAIN if we can't get the needed lock or there are ordered extents
1609 * for * (nowait == true) case
1610 * <0 if other error happened
1611 *
1612 * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1613 */
1614int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1615 size_t *write_bytes)
1616{
1617 return check_can_nocow(inode, pos, write_bytes, false);
1618}
1619
1620void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1621{
1622 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1623}
1624
1625static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1626 struct iov_iter *i)
1627{
1628 struct file *file = iocb->ki_filp;
1629 loff_t pos = iocb->ki_pos;
1630 struct inode *inode = file_inode(file);
1631 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1632 struct page **pages = NULL;
1633 struct extent_changeset *data_reserved = NULL;
1634 u64 release_bytes = 0;
1635 u64 lockstart;
1636 u64 lockend;
1637 size_t num_written = 0;
1638 int nrptrs;
1639 int ret = 0;
1640 bool only_release_metadata = false;
1641 bool force_page_uptodate = false;
1642
1643 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1644 PAGE_SIZE / (sizeof(struct page *)));
1645 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1646 nrptrs = max(nrptrs, 8);
1647 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1648 if (!pages)
1649 return -ENOMEM;
1650
1651 while (iov_iter_count(i) > 0) {
1652 struct extent_state *cached_state = NULL;
1653 size_t offset = offset_in_page(pos);
1654 size_t sector_offset;
1655 size_t write_bytes = min(iov_iter_count(i),
1656 nrptrs * (size_t)PAGE_SIZE -
1657 offset);
1658 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1659 PAGE_SIZE);
1660 size_t reserve_bytes;
1661 size_t dirty_pages;
1662 size_t copied;
1663 size_t dirty_sectors;
1664 size_t num_sectors;
1665 int extents_locked;
1666
1667 WARN_ON(num_pages > nrptrs);
1668
1669 /*
1670 * Fault pages before locking them in prepare_pages
1671 * to avoid recursive lock
1672 */
1673 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1674 ret = -EFAULT;
1675 break;
1676 }
1677
1678 only_release_metadata = false;
1679 sector_offset = pos & (fs_info->sectorsize - 1);
1680 reserve_bytes = round_up(write_bytes + sector_offset,
1681 fs_info->sectorsize);
1682
1683 extent_changeset_release(data_reserved);
1684 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1685 &data_reserved, pos,
1686 write_bytes);
1687 if (ret < 0) {
1688 if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1689 &write_bytes) > 0) {
1690 /*
1691 * For nodata cow case, no need to reserve
1692 * data space.
1693 */
1694 only_release_metadata = true;
1695 /*
1696 * our prealloc extent may be smaller than
1697 * write_bytes, so scale down.
1698 */
1699 num_pages = DIV_ROUND_UP(write_bytes + offset,
1700 PAGE_SIZE);
1701 reserve_bytes = round_up(write_bytes +
1702 sector_offset,
1703 fs_info->sectorsize);
1704 } else {
1705 break;
1706 }
1707 }
1708
1709 WARN_ON(reserve_bytes == 0);
1710 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1711 reserve_bytes);
1712 if (ret) {
1713 if (!only_release_metadata)
1714 btrfs_free_reserved_data_space(BTRFS_I(inode),
1715 data_reserved, pos,
1716 write_bytes);
1717 else
1718 btrfs_check_nocow_unlock(BTRFS_I(inode));
1719 break;
1720 }
1721
1722 release_bytes = reserve_bytes;
1723again:
1724 /*
1725 * This is going to setup the pages array with the number of
1726 * pages we want, so we don't really need to worry about the
1727 * contents of pages from loop to loop
1728 */
1729 ret = prepare_pages(inode, pages, num_pages,
1730 pos, write_bytes,
1731 force_page_uptodate);
1732 if (ret) {
1733 btrfs_delalloc_release_extents(BTRFS_I(inode),
1734 reserve_bytes);
1735 break;
1736 }
1737
1738 extents_locked = lock_and_cleanup_extent_if_need(
1739 BTRFS_I(inode), pages,
1740 num_pages, pos, write_bytes, &lockstart,
1741 &lockend, &cached_state);
1742 if (extents_locked < 0) {
1743 if (extents_locked == -EAGAIN)
1744 goto again;
1745 btrfs_delalloc_release_extents(BTRFS_I(inode),
1746 reserve_bytes);
1747 ret = extents_locked;
1748 break;
1749 }
1750
1751 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1752
1753 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1754 dirty_sectors = round_up(copied + sector_offset,
1755 fs_info->sectorsize);
1756 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1757
1758 /*
1759 * if we have trouble faulting in the pages, fall
1760 * back to one page at a time
1761 */
1762 if (copied < write_bytes)
1763 nrptrs = 1;
1764
1765 if (copied == 0) {
1766 force_page_uptodate = true;
1767 dirty_sectors = 0;
1768 dirty_pages = 0;
1769 } else {
1770 force_page_uptodate = false;
1771 dirty_pages = DIV_ROUND_UP(copied + offset,
1772 PAGE_SIZE);
1773 }
1774
1775 if (num_sectors > dirty_sectors) {
1776 /* release everything except the sectors we dirtied */
1777 release_bytes -= dirty_sectors <<
1778 fs_info->sb->s_blocksize_bits;
1779 if (only_release_metadata) {
1780 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1781 release_bytes, true);
1782 } else {
1783 u64 __pos;
1784
1785 __pos = round_down(pos,
1786 fs_info->sectorsize) +
1787 (dirty_pages << PAGE_SHIFT);
1788 btrfs_delalloc_release_space(BTRFS_I(inode),
1789 data_reserved, __pos,
1790 release_bytes, true);
1791 }
1792 }
1793
1794 release_bytes = round_up(copied + sector_offset,
1795 fs_info->sectorsize);
1796
1797 if (copied > 0)
1798 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1799 dirty_pages, pos, copied,
1800 &cached_state);
1801
1802 /*
1803 * If we have not locked the extent range, because the range's
1804 * start offset is >= i_size, we might still have a non-NULL
1805 * cached extent state, acquired while marking the extent range
1806 * as delalloc through btrfs_dirty_pages(). Therefore free any
1807 * possible cached extent state to avoid a memory leak.
1808 */
1809 if (extents_locked)
1810 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1811 lockstart, lockend, &cached_state);
1812 else
1813 free_extent_state(cached_state);
1814
1815 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1816 if (ret) {
1817 btrfs_drop_pages(pages, num_pages);
1818 break;
1819 }
1820
1821 release_bytes = 0;
1822 if (only_release_metadata)
1823 btrfs_check_nocow_unlock(BTRFS_I(inode));
1824
1825 if (only_release_metadata && copied > 0) {
1826 lockstart = round_down(pos,
1827 fs_info->sectorsize);
1828 lockend = round_up(pos + copied,
1829 fs_info->sectorsize) - 1;
1830
1831 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1832 lockend, EXTENT_NORESERVE, NULL,
1833 NULL, GFP_NOFS);
1834 }
1835
1836 btrfs_drop_pages(pages, num_pages);
1837
1838 cond_resched();
1839
1840 balance_dirty_pages_ratelimited(inode->i_mapping);
1841
1842 pos += copied;
1843 num_written += copied;
1844 }
1845
1846 kfree(pages);
1847
1848 if (release_bytes) {
1849 if (only_release_metadata) {
1850 btrfs_check_nocow_unlock(BTRFS_I(inode));
1851 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1852 release_bytes, true);
1853 } else {
1854 btrfs_delalloc_release_space(BTRFS_I(inode),
1855 data_reserved,
1856 round_down(pos, fs_info->sectorsize),
1857 release_bytes, true);
1858 }
1859 }
1860
1861 extent_changeset_free(data_reserved);
1862 return num_written ? num_written : ret;
1863}
1864
1865static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1866{
1867 struct file *file = iocb->ki_filp;
1868 struct inode *inode = file_inode(file);
1869 loff_t pos;
1870 ssize_t written;
1871 ssize_t written_buffered;
1872 loff_t endbyte;
1873 int err;
1874
1875 written = generic_file_direct_write(iocb, from);
1876
1877 if (written < 0 || !iov_iter_count(from))
1878 return written;
1879
1880 pos = iocb->ki_pos;
1881 written_buffered = btrfs_buffered_write(iocb, from);
1882 if (written_buffered < 0) {
1883 err = written_buffered;
1884 goto out;
1885 }
1886 /*
1887 * Ensure all data is persisted. We want the next direct IO read to be
1888 * able to read what was just written.
1889 */
1890 endbyte = pos + written_buffered - 1;
1891 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1892 if (err)
1893 goto out;
1894 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1895 if (err)
1896 goto out;
1897 written += written_buffered;
1898 iocb->ki_pos = pos + written_buffered;
1899 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1900 endbyte >> PAGE_SHIFT);
1901out:
1902 return written ? written : err;
1903}
1904
1905static void update_time_for_write(struct inode *inode)
1906{
1907 struct timespec64 now;
1908
1909 if (IS_NOCMTIME(inode))
1910 return;
1911
1912 now = current_time(inode);
1913 if (!timespec64_equal(&inode->i_mtime, &now))
1914 inode->i_mtime = now;
1915
1916 if (!timespec64_equal(&inode->i_ctime, &now))
1917 inode->i_ctime = now;
1918
1919 if (IS_I_VERSION(inode))
1920 inode_inc_iversion(inode);
1921}
1922
1923static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1924 struct iov_iter *from)
1925{
1926 struct file *file = iocb->ki_filp;
1927 struct inode *inode = file_inode(file);
1928 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1929 struct btrfs_root *root = BTRFS_I(inode)->root;
1930 u64 start_pos;
1931 u64 end_pos;
1932 ssize_t num_written = 0;
1933 const bool sync = iocb->ki_flags & IOCB_DSYNC;
1934 ssize_t err;
1935 loff_t pos;
1936 size_t count;
1937 loff_t oldsize;
1938 int clean_page = 0;
1939
1940 if (!(iocb->ki_flags & IOCB_DIRECT) &&
1941 (iocb->ki_flags & IOCB_NOWAIT))
1942 return -EOPNOTSUPP;
1943
1944 if (iocb->ki_flags & IOCB_NOWAIT) {
1945 if (!inode_trylock(inode))
1946 return -EAGAIN;
1947 } else {
1948 inode_lock(inode);
1949 }
1950
1951 err = generic_write_checks(iocb, from);
1952 if (err <= 0) {
1953 inode_unlock(inode);
1954 return err;
1955 }
1956
1957 pos = iocb->ki_pos;
1958 count = iov_iter_count(from);
1959 if (iocb->ki_flags & IOCB_NOWAIT) {
1960 size_t nocow_bytes = count;
1961
1962 /*
1963 * We will allocate space in case nodatacow is not set,
1964 * so bail
1965 */
1966 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes)
1967 <= 0) {
1968 inode_unlock(inode);
1969 return -EAGAIN;
1970 }
1971 /*
1972 * There are holes in the range or parts of the range that must
1973 * be COWed (shared extents, RO block groups, etc), so just bail
1974 * out.
1975 */
1976 if (nocow_bytes < count) {
1977 inode_unlock(inode);
1978 return -EAGAIN;
1979 }
1980 }
1981
1982 current->backing_dev_info = inode_to_bdi(inode);
1983 err = file_remove_privs(file);
1984 if (err) {
1985 inode_unlock(inode);
1986 goto out;
1987 }
1988
1989 /*
1990 * If BTRFS flips readonly due to some impossible error
1991 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1992 * although we have opened a file as writable, we have
1993 * to stop this write operation to ensure FS consistency.
1994 */
1995 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1996 inode_unlock(inode);
1997 err = -EROFS;
1998 goto out;
1999 }
2000
2001 /*
2002 * We reserve space for updating the inode when we reserve space for the
2003 * extent we are going to write, so we will enospc out there. We don't
2004 * need to start yet another transaction to update the inode as we will
2005 * update the inode when we finish writing whatever data we write.
2006 */
2007 update_time_for_write(inode);
2008
2009 start_pos = round_down(pos, fs_info->sectorsize);
2010 oldsize = i_size_read(inode);
2011 if (start_pos > oldsize) {
2012 /* Expand hole size to cover write data, preventing empty gap */
2013 end_pos = round_up(pos + count,
2014 fs_info->sectorsize);
2015 err = btrfs_cont_expand(inode, oldsize, end_pos);
2016 if (err) {
2017 inode_unlock(inode);
2018 goto out;
2019 }
2020 if (start_pos > round_up(oldsize, fs_info->sectorsize))
2021 clean_page = 1;
2022 }
2023
2024 if (sync)
2025 atomic_inc(&BTRFS_I(inode)->sync_writers);
2026
2027 if (iocb->ki_flags & IOCB_DIRECT) {
2028 num_written = __btrfs_direct_write(iocb, from);
2029 } else {
2030 num_written = btrfs_buffered_write(iocb, from);
2031 if (num_written > 0)
2032 iocb->ki_pos = pos + num_written;
2033 if (clean_page)
2034 pagecache_isize_extended(inode, oldsize,
2035 i_size_read(inode));
2036 }
2037
2038 inode_unlock(inode);
2039
2040 /*
2041 * We also have to set last_sub_trans to the current log transid,
2042 * otherwise subsequent syncs to a file that's been synced in this
2043 * transaction will appear to have already occurred.
2044 */
2045 spin_lock(&BTRFS_I(inode)->lock);
2046 BTRFS_I(inode)->last_sub_trans = root->log_transid;
2047 spin_unlock(&BTRFS_I(inode)->lock);
2048 if (num_written > 0)
2049 num_written = generic_write_sync(iocb, num_written);
2050
2051 if (sync)
2052 atomic_dec(&BTRFS_I(inode)->sync_writers);
2053out:
2054 current->backing_dev_info = NULL;
2055 return num_written ? num_written : err;
2056}
2057
2058int btrfs_release_file(struct inode *inode, struct file *filp)
2059{
2060 struct btrfs_file_private *private = filp->private_data;
2061
2062 if (private && private->filldir_buf)
2063 kfree(private->filldir_buf);
2064 kfree(private);
2065 filp->private_data = NULL;
2066
2067 /*
2068 * ordered_data_close is set by setattr when we are about to truncate
2069 * a file from a non-zero size to a zero size. This tries to
2070 * flush down new bytes that may have been written if the
2071 * application were using truncate to replace a file in place.
2072 */
2073 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2074 &BTRFS_I(inode)->runtime_flags))
2075 filemap_flush(inode->i_mapping);
2076 return 0;
2077}
2078
2079static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2080{
2081 int ret;
2082 struct blk_plug plug;
2083
2084 /*
2085 * This is only called in fsync, which would do synchronous writes, so
2086 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2087 * multiple disks using raid profile, a large IO can be split to
2088 * several segments of stripe length (currently 64K).
2089 */
2090 blk_start_plug(&plug);
2091 atomic_inc(&BTRFS_I(inode)->sync_writers);
2092 ret = btrfs_fdatawrite_range(inode, start, end);
2093 atomic_dec(&BTRFS_I(inode)->sync_writers);
2094 blk_finish_plug(&plug);
2095
2096 return ret;
2097}
2098
2099/*
2100 * fsync call for both files and directories. This logs the inode into
2101 * the tree log instead of forcing full commits whenever possible.
2102 *
2103 * It needs to call filemap_fdatawait so that all ordered extent updates are
2104 * in the metadata btree are up to date for copying to the log.
2105 *
2106 * It drops the inode mutex before doing the tree log commit. This is an
2107 * important optimization for directories because holding the mutex prevents
2108 * new operations on the dir while we write to disk.
2109 */
2110int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2111{
2112 struct dentry *dentry = file_dentry(file);
2113 struct inode *inode = d_inode(dentry);
2114 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2115 struct btrfs_root *root = BTRFS_I(inode)->root;
2116 struct btrfs_trans_handle *trans;
2117 struct btrfs_log_ctx ctx;
2118 int ret = 0, err;
2119
2120 trace_btrfs_sync_file(file, datasync);
2121
2122 btrfs_init_log_ctx(&ctx, inode);
2123
2124 /*
2125 * Set the range to full if the NO_HOLES feature is not enabled.
2126 * This is to avoid missing file extent items representing holes after
2127 * replaying the log.
2128 */
2129 if (!btrfs_fs_incompat(fs_info, NO_HOLES)) {
2130 start = 0;
2131 end = LLONG_MAX;
2132 }
2133
2134 /*
2135 * We write the dirty pages in the range and wait until they complete
2136 * out of the ->i_mutex. If so, we can flush the dirty pages by
2137 * multi-task, and make the performance up. See
2138 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2139 */
2140 ret = start_ordered_ops(inode, start, end);
2141 if (ret)
2142 goto out;
2143
2144 inode_lock(inode);
2145
2146 /*
2147 * We take the dio_sem here because the tree log stuff can race with
2148 * lockless dio writes and get an extent map logged for an extent we
2149 * never waited on. We need it this high up for lockdep reasons.
2150 */
2151 down_write(&BTRFS_I(inode)->dio_sem);
2152
2153 atomic_inc(&root->log_batch);
2154
2155 /*
2156 * If the inode needs a full sync, make sure we use a full range to
2157 * avoid log tree corruption, due to hole detection racing with ordered
2158 * extent completion for adjacent ranges and races between logging and
2159 * completion of ordered extents for adjancent ranges - both races
2160 * could lead to file extent items in the log with overlapping ranges.
2161 * Do this while holding the inode lock, to avoid races with other
2162 * tasks.
2163 */
2164 if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2165 &BTRFS_I(inode)->runtime_flags)) {
2166 start = 0;
2167 end = LLONG_MAX;
2168 }
2169
2170 /*
2171 * Before we acquired the inode's lock, someone may have dirtied more
2172 * pages in the target range. We need to make sure that writeback for
2173 * any such pages does not start while we are logging the inode, because
2174 * if it does, any of the following might happen when we are not doing a
2175 * full inode sync:
2176 *
2177 * 1) We log an extent after its writeback finishes but before its
2178 * checksums are added to the csum tree, leading to -EIO errors
2179 * when attempting to read the extent after a log replay.
2180 *
2181 * 2) We can end up logging an extent before its writeback finishes.
2182 * Therefore after the log replay we will have a file extent item
2183 * pointing to an unwritten extent (and no data checksums as well).
2184 *
2185 * So trigger writeback for any eventual new dirty pages and then we
2186 * wait for all ordered extents to complete below.
2187 */
2188 ret = start_ordered_ops(inode, start, end);
2189 if (ret) {
2190 up_write(&BTRFS_I(inode)->dio_sem);
2191 inode_unlock(inode);
2192 goto out;
2193 }
2194
2195 /*
2196 * We have to do this here to avoid the priority inversion of waiting on
2197 * IO of a lower priority task while holding a transaction open.
2198 *
2199 * Also, the range length can be represented by u64, we have to do the
2200 * typecasts to avoid signed overflow if it's [0, LLONG_MAX].
2201 */
2202 ret = btrfs_wait_ordered_range(inode, start, (u64)end - (u64)start + 1);
2203 if (ret) {
2204 up_write(&BTRFS_I(inode)->dio_sem);
2205 inode_unlock(inode);
2206 goto out;
2207 }
2208 atomic_inc(&root->log_batch);
2209
2210 smp_mb();
2211 if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2212 BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2213 /*
2214 * We've had everything committed since the last time we were
2215 * modified so clear this flag in case it was set for whatever
2216 * reason, it's no longer relevant.
2217 */
2218 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2219 &BTRFS_I(inode)->runtime_flags);
2220 /*
2221 * An ordered extent might have started before and completed
2222 * already with io errors, in which case the inode was not
2223 * updated and we end up here. So check the inode's mapping
2224 * for any errors that might have happened since we last
2225 * checked called fsync.
2226 */
2227 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2228 up_write(&BTRFS_I(inode)->dio_sem);
2229 inode_unlock(inode);
2230 goto out;
2231 }
2232
2233 /*
2234 * We use start here because we will need to wait on the IO to complete
2235 * in btrfs_sync_log, which could require joining a transaction (for
2236 * example checking cross references in the nocow path). If we use join
2237 * here we could get into a situation where we're waiting on IO to
2238 * happen that is blocked on a transaction trying to commit. With start
2239 * we inc the extwriter counter, so we wait for all extwriters to exit
2240 * before we start blocking joiners. This comment is to keep somebody
2241 * from thinking they are super smart and changing this to
2242 * btrfs_join_transaction *cough*Josef*cough*.
2243 */
2244 trans = btrfs_start_transaction(root, 0);
2245 if (IS_ERR(trans)) {
2246 ret = PTR_ERR(trans);
2247 up_write(&BTRFS_I(inode)->dio_sem);
2248 inode_unlock(inode);
2249 goto out;
2250 }
2251
2252 ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2253 if (ret < 0) {
2254 /* Fallthrough and commit/free transaction. */
2255 ret = 1;
2256 }
2257
2258 /* we've logged all the items and now have a consistent
2259 * version of the file in the log. It is possible that
2260 * someone will come in and modify the file, but that's
2261 * fine because the log is consistent on disk, and we
2262 * have references to all of the file's extents
2263 *
2264 * It is possible that someone will come in and log the
2265 * file again, but that will end up using the synchronization
2266 * inside btrfs_sync_log to keep things safe.
2267 */
2268 up_write(&BTRFS_I(inode)->dio_sem);
2269 inode_unlock(inode);
2270
2271 if (ret != BTRFS_NO_LOG_SYNC) {
2272 if (!ret) {
2273 ret = btrfs_sync_log(trans, root, &ctx);
2274 if (!ret) {
2275 ret = btrfs_end_transaction(trans);
2276 goto out;
2277 }
2278 }
2279 ret = btrfs_commit_transaction(trans);
2280 } else {
2281 ret = btrfs_end_transaction(trans);
2282 }
2283out:
2284 ASSERT(list_empty(&ctx.list));
2285 err = file_check_and_advance_wb_err(file);
2286 if (!ret)
2287 ret = err;
2288 return ret > 0 ? -EIO : ret;
2289}
2290
2291static const struct vm_operations_struct btrfs_file_vm_ops = {
2292 .fault = filemap_fault,
2293 .map_pages = filemap_map_pages,
2294 .page_mkwrite = btrfs_page_mkwrite,
2295};
2296
2297static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2298{
2299 struct address_space *mapping = filp->f_mapping;
2300
2301 if (!mapping->a_ops->readpage)
2302 return -ENOEXEC;
2303
2304 file_accessed(filp);
2305 vma->vm_ops = &btrfs_file_vm_ops;
2306
2307 return 0;
2308}
2309
2310static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2311 int slot, u64 start, u64 end)
2312{
2313 struct btrfs_file_extent_item *fi;
2314 struct btrfs_key key;
2315
2316 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2317 return 0;
2318
2319 btrfs_item_key_to_cpu(leaf, &key, slot);
2320 if (key.objectid != btrfs_ino(inode) ||
2321 key.type != BTRFS_EXTENT_DATA_KEY)
2322 return 0;
2323
2324 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2325
2326 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2327 return 0;
2328
2329 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2330 return 0;
2331
2332 if (key.offset == end)
2333 return 1;
2334 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2335 return 1;
2336 return 0;
2337}
2338
2339static int fill_holes(struct btrfs_trans_handle *trans,
2340 struct btrfs_inode *inode,
2341 struct btrfs_path *path, u64 offset, u64 end)
2342{
2343 struct btrfs_fs_info *fs_info = trans->fs_info;
2344 struct btrfs_root *root = inode->root;
2345 struct extent_buffer *leaf;
2346 struct btrfs_file_extent_item *fi;
2347 struct extent_map *hole_em;
2348 struct extent_map_tree *em_tree = &inode->extent_tree;
2349 struct btrfs_key key;
2350 int ret;
2351
2352 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2353 goto out;
2354
2355 key.objectid = btrfs_ino(inode);
2356 key.type = BTRFS_EXTENT_DATA_KEY;
2357 key.offset = offset;
2358
2359 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2360 if (ret <= 0) {
2361 /*
2362 * We should have dropped this offset, so if we find it then
2363 * something has gone horribly wrong.
2364 */
2365 if (ret == 0)
2366 ret = -EINVAL;
2367 return ret;
2368 }
2369
2370 leaf = path->nodes[0];
2371 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2372 u64 num_bytes;
2373
2374 path->slots[0]--;
2375 fi = btrfs_item_ptr(leaf, path->slots[0],
2376 struct btrfs_file_extent_item);
2377 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2378 end - offset;
2379 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2380 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2381 btrfs_set_file_extent_offset(leaf, fi, 0);
2382 btrfs_mark_buffer_dirty(leaf);
2383 goto out;
2384 }
2385
2386 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2387 u64 num_bytes;
2388
2389 key.offset = offset;
2390 btrfs_set_item_key_safe(fs_info, path, &key);
2391 fi = btrfs_item_ptr(leaf, path->slots[0],
2392 struct btrfs_file_extent_item);
2393 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2394 offset;
2395 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2396 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2397 btrfs_set_file_extent_offset(leaf, fi, 0);
2398 btrfs_mark_buffer_dirty(leaf);
2399 goto out;
2400 }
2401 btrfs_release_path(path);
2402
2403 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2404 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2405 if (ret)
2406 return ret;
2407
2408out:
2409 btrfs_release_path(path);
2410
2411 hole_em = alloc_extent_map();
2412 if (!hole_em) {
2413 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2414 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2415 } else {
2416 hole_em->start = offset;
2417 hole_em->len = end - offset;
2418 hole_em->ram_bytes = hole_em->len;
2419 hole_em->orig_start = offset;
2420
2421 hole_em->block_start = EXTENT_MAP_HOLE;
2422 hole_em->block_len = 0;
2423 hole_em->orig_block_len = 0;
2424 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2425 hole_em->generation = trans->transid;
2426
2427 do {
2428 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2429 write_lock(&em_tree->lock);
2430 ret = add_extent_mapping(em_tree, hole_em, 1);
2431 write_unlock(&em_tree->lock);
2432 } while (ret == -EEXIST);
2433 free_extent_map(hole_em);
2434 if (ret)
2435 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2436 &inode->runtime_flags);
2437 }
2438
2439 return 0;
2440}
2441
2442/*
2443 * Find a hole extent on given inode and change start/len to the end of hole
2444 * extent.(hole/vacuum extent whose em->start <= start &&
2445 * em->start + em->len > start)
2446 * When a hole extent is found, return 1 and modify start/len.
2447 */
2448static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2449{
2450 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2451 struct extent_map *em;
2452 int ret = 0;
2453
2454 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2455 round_down(*start, fs_info->sectorsize),
2456 round_up(*len, fs_info->sectorsize));
2457 if (IS_ERR(em))
2458 return PTR_ERR(em);
2459
2460 /* Hole or vacuum extent(only exists in no-hole mode) */
2461 if (em->block_start == EXTENT_MAP_HOLE) {
2462 ret = 1;
2463 *len = em->start + em->len > *start + *len ?
2464 0 : *start + *len - em->start - em->len;
2465 *start = em->start + em->len;
2466 }
2467 free_extent_map(em);
2468 return ret;
2469}
2470
2471static int btrfs_punch_hole_lock_range(struct inode *inode,
2472 const u64 lockstart,
2473 const u64 lockend,
2474 struct extent_state **cached_state)
2475{
2476 while (1) {
2477 struct btrfs_ordered_extent *ordered;
2478 int ret;
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->num_bytes <= lockstart ||
2493 ordered->file_offset > lockend)) &&
2494 !filemap_range_has_page(inode->i_mapping,
2495 lockstart, lockend)) {
2496 if (ordered)
2497 btrfs_put_ordered_extent(ordered);
2498 break;
2499 }
2500 if (ordered)
2501 btrfs_put_ordered_extent(ordered);
2502 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2503 lockend, cached_state);
2504 ret = btrfs_wait_ordered_range(inode, lockstart,
2505 lockend - lockstart + 1);
2506 if (ret)
2507 return ret;
2508 }
2509 return 0;
2510}
2511
2512static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2513 struct inode *inode,
2514 struct btrfs_path *path,
2515 struct btrfs_clone_extent_info *clone_info,
2516 const u64 clone_len)
2517{
2518 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2519 struct btrfs_root *root = BTRFS_I(inode)->root;
2520 struct btrfs_file_extent_item *extent;
2521 struct extent_buffer *leaf;
2522 struct btrfs_key key;
2523 int slot;
2524 struct btrfs_ref ref = { 0 };
2525 u64 ref_offset;
2526 int ret;
2527
2528 if (clone_len == 0)
2529 return 0;
2530
2531 if (clone_info->disk_offset == 0 &&
2532 btrfs_fs_incompat(fs_info, NO_HOLES))
2533 return 0;
2534
2535 key.objectid = btrfs_ino(BTRFS_I(inode));
2536 key.type = BTRFS_EXTENT_DATA_KEY;
2537 key.offset = clone_info->file_offset;
2538 ret = btrfs_insert_empty_item(trans, root, path, &key,
2539 clone_info->item_size);
2540 if (ret)
2541 return ret;
2542 leaf = path->nodes[0];
2543 slot = path->slots[0];
2544 write_extent_buffer(leaf, clone_info->extent_buf,
2545 btrfs_item_ptr_offset(leaf, slot),
2546 clone_info->item_size);
2547 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2548 btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2549 btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2550 btrfs_mark_buffer_dirty(leaf);
2551 btrfs_release_path(path);
2552
2553 ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
2554 clone_info->file_offset, clone_len);
2555 if (ret)
2556 return ret;
2557
2558 /* If it's a hole, nothing more needs to be done. */
2559 if (clone_info->disk_offset == 0)
2560 return 0;
2561
2562 inode_add_bytes(inode, clone_len);
2563 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2564 clone_info->disk_offset,
2565 clone_info->disk_len, 0);
2566 ref_offset = clone_info->file_offset - clone_info->data_offset;
2567 btrfs_init_data_ref(&ref, root->root_key.objectid,
2568 btrfs_ino(BTRFS_I(inode)), ref_offset);
2569 ret = btrfs_inc_extent_ref(trans, &ref);
2570
2571 return ret;
2572}
2573
2574/*
2575 * The respective range must have been previously locked, as well as the inode.
2576 * The end offset is inclusive (last byte of the range).
2577 * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2578 * cloning.
2579 * When cloning, we don't want to end up in a state where we dropped extents
2580 * without inserting a new one, so we must abort the transaction to avoid a
2581 * corruption.
2582 */
2583int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2584 const u64 start, const u64 end,
2585 struct btrfs_clone_extent_info *clone_info,
2586 struct btrfs_trans_handle **trans_out)
2587{
2588 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2589 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2590 u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2591 struct btrfs_root *root = BTRFS_I(inode)->root;
2592 struct btrfs_trans_handle *trans = NULL;
2593 struct btrfs_block_rsv *rsv;
2594 unsigned int rsv_count;
2595 u64 cur_offset;
2596 u64 drop_end;
2597 u64 len = end - start;
2598 int ret = 0;
2599
2600 if (end <= start)
2601 return -EINVAL;
2602
2603 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2604 if (!rsv) {
2605 ret = -ENOMEM;
2606 goto out;
2607 }
2608 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2609 rsv->failfast = 1;
2610
2611 /*
2612 * 1 - update the inode
2613 * 1 - removing the extents in the range
2614 * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2615 * an extent
2616 */
2617 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2618 rsv_count = 3;
2619 else
2620 rsv_count = 2;
2621
2622 trans = btrfs_start_transaction(root, rsv_count);
2623 if (IS_ERR(trans)) {
2624 ret = PTR_ERR(trans);
2625 trans = NULL;
2626 goto out_free;
2627 }
2628
2629 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2630 min_size, false);
2631 BUG_ON(ret);
2632 trans->block_rsv = rsv;
2633
2634 cur_offset = start;
2635 while (cur_offset < end) {
2636 ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path,
2637 cur_offset, end + 1, &drop_end,
2638 1, 0, 0, NULL);
2639 if (ret != -ENOSPC) {
2640 /*
2641 * When cloning we want to avoid transaction aborts when
2642 * nothing was done and we are attempting to clone parts
2643 * of inline extents, in such cases -EOPNOTSUPP is
2644 * returned by __btrfs_drop_extents() without having
2645 * changed anything in the file.
2646 */
2647 if (clone_info && ret && ret != -EOPNOTSUPP)
2648 btrfs_abort_transaction(trans, ret);
2649 break;
2650 }
2651
2652 trans->block_rsv = &fs_info->trans_block_rsv;
2653
2654 if (!clone_info && cur_offset < drop_end &&
2655 cur_offset < ino_size) {
2656 ret = fill_holes(trans, BTRFS_I(inode), path,
2657 cur_offset, drop_end);
2658 if (ret) {
2659 /*
2660 * If we failed then we didn't insert our hole
2661 * entries for the area we dropped, so now the
2662 * fs is corrupted, so we must abort the
2663 * transaction.
2664 */
2665 btrfs_abort_transaction(trans, ret);
2666 break;
2667 }
2668 } else if (!clone_info && cur_offset < drop_end) {
2669 /*
2670 * We are past the i_size here, but since we didn't
2671 * insert holes we need to clear the mapped area so we
2672 * know to not set disk_i_size in this area until a new
2673 * file extent is inserted here.
2674 */
2675 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2676 cur_offset, drop_end - cur_offset);
2677 if (ret) {
2678 /*
2679 * We couldn't clear our area, so we could
2680 * presumably adjust up and corrupt the fs, so
2681 * we need to abort.
2682 */
2683 btrfs_abort_transaction(trans, ret);
2684 break;
2685 }
2686 }
2687
2688 if (clone_info && drop_end > clone_info->file_offset) {
2689 u64 clone_len = drop_end - clone_info->file_offset;
2690
2691 ret = btrfs_insert_clone_extent(trans, inode, path,
2692 clone_info, clone_len);
2693 if (ret) {
2694 btrfs_abort_transaction(trans, ret);
2695 break;
2696 }
2697 clone_info->data_len -= clone_len;
2698 clone_info->data_offset += clone_len;
2699 clone_info->file_offset += clone_len;
2700 }
2701
2702 cur_offset = drop_end;
2703
2704 ret = btrfs_update_inode(trans, root, inode);
2705 if (ret)
2706 break;
2707
2708 btrfs_end_transaction(trans);
2709 btrfs_btree_balance_dirty(fs_info);
2710
2711 trans = btrfs_start_transaction(root, rsv_count);
2712 if (IS_ERR(trans)) {
2713 ret = PTR_ERR(trans);
2714 trans = NULL;
2715 break;
2716 }
2717
2718 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2719 rsv, min_size, false);
2720 BUG_ON(ret); /* shouldn't happen */
2721 trans->block_rsv = rsv;
2722
2723 if (!clone_info) {
2724 ret = find_first_non_hole(inode, &cur_offset, &len);
2725 if (unlikely(ret < 0))
2726 break;
2727 if (ret && !len) {
2728 ret = 0;
2729 break;
2730 }
2731 }
2732 }
2733
2734 /*
2735 * If we were cloning, force the next fsync to be a full one since we
2736 * we replaced (or just dropped in the case of cloning holes when
2737 * NO_HOLES is enabled) extents and extent maps.
2738 * This is for the sake of simplicity, and cloning into files larger
2739 * than 16Mb would force the full fsync any way (when
2740 * try_release_extent_mapping() is invoked during page cache truncation.
2741 */
2742 if (clone_info)
2743 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2744 &BTRFS_I(inode)->runtime_flags);
2745
2746 if (ret)
2747 goto out_trans;
2748
2749 trans->block_rsv = &fs_info->trans_block_rsv;
2750 /*
2751 * If we are using the NO_HOLES feature we might have had already an
2752 * hole that overlaps a part of the region [lockstart, lockend] and
2753 * ends at (or beyond) lockend. Since we have no file extent items to
2754 * represent holes, drop_end can be less than lockend and so we must
2755 * make sure we have an extent map representing the existing hole (the
2756 * call to __btrfs_drop_extents() might have dropped the existing extent
2757 * map representing the existing hole), otherwise the fast fsync path
2758 * will not record the existence of the hole region
2759 * [existing_hole_start, lockend].
2760 */
2761 if (drop_end <= end)
2762 drop_end = end + 1;
2763 /*
2764 * Don't insert file hole extent item if it's for a range beyond eof
2765 * (because it's useless) or if it represents a 0 bytes range (when
2766 * cur_offset == drop_end).
2767 */
2768 if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2769 ret = fill_holes(trans, BTRFS_I(inode), path,
2770 cur_offset, drop_end);
2771 if (ret) {
2772 /* Same comment as above. */
2773 btrfs_abort_transaction(trans, ret);
2774 goto out_trans;
2775 }
2776 } else if (!clone_info && cur_offset < drop_end) {
2777 /* See the comment in the loop above for the reasoning here. */
2778 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2779 cur_offset, drop_end - cur_offset);
2780 if (ret) {
2781 btrfs_abort_transaction(trans, ret);
2782 goto out_trans;
2783 }
2784
2785 }
2786 if (clone_info) {
2787 ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2788 clone_info->data_len);
2789 if (ret) {
2790 btrfs_abort_transaction(trans, ret);
2791 goto out_trans;
2792 }
2793 }
2794
2795out_trans:
2796 if (!trans)
2797 goto out_free;
2798
2799 trans->block_rsv = &fs_info->trans_block_rsv;
2800 if (ret)
2801 btrfs_end_transaction(trans);
2802 else
2803 *trans_out = trans;
2804out_free:
2805 btrfs_free_block_rsv(fs_info, rsv);
2806out:
2807 return ret;
2808}
2809
2810static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2811{
2812 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2813 struct btrfs_root *root = BTRFS_I(inode)->root;
2814 struct extent_state *cached_state = NULL;
2815 struct btrfs_path *path;
2816 struct btrfs_trans_handle *trans = NULL;
2817 u64 lockstart;
2818 u64 lockend;
2819 u64 tail_start;
2820 u64 tail_len;
2821 u64 orig_start = offset;
2822 int ret = 0;
2823 bool same_block;
2824 u64 ino_size;
2825 bool truncated_block = false;
2826 bool updated_inode = false;
2827
2828 ret = btrfs_wait_ordered_range(inode, offset, len);
2829 if (ret)
2830 return ret;
2831
2832 inode_lock(inode);
2833 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2834 ret = find_first_non_hole(inode, &offset, &len);
2835 if (ret < 0)
2836 goto out_only_mutex;
2837 if (ret && !len) {
2838 /* Already in a large hole */
2839 ret = 0;
2840 goto out_only_mutex;
2841 }
2842
2843 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2844 lockend = round_down(offset + len,
2845 btrfs_inode_sectorsize(inode)) - 1;
2846 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2847 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2848 /*
2849 * We needn't truncate any block which is beyond the end of the file
2850 * because we are sure there is no data there.
2851 */
2852 /*
2853 * Only do this if we are in the same block and we aren't doing the
2854 * entire block.
2855 */
2856 if (same_block && len < fs_info->sectorsize) {
2857 if (offset < ino_size) {
2858 truncated_block = true;
2859 ret = btrfs_truncate_block(inode, offset, len, 0);
2860 } else {
2861 ret = 0;
2862 }
2863 goto out_only_mutex;
2864 }
2865
2866 /* zero back part of the first block */
2867 if (offset < ino_size) {
2868 truncated_block = true;
2869 ret = btrfs_truncate_block(inode, offset, 0, 0);
2870 if (ret) {
2871 inode_unlock(inode);
2872 return ret;
2873 }
2874 }
2875
2876 /* Check the aligned pages after the first unaligned page,
2877 * if offset != orig_start, which means the first unaligned page
2878 * including several following pages are already in holes,
2879 * the extra check can be skipped */
2880 if (offset == orig_start) {
2881 /* after truncate page, check hole again */
2882 len = offset + len - lockstart;
2883 offset = lockstart;
2884 ret = find_first_non_hole(inode, &offset, &len);
2885 if (ret < 0)
2886 goto out_only_mutex;
2887 if (ret && !len) {
2888 ret = 0;
2889 goto out_only_mutex;
2890 }
2891 lockstart = offset;
2892 }
2893
2894 /* Check the tail unaligned part is in a hole */
2895 tail_start = lockend + 1;
2896 tail_len = offset + len - tail_start;
2897 if (tail_len) {
2898 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2899 if (unlikely(ret < 0))
2900 goto out_only_mutex;
2901 if (!ret) {
2902 /* zero the front end of the last page */
2903 if (tail_start + tail_len < ino_size) {
2904 truncated_block = true;
2905 ret = btrfs_truncate_block(inode,
2906 tail_start + tail_len,
2907 0, 1);
2908 if (ret)
2909 goto out_only_mutex;
2910 }
2911 }
2912 }
2913
2914 if (lockend < lockstart) {
2915 ret = 0;
2916 goto out_only_mutex;
2917 }
2918
2919 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2920 &cached_state);
2921 if (ret)
2922 goto out_only_mutex;
2923
2924 path = btrfs_alloc_path();
2925 if (!path) {
2926 ret = -ENOMEM;
2927 goto out;
2928 }
2929
2930 ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2931 &trans);
2932 btrfs_free_path(path);
2933 if (ret)
2934 goto out;
2935
2936 ASSERT(trans != NULL);
2937 inode_inc_iversion(inode);
2938 inode->i_mtime = inode->i_ctime = current_time(inode);
2939 ret = btrfs_update_inode(trans, root, inode);
2940 updated_inode = true;
2941 btrfs_end_transaction(trans);
2942 btrfs_btree_balance_dirty(fs_info);
2943out:
2944 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2945 &cached_state);
2946out_only_mutex:
2947 if (!updated_inode && truncated_block && !ret) {
2948 /*
2949 * If we only end up zeroing part of a page, we still need to
2950 * update the inode item, so that all the time fields are
2951 * updated as well as the necessary btrfs inode in memory fields
2952 * for detecting, at fsync time, if the inode isn't yet in the
2953 * log tree or it's there but not up to date.
2954 */
2955 struct timespec64 now = current_time(inode);
2956
2957 inode_inc_iversion(inode);
2958 inode->i_mtime = now;
2959 inode->i_ctime = now;
2960 trans = btrfs_start_transaction(root, 1);
2961 if (IS_ERR(trans)) {
2962 ret = PTR_ERR(trans);
2963 } else {
2964 int ret2;
2965
2966 ret = btrfs_update_inode(trans, root, inode);
2967 ret2 = btrfs_end_transaction(trans);
2968 if (!ret)
2969 ret = ret2;
2970 }
2971 }
2972 inode_unlock(inode);
2973 return ret;
2974}
2975
2976/* Helper structure to record which range is already reserved */
2977struct falloc_range {
2978 struct list_head list;
2979 u64 start;
2980 u64 len;
2981};
2982
2983/*
2984 * Helper function to add falloc range
2985 *
2986 * Caller should have locked the larger range of extent containing
2987 * [start, len)
2988 */
2989static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2990{
2991 struct falloc_range *prev = NULL;
2992 struct falloc_range *range = NULL;
2993
2994 if (list_empty(head))
2995 goto insert;
2996
2997 /*
2998 * As fallocate iterate by bytenr order, we only need to check
2999 * the last range.
3000 */
3001 prev = list_entry(head->prev, struct falloc_range, list);
3002 if (prev->start + prev->len == start) {
3003 prev->len += len;
3004 return 0;
3005 }
3006insert:
3007 range = kmalloc(sizeof(*range), GFP_KERNEL);
3008 if (!range)
3009 return -ENOMEM;
3010 range->start = start;
3011 range->len = len;
3012 list_add_tail(&range->list, head);
3013 return 0;
3014}
3015
3016static int btrfs_fallocate_update_isize(struct inode *inode,
3017 const u64 end,
3018 const int mode)
3019{
3020 struct btrfs_trans_handle *trans;
3021 struct btrfs_root *root = BTRFS_I(inode)->root;
3022 int ret;
3023 int ret2;
3024
3025 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3026 return 0;
3027
3028 trans = btrfs_start_transaction(root, 1);
3029 if (IS_ERR(trans))
3030 return PTR_ERR(trans);
3031
3032 inode->i_ctime = current_time(inode);
3033 i_size_write(inode, end);
3034 btrfs_inode_safe_disk_i_size_write(inode, 0);
3035 ret = btrfs_update_inode(trans, root, inode);
3036 ret2 = btrfs_end_transaction(trans);
3037
3038 return ret ? ret : ret2;
3039}
3040
3041enum {
3042 RANGE_BOUNDARY_WRITTEN_EXTENT,
3043 RANGE_BOUNDARY_PREALLOC_EXTENT,
3044 RANGE_BOUNDARY_HOLE,
3045};
3046
3047static int btrfs_zero_range_check_range_boundary(struct inode *inode,
3048 u64 offset)
3049{
3050 const u64 sectorsize = btrfs_inode_sectorsize(inode);
3051 struct extent_map *em;
3052 int ret;
3053
3054 offset = round_down(offset, sectorsize);
3055 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize);
3056 if (IS_ERR(em))
3057 return PTR_ERR(em);
3058
3059 if (em->block_start == EXTENT_MAP_HOLE)
3060 ret = RANGE_BOUNDARY_HOLE;
3061 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3062 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3063 else
3064 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3065
3066 free_extent_map(em);
3067 return ret;
3068}
3069
3070static int btrfs_zero_range(struct inode *inode,
3071 loff_t offset,
3072 loff_t len,
3073 const int mode)
3074{
3075 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3076 struct extent_map *em;
3077 struct extent_changeset *data_reserved = NULL;
3078 int ret;
3079 u64 alloc_hint = 0;
3080 const u64 sectorsize = btrfs_inode_sectorsize(inode);
3081 u64 alloc_start = round_down(offset, sectorsize);
3082 u64 alloc_end = round_up(offset + len, sectorsize);
3083 u64 bytes_to_reserve = 0;
3084 bool space_reserved = false;
3085
3086 inode_dio_wait(inode);
3087
3088 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3089 alloc_end - alloc_start);
3090 if (IS_ERR(em)) {
3091 ret = PTR_ERR(em);
3092 goto out;
3093 }
3094
3095 /*
3096 * Avoid hole punching and extent allocation for some cases. More cases
3097 * could be considered, but these are unlikely common and we keep things
3098 * as simple as possible for now. Also, intentionally, if the target
3099 * range contains one or more prealloc extents together with regular
3100 * extents and holes, we drop all the existing extents and allocate a
3101 * new prealloc extent, so that we get a larger contiguous disk extent.
3102 */
3103 if (em->start <= alloc_start &&
3104 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3105 const u64 em_end = em->start + em->len;
3106
3107 if (em_end >= offset + len) {
3108 /*
3109 * The whole range is already a prealloc extent,
3110 * do nothing except updating the inode's i_size if
3111 * needed.
3112 */
3113 free_extent_map(em);
3114 ret = btrfs_fallocate_update_isize(inode, offset + len,
3115 mode);
3116 goto out;
3117 }
3118 /*
3119 * Part of the range is already a prealloc extent, so operate
3120 * only on the remaining part of the range.
3121 */
3122 alloc_start = em_end;
3123 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3124 len = offset + len - alloc_start;
3125 offset = alloc_start;
3126 alloc_hint = em->block_start + em->len;
3127 }
3128 free_extent_map(em);
3129
3130 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3131 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3132 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3133 sectorsize);
3134 if (IS_ERR(em)) {
3135 ret = PTR_ERR(em);
3136 goto out;
3137 }
3138
3139 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3140 free_extent_map(em);
3141 ret = btrfs_fallocate_update_isize(inode, offset + len,
3142 mode);
3143 goto out;
3144 }
3145 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3146 free_extent_map(em);
3147 ret = btrfs_truncate_block(inode, offset, len, 0);
3148 if (!ret)
3149 ret = btrfs_fallocate_update_isize(inode,
3150 offset + len,
3151 mode);
3152 return ret;
3153 }
3154 free_extent_map(em);
3155 alloc_start = round_down(offset, sectorsize);
3156 alloc_end = alloc_start + sectorsize;
3157 goto reserve_space;
3158 }
3159
3160 alloc_start = round_up(offset, sectorsize);
3161 alloc_end = round_down(offset + len, sectorsize);
3162
3163 /*
3164 * For unaligned ranges, check the pages at the boundaries, they might
3165 * map to an extent, in which case we need to partially zero them, or
3166 * they might map to a hole, in which case we need our allocation range
3167 * to cover them.
3168 */
3169 if (!IS_ALIGNED(offset, sectorsize)) {
3170 ret = btrfs_zero_range_check_range_boundary(inode, offset);
3171 if (ret < 0)
3172 goto out;
3173 if (ret == RANGE_BOUNDARY_HOLE) {
3174 alloc_start = round_down(offset, sectorsize);
3175 ret = 0;
3176 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3177 ret = btrfs_truncate_block(inode, offset, 0, 0);
3178 if (ret)
3179 goto out;
3180 } else {
3181 ret = 0;
3182 }
3183 }
3184
3185 if (!IS_ALIGNED(offset + len, sectorsize)) {
3186 ret = btrfs_zero_range_check_range_boundary(inode,
3187 offset + len);
3188 if (ret < 0)
3189 goto out;
3190 if (ret == RANGE_BOUNDARY_HOLE) {
3191 alloc_end = round_up(offset + len, sectorsize);
3192 ret = 0;
3193 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3194 ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3195 if (ret)
3196 goto out;
3197 } else {
3198 ret = 0;
3199 }
3200 }
3201
3202reserve_space:
3203 if (alloc_start < alloc_end) {
3204 struct extent_state *cached_state = NULL;
3205 const u64 lockstart = alloc_start;
3206 const u64 lockend = alloc_end - 1;
3207
3208 bytes_to_reserve = alloc_end - alloc_start;
3209 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3210 bytes_to_reserve);
3211 if (ret < 0)
3212 goto out;
3213 space_reserved = true;
3214 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3215 &cached_state);
3216 if (ret)
3217 goto out;
3218 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3219 alloc_start, bytes_to_reserve);
3220 if (ret)
3221 goto out;
3222 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3223 alloc_end - alloc_start,
3224 i_blocksize(inode),
3225 offset + len, &alloc_hint);
3226 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3227 lockend, &cached_state);
3228 /* btrfs_prealloc_file_range releases reserved space on error */
3229 if (ret) {
3230 space_reserved = false;
3231 goto out;
3232 }
3233 }
3234 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3235 out:
3236 if (ret && space_reserved)
3237 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3238 alloc_start, bytes_to_reserve);
3239 extent_changeset_free(data_reserved);
3240
3241 return ret;
3242}
3243
3244static long btrfs_fallocate(struct file *file, int mode,
3245 loff_t offset, loff_t len)
3246{
3247 struct inode *inode = file_inode(file);
3248 struct extent_state *cached_state = NULL;
3249 struct extent_changeset *data_reserved = NULL;
3250 struct falloc_range *range;
3251 struct falloc_range *tmp;
3252 struct list_head reserve_list;
3253 u64 cur_offset;
3254 u64 last_byte;
3255 u64 alloc_start;
3256 u64 alloc_end;
3257 u64 alloc_hint = 0;
3258 u64 locked_end;
3259 u64 actual_end = 0;
3260 struct extent_map *em;
3261 int blocksize = btrfs_inode_sectorsize(inode);
3262 int ret;
3263
3264 alloc_start = round_down(offset, blocksize);
3265 alloc_end = round_up(offset + len, blocksize);
3266 cur_offset = alloc_start;
3267
3268 /* Make sure we aren't being give some crap mode */
3269 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3270 FALLOC_FL_ZERO_RANGE))
3271 return -EOPNOTSUPP;
3272
3273 if (mode & FALLOC_FL_PUNCH_HOLE)
3274 return btrfs_punch_hole(inode, offset, len);
3275
3276 /*
3277 * Only trigger disk allocation, don't trigger qgroup reserve
3278 *
3279 * For qgroup space, it will be checked later.
3280 */
3281 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3282 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3283 alloc_end - alloc_start);
3284 if (ret < 0)
3285 return ret;
3286 }
3287
3288 inode_lock(inode);
3289
3290 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3291 ret = inode_newsize_ok(inode, offset + len);
3292 if (ret)
3293 goto out;
3294 }
3295
3296 /*
3297 * TODO: Move these two operations after we have checked
3298 * accurate reserved space, or fallocate can still fail but
3299 * with page truncated or size expanded.
3300 *
3301 * But that's a minor problem and won't do much harm BTW.
3302 */
3303 if (alloc_start > inode->i_size) {
3304 ret = btrfs_cont_expand(inode, i_size_read(inode),
3305 alloc_start);
3306 if (ret)
3307 goto out;
3308 } else if (offset + len > inode->i_size) {
3309 /*
3310 * If we are fallocating from the end of the file onward we
3311 * need to zero out the end of the block if i_size lands in the
3312 * middle of a block.
3313 */
3314 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3315 if (ret)
3316 goto out;
3317 }
3318
3319 /*
3320 * wait for ordered IO before we have any locks. We'll loop again
3321 * below with the locks held.
3322 */
3323 ret = btrfs_wait_ordered_range(inode, alloc_start,
3324 alloc_end - alloc_start);
3325 if (ret)
3326 goto out;
3327
3328 if (mode & FALLOC_FL_ZERO_RANGE) {
3329 ret = btrfs_zero_range(inode, offset, len, mode);
3330 inode_unlock(inode);
3331 return ret;
3332 }
3333
3334 locked_end = alloc_end - 1;
3335 while (1) {
3336 struct btrfs_ordered_extent *ordered;
3337
3338 /* the extent lock is ordered inside the running
3339 * transaction
3340 */
3341 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3342 locked_end, &cached_state);
3343 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3344
3345 if (ordered &&
3346 ordered->file_offset + ordered->num_bytes > alloc_start &&
3347 ordered->file_offset < alloc_end) {
3348 btrfs_put_ordered_extent(ordered);
3349 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3350 alloc_start, locked_end,
3351 &cached_state);
3352 /*
3353 * we can't wait on the range with the transaction
3354 * running or with the extent lock held
3355 */
3356 ret = btrfs_wait_ordered_range(inode, alloc_start,
3357 alloc_end - alloc_start);
3358 if (ret)
3359 goto out;
3360 } else {
3361 if (ordered)
3362 btrfs_put_ordered_extent(ordered);
3363 break;
3364 }
3365 }
3366
3367 /* First, check if we exceed the qgroup limit */
3368 INIT_LIST_HEAD(&reserve_list);
3369 while (cur_offset < alloc_end) {
3370 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3371 alloc_end - cur_offset);
3372 if (IS_ERR(em)) {
3373 ret = PTR_ERR(em);
3374 break;
3375 }
3376 last_byte = min(extent_map_end(em), alloc_end);
3377 actual_end = min_t(u64, extent_map_end(em), offset + len);
3378 last_byte = ALIGN(last_byte, blocksize);
3379 if (em->block_start == EXTENT_MAP_HOLE ||
3380 (cur_offset >= inode->i_size &&
3381 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3382 ret = add_falloc_range(&reserve_list, cur_offset,
3383 last_byte - cur_offset);
3384 if (ret < 0) {
3385 free_extent_map(em);
3386 break;
3387 }
3388 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3389 &data_reserved, cur_offset,
3390 last_byte - cur_offset);
3391 if (ret < 0) {
3392 cur_offset = last_byte;
3393 free_extent_map(em);
3394 break;
3395 }
3396 } else {
3397 /*
3398 * Do not need to reserve unwritten extent for this
3399 * range, free reserved data space first, otherwise
3400 * it'll result in false ENOSPC error.
3401 */
3402 btrfs_free_reserved_data_space(BTRFS_I(inode),
3403 data_reserved, cur_offset,
3404 last_byte - cur_offset);
3405 }
3406 free_extent_map(em);
3407 cur_offset = last_byte;
3408 }
3409
3410 /*
3411 * If ret is still 0, means we're OK to fallocate.
3412 * Or just cleanup the list and exit.
3413 */
3414 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3415 if (!ret)
3416 ret = btrfs_prealloc_file_range(inode, mode,
3417 range->start,
3418 range->len, i_blocksize(inode),
3419 offset + len, &alloc_hint);
3420 else
3421 btrfs_free_reserved_data_space(BTRFS_I(inode),
3422 data_reserved, range->start,
3423 range->len);
3424 list_del(&range->list);
3425 kfree(range);
3426 }
3427 if (ret < 0)
3428 goto out_unlock;
3429
3430 /*
3431 * We didn't need to allocate any more space, but we still extended the
3432 * size of the file so we need to update i_size and the inode item.
3433 */
3434 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3435out_unlock:
3436 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3437 &cached_state);
3438out:
3439 inode_unlock(inode);
3440 /* Let go of our reservation. */
3441 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3442 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3443 cur_offset, alloc_end - cur_offset);
3444 extent_changeset_free(data_reserved);
3445 return ret;
3446}
3447
3448static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3449 int whence)
3450{
3451 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3452 struct extent_map *em = NULL;
3453 struct extent_state *cached_state = NULL;
3454 loff_t i_size = inode->i_size;
3455 u64 lockstart;
3456 u64 lockend;
3457 u64 start;
3458 u64 len;
3459 int ret = 0;
3460
3461 if (i_size == 0 || offset >= i_size)
3462 return -ENXIO;
3463
3464 /*
3465 * offset can be negative, in this case we start finding DATA/HOLE from
3466 * the very start of the file.
3467 */
3468 start = max_t(loff_t, 0, offset);
3469
3470 lockstart = round_down(start, fs_info->sectorsize);
3471 lockend = round_up(i_size, fs_info->sectorsize);
3472 if (lockend <= lockstart)
3473 lockend = lockstart + fs_info->sectorsize;
3474 lockend--;
3475 len = lockend - lockstart + 1;
3476
3477 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3478 &cached_state);
3479
3480 while (start < i_size) {
3481 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3482 if (IS_ERR(em)) {
3483 ret = PTR_ERR(em);
3484 em = NULL;
3485 break;
3486 }
3487
3488 if (whence == SEEK_HOLE &&
3489 (em->block_start == EXTENT_MAP_HOLE ||
3490 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3491 break;
3492 else if (whence == SEEK_DATA &&
3493 (em->block_start != EXTENT_MAP_HOLE &&
3494 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3495 break;
3496
3497 start = em->start + em->len;
3498 free_extent_map(em);
3499 em = NULL;
3500 cond_resched();
3501 }
3502 free_extent_map(em);
3503 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3504 &cached_state);
3505 if (ret) {
3506 offset = ret;
3507 } else {
3508 if (whence == SEEK_DATA && start >= i_size)
3509 offset = -ENXIO;
3510 else
3511 offset = min_t(loff_t, start, i_size);
3512 }
3513
3514 return offset;
3515}
3516
3517static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3518{
3519 struct inode *inode = file->f_mapping->host;
3520
3521 switch (whence) {
3522 default:
3523 return generic_file_llseek(file, offset, whence);
3524 case SEEK_DATA:
3525 case SEEK_HOLE:
3526 inode_lock_shared(inode);
3527 offset = find_desired_extent(inode, offset, whence);
3528 inode_unlock_shared(inode);
3529 break;
3530 }
3531
3532 if (offset < 0)
3533 return offset;
3534
3535 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3536}
3537
3538static int btrfs_file_open(struct inode *inode, struct file *filp)
3539{
3540 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3541 return generic_file_open(inode, filp);
3542}
3543
3544const struct file_operations btrfs_file_operations = {
3545 .llseek = btrfs_file_llseek,
3546 .read_iter = generic_file_read_iter,
3547 .splice_read = generic_file_splice_read,
3548 .write_iter = btrfs_file_write_iter,
3549 .splice_write = iter_file_splice_write,
3550 .mmap = btrfs_file_mmap,
3551 .open = btrfs_file_open,
3552 .release = btrfs_release_file,
3553 .fsync = btrfs_sync_file,
3554 .fallocate = btrfs_fallocate,
3555 .unlocked_ioctl = btrfs_ioctl,
3556#ifdef CONFIG_COMPAT
3557 .compat_ioctl = btrfs_compat_ioctl,
3558#endif
3559 .remap_file_range = btrfs_remap_file_range,
3560};
3561
3562void __cold btrfs_auto_defrag_exit(void)
3563{
3564 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3565}
3566
3567int __init btrfs_auto_defrag_init(void)
3568{
3569 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3570 sizeof(struct inode_defrag), 0,
3571 SLAB_MEM_SPREAD,
3572 NULL);
3573 if (!btrfs_inode_defrag_cachep)
3574 return -ENOMEM;
3575
3576 return 0;
3577}
3578
3579int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3580{
3581 int ret;
3582
3583 /*
3584 * So with compression we will find and lock a dirty page and clear the
3585 * first one as dirty, setup an async extent, and immediately return
3586 * with the entire range locked but with nobody actually marked with
3587 * writeback. So we can't just filemap_write_and_wait_range() and
3588 * expect it to work since it will just kick off a thread to do the
3589 * actual work. So we need to call filemap_fdatawrite_range _again_
3590 * since it will wait on the page lock, which won't be unlocked until
3591 * after the pages have been marked as writeback and so we're good to go
3592 * from there. We have to do this otherwise we'll miss the ordered
3593 * extents and that results in badness. Please Josef, do not think you
3594 * know better and pull this out at some point in the future, it is
3595 * right and you are wrong.
3596 */
3597 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3598 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3599 &BTRFS_I(inode)->runtime_flags))
3600 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3601
3602 return ret;
3603}