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