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