xref: /openbmc/linux/fs/btrfs/file.c (revision a20eefae)
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 
31 static struct kmem_cache *btrfs_inode_defrag_cachep;
32 /*
33  * when auto defrag is enabled we
34  * queue up these defrag structs to remember which
35  * inodes need defragging passes
36  */
37 struct inode_defrag {
38 	struct rb_node rb_node;
39 	/* objectid */
40 	u64 ino;
41 	/*
42 	 * transid where the defrag was added, we search for
43 	 * extents newer than this
44 	 */
45 	u64 transid;
46 
47 	/* root objectid */
48 	u64 root;
49 
50 	/* last offset we were able to defrag */
51 	u64 last_offset;
52 
53 	/* if we've wrapped around back to zero once already */
54 	int cycled;
55 };
56 
57 static int __compare_inode_defrag(struct inode_defrag *defrag1,
58 				  struct inode_defrag *defrag2)
59 {
60 	if (defrag1->root > defrag2->root)
61 		return 1;
62 	else if (defrag1->root < defrag2->root)
63 		return -1;
64 	else if (defrag1->ino > defrag2->ino)
65 		return 1;
66 	else if (defrag1->ino < defrag2->ino)
67 		return -1;
68 	else
69 		return 0;
70 }
71 
72 /* pop a record for an inode into the defrag tree.  The lock
73  * must be held already
74  *
75  * If you're inserting a record for an older transid than an
76  * existing record, the transid already in the tree is lowered
77  *
78  * If an existing record is found the defrag item you
79  * pass in is freed
80  */
81 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
82 				    struct inode_defrag *defrag)
83 {
84 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
85 	struct inode_defrag *entry;
86 	struct rb_node **p;
87 	struct rb_node *parent = NULL;
88 	int ret;
89 
90 	p = &fs_info->defrag_inodes.rb_node;
91 	while (*p) {
92 		parent = *p;
93 		entry = rb_entry(parent, struct inode_defrag, rb_node);
94 
95 		ret = __compare_inode_defrag(defrag, entry);
96 		if (ret < 0)
97 			p = &parent->rb_left;
98 		else if (ret > 0)
99 			p = &parent->rb_right;
100 		else {
101 			/* if we're reinserting an entry for
102 			 * an old defrag run, make sure to
103 			 * lower the transid of our existing record
104 			 */
105 			if (defrag->transid < entry->transid)
106 				entry->transid = defrag->transid;
107 			if (defrag->last_offset > entry->last_offset)
108 				entry->last_offset = defrag->last_offset;
109 			return -EEXIST;
110 		}
111 	}
112 	set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
113 	rb_link_node(&defrag->rb_node, parent, p);
114 	rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
115 	return 0;
116 }
117 
118 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
119 {
120 	if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
121 		return 0;
122 
123 	if (btrfs_fs_closing(fs_info))
124 		return 0;
125 
126 	return 1;
127 }
128 
129 /*
130  * insert a defrag record for this inode if auto defrag is
131  * enabled
132  */
133 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
134 			   struct btrfs_inode *inode)
135 {
136 	struct btrfs_root *root = inode->root;
137 	struct btrfs_fs_info *fs_info = root->fs_info;
138 	struct inode_defrag *defrag;
139 	u64 transid;
140 	int ret;
141 
142 	if (!__need_auto_defrag(fs_info))
143 		return 0;
144 
145 	if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
146 		return 0;
147 
148 	if (trans)
149 		transid = trans->transid;
150 	else
151 		transid = inode->root->last_trans;
152 
153 	defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
154 	if (!defrag)
155 		return -ENOMEM;
156 
157 	defrag->ino = btrfs_ino(inode);
158 	defrag->transid = transid;
159 	defrag->root = root->root_key.objectid;
160 
161 	spin_lock(&fs_info->defrag_inodes_lock);
162 	if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
163 		/*
164 		 * If we set IN_DEFRAG flag and evict the inode from memory,
165 		 * and then re-read this inode, this new inode doesn't have
166 		 * IN_DEFRAG flag. At the case, we may find the existed defrag.
167 		 */
168 		ret = __btrfs_add_inode_defrag(inode, defrag);
169 		if (ret)
170 			kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
171 	} else {
172 		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
173 	}
174 	spin_unlock(&fs_info->defrag_inodes_lock);
175 	return 0;
176 }
177 
178 /*
179  * Requeue the defrag object. If there is a defrag object that points to
180  * the same inode in the tree, we will merge them together (by
181  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
182  */
183 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
184 				       struct inode_defrag *defrag)
185 {
186 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
187 	int ret;
188 
189 	if (!__need_auto_defrag(fs_info))
190 		goto out;
191 
192 	/*
193 	 * Here we don't check the IN_DEFRAG flag, because we need merge
194 	 * them together.
195 	 */
196 	spin_lock(&fs_info->defrag_inodes_lock);
197 	ret = __btrfs_add_inode_defrag(inode, defrag);
198 	spin_unlock(&fs_info->defrag_inodes_lock);
199 	if (ret)
200 		goto out;
201 	return;
202 out:
203 	kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
204 }
205 
206 /*
207  * pick the defragable inode that we want, if it doesn't exist, we will get
208  * the next one.
209  */
210 static struct inode_defrag *
211 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
212 {
213 	struct inode_defrag *entry = NULL;
214 	struct inode_defrag tmp;
215 	struct rb_node *p;
216 	struct rb_node *parent = NULL;
217 	int ret;
218 
219 	tmp.ino = ino;
220 	tmp.root = root;
221 
222 	spin_lock(&fs_info->defrag_inodes_lock);
223 	p = fs_info->defrag_inodes.rb_node;
224 	while (p) {
225 		parent = p;
226 		entry = rb_entry(parent, struct inode_defrag, rb_node);
227 
228 		ret = __compare_inode_defrag(&tmp, entry);
229 		if (ret < 0)
230 			p = parent->rb_left;
231 		else if (ret > 0)
232 			p = parent->rb_right;
233 		else
234 			goto out;
235 	}
236 
237 	if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
238 		parent = rb_next(parent);
239 		if (parent)
240 			entry = rb_entry(parent, struct inode_defrag, rb_node);
241 		else
242 			entry = NULL;
243 	}
244 out:
245 	if (entry)
246 		rb_erase(parent, &fs_info->defrag_inodes);
247 	spin_unlock(&fs_info->defrag_inodes_lock);
248 	return entry;
249 }
250 
251 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
252 {
253 	struct inode_defrag *defrag;
254 	struct rb_node *node;
255 
256 	spin_lock(&fs_info->defrag_inodes_lock);
257 	node = rb_first(&fs_info->defrag_inodes);
258 	while (node) {
259 		rb_erase(node, &fs_info->defrag_inodes);
260 		defrag = rb_entry(node, struct inode_defrag, rb_node);
261 		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
262 
263 		cond_resched_lock(&fs_info->defrag_inodes_lock);
264 
265 		node = rb_first(&fs_info->defrag_inodes);
266 	}
267 	spin_unlock(&fs_info->defrag_inodes_lock);
268 }
269 
270 #define BTRFS_DEFRAG_BATCH	1024
271 
272 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
273 				    struct inode_defrag *defrag)
274 {
275 	struct btrfs_root *inode_root;
276 	struct inode *inode;
277 	struct btrfs_key key;
278 	struct btrfs_ioctl_defrag_range_args range;
279 	int num_defrag;
280 	int index;
281 	int ret;
282 
283 	/* get the inode */
284 	key.objectid = defrag->root;
285 	key.type = BTRFS_ROOT_ITEM_KEY;
286 	key.offset = (u64)-1;
287 
288 	index = srcu_read_lock(&fs_info->subvol_srcu);
289 
290 	inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
291 	if (IS_ERR(inode_root)) {
292 		ret = PTR_ERR(inode_root);
293 		goto cleanup;
294 	}
295 
296 	key.objectid = defrag->ino;
297 	key.type = BTRFS_INODE_ITEM_KEY;
298 	key.offset = 0;
299 	inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
300 	if (IS_ERR(inode)) {
301 		ret = PTR_ERR(inode);
302 		goto cleanup;
303 	}
304 	srcu_read_unlock(&fs_info->subvol_srcu, index);
305 
306 	/* do a chunk of defrag */
307 	clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
308 	memset(&range, 0, sizeof(range));
309 	range.len = (u64)-1;
310 	range.start = defrag->last_offset;
311 
312 	sb_start_write(fs_info->sb);
313 	num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
314 				       BTRFS_DEFRAG_BATCH);
315 	sb_end_write(fs_info->sb);
316 	/*
317 	 * if we filled the whole defrag batch, there
318 	 * must be more work to do.  Queue this defrag
319 	 * again
320 	 */
321 	if (num_defrag == BTRFS_DEFRAG_BATCH) {
322 		defrag->last_offset = range.start;
323 		btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
324 	} else if (defrag->last_offset && !defrag->cycled) {
325 		/*
326 		 * we didn't fill our defrag batch, but
327 		 * we didn't start at zero.  Make sure we loop
328 		 * around to the start of the file.
329 		 */
330 		defrag->last_offset = 0;
331 		defrag->cycled = 1;
332 		btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
333 	} else {
334 		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
335 	}
336 
337 	iput(inode);
338 	return 0;
339 cleanup:
340 	srcu_read_unlock(&fs_info->subvol_srcu, index);
341 	kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
342 	return ret;
343 }
344 
345 /*
346  * run through the list of inodes in the FS that need
347  * defragging
348  */
349 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
350 {
351 	struct inode_defrag *defrag;
352 	u64 first_ino = 0;
353 	u64 root_objectid = 0;
354 
355 	atomic_inc(&fs_info->defrag_running);
356 	while (1) {
357 		/* Pause the auto defragger. */
358 		if (test_bit(BTRFS_FS_STATE_REMOUNTING,
359 			     &fs_info->fs_state))
360 			break;
361 
362 		if (!__need_auto_defrag(fs_info))
363 			break;
364 
365 		/* find an inode to defrag */
366 		defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
367 						 first_ino);
368 		if (!defrag) {
369 			if (root_objectid || first_ino) {
370 				root_objectid = 0;
371 				first_ino = 0;
372 				continue;
373 			} else {
374 				break;
375 			}
376 		}
377 
378 		first_ino = defrag->ino + 1;
379 		root_objectid = defrag->root;
380 
381 		__btrfs_run_defrag_inode(fs_info, defrag);
382 	}
383 	atomic_dec(&fs_info->defrag_running);
384 
385 	/*
386 	 * during unmount, we use the transaction_wait queue to
387 	 * wait for the defragger to stop
388 	 */
389 	wake_up(&fs_info->transaction_wait);
390 	return 0;
391 }
392 
393 /* simple helper to fault in pages and copy.  This should go away
394  * and be replaced with calls into generic code.
395  */
396 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
397 					 struct page **prepared_pages,
398 					 struct iov_iter *i)
399 {
400 	size_t copied = 0;
401 	size_t total_copied = 0;
402 	int pg = 0;
403 	int offset = offset_in_page(pos);
404 
405 	while (write_bytes > 0) {
406 		size_t count = min_t(size_t,
407 				     PAGE_SIZE - offset, write_bytes);
408 		struct page *page = prepared_pages[pg];
409 		/*
410 		 * Copy data from userspace to the current page
411 		 */
412 		copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
413 
414 		/* Flush processor's dcache for this page */
415 		flush_dcache_page(page);
416 
417 		/*
418 		 * if we get a partial write, we can end up with
419 		 * partially up to date pages.  These add
420 		 * a lot of complexity, so make sure they don't
421 		 * happen by forcing this copy to be retried.
422 		 *
423 		 * The rest of the btrfs_file_write code will fall
424 		 * back to page at a time copies after we return 0.
425 		 */
426 		if (!PageUptodate(page) && copied < count)
427 			copied = 0;
428 
429 		iov_iter_advance(i, copied);
430 		write_bytes -= copied;
431 		total_copied += copied;
432 
433 		/* Return to btrfs_file_write_iter to fault page */
434 		if (unlikely(copied == 0))
435 			break;
436 
437 		if (copied < PAGE_SIZE - offset) {
438 			offset += copied;
439 		} else {
440 			pg++;
441 			offset = 0;
442 		}
443 	}
444 	return total_copied;
445 }
446 
447 /*
448  * unlocks pages after btrfs_file_write is done with them
449  */
450 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
451 {
452 	size_t i;
453 	for (i = 0; i < num_pages; i++) {
454 		/* page checked is some magic around finding pages that
455 		 * have been modified without going through btrfs_set_page_dirty
456 		 * clear it here. There should be no need to mark the pages
457 		 * accessed as prepare_pages should have marked them accessed
458 		 * in prepare_pages via find_or_create_page()
459 		 */
460 		ClearPageChecked(pages[i]);
461 		unlock_page(pages[i]);
462 		put_page(pages[i]);
463 	}
464 }
465 
466 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
467 					 const u64 start,
468 					 const u64 len,
469 					 struct extent_state **cached_state)
470 {
471 	u64 search_start = start;
472 	const u64 end = start + len - 1;
473 
474 	while (search_start < end) {
475 		const u64 search_len = end - search_start + 1;
476 		struct extent_map *em;
477 		u64 em_len;
478 		int ret = 0;
479 
480 		em = btrfs_get_extent(inode, NULL, 0, search_start,
481 				      search_len, 0);
482 		if (IS_ERR(em))
483 			return PTR_ERR(em);
484 
485 		if (em->block_start != EXTENT_MAP_HOLE)
486 			goto next;
487 
488 		em_len = em->len;
489 		if (em->start < search_start)
490 			em_len -= search_start - em->start;
491 		if (em_len > search_len)
492 			em_len = search_len;
493 
494 		ret = set_extent_bit(&inode->io_tree, search_start,
495 				     search_start + em_len - 1,
496 				     EXTENT_DELALLOC_NEW,
497 				     NULL, cached_state, GFP_NOFS);
498 next:
499 		search_start = extent_map_end(em);
500 		free_extent_map(em);
501 		if (ret)
502 			return ret;
503 	}
504 	return 0;
505 }
506 
507 /*
508  * after copy_from_user, pages need to be dirtied and we need to make
509  * sure holes are created between the current EOF and the start of
510  * any next extents (if required).
511  *
512  * this also makes the decision about creating an inline extent vs
513  * doing real data extents, marking pages dirty and delalloc as required.
514  */
515 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
516 		      size_t num_pages, loff_t pos, size_t write_bytes,
517 		      struct extent_state **cached)
518 {
519 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
520 	int err = 0;
521 	int i;
522 	u64 num_bytes;
523 	u64 start_pos;
524 	u64 end_of_last_block;
525 	u64 end_pos = pos + write_bytes;
526 	loff_t isize = i_size_read(inode);
527 	unsigned int extra_bits = 0;
528 
529 	start_pos = pos & ~((u64) fs_info->sectorsize - 1);
530 	num_bytes = round_up(write_bytes + pos - start_pos,
531 			     fs_info->sectorsize);
532 
533 	end_of_last_block = start_pos + num_bytes - 1;
534 
535 	/*
536 	 * The pages may have already been dirty, clear out old accounting so
537 	 * we can set things up properly
538 	 */
539 	clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
540 			 EXTENT_DIRTY | EXTENT_DELALLOC |
541 			 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 0, 0, cached);
542 
543 	if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
544 		if (start_pos >= isize &&
545 		    !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
546 			/*
547 			 * There can't be any extents following eof in this case
548 			 * so just set the delalloc new bit for the range
549 			 * directly.
550 			 */
551 			extra_bits |= EXTENT_DELALLOC_NEW;
552 		} else {
553 			err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
554 							    start_pos,
555 							    num_bytes, cached);
556 			if (err)
557 				return err;
558 		}
559 	}
560 
561 	err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
562 					extra_bits, cached, 0);
563 	if (err)
564 		return err;
565 
566 	for (i = 0; i < num_pages; i++) {
567 		struct page *p = pages[i];
568 		SetPageUptodate(p);
569 		ClearPageChecked(p);
570 		set_page_dirty(p);
571 	}
572 
573 	/*
574 	 * we've only changed i_size in ram, and we haven't updated
575 	 * the disk i_size.  There is no need to log the inode
576 	 * at this time.
577 	 */
578 	if (end_pos > isize)
579 		i_size_write(inode, end_pos);
580 	return 0;
581 }
582 
583 /*
584  * this drops all the extents in the cache that intersect the range
585  * [start, end].  Existing extents are split as required.
586  */
587 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
588 			     int skip_pinned)
589 {
590 	struct extent_map *em;
591 	struct extent_map *split = NULL;
592 	struct extent_map *split2 = NULL;
593 	struct extent_map_tree *em_tree = &inode->extent_tree;
594 	u64 len = end - start + 1;
595 	u64 gen;
596 	int ret;
597 	int testend = 1;
598 	unsigned long flags;
599 	int compressed = 0;
600 	bool modified;
601 
602 	WARN_ON(end < start);
603 	if (end == (u64)-1) {
604 		len = (u64)-1;
605 		testend = 0;
606 	}
607 	while (1) {
608 		int no_splits = 0;
609 
610 		modified = false;
611 		if (!split)
612 			split = alloc_extent_map();
613 		if (!split2)
614 			split2 = alloc_extent_map();
615 		if (!split || !split2)
616 			no_splits = 1;
617 
618 		write_lock(&em_tree->lock);
619 		em = lookup_extent_mapping(em_tree, start, len);
620 		if (!em) {
621 			write_unlock(&em_tree->lock);
622 			break;
623 		}
624 		flags = em->flags;
625 		gen = em->generation;
626 		if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
627 			if (testend && em->start + em->len >= start + len) {
628 				free_extent_map(em);
629 				write_unlock(&em_tree->lock);
630 				break;
631 			}
632 			start = em->start + em->len;
633 			if (testend)
634 				len = start + len - (em->start + em->len);
635 			free_extent_map(em);
636 			write_unlock(&em_tree->lock);
637 			continue;
638 		}
639 		compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
640 		clear_bit(EXTENT_FLAG_PINNED, &em->flags);
641 		clear_bit(EXTENT_FLAG_LOGGING, &flags);
642 		modified = !list_empty(&em->list);
643 		if (no_splits)
644 			goto next;
645 
646 		if (em->start < start) {
647 			split->start = em->start;
648 			split->len = start - em->start;
649 
650 			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
651 				split->orig_start = em->orig_start;
652 				split->block_start = em->block_start;
653 
654 				if (compressed)
655 					split->block_len = em->block_len;
656 				else
657 					split->block_len = split->len;
658 				split->orig_block_len = max(split->block_len,
659 						em->orig_block_len);
660 				split->ram_bytes = em->ram_bytes;
661 			} else {
662 				split->orig_start = split->start;
663 				split->block_len = 0;
664 				split->block_start = em->block_start;
665 				split->orig_block_len = 0;
666 				split->ram_bytes = split->len;
667 			}
668 
669 			split->generation = gen;
670 			split->bdev = em->bdev;
671 			split->flags = flags;
672 			split->compress_type = em->compress_type;
673 			replace_extent_mapping(em_tree, em, split, modified);
674 			free_extent_map(split);
675 			split = split2;
676 			split2 = NULL;
677 		}
678 		if (testend && em->start + em->len > start + len) {
679 			u64 diff = start + len - em->start;
680 
681 			split->start = start + len;
682 			split->len = em->start + em->len - (start + len);
683 			split->bdev = em->bdev;
684 			split->flags = flags;
685 			split->compress_type = em->compress_type;
686 			split->generation = gen;
687 
688 			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
689 				split->orig_block_len = max(em->block_len,
690 						    em->orig_block_len);
691 
692 				split->ram_bytes = em->ram_bytes;
693 				if (compressed) {
694 					split->block_len = em->block_len;
695 					split->block_start = em->block_start;
696 					split->orig_start = em->orig_start;
697 				} else {
698 					split->block_len = split->len;
699 					split->block_start = em->block_start
700 						+ diff;
701 					split->orig_start = em->orig_start;
702 				}
703 			} else {
704 				split->ram_bytes = split->len;
705 				split->orig_start = split->start;
706 				split->block_len = 0;
707 				split->block_start = em->block_start;
708 				split->orig_block_len = 0;
709 			}
710 
711 			if (extent_map_in_tree(em)) {
712 				replace_extent_mapping(em_tree, em, split,
713 						       modified);
714 			} else {
715 				ret = add_extent_mapping(em_tree, split,
716 							 modified);
717 				ASSERT(ret == 0); /* Logic error */
718 			}
719 			free_extent_map(split);
720 			split = NULL;
721 		}
722 next:
723 		if (extent_map_in_tree(em))
724 			remove_extent_mapping(em_tree, em);
725 		write_unlock(&em_tree->lock);
726 
727 		/* once for us */
728 		free_extent_map(em);
729 		/* once for the tree*/
730 		free_extent_map(em);
731 	}
732 	if (split)
733 		free_extent_map(split);
734 	if (split2)
735 		free_extent_map(split2);
736 }
737 
738 /*
739  * this is very complex, but the basic idea is to drop all extents
740  * in the range start - end.  hint_block is filled in with a block number
741  * that would be a good hint to the block allocator for this file.
742  *
743  * If an extent intersects the range but is not entirely inside the range
744  * it is either truncated or split.  Anything entirely inside the range
745  * is deleted from the tree.
746  */
747 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
748 			 struct btrfs_root *root, struct inode *inode,
749 			 struct btrfs_path *path, u64 start, u64 end,
750 			 u64 *drop_end, int drop_cache,
751 			 int replace_extent,
752 			 u32 extent_item_size,
753 			 int *key_inserted)
754 {
755 	struct btrfs_fs_info *fs_info = root->fs_info;
756 	struct extent_buffer *leaf;
757 	struct btrfs_file_extent_item *fi;
758 	struct btrfs_ref ref = { 0 };
759 	struct btrfs_key key;
760 	struct btrfs_key new_key;
761 	u64 ino = btrfs_ino(BTRFS_I(inode));
762 	u64 search_start = start;
763 	u64 disk_bytenr = 0;
764 	u64 num_bytes = 0;
765 	u64 extent_offset = 0;
766 	u64 extent_end = 0;
767 	u64 last_end = start;
768 	int del_nr = 0;
769 	int del_slot = 0;
770 	int extent_type;
771 	int recow;
772 	int ret;
773 	int modify_tree = -1;
774 	int update_refs;
775 	int found = 0;
776 	int leafs_visited = 0;
777 
778 	if (drop_cache)
779 		btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
780 
781 	if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
782 		modify_tree = 0;
783 
784 	update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
785 		       root == fs_info->tree_root);
786 	while (1) {
787 		recow = 0;
788 		ret = btrfs_lookup_file_extent(trans, root, path, ino,
789 					       search_start, modify_tree);
790 		if (ret < 0)
791 			break;
792 		if (ret > 0 && path->slots[0] > 0 && search_start == start) {
793 			leaf = path->nodes[0];
794 			btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
795 			if (key.objectid == ino &&
796 			    key.type == BTRFS_EXTENT_DATA_KEY)
797 				path->slots[0]--;
798 		}
799 		ret = 0;
800 		leafs_visited++;
801 next_slot:
802 		leaf = path->nodes[0];
803 		if (path->slots[0] >= btrfs_header_nritems(leaf)) {
804 			BUG_ON(del_nr > 0);
805 			ret = btrfs_next_leaf(root, path);
806 			if (ret < 0)
807 				break;
808 			if (ret > 0) {
809 				ret = 0;
810 				break;
811 			}
812 			leafs_visited++;
813 			leaf = path->nodes[0];
814 			recow = 1;
815 		}
816 
817 		btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
818 
819 		if (key.objectid > ino)
820 			break;
821 		if (WARN_ON_ONCE(key.objectid < ino) ||
822 		    key.type < BTRFS_EXTENT_DATA_KEY) {
823 			ASSERT(del_nr == 0);
824 			path->slots[0]++;
825 			goto next_slot;
826 		}
827 		if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
828 			break;
829 
830 		fi = btrfs_item_ptr(leaf, path->slots[0],
831 				    struct btrfs_file_extent_item);
832 		extent_type = btrfs_file_extent_type(leaf, fi);
833 
834 		if (extent_type == BTRFS_FILE_EXTENT_REG ||
835 		    extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
836 			disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
837 			num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
838 			extent_offset = btrfs_file_extent_offset(leaf, fi);
839 			extent_end = key.offset +
840 				btrfs_file_extent_num_bytes(leaf, fi);
841 		} else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
842 			extent_end = key.offset +
843 				btrfs_file_extent_ram_bytes(leaf, fi);
844 		} else {
845 			/* can't happen */
846 			BUG();
847 		}
848 
849 		/*
850 		 * Don't skip extent items representing 0 byte lengths. They
851 		 * used to be created (bug) if while punching holes we hit
852 		 * -ENOSPC condition. So if we find one here, just ensure we
853 		 * delete it, otherwise we would insert a new file extent item
854 		 * with the same key (offset) as that 0 bytes length file
855 		 * extent item in the call to setup_items_for_insert() later
856 		 * in this function.
857 		 */
858 		if (extent_end == key.offset && extent_end >= search_start) {
859 			last_end = extent_end;
860 			goto delete_extent_item;
861 		}
862 
863 		if (extent_end <= search_start) {
864 			path->slots[0]++;
865 			goto next_slot;
866 		}
867 
868 		found = 1;
869 		search_start = max(key.offset, start);
870 		if (recow || !modify_tree) {
871 			modify_tree = -1;
872 			btrfs_release_path(path);
873 			continue;
874 		}
875 
876 		/*
877 		 *     | - range to drop - |
878 		 *  | -------- extent -------- |
879 		 */
880 		if (start > key.offset && end < extent_end) {
881 			BUG_ON(del_nr > 0);
882 			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
883 				ret = -EOPNOTSUPP;
884 				break;
885 			}
886 
887 			memcpy(&new_key, &key, sizeof(new_key));
888 			new_key.offset = start;
889 			ret = btrfs_duplicate_item(trans, root, path,
890 						   &new_key);
891 			if (ret == -EAGAIN) {
892 				btrfs_release_path(path);
893 				continue;
894 			}
895 			if (ret < 0)
896 				break;
897 
898 			leaf = path->nodes[0];
899 			fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
900 					    struct btrfs_file_extent_item);
901 			btrfs_set_file_extent_num_bytes(leaf, fi,
902 							start - key.offset);
903 
904 			fi = btrfs_item_ptr(leaf, path->slots[0],
905 					    struct btrfs_file_extent_item);
906 
907 			extent_offset += start - key.offset;
908 			btrfs_set_file_extent_offset(leaf, fi, extent_offset);
909 			btrfs_set_file_extent_num_bytes(leaf, fi,
910 							extent_end - start);
911 			btrfs_mark_buffer_dirty(leaf);
912 
913 			if (update_refs && disk_bytenr > 0) {
914 				btrfs_init_generic_ref(&ref,
915 						BTRFS_ADD_DELAYED_REF,
916 						disk_bytenr, num_bytes, 0);
917 				btrfs_init_data_ref(&ref,
918 						root->root_key.objectid,
919 						new_key.objectid,
920 						start - extent_offset);
921 				ret = btrfs_inc_extent_ref(trans, &ref);
922 				BUG_ON(ret); /* -ENOMEM */
923 			}
924 			key.offset = start;
925 		}
926 		/*
927 		 * From here on out we will have actually dropped something, so
928 		 * last_end can be updated.
929 		 */
930 		last_end = extent_end;
931 
932 		/*
933 		 *  | ---- range to drop ----- |
934 		 *      | -------- extent -------- |
935 		 */
936 		if (start <= key.offset && end < extent_end) {
937 			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
938 				ret = -EOPNOTSUPP;
939 				break;
940 			}
941 
942 			memcpy(&new_key, &key, sizeof(new_key));
943 			new_key.offset = end;
944 			btrfs_set_item_key_safe(fs_info, path, &new_key);
945 
946 			extent_offset += end - key.offset;
947 			btrfs_set_file_extent_offset(leaf, fi, extent_offset);
948 			btrfs_set_file_extent_num_bytes(leaf, fi,
949 							extent_end - end);
950 			btrfs_mark_buffer_dirty(leaf);
951 			if (update_refs && disk_bytenr > 0)
952 				inode_sub_bytes(inode, end - key.offset);
953 			break;
954 		}
955 
956 		search_start = extent_end;
957 		/*
958 		 *       | ---- range to drop ----- |
959 		 *  | -------- extent -------- |
960 		 */
961 		if (start > key.offset && end >= extent_end) {
962 			BUG_ON(del_nr > 0);
963 			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
964 				ret = -EOPNOTSUPP;
965 				break;
966 			}
967 
968 			btrfs_set_file_extent_num_bytes(leaf, fi,
969 							start - key.offset);
970 			btrfs_mark_buffer_dirty(leaf);
971 			if (update_refs && disk_bytenr > 0)
972 				inode_sub_bytes(inode, extent_end - start);
973 			if (end == extent_end)
974 				break;
975 
976 			path->slots[0]++;
977 			goto next_slot;
978 		}
979 
980 		/*
981 		 *  | ---- range to drop ----- |
982 		 *    | ------ extent ------ |
983 		 */
984 		if (start <= key.offset && end >= extent_end) {
985 delete_extent_item:
986 			if (del_nr == 0) {
987 				del_slot = path->slots[0];
988 				del_nr = 1;
989 			} else {
990 				BUG_ON(del_slot + del_nr != path->slots[0]);
991 				del_nr++;
992 			}
993 
994 			if (update_refs &&
995 			    extent_type == BTRFS_FILE_EXTENT_INLINE) {
996 				inode_sub_bytes(inode,
997 						extent_end - key.offset);
998 				extent_end = ALIGN(extent_end,
999 						   fs_info->sectorsize);
1000 			} else if (update_refs && disk_bytenr > 0) {
1001 				btrfs_init_generic_ref(&ref,
1002 						BTRFS_DROP_DELAYED_REF,
1003 						disk_bytenr, num_bytes, 0);
1004 				btrfs_init_data_ref(&ref,
1005 						root->root_key.objectid,
1006 						key.objectid,
1007 						key.offset - extent_offset);
1008 				ret = btrfs_free_extent(trans, &ref);
1009 				BUG_ON(ret); /* -ENOMEM */
1010 				inode_sub_bytes(inode,
1011 						extent_end - key.offset);
1012 			}
1013 
1014 			if (end == extent_end)
1015 				break;
1016 
1017 			if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1018 				path->slots[0]++;
1019 				goto next_slot;
1020 			}
1021 
1022 			ret = btrfs_del_items(trans, root, path, del_slot,
1023 					      del_nr);
1024 			if (ret) {
1025 				btrfs_abort_transaction(trans, ret);
1026 				break;
1027 			}
1028 
1029 			del_nr = 0;
1030 			del_slot = 0;
1031 
1032 			btrfs_release_path(path);
1033 			continue;
1034 		}
1035 
1036 		BUG();
1037 	}
1038 
1039 	if (!ret && del_nr > 0) {
1040 		/*
1041 		 * Set path->slots[0] to first slot, so that after the delete
1042 		 * if items are move off from our leaf to its immediate left or
1043 		 * right neighbor leafs, we end up with a correct and adjusted
1044 		 * path->slots[0] for our insertion (if replace_extent != 0).
1045 		 */
1046 		path->slots[0] = del_slot;
1047 		ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1048 		if (ret)
1049 			btrfs_abort_transaction(trans, ret);
1050 	}
1051 
1052 	leaf = path->nodes[0];
1053 	/*
1054 	 * If btrfs_del_items() was called, it might have deleted a leaf, in
1055 	 * which case it unlocked our path, so check path->locks[0] matches a
1056 	 * write lock.
1057 	 */
1058 	if (!ret && replace_extent && leafs_visited == 1 &&
1059 	    (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1060 	     path->locks[0] == BTRFS_WRITE_LOCK) &&
1061 	    btrfs_leaf_free_space(leaf) >=
1062 	    sizeof(struct btrfs_item) + extent_item_size) {
1063 
1064 		key.objectid = ino;
1065 		key.type = BTRFS_EXTENT_DATA_KEY;
1066 		key.offset = start;
1067 		if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1068 			struct btrfs_key slot_key;
1069 
1070 			btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1071 			if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1072 				path->slots[0]++;
1073 		}
1074 		setup_items_for_insert(root, path, &key,
1075 				       &extent_item_size,
1076 				       extent_item_size,
1077 				       sizeof(struct btrfs_item) +
1078 				       extent_item_size, 1);
1079 		*key_inserted = 1;
1080 	}
1081 
1082 	if (!replace_extent || !(*key_inserted))
1083 		btrfs_release_path(path);
1084 	if (drop_end)
1085 		*drop_end = found ? min(end, last_end) : end;
1086 	return ret;
1087 }
1088 
1089 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1090 		       struct btrfs_root *root, struct inode *inode, u64 start,
1091 		       u64 end, int drop_cache)
1092 {
1093 	struct btrfs_path *path;
1094 	int ret;
1095 
1096 	path = btrfs_alloc_path();
1097 	if (!path)
1098 		return -ENOMEM;
1099 	ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1100 				   drop_cache, 0, 0, NULL);
1101 	btrfs_free_path(path);
1102 	return ret;
1103 }
1104 
1105 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1106 			    u64 objectid, u64 bytenr, u64 orig_offset,
1107 			    u64 *start, u64 *end)
1108 {
1109 	struct btrfs_file_extent_item *fi;
1110 	struct btrfs_key key;
1111 	u64 extent_end;
1112 
1113 	if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1114 		return 0;
1115 
1116 	btrfs_item_key_to_cpu(leaf, &key, slot);
1117 	if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1118 		return 0;
1119 
1120 	fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1121 	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1122 	    btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1123 	    btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1124 	    btrfs_file_extent_compression(leaf, fi) ||
1125 	    btrfs_file_extent_encryption(leaf, fi) ||
1126 	    btrfs_file_extent_other_encoding(leaf, fi))
1127 		return 0;
1128 
1129 	extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1130 	if ((*start && *start != key.offset) || (*end && *end != extent_end))
1131 		return 0;
1132 
1133 	*start = key.offset;
1134 	*end = extent_end;
1135 	return 1;
1136 }
1137 
1138 /*
1139  * Mark extent in the range start - end as written.
1140  *
1141  * This changes extent type from 'pre-allocated' to 'regular'. If only
1142  * part of extent is marked as written, the extent will be split into
1143  * two or three.
1144  */
1145 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1146 			      struct btrfs_inode *inode, u64 start, u64 end)
1147 {
1148 	struct btrfs_fs_info *fs_info = trans->fs_info;
1149 	struct btrfs_root *root = inode->root;
1150 	struct extent_buffer *leaf;
1151 	struct btrfs_path *path;
1152 	struct btrfs_file_extent_item *fi;
1153 	struct btrfs_ref ref = { 0 };
1154 	struct btrfs_key key;
1155 	struct btrfs_key new_key;
1156 	u64 bytenr;
1157 	u64 num_bytes;
1158 	u64 extent_end;
1159 	u64 orig_offset;
1160 	u64 other_start;
1161 	u64 other_end;
1162 	u64 split;
1163 	int del_nr = 0;
1164 	int del_slot = 0;
1165 	int recow;
1166 	int ret;
1167 	u64 ino = btrfs_ino(inode);
1168 
1169 	path = btrfs_alloc_path();
1170 	if (!path)
1171 		return -ENOMEM;
1172 again:
1173 	recow = 0;
1174 	split = start;
1175 	key.objectid = ino;
1176 	key.type = BTRFS_EXTENT_DATA_KEY;
1177 	key.offset = split;
1178 
1179 	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1180 	if (ret < 0)
1181 		goto out;
1182 	if (ret > 0 && path->slots[0] > 0)
1183 		path->slots[0]--;
1184 
1185 	leaf = path->nodes[0];
1186 	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1187 	if (key.objectid != ino ||
1188 	    key.type != BTRFS_EXTENT_DATA_KEY) {
1189 		ret = -EINVAL;
1190 		btrfs_abort_transaction(trans, ret);
1191 		goto out;
1192 	}
1193 	fi = btrfs_item_ptr(leaf, path->slots[0],
1194 			    struct btrfs_file_extent_item);
1195 	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1196 		ret = -EINVAL;
1197 		btrfs_abort_transaction(trans, ret);
1198 		goto out;
1199 	}
1200 	extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1201 	if (key.offset > start || extent_end < end) {
1202 		ret = -EINVAL;
1203 		btrfs_abort_transaction(trans, ret);
1204 		goto out;
1205 	}
1206 
1207 	bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1208 	num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1209 	orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1210 	memcpy(&new_key, &key, sizeof(new_key));
1211 
1212 	if (start == key.offset && end < extent_end) {
1213 		other_start = 0;
1214 		other_end = start;
1215 		if (extent_mergeable(leaf, path->slots[0] - 1,
1216 				     ino, bytenr, orig_offset,
1217 				     &other_start, &other_end)) {
1218 			new_key.offset = end;
1219 			btrfs_set_item_key_safe(fs_info, path, &new_key);
1220 			fi = btrfs_item_ptr(leaf, path->slots[0],
1221 					    struct btrfs_file_extent_item);
1222 			btrfs_set_file_extent_generation(leaf, fi,
1223 							 trans->transid);
1224 			btrfs_set_file_extent_num_bytes(leaf, fi,
1225 							extent_end - end);
1226 			btrfs_set_file_extent_offset(leaf, fi,
1227 						     end - orig_offset);
1228 			fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1229 					    struct btrfs_file_extent_item);
1230 			btrfs_set_file_extent_generation(leaf, fi,
1231 							 trans->transid);
1232 			btrfs_set_file_extent_num_bytes(leaf, fi,
1233 							end - other_start);
1234 			btrfs_mark_buffer_dirty(leaf);
1235 			goto out;
1236 		}
1237 	}
1238 
1239 	if (start > key.offset && end == extent_end) {
1240 		other_start = end;
1241 		other_end = 0;
1242 		if (extent_mergeable(leaf, path->slots[0] + 1,
1243 				     ino, bytenr, orig_offset,
1244 				     &other_start, &other_end)) {
1245 			fi = btrfs_item_ptr(leaf, path->slots[0],
1246 					    struct btrfs_file_extent_item);
1247 			btrfs_set_file_extent_num_bytes(leaf, fi,
1248 							start - key.offset);
1249 			btrfs_set_file_extent_generation(leaf, fi,
1250 							 trans->transid);
1251 			path->slots[0]++;
1252 			new_key.offset = start;
1253 			btrfs_set_item_key_safe(fs_info, path, &new_key);
1254 
1255 			fi = btrfs_item_ptr(leaf, path->slots[0],
1256 					    struct btrfs_file_extent_item);
1257 			btrfs_set_file_extent_generation(leaf, fi,
1258 							 trans->transid);
1259 			btrfs_set_file_extent_num_bytes(leaf, fi,
1260 							other_end - start);
1261 			btrfs_set_file_extent_offset(leaf, fi,
1262 						     start - orig_offset);
1263 			btrfs_mark_buffer_dirty(leaf);
1264 			goto out;
1265 		}
1266 	}
1267 
1268 	while (start > key.offset || end < extent_end) {
1269 		if (key.offset == start)
1270 			split = end;
1271 
1272 		new_key.offset = split;
1273 		ret = btrfs_duplicate_item(trans, root, path, &new_key);
1274 		if (ret == -EAGAIN) {
1275 			btrfs_release_path(path);
1276 			goto again;
1277 		}
1278 		if (ret < 0) {
1279 			btrfs_abort_transaction(trans, ret);
1280 			goto out;
1281 		}
1282 
1283 		leaf = path->nodes[0];
1284 		fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1285 				    struct btrfs_file_extent_item);
1286 		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1287 		btrfs_set_file_extent_num_bytes(leaf, fi,
1288 						split - key.offset);
1289 
1290 		fi = btrfs_item_ptr(leaf, path->slots[0],
1291 				    struct btrfs_file_extent_item);
1292 
1293 		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1294 		btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1295 		btrfs_set_file_extent_num_bytes(leaf, fi,
1296 						extent_end - split);
1297 		btrfs_mark_buffer_dirty(leaf);
1298 
1299 		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1300 				       num_bytes, 0);
1301 		btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1302 				    orig_offset);
1303 		ret = btrfs_inc_extent_ref(trans, &ref);
1304 		if (ret) {
1305 			btrfs_abort_transaction(trans, ret);
1306 			goto out;
1307 		}
1308 
1309 		if (split == start) {
1310 			key.offset = start;
1311 		} else {
1312 			if (start != key.offset) {
1313 				ret = -EINVAL;
1314 				btrfs_abort_transaction(trans, ret);
1315 				goto out;
1316 			}
1317 			path->slots[0]--;
1318 			extent_end = end;
1319 		}
1320 		recow = 1;
1321 	}
1322 
1323 	other_start = end;
1324 	other_end = 0;
1325 	btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1326 			       num_bytes, 0);
1327 	btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1328 	if (extent_mergeable(leaf, path->slots[0] + 1,
1329 			     ino, bytenr, orig_offset,
1330 			     &other_start, &other_end)) {
1331 		if (recow) {
1332 			btrfs_release_path(path);
1333 			goto again;
1334 		}
1335 		extent_end = other_end;
1336 		del_slot = path->slots[0] + 1;
1337 		del_nr++;
1338 		ret = btrfs_free_extent(trans, &ref);
1339 		if (ret) {
1340 			btrfs_abort_transaction(trans, ret);
1341 			goto out;
1342 		}
1343 	}
1344 	other_start = 0;
1345 	other_end = start;
1346 	if (extent_mergeable(leaf, path->slots[0] - 1,
1347 			     ino, bytenr, orig_offset,
1348 			     &other_start, &other_end)) {
1349 		if (recow) {
1350 			btrfs_release_path(path);
1351 			goto again;
1352 		}
1353 		key.offset = other_start;
1354 		del_slot = path->slots[0];
1355 		del_nr++;
1356 		ret = btrfs_free_extent(trans, &ref);
1357 		if (ret) {
1358 			btrfs_abort_transaction(trans, ret);
1359 			goto out;
1360 		}
1361 	}
1362 	if (del_nr == 0) {
1363 		fi = btrfs_item_ptr(leaf, path->slots[0],
1364 			   struct btrfs_file_extent_item);
1365 		btrfs_set_file_extent_type(leaf, fi,
1366 					   BTRFS_FILE_EXTENT_REG);
1367 		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1368 		btrfs_mark_buffer_dirty(leaf);
1369 	} else {
1370 		fi = btrfs_item_ptr(leaf, del_slot - 1,
1371 			   struct btrfs_file_extent_item);
1372 		btrfs_set_file_extent_type(leaf, fi,
1373 					   BTRFS_FILE_EXTENT_REG);
1374 		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1375 		btrfs_set_file_extent_num_bytes(leaf, fi,
1376 						extent_end - key.offset);
1377 		btrfs_mark_buffer_dirty(leaf);
1378 
1379 		ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1380 		if (ret < 0) {
1381 			btrfs_abort_transaction(trans, ret);
1382 			goto out;
1383 		}
1384 	}
1385 out:
1386 	btrfs_free_path(path);
1387 	return 0;
1388 }
1389 
1390 /*
1391  * on error we return an unlocked page and the error value
1392  * on success we return a locked page and 0
1393  */
1394 static int prepare_uptodate_page(struct inode *inode,
1395 				 struct page *page, u64 pos,
1396 				 bool force_uptodate)
1397 {
1398 	int ret = 0;
1399 
1400 	if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1401 	    !PageUptodate(page)) {
1402 		ret = btrfs_readpage(NULL, page);
1403 		if (ret)
1404 			return ret;
1405 		lock_page(page);
1406 		if (!PageUptodate(page)) {
1407 			unlock_page(page);
1408 			return -EIO;
1409 		}
1410 		if (page->mapping != inode->i_mapping) {
1411 			unlock_page(page);
1412 			return -EAGAIN;
1413 		}
1414 	}
1415 	return 0;
1416 }
1417 
1418 /*
1419  * this just gets pages into the page cache and locks them down.
1420  */
1421 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1422 				  size_t num_pages, loff_t pos,
1423 				  size_t write_bytes, bool force_uptodate)
1424 {
1425 	int i;
1426 	unsigned long index = pos >> PAGE_SHIFT;
1427 	gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1428 	int err = 0;
1429 	int faili;
1430 
1431 	for (i = 0; i < num_pages; i++) {
1432 again:
1433 		pages[i] = find_or_create_page(inode->i_mapping, index + i,
1434 					       mask | __GFP_WRITE);
1435 		if (!pages[i]) {
1436 			faili = i - 1;
1437 			err = -ENOMEM;
1438 			goto fail;
1439 		}
1440 
1441 		if (i == 0)
1442 			err = prepare_uptodate_page(inode, pages[i], pos,
1443 						    force_uptodate);
1444 		if (!err && i == num_pages - 1)
1445 			err = prepare_uptodate_page(inode, pages[i],
1446 						    pos + write_bytes, false);
1447 		if (err) {
1448 			put_page(pages[i]);
1449 			if (err == -EAGAIN) {
1450 				err = 0;
1451 				goto again;
1452 			}
1453 			faili = i - 1;
1454 			goto fail;
1455 		}
1456 		wait_on_page_writeback(pages[i]);
1457 	}
1458 
1459 	return 0;
1460 fail:
1461 	while (faili >= 0) {
1462 		unlock_page(pages[faili]);
1463 		put_page(pages[faili]);
1464 		faili--;
1465 	}
1466 	return err;
1467 
1468 }
1469 
1470 /*
1471  * This function locks the extent and properly waits for data=ordered extents
1472  * to finish before allowing the pages to be modified if need.
1473  *
1474  * The return value:
1475  * 1 - the extent is locked
1476  * 0 - the extent is not locked, and everything is OK
1477  * -EAGAIN - need re-prepare the pages
1478  * the other < 0 number - Something wrong happens
1479  */
1480 static noinline int
1481 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1482 				size_t num_pages, loff_t pos,
1483 				size_t write_bytes,
1484 				u64 *lockstart, u64 *lockend,
1485 				struct extent_state **cached_state)
1486 {
1487 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1488 	u64 start_pos;
1489 	u64 last_pos;
1490 	int i;
1491 	int ret = 0;
1492 
1493 	start_pos = round_down(pos, fs_info->sectorsize);
1494 	last_pos = start_pos
1495 		+ round_up(pos + write_bytes - start_pos,
1496 			   fs_info->sectorsize) - 1;
1497 
1498 	if (start_pos < inode->vfs_inode.i_size) {
1499 		struct btrfs_ordered_extent *ordered;
1500 
1501 		lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1502 				cached_state);
1503 		ordered = btrfs_lookup_ordered_range(inode, start_pos,
1504 						     last_pos - start_pos + 1);
1505 		if (ordered &&
1506 		    ordered->file_offset + ordered->len > start_pos &&
1507 		    ordered->file_offset <= last_pos) {
1508 			unlock_extent_cached(&inode->io_tree, start_pos,
1509 					last_pos, cached_state);
1510 			for (i = 0; i < num_pages; i++) {
1511 				unlock_page(pages[i]);
1512 				put_page(pages[i]);
1513 			}
1514 			btrfs_start_ordered_extent(&inode->vfs_inode,
1515 					ordered, 1);
1516 			btrfs_put_ordered_extent(ordered);
1517 			return -EAGAIN;
1518 		}
1519 		if (ordered)
1520 			btrfs_put_ordered_extent(ordered);
1521 
1522 		*lockstart = start_pos;
1523 		*lockend = last_pos;
1524 		ret = 1;
1525 	}
1526 
1527 	/*
1528 	 * It's possible the pages are dirty right now, but we don't want
1529 	 * to clean them yet because copy_from_user may catch a page fault
1530 	 * and we might have to fall back to one page at a time.  If that
1531 	 * happens, we'll unlock these pages and we'd have a window where
1532 	 * reclaim could sneak in and drop the once-dirty page on the floor
1533 	 * without writing it.
1534 	 *
1535 	 * We have the pages locked and the extent range locked, so there's
1536 	 * no way someone can start IO on any dirty pages in this range.
1537 	 *
1538 	 * We'll call btrfs_dirty_pages() later on, and that will flip around
1539 	 * delalloc bits and dirty the pages as required.
1540 	 */
1541 	for (i = 0; i < num_pages; i++) {
1542 		set_page_extent_mapped(pages[i]);
1543 		WARN_ON(!PageLocked(pages[i]));
1544 	}
1545 
1546 	return ret;
1547 }
1548 
1549 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1550 				    size_t *write_bytes)
1551 {
1552 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1553 	struct btrfs_root *root = inode->root;
1554 	u64 lockstart, lockend;
1555 	u64 num_bytes;
1556 	int ret;
1557 
1558 	ret = btrfs_start_write_no_snapshotting(root);
1559 	if (!ret)
1560 		return -EAGAIN;
1561 
1562 	lockstart = round_down(pos, fs_info->sectorsize);
1563 	lockend = round_up(pos + *write_bytes,
1564 			   fs_info->sectorsize) - 1;
1565 
1566 	btrfs_lock_and_flush_ordered_range(&inode->io_tree, inode, lockstart,
1567 					   lockend, NULL);
1568 
1569 	num_bytes = lockend - lockstart + 1;
1570 	ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1571 			NULL, NULL, NULL);
1572 	if (ret <= 0) {
1573 		ret = 0;
1574 		btrfs_end_write_no_snapshotting(root);
1575 	} else {
1576 		*write_bytes = min_t(size_t, *write_bytes ,
1577 				     num_bytes - pos + lockstart);
1578 	}
1579 
1580 	unlock_extent(&inode->io_tree, lockstart, lockend);
1581 
1582 	return ret;
1583 }
1584 
1585 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1586 					       struct iov_iter *i)
1587 {
1588 	struct file *file = iocb->ki_filp;
1589 	loff_t pos = iocb->ki_pos;
1590 	struct inode *inode = file_inode(file);
1591 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1592 	struct btrfs_root *root = BTRFS_I(inode)->root;
1593 	struct page **pages = NULL;
1594 	struct extent_state *cached_state = NULL;
1595 	struct extent_changeset *data_reserved = NULL;
1596 	u64 release_bytes = 0;
1597 	u64 lockstart;
1598 	u64 lockend;
1599 	size_t num_written = 0;
1600 	int nrptrs;
1601 	int ret = 0;
1602 	bool only_release_metadata = false;
1603 	bool force_page_uptodate = false;
1604 
1605 	nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1606 			PAGE_SIZE / (sizeof(struct page *)));
1607 	nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1608 	nrptrs = max(nrptrs, 8);
1609 	pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1610 	if (!pages)
1611 		return -ENOMEM;
1612 
1613 	while (iov_iter_count(i) > 0) {
1614 		size_t offset = offset_in_page(pos);
1615 		size_t sector_offset;
1616 		size_t write_bytes = min(iov_iter_count(i),
1617 					 nrptrs * (size_t)PAGE_SIZE -
1618 					 offset);
1619 		size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1620 						PAGE_SIZE);
1621 		size_t reserve_bytes;
1622 		size_t dirty_pages;
1623 		size_t copied;
1624 		size_t dirty_sectors;
1625 		size_t num_sectors;
1626 		int extents_locked;
1627 
1628 		WARN_ON(num_pages > nrptrs);
1629 
1630 		/*
1631 		 * Fault pages before locking them in prepare_pages
1632 		 * to avoid recursive lock
1633 		 */
1634 		if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1635 			ret = -EFAULT;
1636 			break;
1637 		}
1638 
1639 		sector_offset = pos & (fs_info->sectorsize - 1);
1640 		reserve_bytes = round_up(write_bytes + sector_offset,
1641 				fs_info->sectorsize);
1642 
1643 		extent_changeset_release(data_reserved);
1644 		ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1645 						  write_bytes);
1646 		if (ret < 0) {
1647 			if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1648 						      BTRFS_INODE_PREALLOC)) &&
1649 			    check_can_nocow(BTRFS_I(inode), pos,
1650 					&write_bytes) > 0) {
1651 				/*
1652 				 * For nodata cow case, no need to reserve
1653 				 * data space.
1654 				 */
1655 				only_release_metadata = true;
1656 				/*
1657 				 * our prealloc extent may be smaller than
1658 				 * write_bytes, so scale down.
1659 				 */
1660 				num_pages = DIV_ROUND_UP(write_bytes + offset,
1661 							 PAGE_SIZE);
1662 				reserve_bytes = round_up(write_bytes +
1663 							 sector_offset,
1664 							 fs_info->sectorsize);
1665 			} else {
1666 				break;
1667 			}
1668 		}
1669 
1670 		WARN_ON(reserve_bytes == 0);
1671 		ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1672 				reserve_bytes);
1673 		if (ret) {
1674 			if (!only_release_metadata)
1675 				btrfs_free_reserved_data_space(inode,
1676 						data_reserved, pos,
1677 						write_bytes);
1678 			else
1679 				btrfs_end_write_no_snapshotting(root);
1680 			break;
1681 		}
1682 
1683 		release_bytes = reserve_bytes;
1684 again:
1685 		/*
1686 		 * This is going to setup the pages array with the number of
1687 		 * pages we want, so we don't really need to worry about the
1688 		 * contents of pages from loop to loop
1689 		 */
1690 		ret = prepare_pages(inode, pages, num_pages,
1691 				    pos, write_bytes,
1692 				    force_page_uptodate);
1693 		if (ret) {
1694 			btrfs_delalloc_release_extents(BTRFS_I(inode),
1695 						       reserve_bytes, true);
1696 			break;
1697 		}
1698 
1699 		extents_locked = lock_and_cleanup_extent_if_need(
1700 				BTRFS_I(inode), pages,
1701 				num_pages, pos, write_bytes, &lockstart,
1702 				&lockend, &cached_state);
1703 		if (extents_locked < 0) {
1704 			if (extents_locked == -EAGAIN)
1705 				goto again;
1706 			btrfs_delalloc_release_extents(BTRFS_I(inode),
1707 						       reserve_bytes, true);
1708 			ret = extents_locked;
1709 			break;
1710 		}
1711 
1712 		copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1713 
1714 		num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1715 		dirty_sectors = round_up(copied + sector_offset,
1716 					fs_info->sectorsize);
1717 		dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1718 
1719 		/*
1720 		 * if we have trouble faulting in the pages, fall
1721 		 * back to one page at a time
1722 		 */
1723 		if (copied < write_bytes)
1724 			nrptrs = 1;
1725 
1726 		if (copied == 0) {
1727 			force_page_uptodate = true;
1728 			dirty_sectors = 0;
1729 			dirty_pages = 0;
1730 		} else {
1731 			force_page_uptodate = false;
1732 			dirty_pages = DIV_ROUND_UP(copied + offset,
1733 						   PAGE_SIZE);
1734 		}
1735 
1736 		if (num_sectors > dirty_sectors) {
1737 			/* release everything except the sectors we dirtied */
1738 			release_bytes -= dirty_sectors <<
1739 						fs_info->sb->s_blocksize_bits;
1740 			if (only_release_metadata) {
1741 				btrfs_delalloc_release_metadata(BTRFS_I(inode),
1742 							release_bytes, true);
1743 			} else {
1744 				u64 __pos;
1745 
1746 				__pos = round_down(pos,
1747 						   fs_info->sectorsize) +
1748 					(dirty_pages << PAGE_SHIFT);
1749 				btrfs_delalloc_release_space(inode,
1750 						data_reserved, __pos,
1751 						release_bytes, true);
1752 			}
1753 		}
1754 
1755 		release_bytes = round_up(copied + sector_offset,
1756 					fs_info->sectorsize);
1757 
1758 		if (copied > 0)
1759 			ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1760 						pos, copied, &cached_state);
1761 		if (extents_locked)
1762 			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1763 					     lockstart, lockend, &cached_state);
1764 		btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes,
1765 					       true);
1766 		if (ret) {
1767 			btrfs_drop_pages(pages, num_pages);
1768 			break;
1769 		}
1770 
1771 		release_bytes = 0;
1772 		if (only_release_metadata)
1773 			btrfs_end_write_no_snapshotting(root);
1774 
1775 		if (only_release_metadata && copied > 0) {
1776 			lockstart = round_down(pos,
1777 					       fs_info->sectorsize);
1778 			lockend = round_up(pos + copied,
1779 					   fs_info->sectorsize) - 1;
1780 
1781 			set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1782 				       lockend, EXTENT_NORESERVE, NULL,
1783 				       NULL, GFP_NOFS);
1784 			only_release_metadata = false;
1785 		}
1786 
1787 		btrfs_drop_pages(pages, num_pages);
1788 
1789 		cond_resched();
1790 
1791 		balance_dirty_pages_ratelimited(inode->i_mapping);
1792 		if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1793 			btrfs_btree_balance_dirty(fs_info);
1794 
1795 		pos += copied;
1796 		num_written += copied;
1797 	}
1798 
1799 	kfree(pages);
1800 
1801 	if (release_bytes) {
1802 		if (only_release_metadata) {
1803 			btrfs_end_write_no_snapshotting(root);
1804 			btrfs_delalloc_release_metadata(BTRFS_I(inode),
1805 					release_bytes, true);
1806 		} else {
1807 			btrfs_delalloc_release_space(inode, data_reserved,
1808 					round_down(pos, fs_info->sectorsize),
1809 					release_bytes, true);
1810 		}
1811 	}
1812 
1813 	extent_changeset_free(data_reserved);
1814 	return num_written ? num_written : ret;
1815 }
1816 
1817 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1818 {
1819 	struct file *file = iocb->ki_filp;
1820 	struct inode *inode = file_inode(file);
1821 	loff_t pos;
1822 	ssize_t written;
1823 	ssize_t written_buffered;
1824 	loff_t endbyte;
1825 	int err;
1826 
1827 	written = generic_file_direct_write(iocb, from);
1828 
1829 	if (written < 0 || !iov_iter_count(from))
1830 		return written;
1831 
1832 	pos = iocb->ki_pos;
1833 	written_buffered = btrfs_buffered_write(iocb, from);
1834 	if (written_buffered < 0) {
1835 		err = written_buffered;
1836 		goto out;
1837 	}
1838 	/*
1839 	 * Ensure all data is persisted. We want the next direct IO read to be
1840 	 * able to read what was just written.
1841 	 */
1842 	endbyte = pos + written_buffered - 1;
1843 	err = btrfs_fdatawrite_range(inode, pos, endbyte);
1844 	if (err)
1845 		goto out;
1846 	err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1847 	if (err)
1848 		goto out;
1849 	written += written_buffered;
1850 	iocb->ki_pos = pos + written_buffered;
1851 	invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1852 				 endbyte >> PAGE_SHIFT);
1853 out:
1854 	return written ? written : err;
1855 }
1856 
1857 static void update_time_for_write(struct inode *inode)
1858 {
1859 	struct timespec64 now;
1860 
1861 	if (IS_NOCMTIME(inode))
1862 		return;
1863 
1864 	now = current_time(inode);
1865 	if (!timespec64_equal(&inode->i_mtime, &now))
1866 		inode->i_mtime = now;
1867 
1868 	if (!timespec64_equal(&inode->i_ctime, &now))
1869 		inode->i_ctime = now;
1870 
1871 	if (IS_I_VERSION(inode))
1872 		inode_inc_iversion(inode);
1873 }
1874 
1875 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1876 				    struct iov_iter *from)
1877 {
1878 	struct file *file = iocb->ki_filp;
1879 	struct inode *inode = file_inode(file);
1880 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1881 	struct btrfs_root *root = BTRFS_I(inode)->root;
1882 	u64 start_pos;
1883 	u64 end_pos;
1884 	ssize_t num_written = 0;
1885 	bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1886 	ssize_t err;
1887 	loff_t pos;
1888 	size_t count = iov_iter_count(from);
1889 	loff_t oldsize;
1890 	int clean_page = 0;
1891 
1892 	if (!(iocb->ki_flags & IOCB_DIRECT) &&
1893 	    (iocb->ki_flags & IOCB_NOWAIT))
1894 		return -EOPNOTSUPP;
1895 
1896 	if (!inode_trylock(inode)) {
1897 		if (iocb->ki_flags & IOCB_NOWAIT)
1898 			return -EAGAIN;
1899 		inode_lock(inode);
1900 	}
1901 
1902 	err = generic_write_checks(iocb, from);
1903 	if (err <= 0) {
1904 		inode_unlock(inode);
1905 		return err;
1906 	}
1907 
1908 	pos = iocb->ki_pos;
1909 	if (iocb->ki_flags & IOCB_NOWAIT) {
1910 		/*
1911 		 * We will allocate space in case nodatacow is not set,
1912 		 * so bail
1913 		 */
1914 		if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1915 					      BTRFS_INODE_PREALLOC)) ||
1916 		    check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1917 			inode_unlock(inode);
1918 			return -EAGAIN;
1919 		}
1920 	}
1921 
1922 	current->backing_dev_info = inode_to_bdi(inode);
1923 	err = file_remove_privs(file);
1924 	if (err) {
1925 		inode_unlock(inode);
1926 		goto out;
1927 	}
1928 
1929 	/*
1930 	 * If BTRFS flips readonly due to some impossible error
1931 	 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1932 	 * although we have opened a file as writable, we have
1933 	 * to stop this write operation to ensure FS consistency.
1934 	 */
1935 	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1936 		inode_unlock(inode);
1937 		err = -EROFS;
1938 		goto out;
1939 	}
1940 
1941 	/*
1942 	 * We reserve space for updating the inode when we reserve space for the
1943 	 * extent we are going to write, so we will enospc out there.  We don't
1944 	 * need to start yet another transaction to update the inode as we will
1945 	 * update the inode when we finish writing whatever data we write.
1946 	 */
1947 	update_time_for_write(inode);
1948 
1949 	start_pos = round_down(pos, fs_info->sectorsize);
1950 	oldsize = i_size_read(inode);
1951 	if (start_pos > oldsize) {
1952 		/* Expand hole size to cover write data, preventing empty gap */
1953 		end_pos = round_up(pos + count,
1954 				   fs_info->sectorsize);
1955 		err = btrfs_cont_expand(inode, oldsize, end_pos);
1956 		if (err) {
1957 			inode_unlock(inode);
1958 			goto out;
1959 		}
1960 		if (start_pos > round_up(oldsize, fs_info->sectorsize))
1961 			clean_page = 1;
1962 	}
1963 
1964 	if (sync)
1965 		atomic_inc(&BTRFS_I(inode)->sync_writers);
1966 
1967 	if (iocb->ki_flags & IOCB_DIRECT) {
1968 		num_written = __btrfs_direct_write(iocb, from);
1969 	} else {
1970 		num_written = btrfs_buffered_write(iocb, from);
1971 		if (num_written > 0)
1972 			iocb->ki_pos = pos + num_written;
1973 		if (clean_page)
1974 			pagecache_isize_extended(inode, oldsize,
1975 						i_size_read(inode));
1976 	}
1977 
1978 	inode_unlock(inode);
1979 
1980 	/*
1981 	 * We also have to set last_sub_trans to the current log transid,
1982 	 * otherwise subsequent syncs to a file that's been synced in this
1983 	 * transaction will appear to have already occurred.
1984 	 */
1985 	spin_lock(&BTRFS_I(inode)->lock);
1986 	BTRFS_I(inode)->last_sub_trans = root->log_transid;
1987 	spin_unlock(&BTRFS_I(inode)->lock);
1988 	if (num_written > 0)
1989 		num_written = generic_write_sync(iocb, num_written);
1990 
1991 	if (sync)
1992 		atomic_dec(&BTRFS_I(inode)->sync_writers);
1993 out:
1994 	current->backing_dev_info = NULL;
1995 	return num_written ? num_written : err;
1996 }
1997 
1998 int btrfs_release_file(struct inode *inode, struct file *filp)
1999 {
2000 	struct btrfs_file_private *private = filp->private_data;
2001 
2002 	if (private && private->filldir_buf)
2003 		kfree(private->filldir_buf);
2004 	kfree(private);
2005 	filp->private_data = NULL;
2006 
2007 	/*
2008 	 * ordered_data_close is set by setattr when we are about to truncate
2009 	 * a file from a non-zero size to a zero size.  This tries to
2010 	 * flush down new bytes that may have been written if the
2011 	 * application were using truncate to replace a file in place.
2012 	 */
2013 	if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2014 			       &BTRFS_I(inode)->runtime_flags))
2015 			filemap_flush(inode->i_mapping);
2016 	return 0;
2017 }
2018 
2019 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2020 {
2021 	int ret;
2022 	struct blk_plug plug;
2023 
2024 	/*
2025 	 * This is only called in fsync, which would do synchronous writes, so
2026 	 * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2027 	 * multiple disks using raid profile, a large IO can be split to
2028 	 * several segments of stripe length (currently 64K).
2029 	 */
2030 	blk_start_plug(&plug);
2031 	atomic_inc(&BTRFS_I(inode)->sync_writers);
2032 	ret = btrfs_fdatawrite_range(inode, start, end);
2033 	atomic_dec(&BTRFS_I(inode)->sync_writers);
2034 	blk_finish_plug(&plug);
2035 
2036 	return ret;
2037 }
2038 
2039 /*
2040  * fsync call for both files and directories.  This logs the inode into
2041  * the tree log instead of forcing full commits whenever possible.
2042  *
2043  * It needs to call filemap_fdatawait so that all ordered extent updates are
2044  * in the metadata btree are up to date for copying to the log.
2045  *
2046  * It drops the inode mutex before doing the tree log commit.  This is an
2047  * important optimization for directories because holding the mutex prevents
2048  * new operations on the dir while we write to disk.
2049  */
2050 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2051 {
2052 	struct dentry *dentry = file_dentry(file);
2053 	struct inode *inode = d_inode(dentry);
2054 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2055 	struct btrfs_root *root = BTRFS_I(inode)->root;
2056 	struct btrfs_trans_handle *trans;
2057 	struct btrfs_log_ctx ctx;
2058 	int ret = 0, err;
2059 	u64 len;
2060 
2061 	/*
2062 	 * If the inode needs a full sync, make sure we use a full range to
2063 	 * avoid log tree corruption, due to hole detection racing with ordered
2064 	 * extent completion for adjacent ranges, and assertion failures during
2065 	 * hole detection.
2066 	 */
2067 	if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2068 		     &BTRFS_I(inode)->runtime_flags)) {
2069 		start = 0;
2070 		end = LLONG_MAX;
2071 	}
2072 
2073 	/*
2074 	 * The range length can be represented by u64, we have to do the typecasts
2075 	 * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2076 	 */
2077 	len = (u64)end - (u64)start + 1;
2078 	trace_btrfs_sync_file(file, datasync);
2079 
2080 	btrfs_init_log_ctx(&ctx, inode);
2081 
2082 	/*
2083 	 * We write the dirty pages in the range and wait until they complete
2084 	 * out of the ->i_mutex. If so, we can flush the dirty pages by
2085 	 * multi-task, and make the performance up.  See
2086 	 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2087 	 */
2088 	ret = start_ordered_ops(inode, start, end);
2089 	if (ret)
2090 		goto out;
2091 
2092 	inode_lock(inode);
2093 
2094 	/*
2095 	 * We take the dio_sem here because the tree log stuff can race with
2096 	 * lockless dio writes and get an extent map logged for an extent we
2097 	 * never waited on.  We need it this high up for lockdep reasons.
2098 	 */
2099 	down_write(&BTRFS_I(inode)->dio_sem);
2100 
2101 	atomic_inc(&root->log_batch);
2102 
2103 	/*
2104 	 * Before we acquired the inode's lock, someone may have dirtied more
2105 	 * pages in the target range. We need to make sure that writeback for
2106 	 * any such pages does not start while we are logging the inode, because
2107 	 * if it does, any of the following might happen when we are not doing a
2108 	 * full inode sync:
2109 	 *
2110 	 * 1) We log an extent after its writeback finishes but before its
2111 	 *    checksums are added to the csum tree, leading to -EIO errors
2112 	 *    when attempting to read the extent after a log replay.
2113 	 *
2114 	 * 2) We can end up logging an extent before its writeback finishes.
2115 	 *    Therefore after the log replay we will have a file extent item
2116 	 *    pointing to an unwritten extent (and no data checksums as well).
2117 	 *
2118 	 * So trigger writeback for any eventual new dirty pages and then we
2119 	 * wait for all ordered extents to complete below.
2120 	 */
2121 	ret = start_ordered_ops(inode, start, end);
2122 	if (ret) {
2123 		inode_unlock(inode);
2124 		goto out;
2125 	}
2126 
2127 	/*
2128 	 * We have to do this here to avoid the priority inversion of waiting on
2129 	 * IO of a lower priority task while holding a transaction open.
2130 	 */
2131 	ret = btrfs_wait_ordered_range(inode, start, len);
2132 	if (ret) {
2133 		up_write(&BTRFS_I(inode)->dio_sem);
2134 		inode_unlock(inode);
2135 		goto out;
2136 	}
2137 	atomic_inc(&root->log_batch);
2138 
2139 	smp_mb();
2140 	if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2141 	    BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
2142 		/*
2143 		 * We've had everything committed since the last time we were
2144 		 * modified so clear this flag in case it was set for whatever
2145 		 * reason, it's no longer relevant.
2146 		 */
2147 		clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2148 			  &BTRFS_I(inode)->runtime_flags);
2149 		/*
2150 		 * An ordered extent might have started before and completed
2151 		 * already with io errors, in which case the inode was not
2152 		 * updated and we end up here. So check the inode's mapping
2153 		 * for any errors that might have happened since we last
2154 		 * checked called fsync.
2155 		 */
2156 		ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2157 		up_write(&BTRFS_I(inode)->dio_sem);
2158 		inode_unlock(inode);
2159 		goto out;
2160 	}
2161 
2162 	/*
2163 	 * We use start here because we will need to wait on the IO to complete
2164 	 * in btrfs_sync_log, which could require joining a transaction (for
2165 	 * example checking cross references in the nocow path).  If we use join
2166 	 * here we could get into a situation where we're waiting on IO to
2167 	 * happen that is blocked on a transaction trying to commit.  With start
2168 	 * we inc the extwriter counter, so we wait for all extwriters to exit
2169 	 * before we start blocking joiners.  This comment is to keep somebody
2170 	 * from thinking they are super smart and changing this to
2171 	 * btrfs_join_transaction *cough*Josef*cough*.
2172 	 */
2173 	trans = btrfs_start_transaction(root, 0);
2174 	if (IS_ERR(trans)) {
2175 		ret = PTR_ERR(trans);
2176 		up_write(&BTRFS_I(inode)->dio_sem);
2177 		inode_unlock(inode);
2178 		goto out;
2179 	}
2180 
2181 	ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2182 	if (ret < 0) {
2183 		/* Fallthrough and commit/free transaction. */
2184 		ret = 1;
2185 	}
2186 
2187 	/* we've logged all the items and now have a consistent
2188 	 * version of the file in the log.  It is possible that
2189 	 * someone will come in and modify the file, but that's
2190 	 * fine because the log is consistent on disk, and we
2191 	 * have references to all of the file's extents
2192 	 *
2193 	 * It is possible that someone will come in and log the
2194 	 * file again, but that will end up using the synchronization
2195 	 * inside btrfs_sync_log to keep things safe.
2196 	 */
2197 	up_write(&BTRFS_I(inode)->dio_sem);
2198 	inode_unlock(inode);
2199 
2200 	if (ret != BTRFS_NO_LOG_SYNC) {
2201 		if (!ret) {
2202 			ret = btrfs_sync_log(trans, root, &ctx);
2203 			if (!ret) {
2204 				ret = btrfs_end_transaction(trans);
2205 				goto out;
2206 			}
2207 		}
2208 		ret = btrfs_commit_transaction(trans);
2209 	} else {
2210 		ret = btrfs_end_transaction(trans);
2211 	}
2212 out:
2213 	ASSERT(list_empty(&ctx.list));
2214 	err = file_check_and_advance_wb_err(file);
2215 	if (!ret)
2216 		ret = err;
2217 	return ret > 0 ? -EIO : ret;
2218 }
2219 
2220 static const struct vm_operations_struct btrfs_file_vm_ops = {
2221 	.fault		= filemap_fault,
2222 	.map_pages	= filemap_map_pages,
2223 	.page_mkwrite	= btrfs_page_mkwrite,
2224 };
2225 
2226 static int btrfs_file_mmap(struct file	*filp, struct vm_area_struct *vma)
2227 {
2228 	struct address_space *mapping = filp->f_mapping;
2229 
2230 	if (!mapping->a_ops->readpage)
2231 		return -ENOEXEC;
2232 
2233 	file_accessed(filp);
2234 	vma->vm_ops = &btrfs_file_vm_ops;
2235 
2236 	return 0;
2237 }
2238 
2239 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2240 			  int slot, u64 start, u64 end)
2241 {
2242 	struct btrfs_file_extent_item *fi;
2243 	struct btrfs_key key;
2244 
2245 	if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2246 		return 0;
2247 
2248 	btrfs_item_key_to_cpu(leaf, &key, slot);
2249 	if (key.objectid != btrfs_ino(inode) ||
2250 	    key.type != BTRFS_EXTENT_DATA_KEY)
2251 		return 0;
2252 
2253 	fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2254 
2255 	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2256 		return 0;
2257 
2258 	if (btrfs_file_extent_disk_bytenr(leaf, fi))
2259 		return 0;
2260 
2261 	if (key.offset == end)
2262 		return 1;
2263 	if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2264 		return 1;
2265 	return 0;
2266 }
2267 
2268 static int fill_holes(struct btrfs_trans_handle *trans,
2269 		struct btrfs_inode *inode,
2270 		struct btrfs_path *path, u64 offset, u64 end)
2271 {
2272 	struct btrfs_fs_info *fs_info = trans->fs_info;
2273 	struct btrfs_root *root = inode->root;
2274 	struct extent_buffer *leaf;
2275 	struct btrfs_file_extent_item *fi;
2276 	struct extent_map *hole_em;
2277 	struct extent_map_tree *em_tree = &inode->extent_tree;
2278 	struct btrfs_key key;
2279 	int ret;
2280 
2281 	if (btrfs_fs_incompat(fs_info, NO_HOLES))
2282 		goto out;
2283 
2284 	key.objectid = btrfs_ino(inode);
2285 	key.type = BTRFS_EXTENT_DATA_KEY;
2286 	key.offset = offset;
2287 
2288 	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2289 	if (ret <= 0) {
2290 		/*
2291 		 * We should have dropped this offset, so if we find it then
2292 		 * something has gone horribly wrong.
2293 		 */
2294 		if (ret == 0)
2295 			ret = -EINVAL;
2296 		return ret;
2297 	}
2298 
2299 	leaf = path->nodes[0];
2300 	if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2301 		u64 num_bytes;
2302 
2303 		path->slots[0]--;
2304 		fi = btrfs_item_ptr(leaf, path->slots[0],
2305 				    struct btrfs_file_extent_item);
2306 		num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2307 			end - offset;
2308 		btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2309 		btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2310 		btrfs_set_file_extent_offset(leaf, fi, 0);
2311 		btrfs_mark_buffer_dirty(leaf);
2312 		goto out;
2313 	}
2314 
2315 	if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2316 		u64 num_bytes;
2317 
2318 		key.offset = offset;
2319 		btrfs_set_item_key_safe(fs_info, path, &key);
2320 		fi = btrfs_item_ptr(leaf, path->slots[0],
2321 				    struct btrfs_file_extent_item);
2322 		num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2323 			offset;
2324 		btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2325 		btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2326 		btrfs_set_file_extent_offset(leaf, fi, 0);
2327 		btrfs_mark_buffer_dirty(leaf);
2328 		goto out;
2329 	}
2330 	btrfs_release_path(path);
2331 
2332 	ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2333 			offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2334 	if (ret)
2335 		return ret;
2336 
2337 out:
2338 	btrfs_release_path(path);
2339 
2340 	hole_em = alloc_extent_map();
2341 	if (!hole_em) {
2342 		btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2343 		set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2344 	} else {
2345 		hole_em->start = offset;
2346 		hole_em->len = end - offset;
2347 		hole_em->ram_bytes = hole_em->len;
2348 		hole_em->orig_start = offset;
2349 
2350 		hole_em->block_start = EXTENT_MAP_HOLE;
2351 		hole_em->block_len = 0;
2352 		hole_em->orig_block_len = 0;
2353 		hole_em->bdev = fs_info->fs_devices->latest_bdev;
2354 		hole_em->compress_type = BTRFS_COMPRESS_NONE;
2355 		hole_em->generation = trans->transid;
2356 
2357 		do {
2358 			btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2359 			write_lock(&em_tree->lock);
2360 			ret = add_extent_mapping(em_tree, hole_em, 1);
2361 			write_unlock(&em_tree->lock);
2362 		} while (ret == -EEXIST);
2363 		free_extent_map(hole_em);
2364 		if (ret)
2365 			set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2366 					&inode->runtime_flags);
2367 	}
2368 
2369 	return 0;
2370 }
2371 
2372 /*
2373  * Find a hole extent on given inode and change start/len to the end of hole
2374  * extent.(hole/vacuum extent whose em->start <= start &&
2375  *	   em->start + em->len > start)
2376  * When a hole extent is found, return 1 and modify start/len.
2377  */
2378 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2379 {
2380 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2381 	struct extent_map *em;
2382 	int ret = 0;
2383 
2384 	em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2385 			      round_down(*start, fs_info->sectorsize),
2386 			      round_up(*len, fs_info->sectorsize), 0);
2387 	if (IS_ERR(em))
2388 		return PTR_ERR(em);
2389 
2390 	/* Hole or vacuum extent(only exists in no-hole mode) */
2391 	if (em->block_start == EXTENT_MAP_HOLE) {
2392 		ret = 1;
2393 		*len = em->start + em->len > *start + *len ?
2394 		       0 : *start + *len - em->start - em->len;
2395 		*start = em->start + em->len;
2396 	}
2397 	free_extent_map(em);
2398 	return ret;
2399 }
2400 
2401 static int btrfs_punch_hole_lock_range(struct inode *inode,
2402 				       const u64 lockstart,
2403 				       const u64 lockend,
2404 				       struct extent_state **cached_state)
2405 {
2406 	while (1) {
2407 		struct btrfs_ordered_extent *ordered;
2408 		int ret;
2409 
2410 		truncate_pagecache_range(inode, lockstart, lockend);
2411 
2412 		lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2413 				 cached_state);
2414 		ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2415 
2416 		/*
2417 		 * We need to make sure we have no ordered extents in this range
2418 		 * and nobody raced in and read a page in this range, if we did
2419 		 * we need to try again.
2420 		 */
2421 		if ((!ordered ||
2422 		    (ordered->file_offset + ordered->len <= lockstart ||
2423 		     ordered->file_offset > lockend)) &&
2424 		     !filemap_range_has_page(inode->i_mapping,
2425 					     lockstart, lockend)) {
2426 			if (ordered)
2427 				btrfs_put_ordered_extent(ordered);
2428 			break;
2429 		}
2430 		if (ordered)
2431 			btrfs_put_ordered_extent(ordered);
2432 		unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2433 				     lockend, cached_state);
2434 		ret = btrfs_wait_ordered_range(inode, lockstart,
2435 					       lockend - lockstart + 1);
2436 		if (ret)
2437 			return ret;
2438 	}
2439 	return 0;
2440 }
2441 
2442 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2443 {
2444 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2445 	struct btrfs_root *root = BTRFS_I(inode)->root;
2446 	struct extent_state *cached_state = NULL;
2447 	struct btrfs_path *path;
2448 	struct btrfs_block_rsv *rsv;
2449 	struct btrfs_trans_handle *trans;
2450 	u64 lockstart;
2451 	u64 lockend;
2452 	u64 tail_start;
2453 	u64 tail_len;
2454 	u64 orig_start = offset;
2455 	u64 cur_offset;
2456 	u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2457 	u64 drop_end;
2458 	int ret = 0;
2459 	int err = 0;
2460 	unsigned int rsv_count;
2461 	bool same_block;
2462 	bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2463 	u64 ino_size;
2464 	bool truncated_block = false;
2465 	bool updated_inode = false;
2466 
2467 	ret = btrfs_wait_ordered_range(inode, offset, len);
2468 	if (ret)
2469 		return ret;
2470 
2471 	inode_lock(inode);
2472 	ino_size = round_up(inode->i_size, fs_info->sectorsize);
2473 	ret = find_first_non_hole(inode, &offset, &len);
2474 	if (ret < 0)
2475 		goto out_only_mutex;
2476 	if (ret && !len) {
2477 		/* Already in a large hole */
2478 		ret = 0;
2479 		goto out_only_mutex;
2480 	}
2481 
2482 	lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2483 	lockend = round_down(offset + len,
2484 			     btrfs_inode_sectorsize(inode)) - 1;
2485 	same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2486 		== (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2487 	/*
2488 	 * We needn't truncate any block which is beyond the end of the file
2489 	 * because we are sure there is no data there.
2490 	 */
2491 	/*
2492 	 * Only do this if we are in the same block and we aren't doing the
2493 	 * entire block.
2494 	 */
2495 	if (same_block && len < fs_info->sectorsize) {
2496 		if (offset < ino_size) {
2497 			truncated_block = true;
2498 			ret = btrfs_truncate_block(inode, offset, len, 0);
2499 		} else {
2500 			ret = 0;
2501 		}
2502 		goto out_only_mutex;
2503 	}
2504 
2505 	/* zero back part of the first block */
2506 	if (offset < ino_size) {
2507 		truncated_block = true;
2508 		ret = btrfs_truncate_block(inode, offset, 0, 0);
2509 		if (ret) {
2510 			inode_unlock(inode);
2511 			return ret;
2512 		}
2513 	}
2514 
2515 	/* Check the aligned pages after the first unaligned page,
2516 	 * if offset != orig_start, which means the first unaligned page
2517 	 * including several following pages are already in holes,
2518 	 * the extra check can be skipped */
2519 	if (offset == orig_start) {
2520 		/* after truncate page, check hole again */
2521 		len = offset + len - lockstart;
2522 		offset = lockstart;
2523 		ret = find_first_non_hole(inode, &offset, &len);
2524 		if (ret < 0)
2525 			goto out_only_mutex;
2526 		if (ret && !len) {
2527 			ret = 0;
2528 			goto out_only_mutex;
2529 		}
2530 		lockstart = offset;
2531 	}
2532 
2533 	/* Check the tail unaligned part is in a hole */
2534 	tail_start = lockend + 1;
2535 	tail_len = offset + len - tail_start;
2536 	if (tail_len) {
2537 		ret = find_first_non_hole(inode, &tail_start, &tail_len);
2538 		if (unlikely(ret < 0))
2539 			goto out_only_mutex;
2540 		if (!ret) {
2541 			/* zero the front end of the last page */
2542 			if (tail_start + tail_len < ino_size) {
2543 				truncated_block = true;
2544 				ret = btrfs_truncate_block(inode,
2545 							tail_start + tail_len,
2546 							0, 1);
2547 				if (ret)
2548 					goto out_only_mutex;
2549 			}
2550 		}
2551 	}
2552 
2553 	if (lockend < lockstart) {
2554 		ret = 0;
2555 		goto out_only_mutex;
2556 	}
2557 
2558 	ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2559 					  &cached_state);
2560 	if (ret)
2561 		goto out_only_mutex;
2562 
2563 	path = btrfs_alloc_path();
2564 	if (!path) {
2565 		ret = -ENOMEM;
2566 		goto out;
2567 	}
2568 
2569 	rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2570 	if (!rsv) {
2571 		ret = -ENOMEM;
2572 		goto out_free;
2573 	}
2574 	rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2575 	rsv->failfast = 1;
2576 
2577 	/*
2578 	 * 1 - update the inode
2579 	 * 1 - removing the extents in the range
2580 	 * 1 - adding the hole extent if no_holes isn't set
2581 	 */
2582 	rsv_count = no_holes ? 2 : 3;
2583 	trans = btrfs_start_transaction(root, rsv_count);
2584 	if (IS_ERR(trans)) {
2585 		err = PTR_ERR(trans);
2586 		goto out_free;
2587 	}
2588 
2589 	ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2590 				      min_size, false);
2591 	BUG_ON(ret);
2592 	trans->block_rsv = rsv;
2593 
2594 	cur_offset = lockstart;
2595 	len = lockend - cur_offset;
2596 	while (cur_offset < lockend) {
2597 		ret = __btrfs_drop_extents(trans, root, inode, path,
2598 					   cur_offset, lockend + 1,
2599 					   &drop_end, 1, 0, 0, NULL);
2600 		if (ret != -ENOSPC)
2601 			break;
2602 
2603 		trans->block_rsv = &fs_info->trans_block_rsv;
2604 
2605 		if (cur_offset < drop_end && cur_offset < ino_size) {
2606 			ret = fill_holes(trans, BTRFS_I(inode), path,
2607 					cur_offset, drop_end);
2608 			if (ret) {
2609 				/*
2610 				 * If we failed then we didn't insert our hole
2611 				 * entries for the area we dropped, so now the
2612 				 * fs is corrupted, so we must abort the
2613 				 * transaction.
2614 				 */
2615 				btrfs_abort_transaction(trans, ret);
2616 				err = ret;
2617 				break;
2618 			}
2619 		}
2620 
2621 		cur_offset = drop_end;
2622 
2623 		ret = btrfs_update_inode(trans, root, inode);
2624 		if (ret) {
2625 			err = ret;
2626 			break;
2627 		}
2628 
2629 		btrfs_end_transaction(trans);
2630 		btrfs_btree_balance_dirty(fs_info);
2631 
2632 		trans = btrfs_start_transaction(root, rsv_count);
2633 		if (IS_ERR(trans)) {
2634 			ret = PTR_ERR(trans);
2635 			trans = NULL;
2636 			break;
2637 		}
2638 
2639 		ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2640 					      rsv, min_size, false);
2641 		BUG_ON(ret);	/* shouldn't happen */
2642 		trans->block_rsv = rsv;
2643 
2644 		ret = find_first_non_hole(inode, &cur_offset, &len);
2645 		if (unlikely(ret < 0))
2646 			break;
2647 		if (ret && !len) {
2648 			ret = 0;
2649 			break;
2650 		}
2651 	}
2652 
2653 	if (ret) {
2654 		err = ret;
2655 		goto out_trans;
2656 	}
2657 
2658 	trans->block_rsv = &fs_info->trans_block_rsv;
2659 	/*
2660 	 * If we are using the NO_HOLES feature we might have had already an
2661 	 * hole that overlaps a part of the region [lockstart, lockend] and
2662 	 * ends at (or beyond) lockend. Since we have no file extent items to
2663 	 * represent holes, drop_end can be less than lockend and so we must
2664 	 * make sure we have an extent map representing the existing hole (the
2665 	 * call to __btrfs_drop_extents() might have dropped the existing extent
2666 	 * map representing the existing hole), otherwise the fast fsync path
2667 	 * will not record the existence of the hole region
2668 	 * [existing_hole_start, lockend].
2669 	 */
2670 	if (drop_end <= lockend)
2671 		drop_end = lockend + 1;
2672 	/*
2673 	 * Don't insert file hole extent item if it's for a range beyond eof
2674 	 * (because it's useless) or if it represents a 0 bytes range (when
2675 	 * cur_offset == drop_end).
2676 	 */
2677 	if (cur_offset < ino_size && cur_offset < drop_end) {
2678 		ret = fill_holes(trans, BTRFS_I(inode), path,
2679 				cur_offset, drop_end);
2680 		if (ret) {
2681 			/* Same comment as above. */
2682 			btrfs_abort_transaction(trans, ret);
2683 			err = ret;
2684 			goto out_trans;
2685 		}
2686 	}
2687 
2688 out_trans:
2689 	if (!trans)
2690 		goto out_free;
2691 
2692 	inode_inc_iversion(inode);
2693 	inode->i_mtime = inode->i_ctime = current_time(inode);
2694 
2695 	trans->block_rsv = &fs_info->trans_block_rsv;
2696 	ret = btrfs_update_inode(trans, root, inode);
2697 	updated_inode = true;
2698 	btrfs_end_transaction(trans);
2699 	btrfs_btree_balance_dirty(fs_info);
2700 out_free:
2701 	btrfs_free_path(path);
2702 	btrfs_free_block_rsv(fs_info, rsv);
2703 out:
2704 	unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2705 			     &cached_state);
2706 out_only_mutex:
2707 	if (!updated_inode && truncated_block && !ret && !err) {
2708 		/*
2709 		 * If we only end up zeroing part of a page, we still need to
2710 		 * update the inode item, so that all the time fields are
2711 		 * updated as well as the necessary btrfs inode in memory fields
2712 		 * for detecting, at fsync time, if the inode isn't yet in the
2713 		 * log tree or it's there but not up to date.
2714 		 */
2715 		struct timespec64 now = current_time(inode);
2716 
2717 		inode_inc_iversion(inode);
2718 		inode->i_mtime = now;
2719 		inode->i_ctime = now;
2720 		trans = btrfs_start_transaction(root, 1);
2721 		if (IS_ERR(trans)) {
2722 			err = PTR_ERR(trans);
2723 		} else {
2724 			err = btrfs_update_inode(trans, root, inode);
2725 			ret = btrfs_end_transaction(trans);
2726 		}
2727 	}
2728 	inode_unlock(inode);
2729 	if (ret && !err)
2730 		err = ret;
2731 	return err;
2732 }
2733 
2734 /* Helper structure to record which range is already reserved */
2735 struct falloc_range {
2736 	struct list_head list;
2737 	u64 start;
2738 	u64 len;
2739 };
2740 
2741 /*
2742  * Helper function to add falloc range
2743  *
2744  * Caller should have locked the larger range of extent containing
2745  * [start, len)
2746  */
2747 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2748 {
2749 	struct falloc_range *prev = NULL;
2750 	struct falloc_range *range = NULL;
2751 
2752 	if (list_empty(head))
2753 		goto insert;
2754 
2755 	/*
2756 	 * As fallocate iterate by bytenr order, we only need to check
2757 	 * the last range.
2758 	 */
2759 	prev = list_entry(head->prev, struct falloc_range, list);
2760 	if (prev->start + prev->len == start) {
2761 		prev->len += len;
2762 		return 0;
2763 	}
2764 insert:
2765 	range = kmalloc(sizeof(*range), GFP_KERNEL);
2766 	if (!range)
2767 		return -ENOMEM;
2768 	range->start = start;
2769 	range->len = len;
2770 	list_add_tail(&range->list, head);
2771 	return 0;
2772 }
2773 
2774 static int btrfs_fallocate_update_isize(struct inode *inode,
2775 					const u64 end,
2776 					const int mode)
2777 {
2778 	struct btrfs_trans_handle *trans;
2779 	struct btrfs_root *root = BTRFS_I(inode)->root;
2780 	int ret;
2781 	int ret2;
2782 
2783 	if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2784 		return 0;
2785 
2786 	trans = btrfs_start_transaction(root, 1);
2787 	if (IS_ERR(trans))
2788 		return PTR_ERR(trans);
2789 
2790 	inode->i_ctime = current_time(inode);
2791 	i_size_write(inode, end);
2792 	btrfs_ordered_update_i_size(inode, end, NULL);
2793 	ret = btrfs_update_inode(trans, root, inode);
2794 	ret2 = btrfs_end_transaction(trans);
2795 
2796 	return ret ? ret : ret2;
2797 }
2798 
2799 enum {
2800 	RANGE_BOUNDARY_WRITTEN_EXTENT,
2801 	RANGE_BOUNDARY_PREALLOC_EXTENT,
2802 	RANGE_BOUNDARY_HOLE,
2803 };
2804 
2805 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2806 						 u64 offset)
2807 {
2808 	const u64 sectorsize = btrfs_inode_sectorsize(inode);
2809 	struct extent_map *em;
2810 	int ret;
2811 
2812 	offset = round_down(offset, sectorsize);
2813 	em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2814 	if (IS_ERR(em))
2815 		return PTR_ERR(em);
2816 
2817 	if (em->block_start == EXTENT_MAP_HOLE)
2818 		ret = RANGE_BOUNDARY_HOLE;
2819 	else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2820 		ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2821 	else
2822 		ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2823 
2824 	free_extent_map(em);
2825 	return ret;
2826 }
2827 
2828 static int btrfs_zero_range(struct inode *inode,
2829 			    loff_t offset,
2830 			    loff_t len,
2831 			    const int mode)
2832 {
2833 	struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2834 	struct extent_map *em;
2835 	struct extent_changeset *data_reserved = NULL;
2836 	int ret;
2837 	u64 alloc_hint = 0;
2838 	const u64 sectorsize = btrfs_inode_sectorsize(inode);
2839 	u64 alloc_start = round_down(offset, sectorsize);
2840 	u64 alloc_end = round_up(offset + len, sectorsize);
2841 	u64 bytes_to_reserve = 0;
2842 	bool space_reserved = false;
2843 
2844 	inode_dio_wait(inode);
2845 
2846 	em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2847 			      alloc_start, alloc_end - alloc_start, 0);
2848 	if (IS_ERR(em)) {
2849 		ret = PTR_ERR(em);
2850 		goto out;
2851 	}
2852 
2853 	/*
2854 	 * Avoid hole punching and extent allocation for some cases. More cases
2855 	 * could be considered, but these are unlikely common and we keep things
2856 	 * as simple as possible for now. Also, intentionally, if the target
2857 	 * range contains one or more prealloc extents together with regular
2858 	 * extents and holes, we drop all the existing extents and allocate a
2859 	 * new prealloc extent, so that we get a larger contiguous disk extent.
2860 	 */
2861 	if (em->start <= alloc_start &&
2862 	    test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2863 		const u64 em_end = em->start + em->len;
2864 
2865 		if (em_end >= offset + len) {
2866 			/*
2867 			 * The whole range is already a prealloc extent,
2868 			 * do nothing except updating the inode's i_size if
2869 			 * needed.
2870 			 */
2871 			free_extent_map(em);
2872 			ret = btrfs_fallocate_update_isize(inode, offset + len,
2873 							   mode);
2874 			goto out;
2875 		}
2876 		/*
2877 		 * Part of the range is already a prealloc extent, so operate
2878 		 * only on the remaining part of the range.
2879 		 */
2880 		alloc_start = em_end;
2881 		ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2882 		len = offset + len - alloc_start;
2883 		offset = alloc_start;
2884 		alloc_hint = em->block_start + em->len;
2885 	}
2886 	free_extent_map(em);
2887 
2888 	if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2889 	    BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2890 		em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2891 				      alloc_start, sectorsize, 0);
2892 		if (IS_ERR(em)) {
2893 			ret = PTR_ERR(em);
2894 			goto out;
2895 		}
2896 
2897 		if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2898 			free_extent_map(em);
2899 			ret = btrfs_fallocate_update_isize(inode, offset + len,
2900 							   mode);
2901 			goto out;
2902 		}
2903 		if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2904 			free_extent_map(em);
2905 			ret = btrfs_truncate_block(inode, offset, len, 0);
2906 			if (!ret)
2907 				ret = btrfs_fallocate_update_isize(inode,
2908 								   offset + len,
2909 								   mode);
2910 			return ret;
2911 		}
2912 		free_extent_map(em);
2913 		alloc_start = round_down(offset, sectorsize);
2914 		alloc_end = alloc_start + sectorsize;
2915 		goto reserve_space;
2916 	}
2917 
2918 	alloc_start = round_up(offset, sectorsize);
2919 	alloc_end = round_down(offset + len, sectorsize);
2920 
2921 	/*
2922 	 * For unaligned ranges, check the pages at the boundaries, they might
2923 	 * map to an extent, in which case we need to partially zero them, or
2924 	 * they might map to a hole, in which case we need our allocation range
2925 	 * to cover them.
2926 	 */
2927 	if (!IS_ALIGNED(offset, sectorsize)) {
2928 		ret = btrfs_zero_range_check_range_boundary(inode, offset);
2929 		if (ret < 0)
2930 			goto out;
2931 		if (ret == RANGE_BOUNDARY_HOLE) {
2932 			alloc_start = round_down(offset, sectorsize);
2933 			ret = 0;
2934 		} else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2935 			ret = btrfs_truncate_block(inode, offset, 0, 0);
2936 			if (ret)
2937 				goto out;
2938 		} else {
2939 			ret = 0;
2940 		}
2941 	}
2942 
2943 	if (!IS_ALIGNED(offset + len, sectorsize)) {
2944 		ret = btrfs_zero_range_check_range_boundary(inode,
2945 							    offset + len);
2946 		if (ret < 0)
2947 			goto out;
2948 		if (ret == RANGE_BOUNDARY_HOLE) {
2949 			alloc_end = round_up(offset + len, sectorsize);
2950 			ret = 0;
2951 		} else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2952 			ret = btrfs_truncate_block(inode, offset + len, 0, 1);
2953 			if (ret)
2954 				goto out;
2955 		} else {
2956 			ret = 0;
2957 		}
2958 	}
2959 
2960 reserve_space:
2961 	if (alloc_start < alloc_end) {
2962 		struct extent_state *cached_state = NULL;
2963 		const u64 lockstart = alloc_start;
2964 		const u64 lockend = alloc_end - 1;
2965 
2966 		bytes_to_reserve = alloc_end - alloc_start;
2967 		ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2968 						      bytes_to_reserve);
2969 		if (ret < 0)
2970 			goto out;
2971 		space_reserved = true;
2972 		ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
2973 						alloc_start, bytes_to_reserve);
2974 		if (ret)
2975 			goto out;
2976 		ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2977 						  &cached_state);
2978 		if (ret)
2979 			goto out;
2980 		ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
2981 						alloc_end - alloc_start,
2982 						i_blocksize(inode),
2983 						offset + len, &alloc_hint);
2984 		unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2985 				     lockend, &cached_state);
2986 		/* btrfs_prealloc_file_range releases reserved space on error */
2987 		if (ret) {
2988 			space_reserved = false;
2989 			goto out;
2990 		}
2991 	}
2992 	ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
2993  out:
2994 	if (ret && space_reserved)
2995 		btrfs_free_reserved_data_space(inode, data_reserved,
2996 					       alloc_start, bytes_to_reserve);
2997 	extent_changeset_free(data_reserved);
2998 
2999 	return ret;
3000 }
3001 
3002 static long btrfs_fallocate(struct file *file, int mode,
3003 			    loff_t offset, loff_t len)
3004 {
3005 	struct inode *inode = file_inode(file);
3006 	struct extent_state *cached_state = NULL;
3007 	struct extent_changeset *data_reserved = NULL;
3008 	struct falloc_range *range;
3009 	struct falloc_range *tmp;
3010 	struct list_head reserve_list;
3011 	u64 cur_offset;
3012 	u64 last_byte;
3013 	u64 alloc_start;
3014 	u64 alloc_end;
3015 	u64 alloc_hint = 0;
3016 	u64 locked_end;
3017 	u64 actual_end = 0;
3018 	struct extent_map *em;
3019 	int blocksize = btrfs_inode_sectorsize(inode);
3020 	int ret;
3021 
3022 	alloc_start = round_down(offset, blocksize);
3023 	alloc_end = round_up(offset + len, blocksize);
3024 	cur_offset = alloc_start;
3025 
3026 	/* Make sure we aren't being give some crap mode */
3027 	if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3028 		     FALLOC_FL_ZERO_RANGE))
3029 		return -EOPNOTSUPP;
3030 
3031 	if (mode & FALLOC_FL_PUNCH_HOLE)
3032 		return btrfs_punch_hole(inode, offset, len);
3033 
3034 	/*
3035 	 * Only trigger disk allocation, don't trigger qgroup reserve
3036 	 *
3037 	 * For qgroup space, it will be checked later.
3038 	 */
3039 	if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3040 		ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3041 						      alloc_end - alloc_start);
3042 		if (ret < 0)
3043 			return ret;
3044 	}
3045 
3046 	inode_lock(inode);
3047 
3048 	if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3049 		ret = inode_newsize_ok(inode, offset + len);
3050 		if (ret)
3051 			goto out;
3052 	}
3053 
3054 	/*
3055 	 * TODO: Move these two operations after we have checked
3056 	 * accurate reserved space, or fallocate can still fail but
3057 	 * with page truncated or size expanded.
3058 	 *
3059 	 * But that's a minor problem and won't do much harm BTW.
3060 	 */
3061 	if (alloc_start > inode->i_size) {
3062 		ret = btrfs_cont_expand(inode, i_size_read(inode),
3063 					alloc_start);
3064 		if (ret)
3065 			goto out;
3066 	} else if (offset + len > inode->i_size) {
3067 		/*
3068 		 * If we are fallocating from the end of the file onward we
3069 		 * need to zero out the end of the block if i_size lands in the
3070 		 * middle of a block.
3071 		 */
3072 		ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3073 		if (ret)
3074 			goto out;
3075 	}
3076 
3077 	/*
3078 	 * wait for ordered IO before we have any locks.  We'll loop again
3079 	 * below with the locks held.
3080 	 */
3081 	ret = btrfs_wait_ordered_range(inode, alloc_start,
3082 				       alloc_end - alloc_start);
3083 	if (ret)
3084 		goto out;
3085 
3086 	if (mode & FALLOC_FL_ZERO_RANGE) {
3087 		ret = btrfs_zero_range(inode, offset, len, mode);
3088 		inode_unlock(inode);
3089 		return ret;
3090 	}
3091 
3092 	locked_end = alloc_end - 1;
3093 	while (1) {
3094 		struct btrfs_ordered_extent *ordered;
3095 
3096 		/* the extent lock is ordered inside the running
3097 		 * transaction
3098 		 */
3099 		lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3100 				 locked_end, &cached_state);
3101 		ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3102 
3103 		if (ordered &&
3104 		    ordered->file_offset + ordered->len > alloc_start &&
3105 		    ordered->file_offset < alloc_end) {
3106 			btrfs_put_ordered_extent(ordered);
3107 			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3108 					     alloc_start, locked_end,
3109 					     &cached_state);
3110 			/*
3111 			 * we can't wait on the range with the transaction
3112 			 * running or with the extent lock held
3113 			 */
3114 			ret = btrfs_wait_ordered_range(inode, alloc_start,
3115 						       alloc_end - alloc_start);
3116 			if (ret)
3117 				goto out;
3118 		} else {
3119 			if (ordered)
3120 				btrfs_put_ordered_extent(ordered);
3121 			break;
3122 		}
3123 	}
3124 
3125 	/* First, check if we exceed the qgroup limit */
3126 	INIT_LIST_HEAD(&reserve_list);
3127 	while (cur_offset < alloc_end) {
3128 		em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3129 				      alloc_end - cur_offset, 0);
3130 		if (IS_ERR(em)) {
3131 			ret = PTR_ERR(em);
3132 			break;
3133 		}
3134 		last_byte = min(extent_map_end(em), alloc_end);
3135 		actual_end = min_t(u64, extent_map_end(em), offset + len);
3136 		last_byte = ALIGN(last_byte, blocksize);
3137 		if (em->block_start == EXTENT_MAP_HOLE ||
3138 		    (cur_offset >= inode->i_size &&
3139 		     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3140 			ret = add_falloc_range(&reserve_list, cur_offset,
3141 					       last_byte - cur_offset);
3142 			if (ret < 0) {
3143 				free_extent_map(em);
3144 				break;
3145 			}
3146 			ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3147 					cur_offset, last_byte - cur_offset);
3148 			if (ret < 0) {
3149 				cur_offset = last_byte;
3150 				free_extent_map(em);
3151 				break;
3152 			}
3153 		} else {
3154 			/*
3155 			 * Do not need to reserve unwritten extent for this
3156 			 * range, free reserved data space first, otherwise
3157 			 * it'll result in false ENOSPC error.
3158 			 */
3159 			btrfs_free_reserved_data_space(inode, data_reserved,
3160 					cur_offset, last_byte - cur_offset);
3161 		}
3162 		free_extent_map(em);
3163 		cur_offset = last_byte;
3164 	}
3165 
3166 	/*
3167 	 * If ret is still 0, means we're OK to fallocate.
3168 	 * Or just cleanup the list and exit.
3169 	 */
3170 	list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3171 		if (!ret)
3172 			ret = btrfs_prealloc_file_range(inode, mode,
3173 					range->start,
3174 					range->len, i_blocksize(inode),
3175 					offset + len, &alloc_hint);
3176 		else
3177 			btrfs_free_reserved_data_space(inode,
3178 					data_reserved, range->start,
3179 					range->len);
3180 		list_del(&range->list);
3181 		kfree(range);
3182 	}
3183 	if (ret < 0)
3184 		goto out_unlock;
3185 
3186 	/*
3187 	 * We didn't need to allocate any more space, but we still extended the
3188 	 * size of the file so we need to update i_size and the inode item.
3189 	 */
3190 	ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3191 out_unlock:
3192 	unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3193 			     &cached_state);
3194 out:
3195 	inode_unlock(inode);
3196 	/* Let go of our reservation. */
3197 	if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3198 		btrfs_free_reserved_data_space(inode, data_reserved,
3199 				cur_offset, alloc_end - cur_offset);
3200 	extent_changeset_free(data_reserved);
3201 	return ret;
3202 }
3203 
3204 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3205 {
3206 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3207 	struct extent_map *em = NULL;
3208 	struct extent_state *cached_state = NULL;
3209 	u64 lockstart;
3210 	u64 lockend;
3211 	u64 start;
3212 	u64 len;
3213 	int ret = 0;
3214 
3215 	if (inode->i_size == 0)
3216 		return -ENXIO;
3217 
3218 	/*
3219 	 * *offset can be negative, in this case we start finding DATA/HOLE from
3220 	 * the very start of the file.
3221 	 */
3222 	start = max_t(loff_t, 0, *offset);
3223 
3224 	lockstart = round_down(start, fs_info->sectorsize);
3225 	lockend = round_up(i_size_read(inode),
3226 			   fs_info->sectorsize);
3227 	if (lockend <= lockstart)
3228 		lockend = lockstart + fs_info->sectorsize;
3229 	lockend--;
3230 	len = lockend - lockstart + 1;
3231 
3232 	lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3233 			 &cached_state);
3234 
3235 	while (start < inode->i_size) {
3236 		em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3237 		if (IS_ERR(em)) {
3238 			ret = PTR_ERR(em);
3239 			em = NULL;
3240 			break;
3241 		}
3242 
3243 		if (whence == SEEK_HOLE &&
3244 		    (em->block_start == EXTENT_MAP_HOLE ||
3245 		     test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3246 			break;
3247 		else if (whence == SEEK_DATA &&
3248 			   (em->block_start != EXTENT_MAP_HOLE &&
3249 			    !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3250 			break;
3251 
3252 		start = em->start + em->len;
3253 		free_extent_map(em);
3254 		em = NULL;
3255 		cond_resched();
3256 	}
3257 	free_extent_map(em);
3258 	if (!ret) {
3259 		if (whence == SEEK_DATA && start >= inode->i_size)
3260 			ret = -ENXIO;
3261 		else
3262 			*offset = min_t(loff_t, start, inode->i_size);
3263 	}
3264 	unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3265 			     &cached_state);
3266 	return ret;
3267 }
3268 
3269 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3270 {
3271 	struct inode *inode = file->f_mapping->host;
3272 	int ret;
3273 
3274 	inode_lock(inode);
3275 	switch (whence) {
3276 	case SEEK_END:
3277 	case SEEK_CUR:
3278 		offset = generic_file_llseek(file, offset, whence);
3279 		goto out;
3280 	case SEEK_DATA:
3281 	case SEEK_HOLE:
3282 		if (offset >= i_size_read(inode)) {
3283 			inode_unlock(inode);
3284 			return -ENXIO;
3285 		}
3286 
3287 		ret = find_desired_extent(inode, &offset, whence);
3288 		if (ret) {
3289 			inode_unlock(inode);
3290 			return ret;
3291 		}
3292 	}
3293 
3294 	offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3295 out:
3296 	inode_unlock(inode);
3297 	return offset;
3298 }
3299 
3300 static int btrfs_file_open(struct inode *inode, struct file *filp)
3301 {
3302 	filp->f_mode |= FMODE_NOWAIT;
3303 	return generic_file_open(inode, filp);
3304 }
3305 
3306 const struct file_operations btrfs_file_operations = {
3307 	.llseek		= btrfs_file_llseek,
3308 	.read_iter      = generic_file_read_iter,
3309 	.splice_read	= generic_file_splice_read,
3310 	.write_iter	= btrfs_file_write_iter,
3311 	.mmap		= btrfs_file_mmap,
3312 	.open		= btrfs_file_open,
3313 	.release	= btrfs_release_file,
3314 	.fsync		= btrfs_sync_file,
3315 	.fallocate	= btrfs_fallocate,
3316 	.unlocked_ioctl	= btrfs_ioctl,
3317 #ifdef CONFIG_COMPAT
3318 	.compat_ioctl	= btrfs_compat_ioctl,
3319 #endif
3320 	.remap_file_range = btrfs_remap_file_range,
3321 };
3322 
3323 void __cold btrfs_auto_defrag_exit(void)
3324 {
3325 	kmem_cache_destroy(btrfs_inode_defrag_cachep);
3326 }
3327 
3328 int __init btrfs_auto_defrag_init(void)
3329 {
3330 	btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3331 					sizeof(struct inode_defrag), 0,
3332 					SLAB_MEM_SPREAD,
3333 					NULL);
3334 	if (!btrfs_inode_defrag_cachep)
3335 		return -ENOMEM;
3336 
3337 	return 0;
3338 }
3339 
3340 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3341 {
3342 	int ret;
3343 
3344 	/*
3345 	 * So with compression we will find and lock a dirty page and clear the
3346 	 * first one as dirty, setup an async extent, and immediately return
3347 	 * with the entire range locked but with nobody actually marked with
3348 	 * writeback.  So we can't just filemap_write_and_wait_range() and
3349 	 * expect it to work since it will just kick off a thread to do the
3350 	 * actual work.  So we need to call filemap_fdatawrite_range _again_
3351 	 * since it will wait on the page lock, which won't be unlocked until
3352 	 * after the pages have been marked as writeback and so we're good to go
3353 	 * from there.  We have to do this otherwise we'll miss the ordered
3354 	 * extents and that results in badness.  Please Josef, do not think you
3355 	 * know better and pull this out at some point in the future, it is
3356 	 * right and you are wrong.
3357 	 */
3358 	ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3359 	if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3360 			     &BTRFS_I(inode)->runtime_flags))
3361 		ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3362 
3363 	return ret;
3364 }
3365