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 <linux/fsverity.h>
20#include "ctree.h"
21#include "disk-io.h"
22#include "transaction.h"
23#include "btrfs_inode.h"
24#include "print-tree.h"
25#include "tree-log.h"
26#include "locking.h"
27#include "volumes.h"
28#include "qgroup.h"
29#include "compression.h"
30#include "delalloc-space.h"
31#include "reflink.h"
32#include "subpage.h"
33
34static struct kmem_cache *btrfs_inode_defrag_cachep;
35/*
36 * when auto defrag is enabled we
37 * queue up these defrag structs to remember which
38 * inodes need defragging passes
39 */
40struct inode_defrag {
41	struct rb_node rb_node;
42	/* objectid */
43	u64 ino;
44	/*
45	 * transid where the defrag was added, we search for
46	 * extents newer than this
47	 */
48	u64 transid;
49
50	/* root objectid */
51	u64 root;
52
53	/* last offset we were able to defrag */
54	u64 last_offset;
55
56	/* if we've wrapped around back to zero once already */
57	int cycled;
58};
59
60static int __compare_inode_defrag(struct inode_defrag *defrag1,
61				  struct inode_defrag *defrag2)
62{
63	if (defrag1->root > defrag2->root)
64		return 1;
65	else if (defrag1->root < defrag2->root)
66		return -1;
67	else if (defrag1->ino > defrag2->ino)
68		return 1;
69	else if (defrag1->ino < defrag2->ino)
70		return -1;
71	else
72		return 0;
73}
74
75/* pop a record for an inode into the defrag tree.  The lock
76 * must be held already
77 *
78 * If you're inserting a record for an older transid than an
79 * existing record, the transid already in the tree is lowered
80 *
81 * If an existing record is found the defrag item you
82 * pass in is freed
83 */
84static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
85				    struct inode_defrag *defrag)
86{
87	struct btrfs_fs_info *fs_info = inode->root->fs_info;
88	struct inode_defrag *entry;
89	struct rb_node **p;
90	struct rb_node *parent = NULL;
91	int ret;
92
93	p = &fs_info->defrag_inodes.rb_node;
94	while (*p) {
95		parent = *p;
96		entry = rb_entry(parent, struct inode_defrag, rb_node);
97
98		ret = __compare_inode_defrag(defrag, entry);
99		if (ret < 0)
100			p = &parent->rb_left;
101		else if (ret > 0)
102			p = &parent->rb_right;
103		else {
104			/* if we're reinserting an entry for
105			 * an old defrag run, make sure to
106			 * lower the transid of our existing record
107			 */
108			if (defrag->transid < entry->transid)
109				entry->transid = defrag->transid;
110			if (defrag->last_offset > entry->last_offset)
111				entry->last_offset = defrag->last_offset;
112			return -EEXIST;
113		}
114	}
115	set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
116	rb_link_node(&defrag->rb_node, parent, p);
117	rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
118	return 0;
119}
120
121static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
122{
123	if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
124		return 0;
125
126	if (btrfs_fs_closing(fs_info))
127		return 0;
128
129	return 1;
130}
131
132/*
133 * insert a defrag record for this inode if auto defrag is
134 * enabled
135 */
136int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
137			   struct btrfs_inode *inode)
138{
139	struct btrfs_root *root = inode->root;
140	struct btrfs_fs_info *fs_info = root->fs_info;
141	struct inode_defrag *defrag;
142	u64 transid;
143	int ret;
144
145	if (!__need_auto_defrag(fs_info))
146		return 0;
147
148	if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
149		return 0;
150
151	if (trans)
152		transid = trans->transid;
153	else
154		transid = inode->root->last_trans;
155
156	defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
157	if (!defrag)
158		return -ENOMEM;
159
160	defrag->ino = btrfs_ino(inode);
161	defrag->transid = transid;
162	defrag->root = root->root_key.objectid;
163
164	spin_lock(&fs_info->defrag_inodes_lock);
165	if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
166		/*
167		 * If we set IN_DEFRAG flag and evict the inode from memory,
168		 * and then re-read this inode, this new inode doesn't have
169		 * IN_DEFRAG flag. At the case, we may find the existed defrag.
170		 */
171		ret = __btrfs_add_inode_defrag(inode, defrag);
172		if (ret)
173			kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
174	} else {
175		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
176	}
177	spin_unlock(&fs_info->defrag_inodes_lock);
178	return 0;
179}
180
181/*
182 * Requeue the defrag object. If there is a defrag object that points to
183 * the same inode in the tree, we will merge them together (by
184 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
185 */
186static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
187				       struct inode_defrag *defrag)
188{
189	struct btrfs_fs_info *fs_info = inode->root->fs_info;
190	int ret;
191
192	if (!__need_auto_defrag(fs_info))
193		goto out;
194
195	/*
196	 * Here we don't check the IN_DEFRAG flag, because we need merge
197	 * them together.
198	 */
199	spin_lock(&fs_info->defrag_inodes_lock);
200	ret = __btrfs_add_inode_defrag(inode, defrag);
201	spin_unlock(&fs_info->defrag_inodes_lock);
202	if (ret)
203		goto out;
204	return;
205out:
206	kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
207}
208
209/*
210 * pick the defragable inode that we want, if it doesn't exist, we will get
211 * the next one.
212 */
213static struct inode_defrag *
214btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
215{
216	struct inode_defrag *entry = NULL;
217	struct inode_defrag tmp;
218	struct rb_node *p;
219	struct rb_node *parent = NULL;
220	int ret;
221
222	tmp.ino = ino;
223	tmp.root = root;
224
225	spin_lock(&fs_info->defrag_inodes_lock);
226	p = fs_info->defrag_inodes.rb_node;
227	while (p) {
228		parent = p;
229		entry = rb_entry(parent, struct inode_defrag, rb_node);
230
231		ret = __compare_inode_defrag(&tmp, entry);
232		if (ret < 0)
233			p = parent->rb_left;
234		else if (ret > 0)
235			p = parent->rb_right;
236		else
237			goto out;
238	}
239
240	if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
241		parent = rb_next(parent);
242		if (parent)
243			entry = rb_entry(parent, struct inode_defrag, rb_node);
244		else
245			entry = NULL;
246	}
247out:
248	if (entry)
249		rb_erase(parent, &fs_info->defrag_inodes);
250	spin_unlock(&fs_info->defrag_inodes_lock);
251	return entry;
252}
253
254void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
255{
256	struct inode_defrag *defrag;
257	struct rb_node *node;
258
259	spin_lock(&fs_info->defrag_inodes_lock);
260	node = rb_first(&fs_info->defrag_inodes);
261	while (node) {
262		rb_erase(node, &fs_info->defrag_inodes);
263		defrag = rb_entry(node, struct inode_defrag, rb_node);
264		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
265
266		cond_resched_lock(&fs_info->defrag_inodes_lock);
267
268		node = rb_first(&fs_info->defrag_inodes);
269	}
270	spin_unlock(&fs_info->defrag_inodes_lock);
271}
272
273#define BTRFS_DEFRAG_BATCH	1024
274
275static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
276				    struct inode_defrag *defrag)
277{
278	struct btrfs_root *inode_root;
279	struct inode *inode;
280	struct btrfs_ioctl_defrag_range_args range;
281	int num_defrag;
282	int ret;
283
284	/* get the inode */
285	inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
286	if (IS_ERR(inode_root)) {
287		ret = PTR_ERR(inode_root);
288		goto cleanup;
289	}
290
291	inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
292	btrfs_put_root(inode_root);
293	if (IS_ERR(inode)) {
294		ret = PTR_ERR(inode);
295		goto cleanup;
296	}
297
298	/* do a chunk of defrag */
299	clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
300	memset(&range, 0, sizeof(range));
301	range.len = (u64)-1;
302	range.start = defrag->last_offset;
303
304	sb_start_write(fs_info->sb);
305	num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
306				       BTRFS_DEFRAG_BATCH);
307	sb_end_write(fs_info->sb);
308	/*
309	 * if we filled the whole defrag batch, there
310	 * must be more work to do.  Queue this defrag
311	 * again
312	 */
313	if (num_defrag == BTRFS_DEFRAG_BATCH) {
314		defrag->last_offset = range.start;
315		btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
316	} else if (defrag->last_offset && !defrag->cycled) {
317		/*
318		 * we didn't fill our defrag batch, but
319		 * we didn't start at zero.  Make sure we loop
320		 * around to the start of the file.
321		 */
322		defrag->last_offset = 0;
323		defrag->cycled = 1;
324		btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
325	} else {
326		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
327	}
328
329	iput(inode);
330	return 0;
331cleanup:
332	kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
333	return ret;
334}
335
336/*
337 * run through the list of inodes in the FS that need
338 * defragging
339 */
340int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
341{
342	struct inode_defrag *defrag;
343	u64 first_ino = 0;
344	u64 root_objectid = 0;
345
346	atomic_inc(&fs_info->defrag_running);
347	while (1) {
348		/* Pause the auto defragger. */
349		if (test_bit(BTRFS_FS_STATE_REMOUNTING,
350			     &fs_info->fs_state))
351			break;
352
353		if (!__need_auto_defrag(fs_info))
354			break;
355
356		/* find an inode to defrag */
357		defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
358						 first_ino);
359		if (!defrag) {
360			if (root_objectid || first_ino) {
361				root_objectid = 0;
362				first_ino = 0;
363				continue;
364			} else {
365				break;
366			}
367		}
368
369		first_ino = defrag->ino + 1;
370		root_objectid = defrag->root;
371
372		__btrfs_run_defrag_inode(fs_info, defrag);
373	}
374	atomic_dec(&fs_info->defrag_running);
375
376	/*
377	 * during unmount, we use the transaction_wait queue to
378	 * wait for the defragger to stop
379	 */
380	wake_up(&fs_info->transaction_wait);
381	return 0;
382}
383
384/* simple helper to fault in pages and copy.  This should go away
385 * and be replaced with calls into generic code.
386 */
387static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
388					 struct page **prepared_pages,
389					 struct iov_iter *i)
390{
391	size_t copied = 0;
392	size_t total_copied = 0;
393	int pg = 0;
394	int offset = offset_in_page(pos);
395
396	while (write_bytes > 0) {
397		size_t count = min_t(size_t,
398				     PAGE_SIZE - offset, write_bytes);
399		struct page *page = prepared_pages[pg];
400		/*
401		 * Copy data from userspace to the current page
402		 */
403		copied = copy_page_from_iter_atomic(page, offset, count, i);
404
405		/* Flush processor's dcache for this page */
406		flush_dcache_page(page);
407
408		/*
409		 * if we get a partial write, we can end up with
410		 * partially up to date pages.  These add
411		 * a lot of complexity, so make sure they don't
412		 * happen by forcing this copy to be retried.
413		 *
414		 * The rest of the btrfs_file_write code will fall
415		 * back to page at a time copies after we return 0.
416		 */
417		if (unlikely(copied < count)) {
418			if (!PageUptodate(page)) {
419				iov_iter_revert(i, copied);
420				copied = 0;
421			}
422			if (!copied)
423				break;
424		}
425
426		write_bytes -= copied;
427		total_copied += copied;
428		offset += copied;
429		if (offset == PAGE_SIZE) {
430			pg++;
431			offset = 0;
432		}
433	}
434	return total_copied;
435}
436
437/*
438 * unlocks pages after btrfs_file_write is done with them
439 */
440static void btrfs_drop_pages(struct btrfs_fs_info *fs_info,
441			     struct page **pages, size_t num_pages,
442			     u64 pos, u64 copied)
443{
444	size_t i;
445	u64 block_start = round_down(pos, fs_info->sectorsize);
446	u64 block_len = round_up(pos + copied, fs_info->sectorsize) - block_start;
447
448	ASSERT(block_len <= U32_MAX);
449	for (i = 0; i < num_pages; i++) {
450		/* page checked is some magic around finding pages that
451		 * have been modified without going through btrfs_set_page_dirty
452		 * clear it here. There should be no need to mark the pages
453		 * accessed as prepare_pages should have marked them accessed
454		 * in prepare_pages via find_or_create_page()
455		 */
456		btrfs_page_clamp_clear_checked(fs_info, pages[i], block_start,
457					       block_len);
458		unlock_page(pages[i]);
459		put_page(pages[i]);
460	}
461}
462
463/*
464 * After btrfs_copy_from_user(), update the following things for delalloc:
465 * - Mark newly dirtied pages as DELALLOC in the io tree.
466 *   Used to advise which range is to be written back.
467 * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
468 * - Update inode size for past EOF write
469 */
470int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
471		      size_t num_pages, loff_t pos, size_t write_bytes,
472		      struct extent_state **cached, bool noreserve)
473{
474	struct btrfs_fs_info *fs_info = inode->root->fs_info;
475	int err = 0;
476	int i;
477	u64 num_bytes;
478	u64 start_pos;
479	u64 end_of_last_block;
480	u64 end_pos = pos + write_bytes;
481	loff_t isize = i_size_read(&inode->vfs_inode);
482	unsigned int extra_bits = 0;
483
484	if (write_bytes == 0)
485		return 0;
486
487	if (noreserve)
488		extra_bits |= EXTENT_NORESERVE;
489
490	start_pos = round_down(pos, fs_info->sectorsize);
491	num_bytes = round_up(write_bytes + pos - start_pos,
492			     fs_info->sectorsize);
493	ASSERT(num_bytes <= U32_MAX);
494
495	end_of_last_block = start_pos + num_bytes - 1;
496
497	/*
498	 * The pages may have already been dirty, clear out old accounting so
499	 * we can set things up properly
500	 */
501	clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
502			 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
503			 0, 0, cached);
504
505	err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
506					extra_bits, cached);
507	if (err)
508		return err;
509
510	for (i = 0; i < num_pages; i++) {
511		struct page *p = pages[i];
512
513		btrfs_page_clamp_set_uptodate(fs_info, p, start_pos, num_bytes);
514		btrfs_page_clamp_clear_checked(fs_info, p, start_pos, num_bytes);
515		btrfs_page_clamp_set_dirty(fs_info, p, start_pos, num_bytes);
516	}
517
518	/*
519	 * we've only changed i_size in ram, and we haven't updated
520	 * the disk i_size.  There is no need to log the inode
521	 * at this time.
522	 */
523	if (end_pos > isize)
524		i_size_write(&inode->vfs_inode, end_pos);
525	return 0;
526}
527
528/*
529 * this drops all the extents in the cache that intersect the range
530 * [start, end].  Existing extents are split as required.
531 */
532void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
533			     int skip_pinned)
534{
535	struct extent_map *em;
536	struct extent_map *split = NULL;
537	struct extent_map *split2 = NULL;
538	struct extent_map_tree *em_tree = &inode->extent_tree;
539	u64 len = end - start + 1;
540	u64 gen;
541	int ret;
542	int testend = 1;
543	unsigned long flags;
544	int compressed = 0;
545	bool modified;
546
547	WARN_ON(end < start);
548	if (end == (u64)-1) {
549		len = (u64)-1;
550		testend = 0;
551	}
552	while (1) {
553		int no_splits = 0;
554
555		modified = false;
556		if (!split)
557			split = alloc_extent_map();
558		if (!split2)
559			split2 = alloc_extent_map();
560		if (!split || !split2)
561			no_splits = 1;
562
563		write_lock(&em_tree->lock);
564		em = lookup_extent_mapping(em_tree, start, len);
565		if (!em) {
566			write_unlock(&em_tree->lock);
567			break;
568		}
569		flags = em->flags;
570		gen = em->generation;
571		if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
572			if (testend && em->start + em->len >= start + len) {
573				free_extent_map(em);
574				write_unlock(&em_tree->lock);
575				break;
576			}
577			start = em->start + em->len;
578			if (testend)
579				len = start + len - (em->start + em->len);
580			free_extent_map(em);
581			write_unlock(&em_tree->lock);
582			continue;
583		}
584		compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
585		clear_bit(EXTENT_FLAG_PINNED, &em->flags);
586		clear_bit(EXTENT_FLAG_LOGGING, &flags);
587		modified = !list_empty(&em->list);
588		if (no_splits)
589			goto next;
590
591		if (em->start < start) {
592			split->start = em->start;
593			split->len = start - em->start;
594
595			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
596				split->orig_start = em->orig_start;
597				split->block_start = em->block_start;
598
599				if (compressed)
600					split->block_len = em->block_len;
601				else
602					split->block_len = split->len;
603				split->orig_block_len = max(split->block_len,
604						em->orig_block_len);
605				split->ram_bytes = em->ram_bytes;
606			} else {
607				split->orig_start = split->start;
608				split->block_len = 0;
609				split->block_start = em->block_start;
610				split->orig_block_len = 0;
611				split->ram_bytes = split->len;
612			}
613
614			split->generation = gen;
615			split->flags = flags;
616			split->compress_type = em->compress_type;
617			replace_extent_mapping(em_tree, em, split, modified);
618			free_extent_map(split);
619			split = split2;
620			split2 = NULL;
621		}
622		if (testend && em->start + em->len > start + len) {
623			u64 diff = start + len - em->start;
624
625			split->start = start + len;
626			split->len = em->start + em->len - (start + len);
627			split->flags = flags;
628			split->compress_type = em->compress_type;
629			split->generation = gen;
630
631			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
632				split->orig_block_len = max(em->block_len,
633						    em->orig_block_len);
634
635				split->ram_bytes = em->ram_bytes;
636				if (compressed) {
637					split->block_len = em->block_len;
638					split->block_start = em->block_start;
639					split->orig_start = em->orig_start;
640				} else {
641					split->block_len = split->len;
642					split->block_start = em->block_start
643						+ diff;
644					split->orig_start = em->orig_start;
645				}
646			} else {
647				split->ram_bytes = split->len;
648				split->orig_start = split->start;
649				split->block_len = 0;
650				split->block_start = em->block_start;
651				split->orig_block_len = 0;
652			}
653
654			if (extent_map_in_tree(em)) {
655				replace_extent_mapping(em_tree, em, split,
656						       modified);
657			} else {
658				ret = add_extent_mapping(em_tree, split,
659							 modified);
660				ASSERT(ret == 0); /* Logic error */
661			}
662			free_extent_map(split);
663			split = NULL;
664		}
665next:
666		if (extent_map_in_tree(em))
667			remove_extent_mapping(em_tree, em);
668		write_unlock(&em_tree->lock);
669
670		/* once for us */
671		free_extent_map(em);
672		/* once for the tree*/
673		free_extent_map(em);
674	}
675	if (split)
676		free_extent_map(split);
677	if (split2)
678		free_extent_map(split2);
679}
680
681/*
682 * this is very complex, but the basic idea is to drop all extents
683 * in the range start - end.  hint_block is filled in with a block number
684 * that would be a good hint to the block allocator for this file.
685 *
686 * If an extent intersects the range but is not entirely inside the range
687 * it is either truncated or split.  Anything entirely inside the range
688 * is deleted from the tree.
689 *
690 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
691 * to deal with that. We set the field 'bytes_found' of the arguments structure
692 * with the number of allocated bytes found in the target range, so that the
693 * caller can update the inode's number of bytes in an atomic way when
694 * replacing extents in a range to avoid races with stat(2).
695 */
696int btrfs_drop_extents(struct btrfs_trans_handle *trans,
697		       struct btrfs_root *root, struct btrfs_inode *inode,
698		       struct btrfs_drop_extents_args *args)
699{
700	struct btrfs_fs_info *fs_info = root->fs_info;
701	struct extent_buffer *leaf;
702	struct btrfs_file_extent_item *fi;
703	struct btrfs_ref ref = { 0 };
704	struct btrfs_key key;
705	struct btrfs_key new_key;
706	u64 ino = btrfs_ino(inode);
707	u64 search_start = args->start;
708	u64 disk_bytenr = 0;
709	u64 num_bytes = 0;
710	u64 extent_offset = 0;
711	u64 extent_end = 0;
712	u64 last_end = args->start;
713	int del_nr = 0;
714	int del_slot = 0;
715	int extent_type;
716	int recow;
717	int ret;
718	int modify_tree = -1;
719	int update_refs;
720	int found = 0;
721	int leafs_visited = 0;
722	struct btrfs_path *path = args->path;
723
724	args->bytes_found = 0;
725	args->extent_inserted = false;
726
727	/* Must always have a path if ->replace_extent is true */
728	ASSERT(!(args->replace_extent && !args->path));
729
730	if (!path) {
731		path = btrfs_alloc_path();
732		if (!path) {
733			ret = -ENOMEM;
734			goto out;
735		}
736	}
737
738	if (args->drop_cache)
739		btrfs_drop_extent_cache(inode, args->start, args->end - 1, 0);
740
741	if (args->start >= inode->disk_i_size && !args->replace_extent)
742		modify_tree = 0;
743
744	update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
745	while (1) {
746		recow = 0;
747		ret = btrfs_lookup_file_extent(trans, root, path, ino,
748					       search_start, modify_tree);
749		if (ret < 0)
750			break;
751		if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
752			leaf = path->nodes[0];
753			btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
754			if (key.objectid == ino &&
755			    key.type == BTRFS_EXTENT_DATA_KEY)
756				path->slots[0]--;
757		}
758		ret = 0;
759		leafs_visited++;
760next_slot:
761		leaf = path->nodes[0];
762		if (path->slots[0] >= btrfs_header_nritems(leaf)) {
763			BUG_ON(del_nr > 0);
764			ret = btrfs_next_leaf(root, path);
765			if (ret < 0)
766				break;
767			if (ret > 0) {
768				ret = 0;
769				break;
770			}
771			leafs_visited++;
772			leaf = path->nodes[0];
773			recow = 1;
774		}
775
776		btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
777
778		if (key.objectid > ino)
779			break;
780		if (WARN_ON_ONCE(key.objectid < ino) ||
781		    key.type < BTRFS_EXTENT_DATA_KEY) {
782			ASSERT(del_nr == 0);
783			path->slots[0]++;
784			goto next_slot;
785		}
786		if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
787			break;
788
789		fi = btrfs_item_ptr(leaf, path->slots[0],
790				    struct btrfs_file_extent_item);
791		extent_type = btrfs_file_extent_type(leaf, fi);
792
793		if (extent_type == BTRFS_FILE_EXTENT_REG ||
794		    extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
795			disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
796			num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
797			extent_offset = btrfs_file_extent_offset(leaf, fi);
798			extent_end = key.offset +
799				btrfs_file_extent_num_bytes(leaf, fi);
800		} else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
801			extent_end = key.offset +
802				btrfs_file_extent_ram_bytes(leaf, fi);
803		} else {
804			/* can't happen */
805			BUG();
806		}
807
808		/*
809		 * Don't skip extent items representing 0 byte lengths. They
810		 * used to be created (bug) if while punching holes we hit
811		 * -ENOSPC condition. So if we find one here, just ensure we
812		 * delete it, otherwise we would insert a new file extent item
813		 * with the same key (offset) as that 0 bytes length file
814		 * extent item in the call to setup_items_for_insert() later
815		 * in this function.
816		 */
817		if (extent_end == key.offset && extent_end >= search_start) {
818			last_end = extent_end;
819			goto delete_extent_item;
820		}
821
822		if (extent_end <= search_start) {
823			path->slots[0]++;
824			goto next_slot;
825		}
826
827		found = 1;
828		search_start = max(key.offset, args->start);
829		if (recow || !modify_tree) {
830			modify_tree = -1;
831			btrfs_release_path(path);
832			continue;
833		}
834
835		/*
836		 *     | - range to drop - |
837		 *  | -------- extent -------- |
838		 */
839		if (args->start > key.offset && args->end < extent_end) {
840			BUG_ON(del_nr > 0);
841			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
842				ret = -EOPNOTSUPP;
843				break;
844			}
845
846			memcpy(&new_key, &key, sizeof(new_key));
847			new_key.offset = args->start;
848			ret = btrfs_duplicate_item(trans, root, path,
849						   &new_key);
850			if (ret == -EAGAIN) {
851				btrfs_release_path(path);
852				continue;
853			}
854			if (ret < 0)
855				break;
856
857			leaf = path->nodes[0];
858			fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
859					    struct btrfs_file_extent_item);
860			btrfs_set_file_extent_num_bytes(leaf, fi,
861							args->start - key.offset);
862
863			fi = btrfs_item_ptr(leaf, path->slots[0],
864					    struct btrfs_file_extent_item);
865
866			extent_offset += args->start - key.offset;
867			btrfs_set_file_extent_offset(leaf, fi, extent_offset);
868			btrfs_set_file_extent_num_bytes(leaf, fi,
869							extent_end - args->start);
870			btrfs_mark_buffer_dirty(leaf);
871
872			if (update_refs && disk_bytenr > 0) {
873				btrfs_init_generic_ref(&ref,
874						BTRFS_ADD_DELAYED_REF,
875						disk_bytenr, num_bytes, 0);
876				btrfs_init_data_ref(&ref,
877						root->root_key.objectid,
878						new_key.objectid,
879						args->start - extent_offset,
880						0, false);
881				ret = btrfs_inc_extent_ref(trans, &ref);
882				BUG_ON(ret); /* -ENOMEM */
883			}
884			key.offset = args->start;
885		}
886		/*
887		 * From here on out we will have actually dropped something, so
888		 * last_end can be updated.
889		 */
890		last_end = extent_end;
891
892		/*
893		 *  | ---- range to drop ----- |
894		 *      | -------- extent -------- |
895		 */
896		if (args->start <= key.offset && args->end < extent_end) {
897			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
898				ret = -EOPNOTSUPP;
899				break;
900			}
901
902			memcpy(&new_key, &key, sizeof(new_key));
903			new_key.offset = args->end;
904			btrfs_set_item_key_safe(fs_info, path, &new_key);
905
906			extent_offset += args->end - key.offset;
907			btrfs_set_file_extent_offset(leaf, fi, extent_offset);
908			btrfs_set_file_extent_num_bytes(leaf, fi,
909							extent_end - args->end);
910			btrfs_mark_buffer_dirty(leaf);
911			if (update_refs && disk_bytenr > 0)
912				args->bytes_found += args->end - key.offset;
913			break;
914		}
915
916		search_start = extent_end;
917		/*
918		 *       | ---- range to drop ----- |
919		 *  | -------- extent -------- |
920		 */
921		if (args->start > key.offset && args->end >= extent_end) {
922			BUG_ON(del_nr > 0);
923			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
924				ret = -EOPNOTSUPP;
925				break;
926			}
927
928			btrfs_set_file_extent_num_bytes(leaf, fi,
929							args->start - key.offset);
930			btrfs_mark_buffer_dirty(leaf);
931			if (update_refs && disk_bytenr > 0)
932				args->bytes_found += extent_end - args->start;
933			if (args->end == extent_end)
934				break;
935
936			path->slots[0]++;
937			goto next_slot;
938		}
939
940		/*
941		 *  | ---- range to drop ----- |
942		 *    | ------ extent ------ |
943		 */
944		if (args->start <= key.offset && args->end >= extent_end) {
945delete_extent_item:
946			if (del_nr == 0) {
947				del_slot = path->slots[0];
948				del_nr = 1;
949			} else {
950				BUG_ON(del_slot + del_nr != path->slots[0]);
951				del_nr++;
952			}
953
954			if (update_refs &&
955			    extent_type == BTRFS_FILE_EXTENT_INLINE) {
956				args->bytes_found += extent_end - key.offset;
957				extent_end = ALIGN(extent_end,
958						   fs_info->sectorsize);
959			} else if (update_refs && disk_bytenr > 0) {
960				btrfs_init_generic_ref(&ref,
961						BTRFS_DROP_DELAYED_REF,
962						disk_bytenr, num_bytes, 0);
963				btrfs_init_data_ref(&ref,
964						root->root_key.objectid,
965						key.objectid,
966						key.offset - extent_offset, 0,
967						false);
968				ret = btrfs_free_extent(trans, &ref);
969				BUG_ON(ret); /* -ENOMEM */
970				args->bytes_found += extent_end - key.offset;
971			}
972
973			if (args->end == extent_end)
974				break;
975
976			if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
977				path->slots[0]++;
978				goto next_slot;
979			}
980
981			ret = btrfs_del_items(trans, root, path, del_slot,
982					      del_nr);
983			if (ret) {
984				btrfs_abort_transaction(trans, ret);
985				break;
986			}
987
988			del_nr = 0;
989			del_slot = 0;
990
991			btrfs_release_path(path);
992			continue;
993		}
994
995		BUG();
996	}
997
998	if (!ret && del_nr > 0) {
999		/*
1000		 * Set path->slots[0] to first slot, so that after the delete
1001		 * if items are move off from our leaf to its immediate left or
1002		 * right neighbor leafs, we end up with a correct and adjusted
1003		 * path->slots[0] for our insertion (if args->replace_extent).
1004		 */
1005		path->slots[0] = del_slot;
1006		ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1007		if (ret)
1008			btrfs_abort_transaction(trans, ret);
1009	}
1010
1011	leaf = path->nodes[0];
1012	/*
1013	 * If btrfs_del_items() was called, it might have deleted a leaf, in
1014	 * which case it unlocked our path, so check path->locks[0] matches a
1015	 * write lock.
1016	 */
1017	if (!ret && args->replace_extent && leafs_visited == 1 &&
1018	    path->locks[0] == BTRFS_WRITE_LOCK &&
1019	    btrfs_leaf_free_space(leaf) >=
1020	    sizeof(struct btrfs_item) + args->extent_item_size) {
1021
1022		key.objectid = ino;
1023		key.type = BTRFS_EXTENT_DATA_KEY;
1024		key.offset = args->start;
1025		if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1026			struct btrfs_key slot_key;
1027
1028			btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1029			if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1030				path->slots[0]++;
1031		}
1032		btrfs_setup_item_for_insert(root, path, &key, args->extent_item_size);
1033		args->extent_inserted = true;
1034	}
1035
1036	if (!args->path)
1037		btrfs_free_path(path);
1038	else if (!args->extent_inserted)
1039		btrfs_release_path(path);
1040out:
1041	args->drop_end = found ? min(args->end, last_end) : args->end;
1042
1043	return ret;
1044}
1045
1046static int extent_mergeable(struct extent_buffer *leaf, int slot,
1047			    u64 objectid, u64 bytenr, u64 orig_offset,
1048			    u64 *start, u64 *end)
1049{
1050	struct btrfs_file_extent_item *fi;
1051	struct btrfs_key key;
1052	u64 extent_end;
1053
1054	if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1055		return 0;
1056
1057	btrfs_item_key_to_cpu(leaf, &key, slot);
1058	if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1059		return 0;
1060
1061	fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1062	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1063	    btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1064	    btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1065	    btrfs_file_extent_compression(leaf, fi) ||
1066	    btrfs_file_extent_encryption(leaf, fi) ||
1067	    btrfs_file_extent_other_encoding(leaf, fi))
1068		return 0;
1069
1070	extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1071	if ((*start && *start != key.offset) || (*end && *end != extent_end))
1072		return 0;
1073
1074	*start = key.offset;
1075	*end = extent_end;
1076	return 1;
1077}
1078
1079/*
1080 * Mark extent in the range start - end as written.
1081 *
1082 * This changes extent type from 'pre-allocated' to 'regular'. If only
1083 * part of extent is marked as written, the extent will be split into
1084 * two or three.
1085 */
1086int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1087			      struct btrfs_inode *inode, u64 start, u64 end)
1088{
1089	struct btrfs_fs_info *fs_info = trans->fs_info;
1090	struct btrfs_root *root = inode->root;
1091	struct extent_buffer *leaf;
1092	struct btrfs_path *path;
1093	struct btrfs_file_extent_item *fi;
1094	struct btrfs_ref ref = { 0 };
1095	struct btrfs_key key;
1096	struct btrfs_key new_key;
1097	u64 bytenr;
1098	u64 num_bytes;
1099	u64 extent_end;
1100	u64 orig_offset;
1101	u64 other_start;
1102	u64 other_end;
1103	u64 split;
1104	int del_nr = 0;
1105	int del_slot = 0;
1106	int recow;
1107	int ret = 0;
1108	u64 ino = btrfs_ino(inode);
1109
1110	path = btrfs_alloc_path();
1111	if (!path)
1112		return -ENOMEM;
1113again:
1114	recow = 0;
1115	split = start;
1116	key.objectid = ino;
1117	key.type = BTRFS_EXTENT_DATA_KEY;
1118	key.offset = split;
1119
1120	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1121	if (ret < 0)
1122		goto out;
1123	if (ret > 0 && path->slots[0] > 0)
1124		path->slots[0]--;
1125
1126	leaf = path->nodes[0];
1127	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1128	if (key.objectid != ino ||
1129	    key.type != BTRFS_EXTENT_DATA_KEY) {
1130		ret = -EINVAL;
1131		btrfs_abort_transaction(trans, ret);
1132		goto out;
1133	}
1134	fi = btrfs_item_ptr(leaf, path->slots[0],
1135			    struct btrfs_file_extent_item);
1136	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1137		ret = -EINVAL;
1138		btrfs_abort_transaction(trans, ret);
1139		goto out;
1140	}
1141	extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1142	if (key.offset > start || extent_end < end) {
1143		ret = -EINVAL;
1144		btrfs_abort_transaction(trans, ret);
1145		goto out;
1146	}
1147
1148	bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1149	num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1150	orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1151	memcpy(&new_key, &key, sizeof(new_key));
1152
1153	if (start == key.offset && end < extent_end) {
1154		other_start = 0;
1155		other_end = start;
1156		if (extent_mergeable(leaf, path->slots[0] - 1,
1157				     ino, bytenr, orig_offset,
1158				     &other_start, &other_end)) {
1159			new_key.offset = end;
1160			btrfs_set_item_key_safe(fs_info, path, &new_key);
1161			fi = btrfs_item_ptr(leaf, path->slots[0],
1162					    struct btrfs_file_extent_item);
1163			btrfs_set_file_extent_generation(leaf, fi,
1164							 trans->transid);
1165			btrfs_set_file_extent_num_bytes(leaf, fi,
1166							extent_end - end);
1167			btrfs_set_file_extent_offset(leaf, fi,
1168						     end - orig_offset);
1169			fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1170					    struct btrfs_file_extent_item);
1171			btrfs_set_file_extent_generation(leaf, fi,
1172							 trans->transid);
1173			btrfs_set_file_extent_num_bytes(leaf, fi,
1174							end - other_start);
1175			btrfs_mark_buffer_dirty(leaf);
1176			goto out;
1177		}
1178	}
1179
1180	if (start > key.offset && end == extent_end) {
1181		other_start = end;
1182		other_end = 0;
1183		if (extent_mergeable(leaf, path->slots[0] + 1,
1184				     ino, bytenr, orig_offset,
1185				     &other_start, &other_end)) {
1186			fi = btrfs_item_ptr(leaf, path->slots[0],
1187					    struct btrfs_file_extent_item);
1188			btrfs_set_file_extent_num_bytes(leaf, fi,
1189							start - key.offset);
1190			btrfs_set_file_extent_generation(leaf, fi,
1191							 trans->transid);
1192			path->slots[0]++;
1193			new_key.offset = start;
1194			btrfs_set_item_key_safe(fs_info, path, &new_key);
1195
1196			fi = btrfs_item_ptr(leaf, path->slots[0],
1197					    struct btrfs_file_extent_item);
1198			btrfs_set_file_extent_generation(leaf, fi,
1199							 trans->transid);
1200			btrfs_set_file_extent_num_bytes(leaf, fi,
1201							other_end - start);
1202			btrfs_set_file_extent_offset(leaf, fi,
1203						     start - orig_offset);
1204			btrfs_mark_buffer_dirty(leaf);
1205			goto out;
1206		}
1207	}
1208
1209	while (start > key.offset || end < extent_end) {
1210		if (key.offset == start)
1211			split = end;
1212
1213		new_key.offset = split;
1214		ret = btrfs_duplicate_item(trans, root, path, &new_key);
1215		if (ret == -EAGAIN) {
1216			btrfs_release_path(path);
1217			goto again;
1218		}
1219		if (ret < 0) {
1220			btrfs_abort_transaction(trans, ret);
1221			goto out;
1222		}
1223
1224		leaf = path->nodes[0];
1225		fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1226				    struct btrfs_file_extent_item);
1227		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1228		btrfs_set_file_extent_num_bytes(leaf, fi,
1229						split - key.offset);
1230
1231		fi = btrfs_item_ptr(leaf, path->slots[0],
1232				    struct btrfs_file_extent_item);
1233
1234		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1235		btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1236		btrfs_set_file_extent_num_bytes(leaf, fi,
1237						extent_end - split);
1238		btrfs_mark_buffer_dirty(leaf);
1239
1240		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1241				       num_bytes, 0);
1242		btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1243				    orig_offset, 0, false);
1244		ret = btrfs_inc_extent_ref(trans, &ref);
1245		if (ret) {
1246			btrfs_abort_transaction(trans, ret);
1247			goto out;
1248		}
1249
1250		if (split == start) {
1251			key.offset = start;
1252		} else {
1253			if (start != key.offset) {
1254				ret = -EINVAL;
1255				btrfs_abort_transaction(trans, ret);
1256				goto out;
1257			}
1258			path->slots[0]--;
1259			extent_end = end;
1260		}
1261		recow = 1;
1262	}
1263
1264	other_start = end;
1265	other_end = 0;
1266	btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1267			       num_bytes, 0);
1268	btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset,
1269			    0, false);
1270	if (extent_mergeable(leaf, path->slots[0] + 1,
1271			     ino, bytenr, orig_offset,
1272			     &other_start, &other_end)) {
1273		if (recow) {
1274			btrfs_release_path(path);
1275			goto again;
1276		}
1277		extent_end = other_end;
1278		del_slot = path->slots[0] + 1;
1279		del_nr++;
1280		ret = btrfs_free_extent(trans, &ref);
1281		if (ret) {
1282			btrfs_abort_transaction(trans, ret);
1283			goto out;
1284		}
1285	}
1286	other_start = 0;
1287	other_end = start;
1288	if (extent_mergeable(leaf, path->slots[0] - 1,
1289			     ino, bytenr, orig_offset,
1290			     &other_start, &other_end)) {
1291		if (recow) {
1292			btrfs_release_path(path);
1293			goto again;
1294		}
1295		key.offset = other_start;
1296		del_slot = path->slots[0];
1297		del_nr++;
1298		ret = btrfs_free_extent(trans, &ref);
1299		if (ret) {
1300			btrfs_abort_transaction(trans, ret);
1301			goto out;
1302		}
1303	}
1304	if (del_nr == 0) {
1305		fi = btrfs_item_ptr(leaf, path->slots[0],
1306			   struct btrfs_file_extent_item);
1307		btrfs_set_file_extent_type(leaf, fi,
1308					   BTRFS_FILE_EXTENT_REG);
1309		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1310		btrfs_mark_buffer_dirty(leaf);
1311	} else {
1312		fi = btrfs_item_ptr(leaf, del_slot - 1,
1313			   struct btrfs_file_extent_item);
1314		btrfs_set_file_extent_type(leaf, fi,
1315					   BTRFS_FILE_EXTENT_REG);
1316		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1317		btrfs_set_file_extent_num_bytes(leaf, fi,
1318						extent_end - key.offset);
1319		btrfs_mark_buffer_dirty(leaf);
1320
1321		ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1322		if (ret < 0) {
1323			btrfs_abort_transaction(trans, ret);
1324			goto out;
1325		}
1326	}
1327out:
1328	btrfs_free_path(path);
1329	return ret;
1330}
1331
1332/*
1333 * on error we return an unlocked page and the error value
1334 * on success we return a locked page and 0
1335 */
1336static int prepare_uptodate_page(struct inode *inode,
1337				 struct page *page, u64 pos,
1338				 bool force_uptodate)
1339{
1340	int ret = 0;
1341
1342	if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1343	    !PageUptodate(page)) {
1344		ret = btrfs_readpage(NULL, page);
1345		if (ret)
1346			return ret;
1347		lock_page(page);
1348		if (!PageUptodate(page)) {
1349			unlock_page(page);
1350			return -EIO;
1351		}
1352
1353		/*
1354		 * Since btrfs_readpage() will unlock the page before it
1355		 * returns, there is a window where btrfs_releasepage() can be
1356		 * called to release the page.  Here we check both inode
1357		 * mapping and PagePrivate() to make sure the page was not
1358		 * released.
1359		 *
1360		 * The private flag check is essential for subpage as we need
1361		 * to store extra bitmap using page->private.
1362		 */
1363		if (page->mapping != inode->i_mapping || !PagePrivate(page)) {
1364			unlock_page(page);
1365			return -EAGAIN;
1366		}
1367	}
1368	return 0;
1369}
1370
1371/*
1372 * this just gets pages into the page cache and locks them down.
1373 */
1374static noinline int prepare_pages(struct inode *inode, struct page **pages,
1375				  size_t num_pages, loff_t pos,
1376				  size_t write_bytes, bool force_uptodate)
1377{
1378	int i;
1379	unsigned long index = pos >> PAGE_SHIFT;
1380	gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1381	int err = 0;
1382	int faili;
1383
1384	for (i = 0; i < num_pages; i++) {
1385again:
1386		pages[i] = find_or_create_page(inode->i_mapping, index + i,
1387					       mask | __GFP_WRITE);
1388		if (!pages[i]) {
1389			faili = i - 1;
1390			err = -ENOMEM;
1391			goto fail;
1392		}
1393
1394		err = set_page_extent_mapped(pages[i]);
1395		if (err < 0) {
1396			faili = i;
1397			goto fail;
1398		}
1399
1400		if (i == 0)
1401			err = prepare_uptodate_page(inode, pages[i], pos,
1402						    force_uptodate);
1403		if (!err && i == num_pages - 1)
1404			err = prepare_uptodate_page(inode, pages[i],
1405						    pos + write_bytes, false);
1406		if (err) {
1407			put_page(pages[i]);
1408			if (err == -EAGAIN) {
1409				err = 0;
1410				goto again;
1411			}
1412			faili = i - 1;
1413			goto fail;
1414		}
1415		wait_on_page_writeback(pages[i]);
1416	}
1417
1418	return 0;
1419fail:
1420	while (faili >= 0) {
1421		unlock_page(pages[faili]);
1422		put_page(pages[faili]);
1423		faili--;
1424	}
1425	return err;
1426
1427}
1428
1429/*
1430 * This function locks the extent and properly waits for data=ordered extents
1431 * to finish before allowing the pages to be modified if need.
1432 *
1433 * The return value:
1434 * 1 - the extent is locked
1435 * 0 - the extent is not locked, and everything is OK
1436 * -EAGAIN - need re-prepare the pages
1437 * the other < 0 number - Something wrong happens
1438 */
1439static noinline int
1440lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1441				size_t num_pages, loff_t pos,
1442				size_t write_bytes,
1443				u64 *lockstart, u64 *lockend,
1444				struct extent_state **cached_state)
1445{
1446	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1447	u64 start_pos;
1448	u64 last_pos;
1449	int i;
1450	int ret = 0;
1451
1452	start_pos = round_down(pos, fs_info->sectorsize);
1453	last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1454
1455	if (start_pos < inode->vfs_inode.i_size) {
1456		struct btrfs_ordered_extent *ordered;
1457
1458		lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1459				cached_state);
1460		ordered = btrfs_lookup_ordered_range(inode, start_pos,
1461						     last_pos - start_pos + 1);
1462		if (ordered &&
1463		    ordered->file_offset + ordered->num_bytes > start_pos &&
1464		    ordered->file_offset <= last_pos) {
1465			unlock_extent_cached(&inode->io_tree, start_pos,
1466					last_pos, cached_state);
1467			for (i = 0; i < num_pages; i++) {
1468				unlock_page(pages[i]);
1469				put_page(pages[i]);
1470			}
1471			btrfs_start_ordered_extent(ordered, 1);
1472			btrfs_put_ordered_extent(ordered);
1473			return -EAGAIN;
1474		}
1475		if (ordered)
1476			btrfs_put_ordered_extent(ordered);
1477
1478		*lockstart = start_pos;
1479		*lockend = last_pos;
1480		ret = 1;
1481	}
1482
1483	/*
1484	 * We should be called after prepare_pages() which should have locked
1485	 * all pages in the range.
1486	 */
1487	for (i = 0; i < num_pages; i++)
1488		WARN_ON(!PageLocked(pages[i]));
1489
1490	return ret;
1491}
1492
1493static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1494			   size_t *write_bytes, bool nowait)
1495{
1496	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1497	struct btrfs_root *root = inode->root;
1498	u64 lockstart, lockend;
1499	u64 num_bytes;
1500	int ret;
1501
1502	if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1503		return 0;
1504
1505	if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1506		return -EAGAIN;
1507
1508	lockstart = round_down(pos, fs_info->sectorsize);
1509	lockend = round_up(pos + *write_bytes,
1510			   fs_info->sectorsize) - 1;
1511	num_bytes = lockend - lockstart + 1;
1512
1513	if (nowait) {
1514		struct btrfs_ordered_extent *ordered;
1515
1516		if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1517			return -EAGAIN;
1518
1519		ordered = btrfs_lookup_ordered_range(inode, lockstart,
1520						     num_bytes);
1521		if (ordered) {
1522			btrfs_put_ordered_extent(ordered);
1523			ret = -EAGAIN;
1524			goto out_unlock;
1525		}
1526	} else {
1527		btrfs_lock_and_flush_ordered_range(inode, lockstart,
1528						   lockend, NULL);
1529	}
1530
1531	ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1532			NULL, NULL, NULL, false);
1533	if (ret <= 0) {
1534		ret = 0;
1535		if (!nowait)
1536			btrfs_drew_write_unlock(&root->snapshot_lock);
1537	} else {
1538		*write_bytes = min_t(size_t, *write_bytes ,
1539				     num_bytes - pos + lockstart);
1540	}
1541out_unlock:
1542	unlock_extent(&inode->io_tree, lockstart, lockend);
1543
1544	return ret;
1545}
1546
1547static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1548			      size_t *write_bytes)
1549{
1550	return check_can_nocow(inode, pos, write_bytes, true);
1551}
1552
1553/*
1554 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1555 *
1556 * @pos:	 File offset
1557 * @write_bytes: The length to write, will be updated to the nocow writeable
1558 *		 range
1559 *
1560 * This function will flush ordered extents in the range to ensure proper
1561 * nocow checks.
1562 *
1563 * Return:
1564 * >0		and update @write_bytes if we can do nocow write
1565 *  0		if we can't do nocow write
1566 * -EAGAIN	if we can't get the needed lock or there are ordered extents
1567 * 		for * (nowait == true) case
1568 * <0		if other error happened
1569 *
1570 * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1571 */
1572int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1573			   size_t *write_bytes)
1574{
1575	return check_can_nocow(inode, pos, write_bytes, false);
1576}
1577
1578void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1579{
1580	btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1581}
1582
1583static void update_time_for_write(struct inode *inode)
1584{
1585	struct timespec64 now;
1586
1587	if (IS_NOCMTIME(inode))
1588		return;
1589
1590	now = current_time(inode);
1591	if (!timespec64_equal(&inode->i_mtime, &now))
1592		inode->i_mtime = now;
1593
1594	if (!timespec64_equal(&inode->i_ctime, &now))
1595		inode->i_ctime = now;
1596
1597	if (IS_I_VERSION(inode))
1598		inode_inc_iversion(inode);
1599}
1600
1601static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1602			     size_t count)
1603{
1604	struct file *file = iocb->ki_filp;
1605	struct inode *inode = file_inode(file);
1606	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1607	loff_t pos = iocb->ki_pos;
1608	int ret;
1609	loff_t oldsize;
1610	loff_t start_pos;
1611
1612	if (iocb->ki_flags & IOCB_NOWAIT) {
1613		size_t nocow_bytes = count;
1614
1615		/* We will allocate space in case nodatacow is not set, so bail */
1616		if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes) <= 0)
1617			return -EAGAIN;
1618		/*
1619		 * There are holes in the range or parts of the range that must
1620		 * be COWed (shared extents, RO block groups, etc), so just bail
1621		 * out.
1622		 */
1623		if (nocow_bytes < count)
1624			return -EAGAIN;
1625	}
1626
1627	current->backing_dev_info = inode_to_bdi(inode);
1628	ret = file_remove_privs(file);
1629	if (ret)
1630		return ret;
1631
1632	/*
1633	 * We reserve space for updating the inode when we reserve space for the
1634	 * extent we are going to write, so we will enospc out there.  We don't
1635	 * need to start yet another transaction to update the inode as we will
1636	 * update the inode when we finish writing whatever data we write.
1637	 */
1638	update_time_for_write(inode);
1639
1640	start_pos = round_down(pos, fs_info->sectorsize);
1641	oldsize = i_size_read(inode);
1642	if (start_pos > oldsize) {
1643		/* Expand hole size to cover write data, preventing empty gap */
1644		loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1645
1646		ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1647		if (ret) {
1648			current->backing_dev_info = NULL;
1649			return ret;
1650		}
1651	}
1652
1653	return 0;
1654}
1655
1656static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1657					       struct iov_iter *i)
1658{
1659	struct file *file = iocb->ki_filp;
1660	loff_t pos;
1661	struct inode *inode = file_inode(file);
1662	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1663	struct page **pages = NULL;
1664	struct extent_changeset *data_reserved = NULL;
1665	u64 release_bytes = 0;
1666	u64 lockstart;
1667	u64 lockend;
1668	size_t num_written = 0;
1669	int nrptrs;
1670	ssize_t ret;
1671	bool only_release_metadata = false;
1672	bool force_page_uptodate = false;
1673	loff_t old_isize = i_size_read(inode);
1674	unsigned int ilock_flags = 0;
1675
1676	if (iocb->ki_flags & IOCB_NOWAIT)
1677		ilock_flags |= BTRFS_ILOCK_TRY;
1678
1679	ret = btrfs_inode_lock(inode, ilock_flags);
1680	if (ret < 0)
1681		return ret;
1682
1683	ret = generic_write_checks(iocb, i);
1684	if (ret <= 0)
1685		goto out;
1686
1687	ret = btrfs_write_check(iocb, i, ret);
1688	if (ret < 0)
1689		goto out;
1690
1691	pos = iocb->ki_pos;
1692	nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1693			PAGE_SIZE / (sizeof(struct page *)));
1694	nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1695	nrptrs = max(nrptrs, 8);
1696	pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1697	if (!pages) {
1698		ret = -ENOMEM;
1699		goto out;
1700	}
1701
1702	while (iov_iter_count(i) > 0) {
1703		struct extent_state *cached_state = NULL;
1704		size_t offset = offset_in_page(pos);
1705		size_t sector_offset;
1706		size_t write_bytes = min(iov_iter_count(i),
1707					 nrptrs * (size_t)PAGE_SIZE -
1708					 offset);
1709		size_t num_pages;
1710		size_t reserve_bytes;
1711		size_t dirty_pages;
1712		size_t copied;
1713		size_t dirty_sectors;
1714		size_t num_sectors;
1715		int extents_locked;
1716
1717		/*
1718		 * Fault pages before locking them in prepare_pages
1719		 * to avoid recursive lock
1720		 */
1721		if (unlikely(fault_in_iov_iter_readable(i, write_bytes))) {
1722			ret = -EFAULT;
1723			break;
1724		}
1725
1726		only_release_metadata = false;
1727		sector_offset = pos & (fs_info->sectorsize - 1);
1728
1729		extent_changeset_release(data_reserved);
1730		ret = btrfs_check_data_free_space(BTRFS_I(inode),
1731						  &data_reserved, pos,
1732						  write_bytes);
1733		if (ret < 0) {
1734			/*
1735			 * If we don't have to COW at the offset, reserve
1736			 * metadata only. write_bytes may get smaller than
1737			 * requested here.
1738			 */
1739			if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1740						   &write_bytes) > 0)
1741				only_release_metadata = true;
1742			else
1743				break;
1744		}
1745
1746		num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1747		WARN_ON(num_pages > nrptrs);
1748		reserve_bytes = round_up(write_bytes + sector_offset,
1749					 fs_info->sectorsize);
1750		WARN_ON(reserve_bytes == 0);
1751		ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1752				reserve_bytes);
1753		if (ret) {
1754			if (!only_release_metadata)
1755				btrfs_free_reserved_data_space(BTRFS_I(inode),
1756						data_reserved, pos,
1757						write_bytes);
1758			else
1759				btrfs_check_nocow_unlock(BTRFS_I(inode));
1760			break;
1761		}
1762
1763		release_bytes = reserve_bytes;
1764again:
1765		/*
1766		 * This is going to setup the pages array with the number of
1767		 * pages we want, so we don't really need to worry about the
1768		 * contents of pages from loop to loop
1769		 */
1770		ret = prepare_pages(inode, pages, num_pages,
1771				    pos, write_bytes,
1772				    force_page_uptodate);
1773		if (ret) {
1774			btrfs_delalloc_release_extents(BTRFS_I(inode),
1775						       reserve_bytes);
1776			break;
1777		}
1778
1779		extents_locked = lock_and_cleanup_extent_if_need(
1780				BTRFS_I(inode), pages,
1781				num_pages, pos, write_bytes, &lockstart,
1782				&lockend, &cached_state);
1783		if (extents_locked < 0) {
1784			if (extents_locked == -EAGAIN)
1785				goto again;
1786			btrfs_delalloc_release_extents(BTRFS_I(inode),
1787						       reserve_bytes);
1788			ret = extents_locked;
1789			break;
1790		}
1791
1792		copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1793
1794		num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1795		dirty_sectors = round_up(copied + sector_offset,
1796					fs_info->sectorsize);
1797		dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1798
1799		/*
1800		 * if we have trouble faulting in the pages, fall
1801		 * back to one page at a time
1802		 */
1803		if (copied < write_bytes)
1804			nrptrs = 1;
1805
1806		if (copied == 0) {
1807			force_page_uptodate = true;
1808			dirty_sectors = 0;
1809			dirty_pages = 0;
1810		} else {
1811			force_page_uptodate = false;
1812			dirty_pages = DIV_ROUND_UP(copied + offset,
1813						   PAGE_SIZE);
1814		}
1815
1816		if (num_sectors > dirty_sectors) {
1817			/* release everything except the sectors we dirtied */
1818			release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1819			if (only_release_metadata) {
1820				btrfs_delalloc_release_metadata(BTRFS_I(inode),
1821							release_bytes, true);
1822			} else {
1823				u64 __pos;
1824
1825				__pos = round_down(pos,
1826						   fs_info->sectorsize) +
1827					(dirty_pages << PAGE_SHIFT);
1828				btrfs_delalloc_release_space(BTRFS_I(inode),
1829						data_reserved, __pos,
1830						release_bytes, true);
1831			}
1832		}
1833
1834		release_bytes = round_up(copied + sector_offset,
1835					fs_info->sectorsize);
1836
1837		ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1838					dirty_pages, pos, copied,
1839					&cached_state, only_release_metadata);
1840
1841		/*
1842		 * If we have not locked the extent range, because the range's
1843		 * start offset is >= i_size, we might still have a non-NULL
1844		 * cached extent state, acquired while marking the extent range
1845		 * as delalloc through btrfs_dirty_pages(). Therefore free any
1846		 * possible cached extent state to avoid a memory leak.
1847		 */
1848		if (extents_locked)
1849			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1850					     lockstart, lockend, &cached_state);
1851		else
1852			free_extent_state(cached_state);
1853
1854		btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1855		if (ret) {
1856			btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1857			break;
1858		}
1859
1860		release_bytes = 0;
1861		if (only_release_metadata)
1862			btrfs_check_nocow_unlock(BTRFS_I(inode));
1863
1864		btrfs_drop_pages(fs_info, pages, num_pages, pos, copied);
1865
1866		cond_resched();
1867
1868		balance_dirty_pages_ratelimited(inode->i_mapping);
1869
1870		pos += copied;
1871		num_written += copied;
1872	}
1873
1874	kfree(pages);
1875
1876	if (release_bytes) {
1877		if (only_release_metadata) {
1878			btrfs_check_nocow_unlock(BTRFS_I(inode));
1879			btrfs_delalloc_release_metadata(BTRFS_I(inode),
1880					release_bytes, true);
1881		} else {
1882			btrfs_delalloc_release_space(BTRFS_I(inode),
1883					data_reserved,
1884					round_down(pos, fs_info->sectorsize),
1885					release_bytes, true);
1886		}
1887	}
1888
1889	extent_changeset_free(data_reserved);
1890	if (num_written > 0) {
1891		pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1892		iocb->ki_pos += num_written;
1893	}
1894out:
1895	btrfs_inode_unlock(inode, ilock_flags);
1896	return num_written ? num_written : ret;
1897}
1898
1899static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1900			       const struct iov_iter *iter, loff_t offset)
1901{
1902	const u32 blocksize_mask = fs_info->sectorsize - 1;
1903
1904	if (offset & blocksize_mask)
1905		return -EINVAL;
1906
1907	if (iov_iter_alignment(iter) & blocksize_mask)
1908		return -EINVAL;
1909
1910	return 0;
1911}
1912
1913static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1914{
1915	const bool is_sync_write = (iocb->ki_flags & IOCB_DSYNC);
1916	struct file *file = iocb->ki_filp;
1917	struct inode *inode = file_inode(file);
1918	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1919	loff_t pos;
1920	ssize_t written = 0;
1921	ssize_t written_buffered;
1922	size_t prev_left = 0;
1923	loff_t endbyte;
1924	ssize_t err;
1925	unsigned int ilock_flags = 0;
1926
1927	if (iocb->ki_flags & IOCB_NOWAIT)
1928		ilock_flags |= BTRFS_ILOCK_TRY;
1929
1930	/* If the write DIO is within EOF, use a shared lock */
1931	if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1932		ilock_flags |= BTRFS_ILOCK_SHARED;
1933
1934relock:
1935	err = btrfs_inode_lock(inode, ilock_flags);
1936	if (err < 0)
1937		return err;
1938
1939	err = generic_write_checks(iocb, from);
1940	if (err <= 0) {
1941		btrfs_inode_unlock(inode, ilock_flags);
1942		return err;
1943	}
1944
1945	err = btrfs_write_check(iocb, from, err);
1946	if (err < 0) {
1947		btrfs_inode_unlock(inode, ilock_flags);
1948		goto out;
1949	}
1950
1951	pos = iocb->ki_pos;
1952	/*
1953	 * Re-check since file size may have changed just before taking the
1954	 * lock or pos may have changed because of O_APPEND in generic_write_check()
1955	 */
1956	if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1957	    pos + iov_iter_count(from) > i_size_read(inode)) {
1958		btrfs_inode_unlock(inode, ilock_flags);
1959		ilock_flags &= ~BTRFS_ILOCK_SHARED;
1960		goto relock;
1961	}
1962
1963	if (check_direct_IO(fs_info, from, pos)) {
1964		btrfs_inode_unlock(inode, ilock_flags);
1965		goto buffered;
1966	}
1967
1968	/*
1969	 * We remove IOCB_DSYNC so that we don't deadlock when iomap_dio_rw()
1970	 * calls generic_write_sync() (through iomap_dio_complete()), because
1971	 * that results in calling fsync (btrfs_sync_file()) which will try to
1972	 * lock the inode in exclusive/write mode.
1973	 */
1974	if (is_sync_write)
1975		iocb->ki_flags &= ~IOCB_DSYNC;
1976
1977	/*
1978	 * The iov_iter can be mapped to the same file range we are writing to.
1979	 * If that's the case, then we will deadlock in the iomap code, because
1980	 * it first calls our callback btrfs_dio_iomap_begin(), which will create
1981	 * an ordered extent, and after that it will fault in the pages that the
1982	 * iov_iter refers to. During the fault in we end up in the readahead
1983	 * pages code (starting at btrfs_readahead()), which will lock the range,
1984	 * find that ordered extent and then wait for it to complete (at
1985	 * btrfs_lock_and_flush_ordered_range()), resulting in a deadlock since
1986	 * obviously the ordered extent can never complete as we didn't submit
1987	 * yet the respective bio(s). This always happens when the buffer is
1988	 * memory mapped to the same file range, since the iomap DIO code always
1989	 * invalidates pages in the target file range (after starting and waiting
1990	 * for any writeback).
1991	 *
1992	 * So here we disable page faults in the iov_iter and then retry if we
1993	 * got -EFAULT, faulting in the pages before the retry.
1994	 */
1995again:
1996	from->nofault = true;
1997	err = iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
1998			   IOMAP_DIO_PARTIAL, written);
1999	from->nofault = false;
2000
2001	/* No increment (+=) because iomap returns a cumulative value. */
2002	if (err > 0)
2003		written = err;
2004
2005	if (iov_iter_count(from) > 0 && (err == -EFAULT || err > 0)) {
2006		const size_t left = iov_iter_count(from);
2007		/*
2008		 * We have more data left to write. Try to fault in as many as
2009		 * possible of the remainder pages and retry. We do this without
2010		 * releasing and locking again the inode, to prevent races with
2011		 * truncate.
2012		 *
2013		 * Also, in case the iov refers to pages in the file range of the
2014		 * file we want to write to (due to a mmap), we could enter an
2015		 * infinite loop if we retry after faulting the pages in, since
2016		 * iomap will invalidate any pages in the range early on, before
2017		 * it tries to fault in the pages of the iov. So we keep track of
2018		 * how much was left of iov in the previous EFAULT and fallback
2019		 * to buffered IO in case we haven't made any progress.
2020		 */
2021		if (left == prev_left) {
2022			err = -ENOTBLK;
2023		} else {
2024			fault_in_iov_iter_readable(from, left);
2025			prev_left = left;
2026			goto again;
2027		}
2028	}
2029
2030	btrfs_inode_unlock(inode, ilock_flags);
2031
2032	/*
2033	 * Add back IOCB_DSYNC. Our caller, btrfs_file_write_iter(), will do
2034	 * the fsync (call generic_write_sync()).
2035	 */
2036	if (is_sync_write)
2037		iocb->ki_flags |= IOCB_DSYNC;
2038
2039	/* If 'err' is -ENOTBLK then it means we must fallback to buffered IO. */
2040	if ((err < 0 && err != -ENOTBLK) || !iov_iter_count(from))
2041		goto out;
2042
2043buffered:
2044	pos = iocb->ki_pos;
2045	written_buffered = btrfs_buffered_write(iocb, from);
2046	if (written_buffered < 0) {
2047		err = written_buffered;
2048		goto out;
2049	}
2050	/*
2051	 * Ensure all data is persisted. We want the next direct IO read to be
2052	 * able to read what was just written.
2053	 */
2054	endbyte = pos + written_buffered - 1;
2055	err = btrfs_fdatawrite_range(inode, pos, endbyte);
2056	if (err)
2057		goto out;
2058	err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
2059	if (err)
2060		goto out;
2061	written += written_buffered;
2062	iocb->ki_pos = pos + written_buffered;
2063	invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
2064				 endbyte >> PAGE_SHIFT);
2065out:
2066	return err < 0 ? err : written;
2067}
2068
2069static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
2070				    struct iov_iter *from)
2071{
2072	struct file *file = iocb->ki_filp;
2073	struct btrfs_inode *inode = BTRFS_I(file_inode(file));
2074	ssize_t num_written = 0;
2075	const bool sync = iocb->ki_flags & IOCB_DSYNC;
2076
2077	/*
2078	 * If the fs flips readonly due to some impossible error, although we
2079	 * have opened a file as writable, we have to stop this write operation
2080	 * to ensure consistency.
2081	 */
2082	if (BTRFS_FS_ERROR(inode->root->fs_info))
2083		return -EROFS;
2084
2085	if (!(iocb->ki_flags & IOCB_DIRECT) &&
2086	    (iocb->ki_flags & IOCB_NOWAIT))
2087		return -EOPNOTSUPP;
2088
2089	if (sync)
2090		atomic_inc(&inode->sync_writers);
2091
2092	if (iocb->ki_flags & IOCB_DIRECT)
2093		num_written = btrfs_direct_write(iocb, from);
2094	else
2095		num_written = btrfs_buffered_write(iocb, from);
2096
2097	btrfs_set_inode_last_sub_trans(inode);
2098
2099	if (num_written > 0)
2100		num_written = generic_write_sync(iocb, num_written);
2101
2102	if (sync)
2103		atomic_dec(&inode->sync_writers);
2104
2105	current->backing_dev_info = NULL;
2106	return num_written;
2107}
2108
2109int btrfs_release_file(struct inode *inode, struct file *filp)
2110{
2111	struct btrfs_file_private *private = filp->private_data;
2112
2113	if (private && private->filldir_buf)
2114		kfree(private->filldir_buf);
2115	kfree(private);
2116	filp->private_data = NULL;
2117
2118	/*
2119	 * Set by setattr when we are about to truncate a file from a non-zero
2120	 * size to a zero size.  This tries to flush down new bytes that may
2121	 * have been written if the application were using truncate to replace
2122	 * a file in place.
2123	 */
2124	if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2125			       &BTRFS_I(inode)->runtime_flags))
2126			filemap_flush(inode->i_mapping);
2127	return 0;
2128}
2129
2130static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2131{
2132	int ret;
2133	struct blk_plug plug;
2134
2135	/*
2136	 * This is only called in fsync, which would do synchronous writes, so
2137	 * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2138	 * multiple disks using raid profile, a large IO can be split to
2139	 * several segments of stripe length (currently 64K).
2140	 */
2141	blk_start_plug(&plug);
2142	atomic_inc(&BTRFS_I(inode)->sync_writers);
2143	ret = btrfs_fdatawrite_range(inode, start, end);
2144	atomic_dec(&BTRFS_I(inode)->sync_writers);
2145	blk_finish_plug(&plug);
2146
2147	return ret;
2148}
2149
2150static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
2151{
2152	struct btrfs_inode *inode = BTRFS_I(ctx->inode);
2153	struct btrfs_fs_info *fs_info = inode->root->fs_info;
2154
2155	if (btrfs_inode_in_log(inode, fs_info->generation) &&
2156	    list_empty(&ctx->ordered_extents))
2157		return true;
2158
2159	/*
2160	 * If we are doing a fast fsync we can not bail out if the inode's
2161	 * last_trans is <= then the last committed transaction, because we only
2162	 * update the last_trans of the inode during ordered extent completion,
2163	 * and for a fast fsync we don't wait for that, we only wait for the
2164	 * writeback to complete.
2165	 */
2166	if (inode->last_trans <= fs_info->last_trans_committed &&
2167	    (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
2168	     list_empty(&ctx->ordered_extents)))
2169		return true;
2170
2171	return false;
2172}
2173
2174/*
2175 * fsync call for both files and directories.  This logs the inode into
2176 * the tree log instead of forcing full commits whenever possible.
2177 *
2178 * It needs to call filemap_fdatawait so that all ordered extent updates are
2179 * in the metadata btree are up to date for copying to the log.
2180 *
2181 * It drops the inode mutex before doing the tree log commit.  This is an
2182 * important optimization for directories because holding the mutex prevents
2183 * new operations on the dir while we write to disk.
2184 */
2185int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2186{
2187	struct dentry *dentry = file_dentry(file);
2188	struct inode *inode = d_inode(dentry);
2189	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2190	struct btrfs_root *root = BTRFS_I(inode)->root;
2191	struct btrfs_trans_handle *trans;
2192	struct btrfs_log_ctx ctx;
2193	int ret = 0, err;
2194	u64 len;
2195	bool full_sync;
2196
2197	trace_btrfs_sync_file(file, datasync);
2198
2199	btrfs_init_log_ctx(&ctx, inode);
2200
2201	/*
2202	 * Always set the range to a full range, otherwise we can get into
2203	 * several problems, from missing file extent items to represent holes
2204	 * when not using the NO_HOLES feature, to log tree corruption due to
2205	 * races between hole detection during logging and completion of ordered
2206	 * extents outside the range, to missing checksums due to ordered extents
2207	 * for which we flushed only a subset of their pages.
2208	 */
2209	start = 0;
2210	end = LLONG_MAX;
2211	len = (u64)LLONG_MAX + 1;
2212
2213	/*
2214	 * We write the dirty pages in the range and wait until they complete
2215	 * out of the ->i_mutex. If so, we can flush the dirty pages by
2216	 * multi-task, and make the performance up.  See
2217	 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2218	 */
2219	ret = start_ordered_ops(inode, start, end);
2220	if (ret)
2221		goto out;
2222
2223	btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2224
2225	atomic_inc(&root->log_batch);
2226
2227	/*
2228	 * Always check for the full sync flag while holding the inode's lock,
2229	 * to avoid races with other tasks. The flag must be either set all the
2230	 * time during logging or always off all the time while logging.
2231	 */
2232	full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2233			     &BTRFS_I(inode)->runtime_flags);
2234
2235	/*
2236	 * Before we acquired the inode's lock and the mmap lock, someone may
2237	 * have dirtied more pages in the target range. We need to make sure
2238	 * that writeback for any such pages does not start while we are logging
2239	 * the inode, because if it does, any of the following might happen when
2240	 * we are not doing a full inode sync:
2241	 *
2242	 * 1) We log an extent after its writeback finishes but before its
2243	 *    checksums are added to the csum tree, leading to -EIO errors
2244	 *    when attempting to read the extent after a log replay.
2245	 *
2246	 * 2) We can end up logging an extent before its writeback finishes.
2247	 *    Therefore after the log replay we will have a file extent item
2248	 *    pointing to an unwritten extent (and no data checksums as well).
2249	 *
2250	 * So trigger writeback for any eventual new dirty pages and then we
2251	 * wait for all ordered extents to complete below.
2252	 */
2253	ret = start_ordered_ops(inode, start, end);
2254	if (ret) {
2255		btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2256		goto out;
2257	}
2258
2259	/*
2260	 * We have to do this here to avoid the priority inversion of waiting on
2261	 * IO of a lower priority task while holding a transaction open.
2262	 *
2263	 * For a full fsync we wait for the ordered extents to complete while
2264	 * for a fast fsync we wait just for writeback to complete, and then
2265	 * attach the ordered extents to the transaction so that a transaction
2266	 * commit waits for their completion, to avoid data loss if we fsync,
2267	 * the current transaction commits before the ordered extents complete
2268	 * and a power failure happens right after that.
2269	 *
2270	 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
2271	 * logical address recorded in the ordered extent may change. We need
2272	 * to wait for the IO to stabilize the logical address.
2273	 */
2274	if (full_sync || btrfs_is_zoned(fs_info)) {
2275		ret = btrfs_wait_ordered_range(inode, start, len);
2276	} else {
2277		/*
2278		 * Get our ordered extents as soon as possible to avoid doing
2279		 * checksum lookups in the csum tree, and use instead the
2280		 * checksums attached to the ordered extents.
2281		 */
2282		btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2283						      &ctx.ordered_extents);
2284		ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2285	}
2286
2287	if (ret)
2288		goto out_release_extents;
2289
2290	atomic_inc(&root->log_batch);
2291
2292	smp_mb();
2293	if (skip_inode_logging(&ctx)) {
2294		/*
2295		 * We've had everything committed since the last time we were
2296		 * modified so clear this flag in case it was set for whatever
2297		 * reason, it's no longer relevant.
2298		 */
2299		clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2300			  &BTRFS_I(inode)->runtime_flags);
2301		/*
2302		 * An ordered extent might have started before and completed
2303		 * already with io errors, in which case the inode was not
2304		 * updated and we end up here. So check the inode's mapping
2305		 * for any errors that might have happened since we last
2306		 * checked called fsync.
2307		 */
2308		ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2309		goto out_release_extents;
2310	}
2311
2312	/*
2313	 * We use start here because we will need to wait on the IO to complete
2314	 * in btrfs_sync_log, which could require joining a transaction (for
2315	 * example checking cross references in the nocow path).  If we use join
2316	 * here we could get into a situation where we're waiting on IO to
2317	 * happen that is blocked on a transaction trying to commit.  With start
2318	 * we inc the extwriter counter, so we wait for all extwriters to exit
2319	 * before we start blocking joiners.  This comment is to keep somebody
2320	 * from thinking they are super smart and changing this to
2321	 * btrfs_join_transaction *cough*Josef*cough*.
2322	 */
2323	trans = btrfs_start_transaction(root, 0);
2324	if (IS_ERR(trans)) {
2325		ret = PTR_ERR(trans);
2326		goto out_release_extents;
2327	}
2328	trans->in_fsync = true;
2329
2330	ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2331	btrfs_release_log_ctx_extents(&ctx);
2332	if (ret < 0) {
2333		/* Fallthrough and commit/free transaction. */
2334		ret = 1;
2335	}
2336
2337	/* we've logged all the items and now have a consistent
2338	 * version of the file in the log.  It is possible that
2339	 * someone will come in and modify the file, but that's
2340	 * fine because the log is consistent on disk, and we
2341	 * have references to all of the file's extents
2342	 *
2343	 * It is possible that someone will come in and log the
2344	 * file again, but that will end up using the synchronization
2345	 * inside btrfs_sync_log to keep things safe.
2346	 */
2347	btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2348
2349	if (ret != BTRFS_NO_LOG_SYNC) {
2350		if (!ret) {
2351			ret = btrfs_sync_log(trans, root, &ctx);
2352			if (!ret) {
2353				ret = btrfs_end_transaction(trans);
2354				goto out;
2355			}
2356		}
2357		if (!full_sync) {
2358			ret = btrfs_wait_ordered_range(inode, start, len);
2359			if (ret) {
2360				btrfs_end_transaction(trans);
2361				goto out;
2362			}
2363		}
2364		ret = btrfs_commit_transaction(trans);
2365	} else {
2366		ret = btrfs_end_transaction(trans);
2367	}
2368out:
2369	ASSERT(list_empty(&ctx.list));
2370	err = file_check_and_advance_wb_err(file);
2371	if (!ret)
2372		ret = err;
2373	return ret > 0 ? -EIO : ret;
2374
2375out_release_extents:
2376	btrfs_release_log_ctx_extents(&ctx);
2377	btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2378	goto out;
2379}
2380
2381static const struct vm_operations_struct btrfs_file_vm_ops = {
2382	.fault		= filemap_fault,
2383	.map_pages	= filemap_map_pages,
2384	.page_mkwrite	= btrfs_page_mkwrite,
2385};
2386
2387static int btrfs_file_mmap(struct file	*filp, struct vm_area_struct *vma)
2388{
2389	struct address_space *mapping = filp->f_mapping;
2390
2391	if (!mapping->a_ops->readpage)
2392		return -ENOEXEC;
2393
2394	file_accessed(filp);
2395	vma->vm_ops = &btrfs_file_vm_ops;
2396
2397	return 0;
2398}
2399
2400static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2401			  int slot, u64 start, u64 end)
2402{
2403	struct btrfs_file_extent_item *fi;
2404	struct btrfs_key key;
2405
2406	if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2407		return 0;
2408
2409	btrfs_item_key_to_cpu(leaf, &key, slot);
2410	if (key.objectid != btrfs_ino(inode) ||
2411	    key.type != BTRFS_EXTENT_DATA_KEY)
2412		return 0;
2413
2414	fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2415
2416	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2417		return 0;
2418
2419	if (btrfs_file_extent_disk_bytenr(leaf, fi))
2420		return 0;
2421
2422	if (key.offset == end)
2423		return 1;
2424	if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2425		return 1;
2426	return 0;
2427}
2428
2429static int fill_holes(struct btrfs_trans_handle *trans,
2430		struct btrfs_inode *inode,
2431		struct btrfs_path *path, u64 offset, u64 end)
2432{
2433	struct btrfs_fs_info *fs_info = trans->fs_info;
2434	struct btrfs_root *root = inode->root;
2435	struct extent_buffer *leaf;
2436	struct btrfs_file_extent_item *fi;
2437	struct extent_map *hole_em;
2438	struct extent_map_tree *em_tree = &inode->extent_tree;
2439	struct btrfs_key key;
2440	int ret;
2441
2442	if (btrfs_fs_incompat(fs_info, NO_HOLES))
2443		goto out;
2444
2445	key.objectid = btrfs_ino(inode);
2446	key.type = BTRFS_EXTENT_DATA_KEY;
2447	key.offset = offset;
2448
2449	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2450	if (ret <= 0) {
2451		/*
2452		 * We should have dropped this offset, so if we find it then
2453		 * something has gone horribly wrong.
2454		 */
2455		if (ret == 0)
2456			ret = -EINVAL;
2457		return ret;
2458	}
2459
2460	leaf = path->nodes[0];
2461	if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2462		u64 num_bytes;
2463
2464		path->slots[0]--;
2465		fi = btrfs_item_ptr(leaf, path->slots[0],
2466				    struct btrfs_file_extent_item);
2467		num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2468			end - offset;
2469		btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2470		btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2471		btrfs_set_file_extent_offset(leaf, fi, 0);
2472		btrfs_mark_buffer_dirty(leaf);
2473		goto out;
2474	}
2475
2476	if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2477		u64 num_bytes;
2478
2479		key.offset = offset;
2480		btrfs_set_item_key_safe(fs_info, path, &key);
2481		fi = btrfs_item_ptr(leaf, path->slots[0],
2482				    struct btrfs_file_extent_item);
2483		num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2484			offset;
2485		btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2486		btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2487		btrfs_set_file_extent_offset(leaf, fi, 0);
2488		btrfs_mark_buffer_dirty(leaf);
2489		goto out;
2490	}
2491	btrfs_release_path(path);
2492
2493	ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2494			offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2495	if (ret)
2496		return ret;
2497
2498out:
2499	btrfs_release_path(path);
2500
2501	hole_em = alloc_extent_map();
2502	if (!hole_em) {
2503		btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2504		set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2505	} else {
2506		hole_em->start = offset;
2507		hole_em->len = end - offset;
2508		hole_em->ram_bytes = hole_em->len;
2509		hole_em->orig_start = offset;
2510
2511		hole_em->block_start = EXTENT_MAP_HOLE;
2512		hole_em->block_len = 0;
2513		hole_em->orig_block_len = 0;
2514		hole_em->compress_type = BTRFS_COMPRESS_NONE;
2515		hole_em->generation = trans->transid;
2516
2517		do {
2518			btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2519			write_lock(&em_tree->lock);
2520			ret = add_extent_mapping(em_tree, hole_em, 1);
2521			write_unlock(&em_tree->lock);
2522		} while (ret == -EEXIST);
2523		free_extent_map(hole_em);
2524		if (ret)
2525			set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2526					&inode->runtime_flags);
2527	}
2528
2529	return 0;
2530}
2531
2532/*
2533 * Find a hole extent on given inode and change start/len to the end of hole
2534 * extent.(hole/vacuum extent whose em->start <= start &&
2535 *	   em->start + em->len > start)
2536 * When a hole extent is found, return 1 and modify start/len.
2537 */
2538static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2539{
2540	struct btrfs_fs_info *fs_info = inode->root->fs_info;
2541	struct extent_map *em;
2542	int ret = 0;
2543
2544	em = btrfs_get_extent(inode, NULL, 0,
2545			      round_down(*start, fs_info->sectorsize),
2546			      round_up(*len, fs_info->sectorsize));
2547	if (IS_ERR(em))
2548		return PTR_ERR(em);
2549
2550	/* Hole or vacuum extent(only exists in no-hole mode) */
2551	if (em->block_start == EXTENT_MAP_HOLE) {
2552		ret = 1;
2553		*len = em->start + em->len > *start + *len ?
2554		       0 : *start + *len - em->start - em->len;
2555		*start = em->start + em->len;
2556	}
2557	free_extent_map(em);
2558	return ret;
2559}
2560
2561static int btrfs_punch_hole_lock_range(struct inode *inode,
2562				       const u64 lockstart,
2563				       const u64 lockend,
2564				       struct extent_state **cached_state)
2565{
2566	/*
2567	 * For subpage case, if the range is not at page boundary, we could
2568	 * have pages at the leading/tailing part of the range.
2569	 * This could lead to dead loop since filemap_range_has_page()
2570	 * will always return true.
2571	 * So here we need to do extra page alignment for
2572	 * filemap_range_has_page().
2573	 */
2574	const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2575	const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2576
2577	while (1) {
2578		struct btrfs_ordered_extent *ordered;
2579		int ret;
2580
2581		truncate_pagecache_range(inode, lockstart, lockend);
2582
2583		lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2584				 cached_state);
2585		ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2586							    lockend);
2587
2588		/*
2589		 * We need to make sure we have no ordered extents in this range
2590		 * and nobody raced in and read a page in this range, if we did
2591		 * we need to try again.
2592		 */
2593		if ((!ordered ||
2594		    (ordered->file_offset + ordered->num_bytes <= lockstart ||
2595		     ordered->file_offset > lockend)) &&
2596		     !filemap_range_has_page(inode->i_mapping,
2597					     page_lockstart, page_lockend)) {
2598			if (ordered)
2599				btrfs_put_ordered_extent(ordered);
2600			break;
2601		}
2602		if (ordered)
2603			btrfs_put_ordered_extent(ordered);
2604		unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2605				     lockend, cached_state);
2606		ret = btrfs_wait_ordered_range(inode, lockstart,
2607					       lockend - lockstart + 1);
2608		if (ret)
2609			return ret;
2610	}
2611	return 0;
2612}
2613
2614static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2615				     struct btrfs_inode *inode,
2616				     struct btrfs_path *path,
2617				     struct btrfs_replace_extent_info *extent_info,
2618				     const u64 replace_len,
2619				     const u64 bytes_to_drop)
2620{
2621	struct btrfs_fs_info *fs_info = trans->fs_info;
2622	struct btrfs_root *root = inode->root;
2623	struct btrfs_file_extent_item *extent;
2624	struct extent_buffer *leaf;
2625	struct btrfs_key key;
2626	int slot;
2627	struct btrfs_ref ref = { 0 };
2628	int ret;
2629
2630	if (replace_len == 0)
2631		return 0;
2632
2633	if (extent_info->disk_offset == 0 &&
2634	    btrfs_fs_incompat(fs_info, NO_HOLES)) {
2635		btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2636		return 0;
2637	}
2638
2639	key.objectid = btrfs_ino(inode);
2640	key.type = BTRFS_EXTENT_DATA_KEY;
2641	key.offset = extent_info->file_offset;
2642	ret = btrfs_insert_empty_item(trans, root, path, &key,
2643				      sizeof(struct btrfs_file_extent_item));
2644	if (ret)
2645		return ret;
2646	leaf = path->nodes[0];
2647	slot = path->slots[0];
2648	write_extent_buffer(leaf, extent_info->extent_buf,
2649			    btrfs_item_ptr_offset(leaf, slot),
2650			    sizeof(struct btrfs_file_extent_item));
2651	extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2652	ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2653	btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2654	btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2655	if (extent_info->is_new_extent)
2656		btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2657	btrfs_mark_buffer_dirty(leaf);
2658	btrfs_release_path(path);
2659
2660	ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2661						replace_len);
2662	if (ret)
2663		return ret;
2664
2665	/* If it's a hole, nothing more needs to be done. */
2666	if (extent_info->disk_offset == 0) {
2667		btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2668		return 0;
2669	}
2670
2671	btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2672
2673	if (extent_info->is_new_extent && extent_info->insertions == 0) {
2674		key.objectid = extent_info->disk_offset;
2675		key.type = BTRFS_EXTENT_ITEM_KEY;
2676		key.offset = extent_info->disk_len;
2677		ret = btrfs_alloc_reserved_file_extent(trans, root,
2678						       btrfs_ino(inode),
2679						       extent_info->file_offset,
2680						       extent_info->qgroup_reserved,
2681						       &key);
2682	} else {
2683		u64 ref_offset;
2684
2685		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2686				       extent_info->disk_offset,
2687				       extent_info->disk_len, 0);
2688		ref_offset = extent_info->file_offset - extent_info->data_offset;
2689		btrfs_init_data_ref(&ref, root->root_key.objectid,
2690				    btrfs_ino(inode), ref_offset, 0, false);
2691		ret = btrfs_inc_extent_ref(trans, &ref);
2692	}
2693
2694	extent_info->insertions++;
2695
2696	return ret;
2697}
2698
2699/*
2700 * The respective range must have been previously locked, as well as the inode.
2701 * The end offset is inclusive (last byte of the range).
2702 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2703 * the file range with an extent.
2704 * When not punching a hole, we don't want to end up in a state where we dropped
2705 * extents without inserting a new one, so we must abort the transaction to avoid
2706 * a corruption.
2707 */
2708int btrfs_replace_file_extents(struct btrfs_inode *inode,
2709			       struct btrfs_path *path, const u64 start,
2710			       const u64 end,
2711			       struct btrfs_replace_extent_info *extent_info,
2712			       struct btrfs_trans_handle **trans_out)
2713{
2714	struct btrfs_drop_extents_args drop_args = { 0 };
2715	struct btrfs_root *root = inode->root;
2716	struct btrfs_fs_info *fs_info = root->fs_info;
2717	u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2718	u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2719	struct btrfs_trans_handle *trans = NULL;
2720	struct btrfs_block_rsv *rsv;
2721	unsigned int rsv_count;
2722	u64 cur_offset;
2723	u64 len = end - start;
2724	int ret = 0;
2725
2726	if (end <= start)
2727		return -EINVAL;
2728
2729	rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2730	if (!rsv) {
2731		ret = -ENOMEM;
2732		goto out;
2733	}
2734	rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2735	rsv->failfast = 1;
2736
2737	/*
2738	 * 1 - update the inode
2739	 * 1 - removing the extents in the range
2740	 * 1 - adding the hole extent if no_holes isn't set or if we are
2741	 *     replacing the range with a new extent
2742	 */
2743	if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2744		rsv_count = 3;
2745	else
2746		rsv_count = 2;
2747
2748	trans = btrfs_start_transaction(root, rsv_count);
2749	if (IS_ERR(trans)) {
2750		ret = PTR_ERR(trans);
2751		trans = NULL;
2752		goto out_free;
2753	}
2754
2755	ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2756				      min_size, false);
2757	BUG_ON(ret);
2758	trans->block_rsv = rsv;
2759
2760	cur_offset = start;
2761	drop_args.path = path;
2762	drop_args.end = end + 1;
2763	drop_args.drop_cache = true;
2764	while (cur_offset < end) {
2765		drop_args.start = cur_offset;
2766		ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2767		/* If we are punching a hole decrement the inode's byte count */
2768		if (!extent_info)
2769			btrfs_update_inode_bytes(inode, 0,
2770						 drop_args.bytes_found);
2771		if (ret != -ENOSPC) {
2772			/*
2773			 * The only time we don't want to abort is if we are
2774			 * attempting to clone a partial inline extent, in which
2775			 * case we'll get EOPNOTSUPP.  However if we aren't
2776			 * clone we need to abort no matter what, because if we
2777			 * got EOPNOTSUPP via prealloc then we messed up and
2778			 * need to abort.
2779			 */
2780			if (ret &&
2781			    (ret != -EOPNOTSUPP ||
2782			     (extent_info && extent_info->is_new_extent)))
2783				btrfs_abort_transaction(trans, ret);
2784			break;
2785		}
2786
2787		trans->block_rsv = &fs_info->trans_block_rsv;
2788
2789		if (!extent_info && cur_offset < drop_args.drop_end &&
2790		    cur_offset < ino_size) {
2791			ret = fill_holes(trans, inode, path, cur_offset,
2792					 drop_args.drop_end);
2793			if (ret) {
2794				/*
2795				 * If we failed then we didn't insert our hole
2796				 * entries for the area we dropped, so now the
2797				 * fs is corrupted, so we must abort the
2798				 * transaction.
2799				 */
2800				btrfs_abort_transaction(trans, ret);
2801				break;
2802			}
2803		} else if (!extent_info && cur_offset < drop_args.drop_end) {
2804			/*
2805			 * We are past the i_size here, but since we didn't
2806			 * insert holes we need to clear the mapped area so we
2807			 * know to not set disk_i_size in this area until a new
2808			 * file extent is inserted here.
2809			 */
2810			ret = btrfs_inode_clear_file_extent_range(inode,
2811					cur_offset,
2812					drop_args.drop_end - cur_offset);
2813			if (ret) {
2814				/*
2815				 * We couldn't clear our area, so we could
2816				 * presumably adjust up and corrupt the fs, so
2817				 * we need to abort.
2818				 */
2819				btrfs_abort_transaction(trans, ret);
2820				break;
2821			}
2822		}
2823
2824		if (extent_info &&
2825		    drop_args.drop_end > extent_info->file_offset) {
2826			u64 replace_len = drop_args.drop_end -
2827					  extent_info->file_offset;
2828
2829			ret = btrfs_insert_replace_extent(trans, inode,	path,
2830					extent_info, replace_len,
2831					drop_args.bytes_found);
2832			if (ret) {
2833				btrfs_abort_transaction(trans, ret);
2834				break;
2835			}
2836			extent_info->data_len -= replace_len;
2837			extent_info->data_offset += replace_len;
2838			extent_info->file_offset += replace_len;
2839		}
2840
2841		ret = btrfs_update_inode(trans, root, inode);
2842		if (ret)
2843			break;
2844
2845		btrfs_end_transaction(trans);
2846		btrfs_btree_balance_dirty(fs_info);
2847
2848		trans = btrfs_start_transaction(root, rsv_count);
2849		if (IS_ERR(trans)) {
2850			ret = PTR_ERR(trans);
2851			trans = NULL;
2852			break;
2853		}
2854
2855		ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2856					      rsv, min_size, false);
2857		BUG_ON(ret);	/* shouldn't happen */
2858		trans->block_rsv = rsv;
2859
2860		cur_offset = drop_args.drop_end;
2861		len = end - cur_offset;
2862		if (!extent_info && len) {
2863			ret = find_first_non_hole(inode, &cur_offset, &len);
2864			if (unlikely(ret < 0))
2865				break;
2866			if (ret && !len) {
2867				ret = 0;
2868				break;
2869			}
2870		}
2871	}
2872
2873	/*
2874	 * If we were cloning, force the next fsync to be a full one since we
2875	 * we replaced (or just dropped in the case of cloning holes when
2876	 * NO_HOLES is enabled) file extent items and did not setup new extent
2877	 * maps for the replacement extents (or holes).
2878	 */
2879	if (extent_info && !extent_info->is_new_extent)
2880		set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2881
2882	if (ret)
2883		goto out_trans;
2884
2885	trans->block_rsv = &fs_info->trans_block_rsv;
2886	/*
2887	 * If we are using the NO_HOLES feature we might have had already an
2888	 * hole that overlaps a part of the region [lockstart, lockend] and
2889	 * ends at (or beyond) lockend. Since we have no file extent items to
2890	 * represent holes, drop_end can be less than lockend and so we must
2891	 * make sure we have an extent map representing the existing hole (the
2892	 * call to __btrfs_drop_extents() might have dropped the existing extent
2893	 * map representing the existing hole), otherwise the fast fsync path
2894	 * will not record the existence of the hole region
2895	 * [existing_hole_start, lockend].
2896	 */
2897	if (drop_args.drop_end <= end)
2898		drop_args.drop_end = end + 1;
2899	/*
2900	 * Don't insert file hole extent item if it's for a range beyond eof
2901	 * (because it's useless) or if it represents a 0 bytes range (when
2902	 * cur_offset == drop_end).
2903	 */
2904	if (!extent_info && cur_offset < ino_size &&
2905	    cur_offset < drop_args.drop_end) {
2906		ret = fill_holes(trans, inode, path, cur_offset,
2907				 drop_args.drop_end);
2908		if (ret) {
2909			/* Same comment as above. */
2910			btrfs_abort_transaction(trans, ret);
2911			goto out_trans;
2912		}
2913	} else if (!extent_info && cur_offset < drop_args.drop_end) {
2914		/* See the comment in the loop above for the reasoning here. */
2915		ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2916					drop_args.drop_end - cur_offset);
2917		if (ret) {
2918			btrfs_abort_transaction(trans, ret);
2919			goto out_trans;
2920		}
2921
2922	}
2923	if (extent_info) {
2924		ret = btrfs_insert_replace_extent(trans, inode, path,
2925				extent_info, extent_info->data_len,
2926				drop_args.bytes_found);
2927		if (ret) {
2928			btrfs_abort_transaction(trans, ret);
2929			goto out_trans;
2930		}
2931	}
2932
2933out_trans:
2934	if (!trans)
2935		goto out_free;
2936
2937	trans->block_rsv = &fs_info->trans_block_rsv;
2938	if (ret)
2939		btrfs_end_transaction(trans);
2940	else
2941		*trans_out = trans;
2942out_free:
2943	btrfs_free_block_rsv(fs_info, rsv);
2944out:
2945	return ret;
2946}
2947
2948static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2949{
2950	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2951	struct btrfs_root *root = BTRFS_I(inode)->root;
2952	struct extent_state *cached_state = NULL;
2953	struct btrfs_path *path;
2954	struct btrfs_trans_handle *trans = NULL;
2955	u64 lockstart;
2956	u64 lockend;
2957	u64 tail_start;
2958	u64 tail_len;
2959	u64 orig_start = offset;
2960	int ret = 0;
2961	bool same_block;
2962	u64 ino_size;
2963	bool truncated_block = false;
2964	bool updated_inode = false;
2965
2966	ret = btrfs_wait_ordered_range(inode, offset, len);
2967	if (ret)
2968		return ret;
2969
2970	btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2971	ino_size = round_up(inode->i_size, fs_info->sectorsize);
2972	ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2973	if (ret < 0)
2974		goto out_only_mutex;
2975	if (ret && !len) {
2976		/* Already in a large hole */
2977		ret = 0;
2978		goto out_only_mutex;
2979	}
2980
2981	lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2982	lockend = round_down(offset + len,
2983			     btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2984	same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2985		== (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2986	/*
2987	 * We needn't truncate any block which is beyond the end of the file
2988	 * because we are sure there is no data there.
2989	 */
2990	/*
2991	 * Only do this if we are in the same block and we aren't doing the
2992	 * entire block.
2993	 */
2994	if (same_block && len < fs_info->sectorsize) {
2995		if (offset < ino_size) {
2996			truncated_block = true;
2997			ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2998						   0);
2999		} else {
3000			ret = 0;
3001		}
3002		goto out_only_mutex;
3003	}
3004
3005	/* zero back part of the first block */
3006	if (offset < ino_size) {
3007		truncated_block = true;
3008		ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3009		if (ret) {
3010			btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3011			return ret;
3012		}
3013	}
3014
3015	/* Check the aligned pages after the first unaligned page,
3016	 * if offset != orig_start, which means the first unaligned page
3017	 * including several following pages are already in holes,
3018	 * the extra check can be skipped */
3019	if (offset == orig_start) {
3020		/* after truncate page, check hole again */
3021		len = offset + len - lockstart;
3022		offset = lockstart;
3023		ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
3024		if (ret < 0)
3025			goto out_only_mutex;
3026		if (ret && !len) {
3027			ret = 0;
3028			goto out_only_mutex;
3029		}
3030		lockstart = offset;
3031	}
3032
3033	/* Check the tail unaligned part is in a hole */
3034	tail_start = lockend + 1;
3035	tail_len = offset + len - tail_start;
3036	if (tail_len) {
3037		ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
3038		if (unlikely(ret < 0))
3039			goto out_only_mutex;
3040		if (!ret) {
3041			/* zero the front end of the last page */
3042			if (tail_start + tail_len < ino_size) {
3043				truncated_block = true;
3044				ret = btrfs_truncate_block(BTRFS_I(inode),
3045							tail_start + tail_len,
3046							0, 1);
3047				if (ret)
3048					goto out_only_mutex;
3049			}
3050		}
3051	}
3052
3053	if (lockend < lockstart) {
3054		ret = 0;
3055		goto out_only_mutex;
3056	}
3057
3058	ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3059					  &cached_state);
3060	if (ret)
3061		goto out_only_mutex;
3062
3063	path = btrfs_alloc_path();
3064	if (!path) {
3065		ret = -ENOMEM;
3066		goto out;
3067	}
3068
3069	ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
3070					 lockend, NULL, &trans);
3071	btrfs_free_path(path);
3072	if (ret)
3073		goto out;
3074
3075	ASSERT(trans != NULL);
3076	inode_inc_iversion(inode);
3077	inode->i_mtime = inode->i_ctime = current_time(inode);
3078	ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3079	updated_inode = true;
3080	btrfs_end_transaction(trans);
3081	btrfs_btree_balance_dirty(fs_info);
3082out:
3083	unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3084			     &cached_state);
3085out_only_mutex:
3086	if (!updated_inode && truncated_block && !ret) {
3087		/*
3088		 * If we only end up zeroing part of a page, we still need to
3089		 * update the inode item, so that all the time fields are
3090		 * updated as well as the necessary btrfs inode in memory fields
3091		 * for detecting, at fsync time, if the inode isn't yet in the
3092		 * log tree or it's there but not up to date.
3093		 */
3094		struct timespec64 now = current_time(inode);
3095
3096		inode_inc_iversion(inode);
3097		inode->i_mtime = now;
3098		inode->i_ctime = now;
3099		trans = btrfs_start_transaction(root, 1);
3100		if (IS_ERR(trans)) {
3101			ret = PTR_ERR(trans);
3102		} else {
3103			int ret2;
3104
3105			ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3106			ret2 = btrfs_end_transaction(trans);
3107			if (!ret)
3108				ret = ret2;
3109		}
3110	}
3111	btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3112	return ret;
3113}
3114
3115/* Helper structure to record which range is already reserved */
3116struct falloc_range {
3117	struct list_head list;
3118	u64 start;
3119	u64 len;
3120};
3121
3122/*
3123 * Helper function to add falloc range
3124 *
3125 * Caller should have locked the larger range of extent containing
3126 * [start, len)
3127 */
3128static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3129{
3130	struct falloc_range *range = NULL;
3131
3132	if (!list_empty(head)) {
3133		/*
3134		 * As fallocate iterates by bytenr order, we only need to check
3135		 * the last range.
3136		 */
3137		range = list_last_entry(head, struct falloc_range, list);
3138		if (range->start + range->len == start) {
3139			range->len += len;
3140			return 0;
3141		}
3142	}
3143
3144	range = kmalloc(sizeof(*range), GFP_KERNEL);
3145	if (!range)
3146		return -ENOMEM;
3147	range->start = start;
3148	range->len = len;
3149	list_add_tail(&range->list, head);
3150	return 0;
3151}
3152
3153static int btrfs_fallocate_update_isize(struct inode *inode,
3154					const u64 end,
3155					const int mode)
3156{
3157	struct btrfs_trans_handle *trans;
3158	struct btrfs_root *root = BTRFS_I(inode)->root;
3159	int ret;
3160	int ret2;
3161
3162	if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3163		return 0;
3164
3165	trans = btrfs_start_transaction(root, 1);
3166	if (IS_ERR(trans))
3167		return PTR_ERR(trans);
3168
3169	inode->i_ctime = current_time(inode);
3170	i_size_write(inode, end);
3171	btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
3172	ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3173	ret2 = btrfs_end_transaction(trans);
3174
3175	return ret ? ret : ret2;
3176}
3177
3178enum {
3179	RANGE_BOUNDARY_WRITTEN_EXTENT,
3180	RANGE_BOUNDARY_PREALLOC_EXTENT,
3181	RANGE_BOUNDARY_HOLE,
3182};
3183
3184static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3185						 u64 offset)
3186{
3187	const u64 sectorsize = btrfs_inode_sectorsize(inode);
3188	struct extent_map *em;
3189	int ret;
3190
3191	offset = round_down(offset, sectorsize);
3192	em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3193	if (IS_ERR(em))
3194		return PTR_ERR(em);
3195
3196	if (em->block_start == EXTENT_MAP_HOLE)
3197		ret = RANGE_BOUNDARY_HOLE;
3198	else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3199		ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3200	else
3201		ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3202
3203	free_extent_map(em);
3204	return ret;
3205}
3206
3207static int btrfs_zero_range(struct inode *inode,
3208			    loff_t offset,
3209			    loff_t len,
3210			    const int mode)
3211{
3212	struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3213	struct extent_map *em;
3214	struct extent_changeset *data_reserved = NULL;
3215	int ret;
3216	u64 alloc_hint = 0;
3217	const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3218	u64 alloc_start = round_down(offset, sectorsize);
3219	u64 alloc_end = round_up(offset + len, sectorsize);
3220	u64 bytes_to_reserve = 0;
3221	bool space_reserved = false;
3222
3223	inode_dio_wait(inode);
3224
3225	em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3226			      alloc_end - alloc_start);
3227	if (IS_ERR(em)) {
3228		ret = PTR_ERR(em);
3229		goto out;
3230	}
3231
3232	/*
3233	 * Avoid hole punching and extent allocation for some cases. More cases
3234	 * could be considered, but these are unlikely common and we keep things
3235	 * as simple as possible for now. Also, intentionally, if the target
3236	 * range contains one or more prealloc extents together with regular
3237	 * extents and holes, we drop all the existing extents and allocate a
3238	 * new prealloc extent, so that we get a larger contiguous disk extent.
3239	 */
3240	if (em->start <= alloc_start &&
3241	    test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3242		const u64 em_end = em->start + em->len;
3243
3244		if (em_end >= offset + len) {
3245			/*
3246			 * The whole range is already a prealloc extent,
3247			 * do nothing except updating the inode's i_size if
3248			 * needed.
3249			 */
3250			free_extent_map(em);
3251			ret = btrfs_fallocate_update_isize(inode, offset + len,
3252							   mode);
3253			goto out;
3254		}
3255		/*
3256		 * Part of the range is already a prealloc extent, so operate
3257		 * only on the remaining part of the range.
3258		 */
3259		alloc_start = em_end;
3260		ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3261		len = offset + len - alloc_start;
3262		offset = alloc_start;
3263		alloc_hint = em->block_start + em->len;
3264	}
3265	free_extent_map(em);
3266
3267	if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3268	    BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3269		em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3270				      sectorsize);
3271		if (IS_ERR(em)) {
3272			ret = PTR_ERR(em);
3273			goto out;
3274		}
3275
3276		if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3277			free_extent_map(em);
3278			ret = btrfs_fallocate_update_isize(inode, offset + len,
3279							   mode);
3280			goto out;
3281		}
3282		if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3283			free_extent_map(em);
3284			ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3285						   0);
3286			if (!ret)
3287				ret = btrfs_fallocate_update_isize(inode,
3288								   offset + len,
3289								   mode);
3290			return ret;
3291		}
3292		free_extent_map(em);
3293		alloc_start = round_down(offset, sectorsize);
3294		alloc_end = alloc_start + sectorsize;
3295		goto reserve_space;
3296	}
3297
3298	alloc_start = round_up(offset, sectorsize);
3299	alloc_end = round_down(offset + len, sectorsize);
3300
3301	/*
3302	 * For unaligned ranges, check the pages at the boundaries, they might
3303	 * map to an extent, in which case we need to partially zero them, or
3304	 * they might map to a hole, in which case we need our allocation range
3305	 * to cover them.
3306	 */
3307	if (!IS_ALIGNED(offset, sectorsize)) {
3308		ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3309							    offset);
3310		if (ret < 0)
3311			goto out;
3312		if (ret == RANGE_BOUNDARY_HOLE) {
3313			alloc_start = round_down(offset, sectorsize);
3314			ret = 0;
3315		} else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3316			ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3317			if (ret)
3318				goto out;
3319		} else {
3320			ret = 0;
3321		}
3322	}
3323
3324	if (!IS_ALIGNED(offset + len, sectorsize)) {
3325		ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3326							    offset + len);
3327		if (ret < 0)
3328			goto out;
3329		if (ret == RANGE_BOUNDARY_HOLE) {
3330			alloc_end = round_up(offset + len, sectorsize);
3331			ret = 0;
3332		} else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3333			ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
3334						   0, 1);
3335			if (ret)
3336				goto out;
3337		} else {
3338			ret = 0;
3339		}
3340	}
3341
3342reserve_space:
3343	if (alloc_start < alloc_end) {
3344		struct extent_state *cached_state = NULL;
3345		const u64 lockstart = alloc_start;
3346		const u64 lockend = alloc_end - 1;
3347
3348		bytes_to_reserve = alloc_end - alloc_start;
3349		ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3350						      bytes_to_reserve);
3351		if (ret < 0)
3352			goto out;
3353		space_reserved = true;
3354		ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3355						  &cached_state);
3356		if (ret)
3357			goto out;
3358		ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3359						alloc_start, bytes_to_reserve);
3360		if (ret) {
3361			unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3362					     lockend, &cached_state);
3363			goto out;
3364		}
3365		ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3366						alloc_end - alloc_start,
3367						i_blocksize(inode),
3368						offset + len, &alloc_hint);
3369		unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3370				     lockend, &cached_state);
3371		/* btrfs_prealloc_file_range releases reserved space on error */
3372		if (ret) {
3373			space_reserved = false;
3374			goto out;
3375		}
3376	}
3377	ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3378 out:
3379	if (ret && space_reserved)
3380		btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3381					       alloc_start, bytes_to_reserve);
3382	extent_changeset_free(data_reserved);
3383
3384	return ret;
3385}
3386
3387static long btrfs_fallocate(struct file *file, int mode,
3388			    loff_t offset, loff_t len)
3389{
3390	struct inode *inode = file_inode(file);
3391	struct extent_state *cached_state = NULL;
3392	struct extent_changeset *data_reserved = NULL;
3393	struct falloc_range *range;
3394	struct falloc_range *tmp;
3395	struct list_head reserve_list;
3396	u64 cur_offset;
3397	u64 last_byte;
3398	u64 alloc_start;
3399	u64 alloc_end;
3400	u64 alloc_hint = 0;
3401	u64 locked_end;
3402	u64 actual_end = 0;
3403	struct extent_map *em;
3404	int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3405	int ret;
3406
3407	/* Do not allow fallocate in ZONED mode */
3408	if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3409		return -EOPNOTSUPP;
3410
3411	alloc_start = round_down(offset, blocksize);
3412	alloc_end = round_up(offset + len, blocksize);
3413	cur_offset = alloc_start;
3414
3415	/* Make sure we aren't being give some crap mode */
3416	if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3417		     FALLOC_FL_ZERO_RANGE))
3418		return -EOPNOTSUPP;
3419
3420	if (mode & FALLOC_FL_PUNCH_HOLE)
3421		return btrfs_punch_hole(inode, offset, len);
3422
3423	/*
3424	 * Only trigger disk allocation, don't trigger qgroup reserve
3425	 *
3426	 * For qgroup space, it will be checked later.
3427	 */
3428	if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3429		ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3430						      alloc_end - alloc_start);
3431		if (ret < 0)
3432			return ret;
3433	}
3434
3435	btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
3436
3437	if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3438		ret = inode_newsize_ok(inode, offset + len);
3439		if (ret)
3440			goto out;
3441	}
3442
3443	/*
3444	 * TODO: Move these two operations after we have checked
3445	 * accurate reserved space, or fallocate can still fail but
3446	 * with page truncated or size expanded.
3447	 *
3448	 * But that's a minor problem and won't do much harm BTW.
3449	 */
3450	if (alloc_start > inode->i_size) {
3451		ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3452					alloc_start);
3453		if (ret)
3454			goto out;
3455	} else if (offset + len > inode->i_size) {
3456		/*
3457		 * If we are fallocating from the end of the file onward we
3458		 * need to zero out the end of the block if i_size lands in the
3459		 * middle of a block.
3460		 */
3461		ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3462		if (ret)
3463			goto out;
3464	}
3465
3466	/*
3467	 * wait for ordered IO before we have any locks.  We'll loop again
3468	 * below with the locks held.
3469	 */
3470	ret = btrfs_wait_ordered_range(inode, alloc_start,
3471				       alloc_end - alloc_start);
3472	if (ret)
3473		goto out;
3474
3475	if (mode & FALLOC_FL_ZERO_RANGE) {
3476		ret = btrfs_zero_range(inode, offset, len, mode);
3477		btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3478		return ret;
3479	}
3480
3481	locked_end = alloc_end - 1;
3482	while (1) {
3483		struct btrfs_ordered_extent *ordered;
3484
3485		/* the extent lock is ordered inside the running
3486		 * transaction
3487		 */
3488		lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3489				 locked_end, &cached_state);
3490		ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3491							    locked_end);
3492
3493		if (ordered &&
3494		    ordered->file_offset + ordered->num_bytes > alloc_start &&
3495		    ordered->file_offset < alloc_end) {
3496			btrfs_put_ordered_extent(ordered);
3497			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3498					     alloc_start, locked_end,
3499					     &cached_state);
3500			/*
3501			 * we can't wait on the range with the transaction
3502			 * running or with the extent lock held
3503			 */
3504			ret = btrfs_wait_ordered_range(inode, alloc_start,
3505						       alloc_end - alloc_start);
3506			if (ret)
3507				goto out;
3508		} else {
3509			if (ordered)
3510				btrfs_put_ordered_extent(ordered);
3511			break;
3512		}
3513	}
3514
3515	/* First, check if we exceed the qgroup limit */
3516	INIT_LIST_HEAD(&reserve_list);
3517	while (cur_offset < alloc_end) {
3518		em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3519				      alloc_end - cur_offset);
3520		if (IS_ERR(em)) {
3521			ret = PTR_ERR(em);
3522			break;
3523		}
3524		last_byte = min(extent_map_end(em), alloc_end);
3525		actual_end = min_t(u64, extent_map_end(em), offset + len);
3526		last_byte = ALIGN(last_byte, blocksize);
3527		if (em->block_start == EXTENT_MAP_HOLE ||
3528		    (cur_offset >= inode->i_size &&
3529		     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3530			ret = add_falloc_range(&reserve_list, cur_offset,
3531					       last_byte - cur_offset);
3532			if (ret < 0) {
3533				free_extent_map(em);
3534				break;
3535			}
3536			ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3537					&data_reserved, cur_offset,
3538					last_byte - cur_offset);
3539			if (ret < 0) {
3540				cur_offset = last_byte;
3541				free_extent_map(em);
3542				break;
3543			}
3544		} else {
3545			/*
3546			 * Do not need to reserve unwritten extent for this
3547			 * range, free reserved data space first, otherwise
3548			 * it'll result in false ENOSPC error.
3549			 */
3550			btrfs_free_reserved_data_space(BTRFS_I(inode),
3551				data_reserved, cur_offset,
3552				last_byte - cur_offset);
3553		}
3554		free_extent_map(em);
3555		cur_offset = last_byte;
3556	}
3557
3558	/*
3559	 * If ret is still 0, means we're OK to fallocate.
3560	 * Or just cleanup the list and exit.
3561	 */
3562	list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3563		if (!ret)
3564			ret = btrfs_prealloc_file_range(inode, mode,
3565					range->start,
3566					range->len, i_blocksize(inode),
3567					offset + len, &alloc_hint);
3568		else
3569			btrfs_free_reserved_data_space(BTRFS_I(inode),
3570					data_reserved, range->start,
3571					range->len);
3572		list_del(&range->list);
3573		kfree(range);
3574	}
3575	if (ret < 0)
3576		goto out_unlock;
3577
3578	/*
3579	 * We didn't need to allocate any more space, but we still extended the
3580	 * size of the file so we need to update i_size and the inode item.
3581	 */
3582	ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3583out_unlock:
3584	unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3585			     &cached_state);
3586out:
3587	btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3588	/* Let go of our reservation. */
3589	if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3590		btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3591				cur_offset, alloc_end - cur_offset);
3592	extent_changeset_free(data_reserved);
3593	return ret;
3594}
3595
3596static loff_t find_desired_extent(struct btrfs_inode *inode, loff_t offset,
3597				  int whence)
3598{
3599	struct btrfs_fs_info *fs_info = inode->root->fs_info;
3600	struct extent_map *em = NULL;
3601	struct extent_state *cached_state = NULL;
3602	loff_t i_size = inode->vfs_inode.i_size;
3603	u64 lockstart;
3604	u64 lockend;
3605	u64 start;
3606	u64 len;
3607	int ret = 0;
3608
3609	if (i_size == 0 || offset >= i_size)
3610		return -ENXIO;
3611
3612	/*
3613	 * offset can be negative, in this case we start finding DATA/HOLE from
3614	 * the very start of the file.
3615	 */
3616	start = max_t(loff_t, 0, offset);
3617
3618	lockstart = round_down(start, fs_info->sectorsize);
3619	lockend = round_up(i_size, fs_info->sectorsize);
3620	if (lockend <= lockstart)
3621		lockend = lockstart + fs_info->sectorsize;
3622	lockend--;
3623	len = lockend - lockstart + 1;
3624
3625	lock_extent_bits(&inode->io_tree, lockstart, lockend, &cached_state);
3626
3627	while (start < i_size) {
3628		em = btrfs_get_extent_fiemap(inode, start, len);
3629		if (IS_ERR(em)) {
3630			ret = PTR_ERR(em);
3631			em = NULL;
3632			break;
3633		}
3634
3635		if (whence == SEEK_HOLE &&
3636		    (em->block_start == EXTENT_MAP_HOLE ||
3637		     test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3638			break;
3639		else if (whence == SEEK_DATA &&
3640			   (em->block_start != EXTENT_MAP_HOLE &&
3641			    !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3642			break;
3643
3644		start = em->start + em->len;
3645		free_extent_map(em);
3646		em = NULL;
3647		cond_resched();
3648	}
3649	free_extent_map(em);
3650	unlock_extent_cached(&inode->io_tree, lockstart, lockend,
3651			     &cached_state);
3652	if (ret) {
3653		offset = ret;
3654	} else {
3655		if (whence == SEEK_DATA && start >= i_size)
3656			offset = -ENXIO;
3657		else
3658			offset = min_t(loff_t, start, i_size);
3659	}
3660
3661	return offset;
3662}
3663
3664static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3665{
3666	struct inode *inode = file->f_mapping->host;
3667
3668	switch (whence) {
3669	default:
3670		return generic_file_llseek(file, offset, whence);
3671	case SEEK_DATA:
3672	case SEEK_HOLE:
3673		btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3674		offset = find_desired_extent(BTRFS_I(inode), offset, whence);
3675		btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3676		break;
3677	}
3678
3679	if (offset < 0)
3680		return offset;
3681
3682	return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3683}
3684
3685static int btrfs_file_open(struct inode *inode, struct file *filp)
3686{
3687	int ret;
3688
3689	filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3690
3691	ret = fsverity_file_open(inode, filp);
3692	if (ret)
3693		return ret;
3694	return generic_file_open(inode, filp);
3695}
3696
3697static int check_direct_read(struct btrfs_fs_info *fs_info,
3698			     const struct iov_iter *iter, loff_t offset)
3699{
3700	int ret;
3701	int i, seg;
3702
3703	ret = check_direct_IO(fs_info, iter, offset);
3704	if (ret < 0)
3705		return ret;
3706
3707	if (!iter_is_iovec(iter))
3708		return 0;
3709
3710	for (seg = 0; seg < iter->nr_segs; seg++)
3711		for (i = seg + 1; i < iter->nr_segs; i++)
3712			if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3713				return -EINVAL;
3714	return 0;
3715}
3716
3717static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3718{
3719	struct inode *inode = file_inode(iocb->ki_filp);
3720	size_t prev_left = 0;
3721	ssize_t read = 0;
3722	ssize_t ret;
3723
3724	if (fsverity_active(inode))
3725		return 0;
3726
3727	if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3728		return 0;
3729
3730	btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3731again:
3732	/*
3733	 * This is similar to what we do for direct IO writes, see the comment
3734	 * at btrfs_direct_write(), but we also disable page faults in addition
3735	 * to disabling them only at the iov_iter level. This is because when
3736	 * reading from a hole or prealloc extent, iomap calls iov_iter_zero(),
3737	 * which can still trigger page fault ins despite having set ->nofault
3738	 * to true of our 'to' iov_iter.
3739	 *
3740	 * The difference to direct IO writes is that we deadlock when trying
3741	 * to lock the extent range in the inode's tree during he page reads
3742	 * triggered by the fault in (while for writes it is due to waiting for
3743	 * our own ordered extent). This is because for direct IO reads,
3744	 * btrfs_dio_iomap_begin() returns with the extent range locked, which
3745	 * is only unlocked in the endio callback (end_bio_extent_readpage()).
3746	 */
3747	pagefault_disable();
3748	to->nofault = true;
3749	ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
3750			   IOMAP_DIO_PARTIAL, read);
3751	to->nofault = false;
3752	pagefault_enable();
3753
3754	/* No increment (+=) because iomap returns a cumulative value. */
3755	if (ret > 0)
3756		read = ret;
3757
3758	if (iov_iter_count(to) > 0 && (ret == -EFAULT || ret > 0)) {
3759		const size_t left = iov_iter_count(to);
3760
3761		if (left == prev_left) {
3762			/*
3763			 * We didn't make any progress since the last attempt,
3764			 * fallback to a buffered read for the remainder of the
3765			 * range. This is just to avoid any possibility of looping
3766			 * for too long.
3767			 */
3768			ret = read;
3769		} else {
3770			/*
3771			 * We made some progress since the last retry or this is
3772			 * the first time we are retrying. Fault in as many pages
3773			 * as possible and retry.
3774			 */
3775			fault_in_iov_iter_writeable(to, left);
3776			prev_left = left;
3777			goto again;
3778		}
3779	}
3780	btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3781	return ret < 0 ? ret : read;
3782}
3783
3784static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3785{
3786	ssize_t ret = 0;
3787
3788	if (iocb->ki_flags & IOCB_DIRECT) {
3789		ret = btrfs_direct_read(iocb, to);
3790		if (ret < 0 || !iov_iter_count(to) ||
3791		    iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3792			return ret;
3793	}
3794
3795	return filemap_read(iocb, to, ret);
3796}
3797
3798const struct file_operations btrfs_file_operations = {
3799	.llseek		= btrfs_file_llseek,
3800	.read_iter      = btrfs_file_read_iter,
3801	.splice_read	= generic_file_splice_read,
3802	.write_iter	= btrfs_file_write_iter,
3803	.splice_write	= iter_file_splice_write,
3804	.mmap		= btrfs_file_mmap,
3805	.open		= btrfs_file_open,
3806	.release	= btrfs_release_file,
3807	.fsync		= btrfs_sync_file,
3808	.fallocate	= btrfs_fallocate,
3809	.unlocked_ioctl	= btrfs_ioctl,
3810#ifdef CONFIG_COMPAT
3811	.compat_ioctl	= btrfs_compat_ioctl,
3812#endif
3813	.remap_file_range = btrfs_remap_file_range,
3814};
3815
3816void __cold btrfs_auto_defrag_exit(void)
3817{
3818	kmem_cache_destroy(btrfs_inode_defrag_cachep);
3819}
3820
3821int __init btrfs_auto_defrag_init(void)
3822{
3823	btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3824					sizeof(struct inode_defrag), 0,
3825					SLAB_MEM_SPREAD,
3826					NULL);
3827	if (!btrfs_inode_defrag_cachep)
3828		return -ENOMEM;
3829
3830	return 0;
3831}
3832
3833int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3834{
3835	int ret;
3836
3837	/*
3838	 * So with compression we will find and lock a dirty page and clear the
3839	 * first one as dirty, setup an async extent, and immediately return
3840	 * with the entire range locked but with nobody actually marked with
3841	 * writeback.  So we can't just filemap_write_and_wait_range() and
3842	 * expect it to work since it will just kick off a thread to do the
3843	 * actual work.  So we need to call filemap_fdatawrite_range _again_
3844	 * since it will wait on the page lock, which won't be unlocked until
3845	 * after the pages have been marked as writeback and so we're good to go
3846	 * from there.  We have to do this otherwise we'll miss the ordered
3847	 * extents and that results in badness.  Please Josef, do not think you
3848	 * know better and pull this out at some point in the future, it is
3849	 * right and you are wrong.
3850	 */
3851	ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3852	if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3853			     &BTRFS_I(inode)->runtime_flags))
3854		ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3855
3856	return ret;
3857}
3858