xref: /openbmc/linux/fs/btrfs/ordered-data.c (revision 62e59c4e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5 
6 #include <linux/slab.h>
7 #include <linux/blkdev.h>
8 #include <linux/writeback.h>
9 #include <linux/sched/mm.h>
10 #include "ctree.h"
11 #include "transaction.h"
12 #include "btrfs_inode.h"
13 #include "extent_io.h"
14 #include "disk-io.h"
15 #include "compression.h"
16 
17 static struct kmem_cache *btrfs_ordered_extent_cache;
18 
19 static u64 entry_end(struct btrfs_ordered_extent *entry)
20 {
21 	if (entry->file_offset + entry->len < entry->file_offset)
22 		return (u64)-1;
23 	return entry->file_offset + entry->len;
24 }
25 
26 /* returns NULL if the insertion worked, or it returns the node it did find
27  * in the tree
28  */
29 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
30 				   struct rb_node *node)
31 {
32 	struct rb_node **p = &root->rb_node;
33 	struct rb_node *parent = NULL;
34 	struct btrfs_ordered_extent *entry;
35 
36 	while (*p) {
37 		parent = *p;
38 		entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
39 
40 		if (file_offset < entry->file_offset)
41 			p = &(*p)->rb_left;
42 		else if (file_offset >= entry_end(entry))
43 			p = &(*p)->rb_right;
44 		else
45 			return parent;
46 	}
47 
48 	rb_link_node(node, parent, p);
49 	rb_insert_color(node, root);
50 	return NULL;
51 }
52 
53 static void ordered_data_tree_panic(struct inode *inode, int errno,
54 					       u64 offset)
55 {
56 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
57 	btrfs_panic(fs_info, errno,
58 		    "Inconsistency in ordered tree at offset %llu", offset);
59 }
60 
61 /*
62  * look for a given offset in the tree, and if it can't be found return the
63  * first lesser offset
64  */
65 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
66 				     struct rb_node **prev_ret)
67 {
68 	struct rb_node *n = root->rb_node;
69 	struct rb_node *prev = NULL;
70 	struct rb_node *test;
71 	struct btrfs_ordered_extent *entry;
72 	struct btrfs_ordered_extent *prev_entry = NULL;
73 
74 	while (n) {
75 		entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
76 		prev = n;
77 		prev_entry = entry;
78 
79 		if (file_offset < entry->file_offset)
80 			n = n->rb_left;
81 		else if (file_offset >= entry_end(entry))
82 			n = n->rb_right;
83 		else
84 			return n;
85 	}
86 	if (!prev_ret)
87 		return NULL;
88 
89 	while (prev && file_offset >= entry_end(prev_entry)) {
90 		test = rb_next(prev);
91 		if (!test)
92 			break;
93 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
94 				      rb_node);
95 		if (file_offset < entry_end(prev_entry))
96 			break;
97 
98 		prev = test;
99 	}
100 	if (prev)
101 		prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
102 				      rb_node);
103 	while (prev && file_offset < entry_end(prev_entry)) {
104 		test = rb_prev(prev);
105 		if (!test)
106 			break;
107 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
108 				      rb_node);
109 		prev = test;
110 	}
111 	*prev_ret = prev;
112 	return NULL;
113 }
114 
115 /*
116  * helper to check if a given offset is inside a given entry
117  */
118 static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
119 {
120 	if (file_offset < entry->file_offset ||
121 	    entry->file_offset + entry->len <= file_offset)
122 		return 0;
123 	return 1;
124 }
125 
126 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
127 			  u64 len)
128 {
129 	if (file_offset + len <= entry->file_offset ||
130 	    entry->file_offset + entry->len <= file_offset)
131 		return 0;
132 	return 1;
133 }
134 
135 /*
136  * look find the first ordered struct that has this offset, otherwise
137  * the first one less than this offset
138  */
139 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
140 					  u64 file_offset)
141 {
142 	struct rb_root *root = &tree->tree;
143 	struct rb_node *prev = NULL;
144 	struct rb_node *ret;
145 	struct btrfs_ordered_extent *entry;
146 
147 	if (tree->last) {
148 		entry = rb_entry(tree->last, struct btrfs_ordered_extent,
149 				 rb_node);
150 		if (offset_in_entry(entry, file_offset))
151 			return tree->last;
152 	}
153 	ret = __tree_search(root, file_offset, &prev);
154 	if (!ret)
155 		ret = prev;
156 	if (ret)
157 		tree->last = ret;
158 	return ret;
159 }
160 
161 /* allocate and add a new ordered_extent into the per-inode tree.
162  * file_offset is the logical offset in the file
163  *
164  * start is the disk block number of an extent already reserved in the
165  * extent allocation tree
166  *
167  * len is the length of the extent
168  *
169  * The tree is given a single reference on the ordered extent that was
170  * inserted.
171  */
172 static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
173 				      u64 start, u64 len, u64 disk_len,
174 				      int type, int dio, int compress_type)
175 {
176 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
177 	struct btrfs_root *root = BTRFS_I(inode)->root;
178 	struct btrfs_ordered_inode_tree *tree;
179 	struct rb_node *node;
180 	struct btrfs_ordered_extent *entry;
181 
182 	tree = &BTRFS_I(inode)->ordered_tree;
183 	entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
184 	if (!entry)
185 		return -ENOMEM;
186 
187 	entry->file_offset = file_offset;
188 	entry->start = start;
189 	entry->len = len;
190 	entry->disk_len = disk_len;
191 	entry->bytes_left = len;
192 	entry->inode = igrab(inode);
193 	entry->compress_type = compress_type;
194 	entry->truncated_len = (u64)-1;
195 	if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
196 		set_bit(type, &entry->flags);
197 
198 	if (dio)
199 		set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
200 
201 	/* one ref for the tree */
202 	refcount_set(&entry->refs, 1);
203 	init_waitqueue_head(&entry->wait);
204 	INIT_LIST_HEAD(&entry->list);
205 	INIT_LIST_HEAD(&entry->root_extent_list);
206 	INIT_LIST_HEAD(&entry->work_list);
207 	init_completion(&entry->completion);
208 	INIT_LIST_HEAD(&entry->log_list);
209 	INIT_LIST_HEAD(&entry->trans_list);
210 
211 	trace_btrfs_ordered_extent_add(inode, entry);
212 
213 	spin_lock_irq(&tree->lock);
214 	node = tree_insert(&tree->tree, file_offset,
215 			   &entry->rb_node);
216 	if (node)
217 		ordered_data_tree_panic(inode, -EEXIST, file_offset);
218 	spin_unlock_irq(&tree->lock);
219 
220 	spin_lock(&root->ordered_extent_lock);
221 	list_add_tail(&entry->root_extent_list,
222 		      &root->ordered_extents);
223 	root->nr_ordered_extents++;
224 	if (root->nr_ordered_extents == 1) {
225 		spin_lock(&fs_info->ordered_root_lock);
226 		BUG_ON(!list_empty(&root->ordered_root));
227 		list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
228 		spin_unlock(&fs_info->ordered_root_lock);
229 	}
230 	spin_unlock(&root->ordered_extent_lock);
231 
232 	/*
233 	 * We don't need the count_max_extents here, we can assume that all of
234 	 * that work has been done at higher layers, so this is truly the
235 	 * smallest the extent is going to get.
236 	 */
237 	spin_lock(&BTRFS_I(inode)->lock);
238 	btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
239 	spin_unlock(&BTRFS_I(inode)->lock);
240 
241 	return 0;
242 }
243 
244 int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
245 			     u64 start, u64 len, u64 disk_len, int type)
246 {
247 	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
248 					  disk_len, type, 0,
249 					  BTRFS_COMPRESS_NONE);
250 }
251 
252 int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
253 				 u64 start, u64 len, u64 disk_len, int type)
254 {
255 	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
256 					  disk_len, type, 1,
257 					  BTRFS_COMPRESS_NONE);
258 }
259 
260 int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
261 				      u64 start, u64 len, u64 disk_len,
262 				      int type, int compress_type)
263 {
264 	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
265 					  disk_len, type, 0,
266 					  compress_type);
267 }
268 
269 /*
270  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
271  * when an ordered extent is finished.  If the list covers more than one
272  * ordered extent, it is split across multiples.
273  */
274 void btrfs_add_ordered_sum(struct inode *inode,
275 			   struct btrfs_ordered_extent *entry,
276 			   struct btrfs_ordered_sum *sum)
277 {
278 	struct btrfs_ordered_inode_tree *tree;
279 
280 	tree = &BTRFS_I(inode)->ordered_tree;
281 	spin_lock_irq(&tree->lock);
282 	list_add_tail(&sum->list, &entry->list);
283 	spin_unlock_irq(&tree->lock);
284 }
285 
286 /*
287  * this is used to account for finished IO across a given range
288  * of the file.  The IO may span ordered extents.  If
289  * a given ordered_extent is completely done, 1 is returned, otherwise
290  * 0.
291  *
292  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
293  * to make sure this function only returns 1 once for a given ordered extent.
294  *
295  * file_offset is updated to one byte past the range that is recorded as
296  * complete.  This allows you to walk forward in the file.
297  */
298 int btrfs_dec_test_first_ordered_pending(struct inode *inode,
299 				   struct btrfs_ordered_extent **cached,
300 				   u64 *file_offset, u64 io_size, int uptodate)
301 {
302 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
303 	struct btrfs_ordered_inode_tree *tree;
304 	struct rb_node *node;
305 	struct btrfs_ordered_extent *entry = NULL;
306 	int ret;
307 	unsigned long flags;
308 	u64 dec_end;
309 	u64 dec_start;
310 	u64 to_dec;
311 
312 	tree = &BTRFS_I(inode)->ordered_tree;
313 	spin_lock_irqsave(&tree->lock, flags);
314 	node = tree_search(tree, *file_offset);
315 	if (!node) {
316 		ret = 1;
317 		goto out;
318 	}
319 
320 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
321 	if (!offset_in_entry(entry, *file_offset)) {
322 		ret = 1;
323 		goto out;
324 	}
325 
326 	dec_start = max(*file_offset, entry->file_offset);
327 	dec_end = min(*file_offset + io_size, entry->file_offset +
328 		      entry->len);
329 	*file_offset = dec_end;
330 	if (dec_start > dec_end) {
331 		btrfs_crit(fs_info, "bad ordering dec_start %llu end %llu",
332 			   dec_start, dec_end);
333 	}
334 	to_dec = dec_end - dec_start;
335 	if (to_dec > entry->bytes_left) {
336 		btrfs_crit(fs_info,
337 			   "bad ordered accounting left %llu size %llu",
338 			   entry->bytes_left, to_dec);
339 	}
340 	entry->bytes_left -= to_dec;
341 	if (!uptodate)
342 		set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
343 
344 	if (entry->bytes_left == 0) {
345 		ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
346 		/* test_and_set_bit implies a barrier */
347 		cond_wake_up_nomb(&entry->wait);
348 	} else {
349 		ret = 1;
350 	}
351 out:
352 	if (!ret && cached && entry) {
353 		*cached = entry;
354 		refcount_inc(&entry->refs);
355 	}
356 	spin_unlock_irqrestore(&tree->lock, flags);
357 	return ret == 0;
358 }
359 
360 /*
361  * this is used to account for finished IO across a given range
362  * of the file.  The IO should not span ordered extents.  If
363  * a given ordered_extent is completely done, 1 is returned, otherwise
364  * 0.
365  *
366  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
367  * to make sure this function only returns 1 once for a given ordered extent.
368  */
369 int btrfs_dec_test_ordered_pending(struct inode *inode,
370 				   struct btrfs_ordered_extent **cached,
371 				   u64 file_offset, u64 io_size, int uptodate)
372 {
373 	struct btrfs_ordered_inode_tree *tree;
374 	struct rb_node *node;
375 	struct btrfs_ordered_extent *entry = NULL;
376 	unsigned long flags;
377 	int ret;
378 
379 	tree = &BTRFS_I(inode)->ordered_tree;
380 	spin_lock_irqsave(&tree->lock, flags);
381 	if (cached && *cached) {
382 		entry = *cached;
383 		goto have_entry;
384 	}
385 
386 	node = tree_search(tree, file_offset);
387 	if (!node) {
388 		ret = 1;
389 		goto out;
390 	}
391 
392 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
393 have_entry:
394 	if (!offset_in_entry(entry, file_offset)) {
395 		ret = 1;
396 		goto out;
397 	}
398 
399 	if (io_size > entry->bytes_left) {
400 		btrfs_crit(BTRFS_I(inode)->root->fs_info,
401 			   "bad ordered accounting left %llu size %llu",
402 		       entry->bytes_left, io_size);
403 	}
404 	entry->bytes_left -= io_size;
405 	if (!uptodate)
406 		set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
407 
408 	if (entry->bytes_left == 0) {
409 		ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
410 		/* test_and_set_bit implies a barrier */
411 		cond_wake_up_nomb(&entry->wait);
412 	} else {
413 		ret = 1;
414 	}
415 out:
416 	if (!ret && cached && entry) {
417 		*cached = entry;
418 		refcount_inc(&entry->refs);
419 	}
420 	spin_unlock_irqrestore(&tree->lock, flags);
421 	return ret == 0;
422 }
423 
424 /*
425  * used to drop a reference on an ordered extent.  This will free
426  * the extent if the last reference is dropped
427  */
428 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
429 {
430 	struct list_head *cur;
431 	struct btrfs_ordered_sum *sum;
432 
433 	trace_btrfs_ordered_extent_put(entry->inode, entry);
434 
435 	if (refcount_dec_and_test(&entry->refs)) {
436 		ASSERT(list_empty(&entry->log_list));
437 		ASSERT(list_empty(&entry->trans_list));
438 		ASSERT(list_empty(&entry->root_extent_list));
439 		ASSERT(RB_EMPTY_NODE(&entry->rb_node));
440 		if (entry->inode)
441 			btrfs_add_delayed_iput(entry->inode);
442 		while (!list_empty(&entry->list)) {
443 			cur = entry->list.next;
444 			sum = list_entry(cur, struct btrfs_ordered_sum, list);
445 			list_del(&sum->list);
446 			kvfree(sum);
447 		}
448 		kmem_cache_free(btrfs_ordered_extent_cache, entry);
449 	}
450 }
451 
452 /*
453  * remove an ordered extent from the tree.  No references are dropped
454  * and waiters are woken up.
455  */
456 void btrfs_remove_ordered_extent(struct inode *inode,
457 				 struct btrfs_ordered_extent *entry)
458 {
459 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
460 	struct btrfs_ordered_inode_tree *tree;
461 	struct btrfs_inode *btrfs_inode = BTRFS_I(inode);
462 	struct btrfs_root *root = btrfs_inode->root;
463 	struct rb_node *node;
464 
465 	/* This is paired with btrfs_add_ordered_extent. */
466 	spin_lock(&btrfs_inode->lock);
467 	btrfs_mod_outstanding_extents(btrfs_inode, -1);
468 	spin_unlock(&btrfs_inode->lock);
469 	if (root != fs_info->tree_root)
470 		btrfs_delalloc_release_metadata(btrfs_inode, entry->len, false);
471 
472 	tree = &btrfs_inode->ordered_tree;
473 	spin_lock_irq(&tree->lock);
474 	node = &entry->rb_node;
475 	rb_erase(node, &tree->tree);
476 	RB_CLEAR_NODE(node);
477 	if (tree->last == node)
478 		tree->last = NULL;
479 	set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
480 	spin_unlock_irq(&tree->lock);
481 
482 	spin_lock(&root->ordered_extent_lock);
483 	list_del_init(&entry->root_extent_list);
484 	root->nr_ordered_extents--;
485 
486 	trace_btrfs_ordered_extent_remove(inode, entry);
487 
488 	if (!root->nr_ordered_extents) {
489 		spin_lock(&fs_info->ordered_root_lock);
490 		BUG_ON(list_empty(&root->ordered_root));
491 		list_del_init(&root->ordered_root);
492 		spin_unlock(&fs_info->ordered_root_lock);
493 	}
494 	spin_unlock(&root->ordered_extent_lock);
495 	wake_up(&entry->wait);
496 }
497 
498 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
499 {
500 	struct btrfs_ordered_extent *ordered;
501 
502 	ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
503 	btrfs_start_ordered_extent(ordered->inode, ordered, 1);
504 	complete(&ordered->completion);
505 }
506 
507 /*
508  * wait for all the ordered extents in a root.  This is done when balancing
509  * space between drives.
510  */
511 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
512 			       const u64 range_start, const u64 range_len)
513 {
514 	struct btrfs_fs_info *fs_info = root->fs_info;
515 	LIST_HEAD(splice);
516 	LIST_HEAD(skipped);
517 	LIST_HEAD(works);
518 	struct btrfs_ordered_extent *ordered, *next;
519 	u64 count = 0;
520 	const u64 range_end = range_start + range_len;
521 
522 	mutex_lock(&root->ordered_extent_mutex);
523 	spin_lock(&root->ordered_extent_lock);
524 	list_splice_init(&root->ordered_extents, &splice);
525 	while (!list_empty(&splice) && nr) {
526 		ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
527 					   root_extent_list);
528 
529 		if (range_end <= ordered->start ||
530 		    ordered->start + ordered->disk_len <= range_start) {
531 			list_move_tail(&ordered->root_extent_list, &skipped);
532 			cond_resched_lock(&root->ordered_extent_lock);
533 			continue;
534 		}
535 
536 		list_move_tail(&ordered->root_extent_list,
537 			       &root->ordered_extents);
538 		refcount_inc(&ordered->refs);
539 		spin_unlock(&root->ordered_extent_lock);
540 
541 		btrfs_init_work(&ordered->flush_work,
542 				btrfs_flush_delalloc_helper,
543 				btrfs_run_ordered_extent_work, NULL, NULL);
544 		list_add_tail(&ordered->work_list, &works);
545 		btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
546 
547 		cond_resched();
548 		spin_lock(&root->ordered_extent_lock);
549 		if (nr != U64_MAX)
550 			nr--;
551 		count++;
552 	}
553 	list_splice_tail(&skipped, &root->ordered_extents);
554 	list_splice_tail(&splice, &root->ordered_extents);
555 	spin_unlock(&root->ordered_extent_lock);
556 
557 	list_for_each_entry_safe(ordered, next, &works, work_list) {
558 		list_del_init(&ordered->work_list);
559 		wait_for_completion(&ordered->completion);
560 		btrfs_put_ordered_extent(ordered);
561 		cond_resched();
562 	}
563 	mutex_unlock(&root->ordered_extent_mutex);
564 
565 	return count;
566 }
567 
568 u64 btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
569 			     const u64 range_start, const u64 range_len)
570 {
571 	struct btrfs_root *root;
572 	struct list_head splice;
573 	u64 total_done = 0;
574 	u64 done;
575 
576 	INIT_LIST_HEAD(&splice);
577 
578 	mutex_lock(&fs_info->ordered_operations_mutex);
579 	spin_lock(&fs_info->ordered_root_lock);
580 	list_splice_init(&fs_info->ordered_roots, &splice);
581 	while (!list_empty(&splice) && nr) {
582 		root = list_first_entry(&splice, struct btrfs_root,
583 					ordered_root);
584 		root = btrfs_grab_fs_root(root);
585 		BUG_ON(!root);
586 		list_move_tail(&root->ordered_root,
587 			       &fs_info->ordered_roots);
588 		spin_unlock(&fs_info->ordered_root_lock);
589 
590 		done = btrfs_wait_ordered_extents(root, nr,
591 						  range_start, range_len);
592 		btrfs_put_fs_root(root);
593 		total_done += done;
594 
595 		spin_lock(&fs_info->ordered_root_lock);
596 		if (nr != U64_MAX) {
597 			nr -= done;
598 		}
599 	}
600 	list_splice_tail(&splice, &fs_info->ordered_roots);
601 	spin_unlock(&fs_info->ordered_root_lock);
602 	mutex_unlock(&fs_info->ordered_operations_mutex);
603 
604 	return total_done;
605 }
606 
607 /*
608  * Used to start IO or wait for a given ordered extent to finish.
609  *
610  * If wait is one, this effectively waits on page writeback for all the pages
611  * in the extent, and it waits on the io completion code to insert
612  * metadata into the btree corresponding to the extent
613  */
614 void btrfs_start_ordered_extent(struct inode *inode,
615 				       struct btrfs_ordered_extent *entry,
616 				       int wait)
617 {
618 	u64 start = entry->file_offset;
619 	u64 end = start + entry->len - 1;
620 
621 	trace_btrfs_ordered_extent_start(inode, entry);
622 
623 	/*
624 	 * pages in the range can be dirty, clean or writeback.  We
625 	 * start IO on any dirty ones so the wait doesn't stall waiting
626 	 * for the flusher thread to find them
627 	 */
628 	if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
629 		filemap_fdatawrite_range(inode->i_mapping, start, end);
630 	if (wait) {
631 		wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
632 						 &entry->flags));
633 	}
634 }
635 
636 /*
637  * Used to wait on ordered extents across a large range of bytes.
638  */
639 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
640 {
641 	int ret = 0;
642 	int ret_wb = 0;
643 	u64 end;
644 	u64 orig_end;
645 	struct btrfs_ordered_extent *ordered;
646 
647 	if (start + len < start) {
648 		orig_end = INT_LIMIT(loff_t);
649 	} else {
650 		orig_end = start + len - 1;
651 		if (orig_end > INT_LIMIT(loff_t))
652 			orig_end = INT_LIMIT(loff_t);
653 	}
654 
655 	/* start IO across the range first to instantiate any delalloc
656 	 * extents
657 	 */
658 	ret = btrfs_fdatawrite_range(inode, start, orig_end);
659 	if (ret)
660 		return ret;
661 
662 	/*
663 	 * If we have a writeback error don't return immediately. Wait first
664 	 * for any ordered extents that haven't completed yet. This is to make
665 	 * sure no one can dirty the same page ranges and call writepages()
666 	 * before the ordered extents complete - to avoid failures (-EEXIST)
667 	 * when adding the new ordered extents to the ordered tree.
668 	 */
669 	ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
670 
671 	end = orig_end;
672 	while (1) {
673 		ordered = btrfs_lookup_first_ordered_extent(inode, end);
674 		if (!ordered)
675 			break;
676 		if (ordered->file_offset > orig_end) {
677 			btrfs_put_ordered_extent(ordered);
678 			break;
679 		}
680 		if (ordered->file_offset + ordered->len <= start) {
681 			btrfs_put_ordered_extent(ordered);
682 			break;
683 		}
684 		btrfs_start_ordered_extent(inode, ordered, 1);
685 		end = ordered->file_offset;
686 		if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
687 			ret = -EIO;
688 		btrfs_put_ordered_extent(ordered);
689 		if (ret || end == 0 || end == start)
690 			break;
691 		end--;
692 	}
693 	return ret_wb ? ret_wb : ret;
694 }
695 
696 /*
697  * find an ordered extent corresponding to file_offset.  return NULL if
698  * nothing is found, otherwise take a reference on the extent and return it
699  */
700 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
701 							 u64 file_offset)
702 {
703 	struct btrfs_ordered_inode_tree *tree;
704 	struct rb_node *node;
705 	struct btrfs_ordered_extent *entry = NULL;
706 
707 	tree = &BTRFS_I(inode)->ordered_tree;
708 	spin_lock_irq(&tree->lock);
709 	node = tree_search(tree, file_offset);
710 	if (!node)
711 		goto out;
712 
713 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
714 	if (!offset_in_entry(entry, file_offset))
715 		entry = NULL;
716 	if (entry)
717 		refcount_inc(&entry->refs);
718 out:
719 	spin_unlock_irq(&tree->lock);
720 	return entry;
721 }
722 
723 /* Since the DIO code tries to lock a wide area we need to look for any ordered
724  * extents that exist in the range, rather than just the start of the range.
725  */
726 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
727 		struct btrfs_inode *inode, u64 file_offset, u64 len)
728 {
729 	struct btrfs_ordered_inode_tree *tree;
730 	struct rb_node *node;
731 	struct btrfs_ordered_extent *entry = NULL;
732 
733 	tree = &inode->ordered_tree;
734 	spin_lock_irq(&tree->lock);
735 	node = tree_search(tree, file_offset);
736 	if (!node) {
737 		node = tree_search(tree, file_offset + len);
738 		if (!node)
739 			goto out;
740 	}
741 
742 	while (1) {
743 		entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
744 		if (range_overlaps(entry, file_offset, len))
745 			break;
746 
747 		if (entry->file_offset >= file_offset + len) {
748 			entry = NULL;
749 			break;
750 		}
751 		entry = NULL;
752 		node = rb_next(node);
753 		if (!node)
754 			break;
755 	}
756 out:
757 	if (entry)
758 		refcount_inc(&entry->refs);
759 	spin_unlock_irq(&tree->lock);
760 	return entry;
761 }
762 
763 /*
764  * lookup and return any extent before 'file_offset'.  NULL is returned
765  * if none is found
766  */
767 struct btrfs_ordered_extent *
768 btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
769 {
770 	struct btrfs_ordered_inode_tree *tree;
771 	struct rb_node *node;
772 	struct btrfs_ordered_extent *entry = NULL;
773 
774 	tree = &BTRFS_I(inode)->ordered_tree;
775 	spin_lock_irq(&tree->lock);
776 	node = tree_search(tree, file_offset);
777 	if (!node)
778 		goto out;
779 
780 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
781 	refcount_inc(&entry->refs);
782 out:
783 	spin_unlock_irq(&tree->lock);
784 	return entry;
785 }
786 
787 /*
788  * After an extent is done, call this to conditionally update the on disk
789  * i_size.  i_size is updated to cover any fully written part of the file.
790  */
791 int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
792 				struct btrfs_ordered_extent *ordered)
793 {
794 	struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
795 	u64 disk_i_size;
796 	u64 new_i_size;
797 	u64 i_size = i_size_read(inode);
798 	struct rb_node *node;
799 	struct rb_node *prev = NULL;
800 	struct btrfs_ordered_extent *test;
801 	int ret = 1;
802 	u64 orig_offset = offset;
803 
804 	spin_lock_irq(&tree->lock);
805 	if (ordered) {
806 		offset = entry_end(ordered);
807 		if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags))
808 			offset = min(offset,
809 				     ordered->file_offset +
810 				     ordered->truncated_len);
811 	} else {
812 		offset = ALIGN(offset, btrfs_inode_sectorsize(inode));
813 	}
814 	disk_i_size = BTRFS_I(inode)->disk_i_size;
815 
816 	/*
817 	 * truncate file.
818 	 * If ordered is not NULL, then this is called from endio and
819 	 * disk_i_size will be updated by either truncate itself or any
820 	 * in-flight IOs which are inside the disk_i_size.
821 	 *
822 	 * Because btrfs_setsize() may set i_size with disk_i_size if truncate
823 	 * fails somehow, we need to make sure we have a precise disk_i_size by
824 	 * updating it as usual.
825 	 *
826 	 */
827 	if (!ordered && disk_i_size > i_size) {
828 		BTRFS_I(inode)->disk_i_size = orig_offset;
829 		ret = 0;
830 		goto out;
831 	}
832 
833 	/*
834 	 * if the disk i_size is already at the inode->i_size, or
835 	 * this ordered extent is inside the disk i_size, we're done
836 	 */
837 	if (disk_i_size == i_size)
838 		goto out;
839 
840 	/*
841 	 * We still need to update disk_i_size if outstanding_isize is greater
842 	 * than disk_i_size.
843 	 */
844 	if (offset <= disk_i_size &&
845 	    (!ordered || ordered->outstanding_isize <= disk_i_size))
846 		goto out;
847 
848 	/*
849 	 * walk backward from this ordered extent to disk_i_size.
850 	 * if we find an ordered extent then we can't update disk i_size
851 	 * yet
852 	 */
853 	if (ordered) {
854 		node = rb_prev(&ordered->rb_node);
855 	} else {
856 		prev = tree_search(tree, offset);
857 		/*
858 		 * we insert file extents without involving ordered struct,
859 		 * so there should be no ordered struct cover this offset
860 		 */
861 		if (prev) {
862 			test = rb_entry(prev, struct btrfs_ordered_extent,
863 					rb_node);
864 			BUG_ON(offset_in_entry(test, offset));
865 		}
866 		node = prev;
867 	}
868 	for (; node; node = rb_prev(node)) {
869 		test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
870 
871 		/* We treat this entry as if it doesn't exist */
872 		if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
873 			continue;
874 
875 		if (entry_end(test) <= disk_i_size)
876 			break;
877 		if (test->file_offset >= i_size)
878 			break;
879 
880 		/*
881 		 * We don't update disk_i_size now, so record this undealt
882 		 * i_size. Or we will not know the real i_size.
883 		 */
884 		if (test->outstanding_isize < offset)
885 			test->outstanding_isize = offset;
886 		if (ordered &&
887 		    ordered->outstanding_isize > test->outstanding_isize)
888 			test->outstanding_isize = ordered->outstanding_isize;
889 		goto out;
890 	}
891 	new_i_size = min_t(u64, offset, i_size);
892 
893 	/*
894 	 * Some ordered extents may completed before the current one, and
895 	 * we hold the real i_size in ->outstanding_isize.
896 	 */
897 	if (ordered && ordered->outstanding_isize > new_i_size)
898 		new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
899 	BTRFS_I(inode)->disk_i_size = new_i_size;
900 	ret = 0;
901 out:
902 	/*
903 	 * We need to do this because we can't remove ordered extents until
904 	 * after the i_disk_size has been updated and then the inode has been
905 	 * updated to reflect the change, so we need to tell anybody who finds
906 	 * this ordered extent that we've already done all the real work, we
907 	 * just haven't completed all the other work.
908 	 */
909 	if (ordered)
910 		set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
911 	spin_unlock_irq(&tree->lock);
912 	return ret;
913 }
914 
915 /*
916  * search the ordered extents for one corresponding to 'offset' and
917  * try to find a checksum.  This is used because we allow pages to
918  * be reclaimed before their checksum is actually put into the btree
919  */
920 int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
921 			   u32 *sum, int len)
922 {
923 	struct btrfs_ordered_sum *ordered_sum;
924 	struct btrfs_ordered_extent *ordered;
925 	struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
926 	unsigned long num_sectors;
927 	unsigned long i;
928 	u32 sectorsize = btrfs_inode_sectorsize(inode);
929 	int index = 0;
930 
931 	ordered = btrfs_lookup_ordered_extent(inode, offset);
932 	if (!ordered)
933 		return 0;
934 
935 	spin_lock_irq(&tree->lock);
936 	list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
937 		if (disk_bytenr >= ordered_sum->bytenr &&
938 		    disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
939 			i = (disk_bytenr - ordered_sum->bytenr) >>
940 			    inode->i_sb->s_blocksize_bits;
941 			num_sectors = ordered_sum->len >>
942 				      inode->i_sb->s_blocksize_bits;
943 			num_sectors = min_t(int, len - index, num_sectors - i);
944 			memcpy(sum + index, ordered_sum->sums + i,
945 			       num_sectors);
946 
947 			index += (int)num_sectors;
948 			if (index == len)
949 				goto out;
950 			disk_bytenr += num_sectors * sectorsize;
951 		}
952 	}
953 out:
954 	spin_unlock_irq(&tree->lock);
955 	btrfs_put_ordered_extent(ordered);
956 	return index;
957 }
958 
959 int __init ordered_data_init(void)
960 {
961 	btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
962 				     sizeof(struct btrfs_ordered_extent), 0,
963 				     SLAB_MEM_SPREAD,
964 				     NULL);
965 	if (!btrfs_ordered_extent_cache)
966 		return -ENOMEM;
967 
968 	return 0;
969 }
970 
971 void __cold ordered_data_exit(void)
972 {
973 	kmem_cache_destroy(btrfs_ordered_extent_cache);
974 }
975