xref: /openbmc/linux/fs/btrfs/ordered-data.c (revision dbf563ee)
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 "misc.h"
11 #include "ctree.h"
12 #include "transaction.h"
13 #include "btrfs_inode.h"
14 #include "extent_io.h"
15 #include "disk-io.h"
16 #include "compression.h"
17 #include "delalloc-space.h"
18 #include "qgroup.h"
19 
20 static struct kmem_cache *btrfs_ordered_extent_cache;
21 
22 static u64 entry_end(struct btrfs_ordered_extent *entry)
23 {
24 	if (entry->file_offset + entry->num_bytes < entry->file_offset)
25 		return (u64)-1;
26 	return entry->file_offset + entry->num_bytes;
27 }
28 
29 /* returns NULL if the insertion worked, or it returns the node it did find
30  * in the tree
31  */
32 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
33 				   struct rb_node *node)
34 {
35 	struct rb_node **p = &root->rb_node;
36 	struct rb_node *parent = NULL;
37 	struct btrfs_ordered_extent *entry;
38 
39 	while (*p) {
40 		parent = *p;
41 		entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
42 
43 		if (file_offset < entry->file_offset)
44 			p = &(*p)->rb_left;
45 		else if (file_offset >= entry_end(entry))
46 			p = &(*p)->rb_right;
47 		else
48 			return parent;
49 	}
50 
51 	rb_link_node(node, parent, p);
52 	rb_insert_color(node, root);
53 	return NULL;
54 }
55 
56 /*
57  * look for a given offset in the tree, and if it can't be found return the
58  * first lesser offset
59  */
60 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
61 				     struct rb_node **prev_ret)
62 {
63 	struct rb_node *n = root->rb_node;
64 	struct rb_node *prev = NULL;
65 	struct rb_node *test;
66 	struct btrfs_ordered_extent *entry;
67 	struct btrfs_ordered_extent *prev_entry = NULL;
68 
69 	while (n) {
70 		entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
71 		prev = n;
72 		prev_entry = entry;
73 
74 		if (file_offset < entry->file_offset)
75 			n = n->rb_left;
76 		else if (file_offset >= entry_end(entry))
77 			n = n->rb_right;
78 		else
79 			return n;
80 	}
81 	if (!prev_ret)
82 		return NULL;
83 
84 	while (prev && file_offset >= entry_end(prev_entry)) {
85 		test = rb_next(prev);
86 		if (!test)
87 			break;
88 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
89 				      rb_node);
90 		if (file_offset < entry_end(prev_entry))
91 			break;
92 
93 		prev = test;
94 	}
95 	if (prev)
96 		prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
97 				      rb_node);
98 	while (prev && file_offset < entry_end(prev_entry)) {
99 		test = rb_prev(prev);
100 		if (!test)
101 			break;
102 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
103 				      rb_node);
104 		prev = test;
105 	}
106 	*prev_ret = prev;
107 	return NULL;
108 }
109 
110 /*
111  * helper to check if a given offset is inside a given entry
112  */
113 static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
114 {
115 	if (file_offset < entry->file_offset ||
116 	    entry->file_offset + entry->num_bytes <= file_offset)
117 		return 0;
118 	return 1;
119 }
120 
121 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
122 			  u64 len)
123 {
124 	if (file_offset + len <= entry->file_offset ||
125 	    entry->file_offset + entry->num_bytes <= file_offset)
126 		return 0;
127 	return 1;
128 }
129 
130 /*
131  * look find the first ordered struct that has this offset, otherwise
132  * the first one less than this offset
133  */
134 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
135 					  u64 file_offset)
136 {
137 	struct rb_root *root = &tree->tree;
138 	struct rb_node *prev = NULL;
139 	struct rb_node *ret;
140 	struct btrfs_ordered_extent *entry;
141 
142 	if (tree->last) {
143 		entry = rb_entry(tree->last, struct btrfs_ordered_extent,
144 				 rb_node);
145 		if (offset_in_entry(entry, file_offset))
146 			return tree->last;
147 	}
148 	ret = __tree_search(root, file_offset, &prev);
149 	if (!ret)
150 		ret = prev;
151 	if (ret)
152 		tree->last = ret;
153 	return ret;
154 }
155 
156 /*
157  * Allocate and add a new ordered_extent into the per-inode tree.
158  *
159  * The tree is given a single reference on the ordered extent that was
160  * inserted.
161  */
162 static int __btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
163 				      u64 disk_bytenr, u64 num_bytes,
164 				      u64 disk_num_bytes, int type, int dio,
165 				      int compress_type)
166 {
167 	struct btrfs_root *root = inode->root;
168 	struct btrfs_fs_info *fs_info = root->fs_info;
169 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
170 	struct rb_node *node;
171 	struct btrfs_ordered_extent *entry;
172 	int ret;
173 
174 	if (type == BTRFS_ORDERED_NOCOW || type == BTRFS_ORDERED_PREALLOC) {
175 		/* For nocow write, we can release the qgroup rsv right now */
176 		ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes);
177 		if (ret < 0)
178 			return ret;
179 		ret = 0;
180 	} else {
181 		/*
182 		 * The ordered extent has reserved qgroup space, release now
183 		 * and pass the reserved number for qgroup_record to free.
184 		 */
185 		ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes);
186 		if (ret < 0)
187 			return ret;
188 	}
189 	entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
190 	if (!entry)
191 		return -ENOMEM;
192 
193 	entry->file_offset = file_offset;
194 	entry->disk_bytenr = disk_bytenr;
195 	entry->num_bytes = num_bytes;
196 	entry->disk_num_bytes = disk_num_bytes;
197 	entry->bytes_left = num_bytes;
198 	entry->inode = igrab(&inode->vfs_inode);
199 	entry->compress_type = compress_type;
200 	entry->truncated_len = (u64)-1;
201 	entry->qgroup_rsv = ret;
202 	if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
203 		set_bit(type, &entry->flags);
204 
205 	if (dio) {
206 		percpu_counter_add_batch(&fs_info->dio_bytes, num_bytes,
207 					 fs_info->delalloc_batch);
208 		set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
209 	}
210 
211 	/* one ref for the tree */
212 	refcount_set(&entry->refs, 1);
213 	init_waitqueue_head(&entry->wait);
214 	INIT_LIST_HEAD(&entry->list);
215 	INIT_LIST_HEAD(&entry->root_extent_list);
216 	INIT_LIST_HEAD(&entry->work_list);
217 	init_completion(&entry->completion);
218 
219 	trace_btrfs_ordered_extent_add(&inode->vfs_inode, entry);
220 
221 	spin_lock_irq(&tree->lock);
222 	node = tree_insert(&tree->tree, file_offset,
223 			   &entry->rb_node);
224 	if (node)
225 		btrfs_panic(fs_info, -EEXIST,
226 				"inconsistency in ordered tree at offset %llu",
227 				file_offset);
228 	spin_unlock_irq(&tree->lock);
229 
230 	spin_lock(&root->ordered_extent_lock);
231 	list_add_tail(&entry->root_extent_list,
232 		      &root->ordered_extents);
233 	root->nr_ordered_extents++;
234 	if (root->nr_ordered_extents == 1) {
235 		spin_lock(&fs_info->ordered_root_lock);
236 		BUG_ON(!list_empty(&root->ordered_root));
237 		list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
238 		spin_unlock(&fs_info->ordered_root_lock);
239 	}
240 	spin_unlock(&root->ordered_extent_lock);
241 
242 	/*
243 	 * We don't need the count_max_extents here, we can assume that all of
244 	 * that work has been done at higher layers, so this is truly the
245 	 * smallest the extent is going to get.
246 	 */
247 	spin_lock(&inode->lock);
248 	btrfs_mod_outstanding_extents(inode, 1);
249 	spin_unlock(&inode->lock);
250 
251 	return 0;
252 }
253 
254 int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
255 			     u64 disk_bytenr, u64 num_bytes, u64 disk_num_bytes,
256 			     int type)
257 {
258 	return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
259 					  num_bytes, disk_num_bytes, type, 0,
260 					  BTRFS_COMPRESS_NONE);
261 }
262 
263 int btrfs_add_ordered_extent_dio(struct btrfs_inode *inode, u64 file_offset,
264 				 u64 disk_bytenr, u64 num_bytes,
265 				 u64 disk_num_bytes, int type)
266 {
267 	return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
268 					  num_bytes, disk_num_bytes, type, 1,
269 					  BTRFS_COMPRESS_NONE);
270 }
271 
272 int btrfs_add_ordered_extent_compress(struct btrfs_inode *inode, u64 file_offset,
273 				      u64 disk_bytenr, u64 num_bytes,
274 				      u64 disk_num_bytes, int type,
275 				      int compress_type)
276 {
277 	return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
278 					  num_bytes, disk_num_bytes, type, 0,
279 					  compress_type);
280 }
281 
282 /*
283  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
284  * when an ordered extent is finished.  If the list covers more than one
285  * ordered extent, it is split across multiples.
286  */
287 void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
288 			   struct btrfs_ordered_sum *sum)
289 {
290 	struct btrfs_ordered_inode_tree *tree;
291 
292 	tree = &BTRFS_I(entry->inode)->ordered_tree;
293 	spin_lock_irq(&tree->lock);
294 	list_add_tail(&sum->list, &entry->list);
295 	spin_unlock_irq(&tree->lock);
296 }
297 
298 /*
299  * this is used to account for finished IO across a given range
300  * of the file.  The IO may span ordered extents.  If
301  * a given ordered_extent is completely done, 1 is returned, otherwise
302  * 0.
303  *
304  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
305  * to make sure this function only returns 1 once for a given ordered extent.
306  *
307  * file_offset is updated to one byte past the range that is recorded as
308  * complete.  This allows you to walk forward in the file.
309  */
310 int btrfs_dec_test_first_ordered_pending(struct btrfs_inode *inode,
311 				   struct btrfs_ordered_extent **cached,
312 				   u64 *file_offset, u64 io_size, int uptodate)
313 {
314 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
315 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
316 	struct rb_node *node;
317 	struct btrfs_ordered_extent *entry = NULL;
318 	int ret;
319 	unsigned long flags;
320 	u64 dec_end;
321 	u64 dec_start;
322 	u64 to_dec;
323 
324 	spin_lock_irqsave(&tree->lock, flags);
325 	node = tree_search(tree, *file_offset);
326 	if (!node) {
327 		ret = 1;
328 		goto out;
329 	}
330 
331 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
332 	if (!offset_in_entry(entry, *file_offset)) {
333 		ret = 1;
334 		goto out;
335 	}
336 
337 	dec_start = max(*file_offset, entry->file_offset);
338 	dec_end = min(*file_offset + io_size,
339 		      entry->file_offset + entry->num_bytes);
340 	*file_offset = dec_end;
341 	if (dec_start > dec_end) {
342 		btrfs_crit(fs_info, "bad ordering dec_start %llu end %llu",
343 			   dec_start, dec_end);
344 	}
345 	to_dec = dec_end - dec_start;
346 	if (to_dec > entry->bytes_left) {
347 		btrfs_crit(fs_info,
348 			   "bad ordered accounting left %llu size %llu",
349 			   entry->bytes_left, to_dec);
350 	}
351 	entry->bytes_left -= to_dec;
352 	if (!uptodate)
353 		set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
354 
355 	if (entry->bytes_left == 0) {
356 		ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
357 		/* test_and_set_bit implies a barrier */
358 		cond_wake_up_nomb(&entry->wait);
359 	} else {
360 		ret = 1;
361 	}
362 out:
363 	if (!ret && cached && entry) {
364 		*cached = entry;
365 		refcount_inc(&entry->refs);
366 	}
367 	spin_unlock_irqrestore(&tree->lock, flags);
368 	return ret == 0;
369 }
370 
371 /*
372  * this is used to account for finished IO across a given range
373  * of the file.  The IO should not span ordered extents.  If
374  * a given ordered_extent is completely done, 1 is returned, otherwise
375  * 0.
376  *
377  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
378  * to make sure this function only returns 1 once for a given ordered extent.
379  */
380 int btrfs_dec_test_ordered_pending(struct inode *inode,
381 				   struct btrfs_ordered_extent **cached,
382 				   u64 file_offset, u64 io_size, int uptodate)
383 {
384 	struct btrfs_ordered_inode_tree *tree;
385 	struct rb_node *node;
386 	struct btrfs_ordered_extent *entry = NULL;
387 	unsigned long flags;
388 	int ret;
389 
390 	tree = &BTRFS_I(inode)->ordered_tree;
391 	spin_lock_irqsave(&tree->lock, flags);
392 	if (cached && *cached) {
393 		entry = *cached;
394 		goto have_entry;
395 	}
396 
397 	node = tree_search(tree, file_offset);
398 	if (!node) {
399 		ret = 1;
400 		goto out;
401 	}
402 
403 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
404 have_entry:
405 	if (!offset_in_entry(entry, file_offset)) {
406 		ret = 1;
407 		goto out;
408 	}
409 
410 	if (io_size > entry->bytes_left) {
411 		btrfs_crit(BTRFS_I(inode)->root->fs_info,
412 			   "bad ordered accounting left %llu size %llu",
413 		       entry->bytes_left, io_size);
414 	}
415 	entry->bytes_left -= io_size;
416 	if (!uptodate)
417 		set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
418 
419 	if (entry->bytes_left == 0) {
420 		ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
421 		/* test_and_set_bit implies a barrier */
422 		cond_wake_up_nomb(&entry->wait);
423 	} else {
424 		ret = 1;
425 	}
426 out:
427 	if (!ret && cached && entry) {
428 		*cached = entry;
429 		refcount_inc(&entry->refs);
430 	}
431 	spin_unlock_irqrestore(&tree->lock, flags);
432 	return ret == 0;
433 }
434 
435 /*
436  * used to drop a reference on an ordered extent.  This will free
437  * the extent if the last reference is dropped
438  */
439 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
440 {
441 	struct list_head *cur;
442 	struct btrfs_ordered_sum *sum;
443 
444 	trace_btrfs_ordered_extent_put(entry->inode, entry);
445 
446 	if (refcount_dec_and_test(&entry->refs)) {
447 		ASSERT(list_empty(&entry->root_extent_list));
448 		ASSERT(RB_EMPTY_NODE(&entry->rb_node));
449 		if (entry->inode)
450 			btrfs_add_delayed_iput(entry->inode);
451 		while (!list_empty(&entry->list)) {
452 			cur = entry->list.next;
453 			sum = list_entry(cur, struct btrfs_ordered_sum, list);
454 			list_del(&sum->list);
455 			kvfree(sum);
456 		}
457 		kmem_cache_free(btrfs_ordered_extent_cache, entry);
458 	}
459 }
460 
461 /*
462  * remove an ordered extent from the tree.  No references are dropped
463  * and waiters are woken up.
464  */
465 void btrfs_remove_ordered_extent(struct inode *inode,
466 				 struct btrfs_ordered_extent *entry)
467 {
468 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
469 	struct btrfs_ordered_inode_tree *tree;
470 	struct btrfs_inode *btrfs_inode = BTRFS_I(inode);
471 	struct btrfs_root *root = btrfs_inode->root;
472 	struct rb_node *node;
473 
474 	/* This is paired with btrfs_add_ordered_extent. */
475 	spin_lock(&btrfs_inode->lock);
476 	btrfs_mod_outstanding_extents(btrfs_inode, -1);
477 	spin_unlock(&btrfs_inode->lock);
478 	if (root != fs_info->tree_root)
479 		btrfs_delalloc_release_metadata(btrfs_inode, entry->num_bytes,
480 						false);
481 
482 	if (test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
483 		percpu_counter_add_batch(&fs_info->dio_bytes, -entry->num_bytes,
484 					 fs_info->delalloc_batch);
485 
486 	tree = &btrfs_inode->ordered_tree;
487 	spin_lock_irq(&tree->lock);
488 	node = &entry->rb_node;
489 	rb_erase(node, &tree->tree);
490 	RB_CLEAR_NODE(node);
491 	if (tree->last == node)
492 		tree->last = NULL;
493 	set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
494 	spin_unlock_irq(&tree->lock);
495 
496 	spin_lock(&root->ordered_extent_lock);
497 	list_del_init(&entry->root_extent_list);
498 	root->nr_ordered_extents--;
499 
500 	trace_btrfs_ordered_extent_remove(inode, entry);
501 
502 	if (!root->nr_ordered_extents) {
503 		spin_lock(&fs_info->ordered_root_lock);
504 		BUG_ON(list_empty(&root->ordered_root));
505 		list_del_init(&root->ordered_root);
506 		spin_unlock(&fs_info->ordered_root_lock);
507 	}
508 	spin_unlock(&root->ordered_extent_lock);
509 	wake_up(&entry->wait);
510 }
511 
512 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
513 {
514 	struct btrfs_ordered_extent *ordered;
515 
516 	ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
517 	btrfs_start_ordered_extent(ordered->inode, ordered, 1);
518 	complete(&ordered->completion);
519 }
520 
521 /*
522  * wait for all the ordered extents in a root.  This is done when balancing
523  * space between drives.
524  */
525 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
526 			       const u64 range_start, const u64 range_len)
527 {
528 	struct btrfs_fs_info *fs_info = root->fs_info;
529 	LIST_HEAD(splice);
530 	LIST_HEAD(skipped);
531 	LIST_HEAD(works);
532 	struct btrfs_ordered_extent *ordered, *next;
533 	u64 count = 0;
534 	const u64 range_end = range_start + range_len;
535 
536 	mutex_lock(&root->ordered_extent_mutex);
537 	spin_lock(&root->ordered_extent_lock);
538 	list_splice_init(&root->ordered_extents, &splice);
539 	while (!list_empty(&splice) && nr) {
540 		ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
541 					   root_extent_list);
542 
543 		if (range_end <= ordered->disk_bytenr ||
544 		    ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
545 			list_move_tail(&ordered->root_extent_list, &skipped);
546 			cond_resched_lock(&root->ordered_extent_lock);
547 			continue;
548 		}
549 
550 		list_move_tail(&ordered->root_extent_list,
551 			       &root->ordered_extents);
552 		refcount_inc(&ordered->refs);
553 		spin_unlock(&root->ordered_extent_lock);
554 
555 		btrfs_init_work(&ordered->flush_work,
556 				btrfs_run_ordered_extent_work, NULL, NULL);
557 		list_add_tail(&ordered->work_list, &works);
558 		btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
559 
560 		cond_resched();
561 		spin_lock(&root->ordered_extent_lock);
562 		if (nr != U64_MAX)
563 			nr--;
564 		count++;
565 	}
566 	list_splice_tail(&skipped, &root->ordered_extents);
567 	list_splice_tail(&splice, &root->ordered_extents);
568 	spin_unlock(&root->ordered_extent_lock);
569 
570 	list_for_each_entry_safe(ordered, next, &works, work_list) {
571 		list_del_init(&ordered->work_list);
572 		wait_for_completion(&ordered->completion);
573 		btrfs_put_ordered_extent(ordered);
574 		cond_resched();
575 	}
576 	mutex_unlock(&root->ordered_extent_mutex);
577 
578 	return count;
579 }
580 
581 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
582 			     const u64 range_start, const u64 range_len)
583 {
584 	struct btrfs_root *root;
585 	struct list_head splice;
586 	u64 done;
587 
588 	INIT_LIST_HEAD(&splice);
589 
590 	mutex_lock(&fs_info->ordered_operations_mutex);
591 	spin_lock(&fs_info->ordered_root_lock);
592 	list_splice_init(&fs_info->ordered_roots, &splice);
593 	while (!list_empty(&splice) && nr) {
594 		root = list_first_entry(&splice, struct btrfs_root,
595 					ordered_root);
596 		root = btrfs_grab_root(root);
597 		BUG_ON(!root);
598 		list_move_tail(&root->ordered_root,
599 			       &fs_info->ordered_roots);
600 		spin_unlock(&fs_info->ordered_root_lock);
601 
602 		done = btrfs_wait_ordered_extents(root, nr,
603 						  range_start, range_len);
604 		btrfs_put_root(root);
605 
606 		spin_lock(&fs_info->ordered_root_lock);
607 		if (nr != U64_MAX) {
608 			nr -= done;
609 		}
610 	}
611 	list_splice_tail(&splice, &fs_info->ordered_roots);
612 	spin_unlock(&fs_info->ordered_root_lock);
613 	mutex_unlock(&fs_info->ordered_operations_mutex);
614 }
615 
616 /*
617  * Used to start IO or wait for a given ordered extent to finish.
618  *
619  * If wait is one, this effectively waits on page writeback for all the pages
620  * in the extent, and it waits on the io completion code to insert
621  * metadata into the btree corresponding to the extent
622  */
623 void btrfs_start_ordered_extent(struct inode *inode,
624 				       struct btrfs_ordered_extent *entry,
625 				       int wait)
626 {
627 	u64 start = entry->file_offset;
628 	u64 end = start + entry->num_bytes - 1;
629 
630 	trace_btrfs_ordered_extent_start(inode, entry);
631 
632 	/*
633 	 * pages in the range can be dirty, clean or writeback.  We
634 	 * start IO on any dirty ones so the wait doesn't stall waiting
635 	 * for the flusher thread to find them
636 	 */
637 	if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
638 		filemap_fdatawrite_range(inode->i_mapping, start, end);
639 	if (wait) {
640 		wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
641 						 &entry->flags));
642 	}
643 }
644 
645 /*
646  * Used to wait on ordered extents across a large range of bytes.
647  */
648 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
649 {
650 	int ret = 0;
651 	int ret_wb = 0;
652 	u64 end;
653 	u64 orig_end;
654 	struct btrfs_ordered_extent *ordered;
655 
656 	if (start + len < start) {
657 		orig_end = INT_LIMIT(loff_t);
658 	} else {
659 		orig_end = start + len - 1;
660 		if (orig_end > INT_LIMIT(loff_t))
661 			orig_end = INT_LIMIT(loff_t);
662 	}
663 
664 	/* start IO across the range first to instantiate any delalloc
665 	 * extents
666 	 */
667 	ret = btrfs_fdatawrite_range(inode, start, orig_end);
668 	if (ret)
669 		return ret;
670 
671 	/*
672 	 * If we have a writeback error don't return immediately. Wait first
673 	 * for any ordered extents that haven't completed yet. This is to make
674 	 * sure no one can dirty the same page ranges and call writepages()
675 	 * before the ordered extents complete - to avoid failures (-EEXIST)
676 	 * when adding the new ordered extents to the ordered tree.
677 	 */
678 	ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
679 
680 	end = orig_end;
681 	while (1) {
682 		ordered = btrfs_lookup_first_ordered_extent(inode, end);
683 		if (!ordered)
684 			break;
685 		if (ordered->file_offset > orig_end) {
686 			btrfs_put_ordered_extent(ordered);
687 			break;
688 		}
689 		if (ordered->file_offset + ordered->num_bytes <= start) {
690 			btrfs_put_ordered_extent(ordered);
691 			break;
692 		}
693 		btrfs_start_ordered_extent(inode, ordered, 1);
694 		end = ordered->file_offset;
695 		/*
696 		 * If the ordered extent had an error save the error but don't
697 		 * exit without waiting first for all other ordered extents in
698 		 * the range to complete.
699 		 */
700 		if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
701 			ret = -EIO;
702 		btrfs_put_ordered_extent(ordered);
703 		if (end == 0 || end == start)
704 			break;
705 		end--;
706 	}
707 	return ret_wb ? ret_wb : ret;
708 }
709 
710 /*
711  * find an ordered extent corresponding to file_offset.  return NULL if
712  * nothing is found, otherwise take a reference on the extent and return it
713  */
714 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
715 							 u64 file_offset)
716 {
717 	struct btrfs_ordered_inode_tree *tree;
718 	struct rb_node *node;
719 	struct btrfs_ordered_extent *entry = NULL;
720 
721 	tree = &inode->ordered_tree;
722 	spin_lock_irq(&tree->lock);
723 	node = tree_search(tree, file_offset);
724 	if (!node)
725 		goto out;
726 
727 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
728 	if (!offset_in_entry(entry, file_offset))
729 		entry = NULL;
730 	if (entry)
731 		refcount_inc(&entry->refs);
732 out:
733 	spin_unlock_irq(&tree->lock);
734 	return entry;
735 }
736 
737 /* Since the DIO code tries to lock a wide area we need to look for any ordered
738  * extents that exist in the range, rather than just the start of the range.
739  */
740 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
741 		struct btrfs_inode *inode, u64 file_offset, u64 len)
742 {
743 	struct btrfs_ordered_inode_tree *tree;
744 	struct rb_node *node;
745 	struct btrfs_ordered_extent *entry = NULL;
746 
747 	tree = &inode->ordered_tree;
748 	spin_lock_irq(&tree->lock);
749 	node = tree_search(tree, file_offset);
750 	if (!node) {
751 		node = tree_search(tree, file_offset + len);
752 		if (!node)
753 			goto out;
754 	}
755 
756 	while (1) {
757 		entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
758 		if (range_overlaps(entry, file_offset, len))
759 			break;
760 
761 		if (entry->file_offset >= file_offset + len) {
762 			entry = NULL;
763 			break;
764 		}
765 		entry = NULL;
766 		node = rb_next(node);
767 		if (!node)
768 			break;
769 	}
770 out:
771 	if (entry)
772 		refcount_inc(&entry->refs);
773 	spin_unlock_irq(&tree->lock);
774 	return entry;
775 }
776 
777 /*
778  * lookup and return any extent before 'file_offset'.  NULL is returned
779  * if none is found
780  */
781 struct btrfs_ordered_extent *
782 btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
783 {
784 	struct btrfs_ordered_inode_tree *tree;
785 	struct rb_node *node;
786 	struct btrfs_ordered_extent *entry = NULL;
787 
788 	tree = &BTRFS_I(inode)->ordered_tree;
789 	spin_lock_irq(&tree->lock);
790 	node = tree_search(tree, file_offset);
791 	if (!node)
792 		goto out;
793 
794 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
795 	refcount_inc(&entry->refs);
796 out:
797 	spin_unlock_irq(&tree->lock);
798 	return entry;
799 }
800 
801 /*
802  * search the ordered extents for one corresponding to 'offset' and
803  * try to find a checksum.  This is used because we allow pages to
804  * be reclaimed before their checksum is actually put into the btree
805  */
806 int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
807 			   u8 *sum, int len)
808 {
809 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
810 	struct btrfs_ordered_sum *ordered_sum;
811 	struct btrfs_ordered_extent *ordered;
812 	struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
813 	unsigned long num_sectors;
814 	unsigned long i;
815 	u32 sectorsize = btrfs_inode_sectorsize(inode);
816 	const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
817 	int index = 0;
818 
819 	ordered = btrfs_lookup_ordered_extent(BTRFS_I(inode), offset);
820 	if (!ordered)
821 		return 0;
822 
823 	spin_lock_irq(&tree->lock);
824 	list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
825 		if (disk_bytenr >= ordered_sum->bytenr &&
826 		    disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
827 			i = (disk_bytenr - ordered_sum->bytenr) >>
828 			    inode->i_sb->s_blocksize_bits;
829 			num_sectors = ordered_sum->len >>
830 				      inode->i_sb->s_blocksize_bits;
831 			num_sectors = min_t(int, len - index, num_sectors - i);
832 			memcpy(sum + index, ordered_sum->sums + i * csum_size,
833 			       num_sectors * csum_size);
834 
835 			index += (int)num_sectors * csum_size;
836 			if (index == len)
837 				goto out;
838 			disk_bytenr += num_sectors * sectorsize;
839 		}
840 	}
841 out:
842 	spin_unlock_irq(&tree->lock);
843 	btrfs_put_ordered_extent(ordered);
844 	return index;
845 }
846 
847 /*
848  * btrfs_flush_ordered_range - Lock the passed range and ensures all pending
849  * ordered extents in it are run to completion.
850  *
851  * @inode:        Inode whose ordered tree is to be searched
852  * @start:        Beginning of range to flush
853  * @end:          Last byte of range to lock
854  * @cached_state: If passed, will return the extent state responsible for the
855  * locked range. It's the caller's responsibility to free the cached state.
856  *
857  * This function always returns with the given range locked, ensuring after it's
858  * called no order extent can be pending.
859  */
860 void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
861 					u64 end,
862 					struct extent_state **cached_state)
863 {
864 	struct btrfs_ordered_extent *ordered;
865 	struct extent_state *cache = NULL;
866 	struct extent_state **cachedp = &cache;
867 
868 	if (cached_state)
869 		cachedp = cached_state;
870 
871 	while (1) {
872 		lock_extent_bits(&inode->io_tree, start, end, cachedp);
873 		ordered = btrfs_lookup_ordered_range(inode, start,
874 						     end - start + 1);
875 		if (!ordered) {
876 			/*
877 			 * If no external cached_state has been passed then
878 			 * decrement the extra ref taken for cachedp since we
879 			 * aren't exposing it outside of this function
880 			 */
881 			if (!cached_state)
882 				refcount_dec(&cache->refs);
883 			break;
884 		}
885 		unlock_extent_cached(&inode->io_tree, start, end, cachedp);
886 		btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
887 		btrfs_put_ordered_extent(ordered);
888 	}
889 }
890 
891 int __init ordered_data_init(void)
892 {
893 	btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
894 				     sizeof(struct btrfs_ordered_extent), 0,
895 				     SLAB_MEM_SPREAD,
896 				     NULL);
897 	if (!btrfs_ordered_extent_cache)
898 		return -ENOMEM;
899 
900 	return 0;
901 }
902 
903 void __cold ordered_data_exit(void)
904 {
905 	kmem_cache_destroy(btrfs_ordered_extent_cache);
906 }
907