xref: /openbmc/linux/fs/btrfs/ordered-data.c (revision abcda807)
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->log_list);
216 	INIT_LIST_HEAD(&entry->root_extent_list);
217 	INIT_LIST_HEAD(&entry->work_list);
218 	init_completion(&entry->completion);
219 
220 	trace_btrfs_ordered_extent_add(inode, entry);
221 
222 	spin_lock_irq(&tree->lock);
223 	node = tree_insert(&tree->tree, file_offset,
224 			   &entry->rb_node);
225 	if (node)
226 		btrfs_panic(fs_info, -EEXIST,
227 				"inconsistency in ordered tree at offset %llu",
228 				file_offset);
229 	spin_unlock_irq(&tree->lock);
230 
231 	spin_lock(&root->ordered_extent_lock);
232 	list_add_tail(&entry->root_extent_list,
233 		      &root->ordered_extents);
234 	root->nr_ordered_extents++;
235 	if (root->nr_ordered_extents == 1) {
236 		spin_lock(&fs_info->ordered_root_lock);
237 		BUG_ON(!list_empty(&root->ordered_root));
238 		list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
239 		spin_unlock(&fs_info->ordered_root_lock);
240 	}
241 	spin_unlock(&root->ordered_extent_lock);
242 
243 	/*
244 	 * We don't need the count_max_extents here, we can assume that all of
245 	 * that work has been done at higher layers, so this is truly the
246 	 * smallest the extent is going to get.
247 	 */
248 	spin_lock(&inode->lock);
249 	btrfs_mod_outstanding_extents(inode, 1);
250 	spin_unlock(&inode->lock);
251 
252 	return 0;
253 }
254 
255 int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
256 			     u64 disk_bytenr, u64 num_bytes, u64 disk_num_bytes,
257 			     int type)
258 {
259 	return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
260 					  num_bytes, disk_num_bytes, type, 0,
261 					  BTRFS_COMPRESS_NONE);
262 }
263 
264 int btrfs_add_ordered_extent_dio(struct btrfs_inode *inode, u64 file_offset,
265 				 u64 disk_bytenr, u64 num_bytes,
266 				 u64 disk_num_bytes, int type)
267 {
268 	return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
269 					  num_bytes, disk_num_bytes, type, 1,
270 					  BTRFS_COMPRESS_NONE);
271 }
272 
273 int btrfs_add_ordered_extent_compress(struct btrfs_inode *inode, u64 file_offset,
274 				      u64 disk_bytenr, u64 num_bytes,
275 				      u64 disk_num_bytes, int type,
276 				      int compress_type)
277 {
278 	return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
279 					  num_bytes, disk_num_bytes, type, 0,
280 					  compress_type);
281 }
282 
283 /*
284  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
285  * when an ordered extent is finished.  If the list covers more than one
286  * ordered extent, it is split across multiples.
287  */
288 void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
289 			   struct btrfs_ordered_sum *sum)
290 {
291 	struct btrfs_ordered_inode_tree *tree;
292 
293 	tree = &BTRFS_I(entry->inode)->ordered_tree;
294 	spin_lock_irq(&tree->lock);
295 	list_add_tail(&sum->list, &entry->list);
296 	spin_unlock_irq(&tree->lock);
297 }
298 
299 /*
300  * this is used to account for finished IO across a given range
301  * of the file.  The IO may span ordered extents.  If
302  * a given ordered_extent is completely done, 1 is returned, otherwise
303  * 0.
304  *
305  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
306  * to make sure this function only returns 1 once for a given ordered extent.
307  *
308  * file_offset is updated to one byte past the range that is recorded as
309  * complete.  This allows you to walk forward in the file.
310  */
311 int btrfs_dec_test_first_ordered_pending(struct btrfs_inode *inode,
312 				   struct btrfs_ordered_extent **cached,
313 				   u64 *file_offset, u64 io_size, int uptodate)
314 {
315 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
316 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
317 	struct rb_node *node;
318 	struct btrfs_ordered_extent *entry = NULL;
319 	int ret;
320 	unsigned long flags;
321 	u64 dec_end;
322 	u64 dec_start;
323 	u64 to_dec;
324 
325 	spin_lock_irqsave(&tree->lock, flags);
326 	node = tree_search(tree, *file_offset);
327 	if (!node) {
328 		ret = 1;
329 		goto out;
330 	}
331 
332 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
333 	if (!offset_in_entry(entry, *file_offset)) {
334 		ret = 1;
335 		goto out;
336 	}
337 
338 	dec_start = max(*file_offset, entry->file_offset);
339 	dec_end = min(*file_offset + io_size,
340 		      entry->file_offset + entry->num_bytes);
341 	*file_offset = dec_end;
342 	if (dec_start > dec_end) {
343 		btrfs_crit(fs_info, "bad ordering dec_start %llu end %llu",
344 			   dec_start, dec_end);
345 	}
346 	to_dec = dec_end - dec_start;
347 	if (to_dec > entry->bytes_left) {
348 		btrfs_crit(fs_info,
349 			   "bad ordered accounting left %llu size %llu",
350 			   entry->bytes_left, to_dec);
351 	}
352 	entry->bytes_left -= to_dec;
353 	if (!uptodate)
354 		set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
355 
356 	if (entry->bytes_left == 0) {
357 		ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
358 		/* test_and_set_bit implies a barrier */
359 		cond_wake_up_nomb(&entry->wait);
360 	} else {
361 		ret = 1;
362 	}
363 out:
364 	if (!ret && cached && entry) {
365 		*cached = entry;
366 		refcount_inc(&entry->refs);
367 	}
368 	spin_unlock_irqrestore(&tree->lock, flags);
369 	return ret == 0;
370 }
371 
372 /*
373  * this is used to account for finished IO across a given range
374  * of the file.  The IO should not span ordered extents.  If
375  * a given ordered_extent is completely done, 1 is returned, otherwise
376  * 0.
377  *
378  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
379  * to make sure this function only returns 1 once for a given ordered extent.
380  */
381 int btrfs_dec_test_ordered_pending(struct btrfs_inode *inode,
382 				   struct btrfs_ordered_extent **cached,
383 				   u64 file_offset, u64 io_size, int uptodate)
384 {
385 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
386 	struct rb_node *node;
387 	struct btrfs_ordered_extent *entry = NULL;
388 	unsigned long flags;
389 	int ret;
390 
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(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(BTRFS_I(entry->inode), entry);
445 
446 	if (refcount_dec_and_test(&entry->refs)) {
447 		ASSERT(list_empty(&entry->root_extent_list));
448 		ASSERT(list_empty(&entry->log_list));
449 		ASSERT(RB_EMPTY_NODE(&entry->rb_node));
450 		if (entry->inode)
451 			btrfs_add_delayed_iput(entry->inode);
452 		while (!list_empty(&entry->list)) {
453 			cur = entry->list.next;
454 			sum = list_entry(cur, struct btrfs_ordered_sum, list);
455 			list_del(&sum->list);
456 			kvfree(sum);
457 		}
458 		kmem_cache_free(btrfs_ordered_extent_cache, entry);
459 	}
460 }
461 
462 /*
463  * remove an ordered extent from the tree.  No references are dropped
464  * and waiters are woken up.
465  */
466 void btrfs_remove_ordered_extent(struct btrfs_inode *btrfs_inode,
467 				 struct btrfs_ordered_extent *entry)
468 {
469 	struct btrfs_ordered_inode_tree *tree;
470 	struct btrfs_root *root = btrfs_inode->root;
471 	struct btrfs_fs_info *fs_info = root->fs_info;
472 	struct rb_node *node;
473 	bool pending;
474 
475 	/* This is paired with btrfs_add_ordered_extent. */
476 	spin_lock(&btrfs_inode->lock);
477 	btrfs_mod_outstanding_extents(btrfs_inode, -1);
478 	spin_unlock(&btrfs_inode->lock);
479 	if (root != fs_info->tree_root)
480 		btrfs_delalloc_release_metadata(btrfs_inode, entry->num_bytes,
481 						false);
482 
483 	if (test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
484 		percpu_counter_add_batch(&fs_info->dio_bytes, -entry->num_bytes,
485 					 fs_info->delalloc_batch);
486 
487 	tree = &btrfs_inode->ordered_tree;
488 	spin_lock_irq(&tree->lock);
489 	node = &entry->rb_node;
490 	rb_erase(node, &tree->tree);
491 	RB_CLEAR_NODE(node);
492 	if (tree->last == node)
493 		tree->last = NULL;
494 	set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
495 	pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
496 	spin_unlock_irq(&tree->lock);
497 
498 	/*
499 	 * The current running transaction is waiting on us, we need to let it
500 	 * know that we're complete and wake it up.
501 	 */
502 	if (pending) {
503 		struct btrfs_transaction *trans;
504 
505 		/*
506 		 * The checks for trans are just a formality, it should be set,
507 		 * but if it isn't we don't want to deref/assert under the spin
508 		 * lock, so be nice and check if trans is set, but ASSERT() so
509 		 * if it isn't set a developer will notice.
510 		 */
511 		spin_lock(&fs_info->trans_lock);
512 		trans = fs_info->running_transaction;
513 		if (trans)
514 			refcount_inc(&trans->use_count);
515 		spin_unlock(&fs_info->trans_lock);
516 
517 		ASSERT(trans);
518 		if (trans) {
519 			if (atomic_dec_and_test(&trans->pending_ordered))
520 				wake_up(&trans->pending_wait);
521 			btrfs_put_transaction(trans);
522 		}
523 	}
524 
525 	spin_lock(&root->ordered_extent_lock);
526 	list_del_init(&entry->root_extent_list);
527 	root->nr_ordered_extents--;
528 
529 	trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
530 
531 	if (!root->nr_ordered_extents) {
532 		spin_lock(&fs_info->ordered_root_lock);
533 		BUG_ON(list_empty(&root->ordered_root));
534 		list_del_init(&root->ordered_root);
535 		spin_unlock(&fs_info->ordered_root_lock);
536 	}
537 	spin_unlock(&root->ordered_extent_lock);
538 	wake_up(&entry->wait);
539 }
540 
541 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
542 {
543 	struct btrfs_ordered_extent *ordered;
544 
545 	ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
546 	btrfs_start_ordered_extent(ordered, 1);
547 	complete(&ordered->completion);
548 }
549 
550 /*
551  * wait for all the ordered extents in a root.  This is done when balancing
552  * space between drives.
553  */
554 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
555 			       const u64 range_start, const u64 range_len)
556 {
557 	struct btrfs_fs_info *fs_info = root->fs_info;
558 	LIST_HEAD(splice);
559 	LIST_HEAD(skipped);
560 	LIST_HEAD(works);
561 	struct btrfs_ordered_extent *ordered, *next;
562 	u64 count = 0;
563 	const u64 range_end = range_start + range_len;
564 
565 	mutex_lock(&root->ordered_extent_mutex);
566 	spin_lock(&root->ordered_extent_lock);
567 	list_splice_init(&root->ordered_extents, &splice);
568 	while (!list_empty(&splice) && nr) {
569 		ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
570 					   root_extent_list);
571 
572 		if (range_end <= ordered->disk_bytenr ||
573 		    ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
574 			list_move_tail(&ordered->root_extent_list, &skipped);
575 			cond_resched_lock(&root->ordered_extent_lock);
576 			continue;
577 		}
578 
579 		list_move_tail(&ordered->root_extent_list,
580 			       &root->ordered_extents);
581 		refcount_inc(&ordered->refs);
582 		spin_unlock(&root->ordered_extent_lock);
583 
584 		btrfs_init_work(&ordered->flush_work,
585 				btrfs_run_ordered_extent_work, NULL, NULL);
586 		list_add_tail(&ordered->work_list, &works);
587 		btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
588 
589 		cond_resched();
590 		spin_lock(&root->ordered_extent_lock);
591 		if (nr != U64_MAX)
592 			nr--;
593 		count++;
594 	}
595 	list_splice_tail(&skipped, &root->ordered_extents);
596 	list_splice_tail(&splice, &root->ordered_extents);
597 	spin_unlock(&root->ordered_extent_lock);
598 
599 	list_for_each_entry_safe(ordered, next, &works, work_list) {
600 		list_del_init(&ordered->work_list);
601 		wait_for_completion(&ordered->completion);
602 		btrfs_put_ordered_extent(ordered);
603 		cond_resched();
604 	}
605 	mutex_unlock(&root->ordered_extent_mutex);
606 
607 	return count;
608 }
609 
610 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
611 			     const u64 range_start, const u64 range_len)
612 {
613 	struct btrfs_root *root;
614 	struct list_head splice;
615 	u64 done;
616 
617 	INIT_LIST_HEAD(&splice);
618 
619 	mutex_lock(&fs_info->ordered_operations_mutex);
620 	spin_lock(&fs_info->ordered_root_lock);
621 	list_splice_init(&fs_info->ordered_roots, &splice);
622 	while (!list_empty(&splice) && nr) {
623 		root = list_first_entry(&splice, struct btrfs_root,
624 					ordered_root);
625 		root = btrfs_grab_root(root);
626 		BUG_ON(!root);
627 		list_move_tail(&root->ordered_root,
628 			       &fs_info->ordered_roots);
629 		spin_unlock(&fs_info->ordered_root_lock);
630 
631 		done = btrfs_wait_ordered_extents(root, nr,
632 						  range_start, range_len);
633 		btrfs_put_root(root);
634 
635 		spin_lock(&fs_info->ordered_root_lock);
636 		if (nr != U64_MAX) {
637 			nr -= done;
638 		}
639 	}
640 	list_splice_tail(&splice, &fs_info->ordered_roots);
641 	spin_unlock(&fs_info->ordered_root_lock);
642 	mutex_unlock(&fs_info->ordered_operations_mutex);
643 }
644 
645 /*
646  * Used to start IO or wait for a given ordered extent to finish.
647  *
648  * If wait is one, this effectively waits on page writeback for all the pages
649  * in the extent, and it waits on the io completion code to insert
650  * metadata into the btree corresponding to the extent
651  */
652 void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry, int wait)
653 {
654 	u64 start = entry->file_offset;
655 	u64 end = start + entry->num_bytes - 1;
656 	struct btrfs_inode *inode = BTRFS_I(entry->inode);
657 
658 	trace_btrfs_ordered_extent_start(inode, entry);
659 
660 	/*
661 	 * pages in the range can be dirty, clean or writeback.  We
662 	 * start IO on any dirty ones so the wait doesn't stall waiting
663 	 * for the flusher thread to find them
664 	 */
665 	if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
666 		filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
667 	if (wait) {
668 		wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
669 						 &entry->flags));
670 	}
671 }
672 
673 /*
674  * Used to wait on ordered extents across a large range of bytes.
675  */
676 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
677 {
678 	int ret = 0;
679 	int ret_wb = 0;
680 	u64 end;
681 	u64 orig_end;
682 	struct btrfs_ordered_extent *ordered;
683 
684 	if (start + len < start) {
685 		orig_end = INT_LIMIT(loff_t);
686 	} else {
687 		orig_end = start + len - 1;
688 		if (orig_end > INT_LIMIT(loff_t))
689 			orig_end = INT_LIMIT(loff_t);
690 	}
691 
692 	/* start IO across the range first to instantiate any delalloc
693 	 * extents
694 	 */
695 	ret = btrfs_fdatawrite_range(inode, start, orig_end);
696 	if (ret)
697 		return ret;
698 
699 	/*
700 	 * If we have a writeback error don't return immediately. Wait first
701 	 * for any ordered extents that haven't completed yet. This is to make
702 	 * sure no one can dirty the same page ranges and call writepages()
703 	 * before the ordered extents complete - to avoid failures (-EEXIST)
704 	 * when adding the new ordered extents to the ordered tree.
705 	 */
706 	ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
707 
708 	end = orig_end;
709 	while (1) {
710 		ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode), end);
711 		if (!ordered)
712 			break;
713 		if (ordered->file_offset > orig_end) {
714 			btrfs_put_ordered_extent(ordered);
715 			break;
716 		}
717 		if (ordered->file_offset + ordered->num_bytes <= start) {
718 			btrfs_put_ordered_extent(ordered);
719 			break;
720 		}
721 		btrfs_start_ordered_extent(ordered, 1);
722 		end = ordered->file_offset;
723 		/*
724 		 * If the ordered extent had an error save the error but don't
725 		 * exit without waiting first for all other ordered extents in
726 		 * the range to complete.
727 		 */
728 		if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
729 			ret = -EIO;
730 		btrfs_put_ordered_extent(ordered);
731 		if (end == 0 || end == start)
732 			break;
733 		end--;
734 	}
735 	return ret_wb ? ret_wb : ret;
736 }
737 
738 /*
739  * find an ordered extent corresponding to file_offset.  return NULL if
740  * nothing is found, otherwise take a reference on the extent and return it
741  */
742 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
743 							 u64 file_offset)
744 {
745 	struct btrfs_ordered_inode_tree *tree;
746 	struct rb_node *node;
747 	struct btrfs_ordered_extent *entry = NULL;
748 
749 	tree = &inode->ordered_tree;
750 	spin_lock_irq(&tree->lock);
751 	node = tree_search(tree, file_offset);
752 	if (!node)
753 		goto out;
754 
755 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
756 	if (!offset_in_entry(entry, file_offset))
757 		entry = NULL;
758 	if (entry)
759 		refcount_inc(&entry->refs);
760 out:
761 	spin_unlock_irq(&tree->lock);
762 	return entry;
763 }
764 
765 /* Since the DIO code tries to lock a wide area we need to look for any ordered
766  * extents that exist in the range, rather than just the start of the range.
767  */
768 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
769 		struct btrfs_inode *inode, u64 file_offset, u64 len)
770 {
771 	struct btrfs_ordered_inode_tree *tree;
772 	struct rb_node *node;
773 	struct btrfs_ordered_extent *entry = NULL;
774 
775 	tree = &inode->ordered_tree;
776 	spin_lock_irq(&tree->lock);
777 	node = tree_search(tree, file_offset);
778 	if (!node) {
779 		node = tree_search(tree, file_offset + len);
780 		if (!node)
781 			goto out;
782 	}
783 
784 	while (1) {
785 		entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
786 		if (range_overlaps(entry, file_offset, len))
787 			break;
788 
789 		if (entry->file_offset >= file_offset + len) {
790 			entry = NULL;
791 			break;
792 		}
793 		entry = NULL;
794 		node = rb_next(node);
795 		if (!node)
796 			break;
797 	}
798 out:
799 	if (entry)
800 		refcount_inc(&entry->refs);
801 	spin_unlock_irq(&tree->lock);
802 	return entry;
803 }
804 
805 /*
806  * Adds all ordered extents to the given list. The list ends up sorted by the
807  * file_offset of the ordered extents.
808  */
809 void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
810 					   struct list_head *list)
811 {
812 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
813 	struct rb_node *n;
814 
815 	ASSERT(inode_is_locked(&inode->vfs_inode));
816 
817 	spin_lock_irq(&tree->lock);
818 	for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
819 		struct btrfs_ordered_extent *ordered;
820 
821 		ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
822 
823 		if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
824 			continue;
825 
826 		ASSERT(list_empty(&ordered->log_list));
827 		list_add_tail(&ordered->log_list, list);
828 		refcount_inc(&ordered->refs);
829 	}
830 	spin_unlock_irq(&tree->lock);
831 }
832 
833 /*
834  * lookup and return any extent before 'file_offset'.  NULL is returned
835  * if none is found
836  */
837 struct btrfs_ordered_extent *
838 btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
839 {
840 	struct btrfs_ordered_inode_tree *tree;
841 	struct rb_node *node;
842 	struct btrfs_ordered_extent *entry = NULL;
843 
844 	tree = &inode->ordered_tree;
845 	spin_lock_irq(&tree->lock);
846 	node = tree_search(tree, file_offset);
847 	if (!node)
848 		goto out;
849 
850 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
851 	refcount_inc(&entry->refs);
852 out:
853 	spin_unlock_irq(&tree->lock);
854 	return entry;
855 }
856 
857 /*
858  * search the ordered extents for one corresponding to 'offset' and
859  * try to find a checksum.  This is used because we allow pages to
860  * be reclaimed before their checksum is actually put into the btree
861  */
862 int btrfs_find_ordered_sum(struct btrfs_inode *inode, u64 offset,
863 			   u64 disk_bytenr, u8 *sum, int len)
864 {
865 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
866 	struct btrfs_ordered_sum *ordered_sum;
867 	struct btrfs_ordered_extent *ordered;
868 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
869 	unsigned long num_sectors;
870 	unsigned long i;
871 	u32 sectorsize = btrfs_inode_sectorsize(inode);
872 	const u8 blocksize_bits = inode->vfs_inode.i_sb->s_blocksize_bits;
873 	const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
874 	int index = 0;
875 
876 	ordered = btrfs_lookup_ordered_extent(inode, offset);
877 	if (!ordered)
878 		return 0;
879 
880 	spin_lock_irq(&tree->lock);
881 	list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
882 		if (disk_bytenr >= ordered_sum->bytenr &&
883 		    disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
884 			i = (disk_bytenr - ordered_sum->bytenr) >> blocksize_bits;
885 			num_sectors = ordered_sum->len >> blocksize_bits;
886 			num_sectors = min_t(int, len - index, num_sectors - i);
887 			memcpy(sum + index, ordered_sum->sums + i * csum_size,
888 			       num_sectors * csum_size);
889 
890 			index += (int)num_sectors * csum_size;
891 			if (index == len)
892 				goto out;
893 			disk_bytenr += num_sectors * sectorsize;
894 		}
895 	}
896 out:
897 	spin_unlock_irq(&tree->lock);
898 	btrfs_put_ordered_extent(ordered);
899 	return index;
900 }
901 
902 /*
903  * btrfs_flush_ordered_range - Lock the passed range and ensures all pending
904  * ordered extents in it are run to completion.
905  *
906  * @inode:        Inode whose ordered tree is to be searched
907  * @start:        Beginning of range to flush
908  * @end:          Last byte of range to lock
909  * @cached_state: If passed, will return the extent state responsible for the
910  * locked range. It's the caller's responsibility to free the cached state.
911  *
912  * This function always returns with the given range locked, ensuring after it's
913  * called no order extent can be pending.
914  */
915 void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
916 					u64 end,
917 					struct extent_state **cached_state)
918 {
919 	struct btrfs_ordered_extent *ordered;
920 	struct extent_state *cache = NULL;
921 	struct extent_state **cachedp = &cache;
922 
923 	if (cached_state)
924 		cachedp = cached_state;
925 
926 	while (1) {
927 		lock_extent_bits(&inode->io_tree, start, end, cachedp);
928 		ordered = btrfs_lookup_ordered_range(inode, start,
929 						     end - start + 1);
930 		if (!ordered) {
931 			/*
932 			 * If no external cached_state has been passed then
933 			 * decrement the extra ref taken for cachedp since we
934 			 * aren't exposing it outside of this function
935 			 */
936 			if (!cached_state)
937 				refcount_dec(&cache->refs);
938 			break;
939 		}
940 		unlock_extent_cached(&inode->io_tree, start, end, cachedp);
941 		btrfs_start_ordered_extent(ordered, 1);
942 		btrfs_put_ordered_extent(ordered);
943 	}
944 }
945 
946 int __init ordered_data_init(void)
947 {
948 	btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
949 				     sizeof(struct btrfs_ordered_extent), 0,
950 				     SLAB_MEM_SPREAD,
951 				     NULL);
952 	if (!btrfs_ordered_extent_cache)
953 		return -ENOMEM;
954 
955 	return 0;
956 }
957 
958 void __cold ordered_data_exit(void)
959 {
960 	kmem_cache_destroy(btrfs_ordered_extent_cache);
961 }
962