xref: /openbmc/linux/fs/btrfs/ordered-data.c (revision 9ad685db)
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 "messages.h"
11 #include "misc.h"
12 #include "ctree.h"
13 #include "transaction.h"
14 #include "btrfs_inode.h"
15 #include "extent_io.h"
16 #include "disk-io.h"
17 #include "compression.h"
18 #include "delalloc-space.h"
19 #include "qgroup.h"
20 #include "subpage.h"
21 #include "file.h"
22 #include "super.h"
23 
24 static struct kmem_cache *btrfs_ordered_extent_cache;
25 
26 static u64 entry_end(struct btrfs_ordered_extent *entry)
27 {
28 	if (entry->file_offset + entry->num_bytes < entry->file_offset)
29 		return (u64)-1;
30 	return entry->file_offset + entry->num_bytes;
31 }
32 
33 /* returns NULL if the insertion worked, or it returns the node it did find
34  * in the tree
35  */
36 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
37 				   struct rb_node *node)
38 {
39 	struct rb_node **p = &root->rb_node;
40 	struct rb_node *parent = NULL;
41 	struct btrfs_ordered_extent *entry;
42 
43 	while (*p) {
44 		parent = *p;
45 		entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
46 
47 		if (file_offset < entry->file_offset)
48 			p = &(*p)->rb_left;
49 		else if (file_offset >= entry_end(entry))
50 			p = &(*p)->rb_right;
51 		else
52 			return parent;
53 	}
54 
55 	rb_link_node(node, parent, p);
56 	rb_insert_color(node, root);
57 	return NULL;
58 }
59 
60 /*
61  * look for a given offset in the tree, and if it can't be found return the
62  * first lesser offset
63  */
64 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
65 				     struct rb_node **prev_ret)
66 {
67 	struct rb_node *n = root->rb_node;
68 	struct rb_node *prev = NULL;
69 	struct rb_node *test;
70 	struct btrfs_ordered_extent *entry;
71 	struct btrfs_ordered_extent *prev_entry = NULL;
72 
73 	while (n) {
74 		entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
75 		prev = n;
76 		prev_entry = entry;
77 
78 		if (file_offset < entry->file_offset)
79 			n = n->rb_left;
80 		else if (file_offset >= entry_end(entry))
81 			n = n->rb_right;
82 		else
83 			return n;
84 	}
85 	if (!prev_ret)
86 		return NULL;
87 
88 	while (prev && file_offset >= entry_end(prev_entry)) {
89 		test = rb_next(prev);
90 		if (!test)
91 			break;
92 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
93 				      rb_node);
94 		if (file_offset < entry_end(prev_entry))
95 			break;
96 
97 		prev = test;
98 	}
99 	if (prev)
100 		prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
101 				      rb_node);
102 	while (prev && file_offset < entry_end(prev_entry)) {
103 		test = rb_prev(prev);
104 		if (!test)
105 			break;
106 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
107 				      rb_node);
108 		prev = test;
109 	}
110 	*prev_ret = prev;
111 	return NULL;
112 }
113 
114 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
115 			  u64 len)
116 {
117 	if (file_offset + len <= entry->file_offset ||
118 	    entry->file_offset + entry->num_bytes <= file_offset)
119 		return 0;
120 	return 1;
121 }
122 
123 /*
124  * look find the first ordered struct that has this offset, otherwise
125  * the first one less than this offset
126  */
127 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
128 					  u64 file_offset)
129 {
130 	struct rb_root *root = &tree->tree;
131 	struct rb_node *prev = NULL;
132 	struct rb_node *ret;
133 	struct btrfs_ordered_extent *entry;
134 
135 	if (tree->last) {
136 		entry = rb_entry(tree->last, struct btrfs_ordered_extent,
137 				 rb_node);
138 		if (in_range(file_offset, entry->file_offset, entry->num_bytes))
139 			return tree->last;
140 	}
141 	ret = __tree_search(root, file_offset, &prev);
142 	if (!ret)
143 		ret = prev;
144 	if (ret)
145 		tree->last = ret;
146 	return ret;
147 }
148 
149 /*
150  * Add an ordered extent to the per-inode tree.
151  *
152  * @inode:           Inode that this extent is for.
153  * @file_offset:     Logical offset in file where the extent starts.
154  * @num_bytes:       Logical length of extent in file.
155  * @ram_bytes:       Full length of unencoded data.
156  * @disk_bytenr:     Offset of extent on disk.
157  * @disk_num_bytes:  Size of extent on disk.
158  * @offset:          Offset into unencoded data where file data starts.
159  * @flags:           Flags specifying type of extent (1 << BTRFS_ORDERED_*).
160  * @compress_type:   Compression algorithm used for data.
161  *
162  * Most of these parameters correspond to &struct btrfs_file_extent_item. The
163  * tree is given a single reference on the ordered extent that was inserted, and
164  * the returned pointer is given a second reference.
165  *
166  * Return: the new ordered extent or error pointer.
167  */
168 struct btrfs_ordered_extent *btrfs_alloc_ordered_extent(
169 			struct btrfs_inode *inode, u64 file_offset,
170 			u64 num_bytes, u64 ram_bytes, u64 disk_bytenr,
171 			u64 disk_num_bytes, u64 offset, unsigned long flags,
172 			int compress_type)
173 {
174 	struct btrfs_root *root = inode->root;
175 	struct btrfs_fs_info *fs_info = root->fs_info;
176 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
177 	struct rb_node *node;
178 	struct btrfs_ordered_extent *entry;
179 	int ret;
180 
181 	if (flags &
182 	    ((1 << BTRFS_ORDERED_NOCOW) | (1 << BTRFS_ORDERED_PREALLOC))) {
183 		/* For nocow write, we can release the qgroup rsv right now */
184 		ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes);
185 		if (ret < 0)
186 			return ERR_PTR(ret);
187 		ret = 0;
188 	} else {
189 		/*
190 		 * The ordered extent has reserved qgroup space, release now
191 		 * and pass the reserved number for qgroup_record to free.
192 		 */
193 		ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes);
194 		if (ret < 0)
195 			return ERR_PTR(ret);
196 	}
197 	entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
198 	if (!entry)
199 		return ERR_PTR(-ENOMEM);
200 
201 	entry->file_offset = file_offset;
202 	entry->num_bytes = num_bytes;
203 	entry->ram_bytes = ram_bytes;
204 	entry->disk_bytenr = disk_bytenr;
205 	entry->disk_num_bytes = disk_num_bytes;
206 	entry->offset = offset;
207 	entry->bytes_left = num_bytes;
208 	entry->inode = igrab(&inode->vfs_inode);
209 	entry->compress_type = compress_type;
210 	entry->truncated_len = (u64)-1;
211 	entry->qgroup_rsv = ret;
212 	entry->physical = (u64)-1;
213 
214 	ASSERT((flags & ~BTRFS_ORDERED_TYPE_FLAGS) == 0);
215 	entry->flags = flags;
216 
217 	percpu_counter_add_batch(&fs_info->ordered_bytes, num_bytes,
218 				 fs_info->delalloc_batch);
219 
220 	/* one ref for the tree */
221 	refcount_set(&entry->refs, 1);
222 	init_waitqueue_head(&entry->wait);
223 	INIT_LIST_HEAD(&entry->list);
224 	INIT_LIST_HEAD(&entry->log_list);
225 	INIT_LIST_HEAD(&entry->root_extent_list);
226 	INIT_LIST_HEAD(&entry->work_list);
227 	init_completion(&entry->completion);
228 
229 	trace_btrfs_ordered_extent_add(inode, entry);
230 
231 	spin_lock_irq(&tree->lock);
232 	node = tree_insert(&tree->tree, file_offset,
233 			   &entry->rb_node);
234 	if (node)
235 		btrfs_panic(fs_info, -EEXIST,
236 				"inconsistency in ordered tree at offset %llu",
237 				file_offset);
238 	spin_unlock_irq(&tree->lock);
239 
240 	spin_lock(&root->ordered_extent_lock);
241 	list_add_tail(&entry->root_extent_list,
242 		      &root->ordered_extents);
243 	root->nr_ordered_extents++;
244 	if (root->nr_ordered_extents == 1) {
245 		spin_lock(&fs_info->ordered_root_lock);
246 		BUG_ON(!list_empty(&root->ordered_root));
247 		list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
248 		spin_unlock(&fs_info->ordered_root_lock);
249 	}
250 	spin_unlock(&root->ordered_extent_lock);
251 
252 	/*
253 	 * We don't need the count_max_extents here, we can assume that all of
254 	 * that work has been done at higher layers, so this is truly the
255 	 * smallest the extent is going to get.
256 	 */
257 	spin_lock(&inode->lock);
258 	btrfs_mod_outstanding_extents(inode, 1);
259 	spin_unlock(&inode->lock);
260 
261 	/* One ref for the returned entry to match semantics of lookup. */
262 	refcount_inc(&entry->refs);
263 
264 	return entry;
265 }
266 
267 /*
268  * Add a new btrfs_ordered_extent for the range, but drop the reference instead
269  * of returning it to the caller.
270  */
271 int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
272 			     u64 num_bytes, u64 ram_bytes, u64 disk_bytenr,
273 			     u64 disk_num_bytes, u64 offset, unsigned long flags,
274 			     int compress_type)
275 {
276 	struct btrfs_ordered_extent *ordered;
277 
278 	ordered = btrfs_alloc_ordered_extent(inode, file_offset, num_bytes,
279 					     ram_bytes, disk_bytenr,
280 					     disk_num_bytes, offset, flags,
281 					     compress_type);
282 
283 	if (IS_ERR(ordered))
284 		return PTR_ERR(ordered);
285 	btrfs_put_ordered_extent(ordered);
286 
287 	return 0;
288 }
289 
290 /*
291  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
292  * when an ordered extent is finished.  If the list covers more than one
293  * ordered extent, it is split across multiples.
294  */
295 void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
296 			   struct btrfs_ordered_sum *sum)
297 {
298 	struct btrfs_ordered_inode_tree *tree;
299 
300 	tree = &BTRFS_I(entry->inode)->ordered_tree;
301 	spin_lock_irq(&tree->lock);
302 	list_add_tail(&sum->list, &entry->list);
303 	spin_unlock_irq(&tree->lock);
304 }
305 
306 static void finish_ordered_fn(struct btrfs_work *work)
307 {
308 	struct btrfs_ordered_extent *ordered_extent;
309 
310 	ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
311 	btrfs_finish_ordered_io(ordered_extent);
312 }
313 
314 /*
315  * Mark all ordered extents io inside the specified range finished.
316  *
317  * @page:	 The involved page for the operation.
318  *		 For uncompressed buffered IO, the page status also needs to be
319  *		 updated to indicate whether the pending ordered io is finished.
320  *		 Can be NULL for direct IO and compressed write.
321  *		 For these cases, callers are ensured they won't execute the
322  *		 endio function twice.
323  *
324  * This function is called for endio, thus the range must have ordered
325  * extent(s) covering it.
326  */
327 void btrfs_mark_ordered_io_finished(struct btrfs_inode *inode,
328 				    struct page *page, u64 file_offset,
329 				    u64 num_bytes, bool uptodate)
330 {
331 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
332 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
333 	struct btrfs_workqueue *wq;
334 	struct rb_node *node;
335 	struct btrfs_ordered_extent *entry = NULL;
336 	unsigned long flags;
337 	u64 cur = file_offset;
338 
339 	if (btrfs_is_free_space_inode(inode))
340 		wq = fs_info->endio_freespace_worker;
341 	else
342 		wq = fs_info->endio_write_workers;
343 
344 	if (page)
345 		ASSERT(page->mapping && page_offset(page) <= file_offset &&
346 		       file_offset + num_bytes <= page_offset(page) + PAGE_SIZE);
347 
348 	spin_lock_irqsave(&tree->lock, flags);
349 	while (cur < file_offset + num_bytes) {
350 		u64 entry_end;
351 		u64 end;
352 		u32 len;
353 
354 		node = tree_search(tree, cur);
355 		/* No ordered extents at all */
356 		if (!node)
357 			break;
358 
359 		entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
360 		entry_end = entry->file_offset + entry->num_bytes;
361 		/*
362 		 * |<-- OE --->|  |
363 		 *		  cur
364 		 * Go to next OE.
365 		 */
366 		if (cur >= entry_end) {
367 			node = rb_next(node);
368 			/* No more ordered extents, exit */
369 			if (!node)
370 				break;
371 			entry = rb_entry(node, struct btrfs_ordered_extent,
372 					 rb_node);
373 
374 			/* Go to next ordered extent and continue */
375 			cur = entry->file_offset;
376 			continue;
377 		}
378 		/*
379 		 * |	|<--- OE --->|
380 		 * cur
381 		 * Go to the start of OE.
382 		 */
383 		if (cur < entry->file_offset) {
384 			cur = entry->file_offset;
385 			continue;
386 		}
387 
388 		/*
389 		 * Now we are definitely inside one ordered extent.
390 		 *
391 		 * |<--- OE --->|
392 		 *	|
393 		 *	cur
394 		 */
395 		end = min(entry->file_offset + entry->num_bytes,
396 			  file_offset + num_bytes) - 1;
397 		ASSERT(end + 1 - cur < U32_MAX);
398 		len = end + 1 - cur;
399 
400 		if (page) {
401 			/*
402 			 * Ordered (Private2) bit indicates whether we still
403 			 * have pending io unfinished for the ordered extent.
404 			 *
405 			 * If there's no such bit, we need to skip to next range.
406 			 */
407 			if (!btrfs_page_test_ordered(fs_info, page, cur, len)) {
408 				cur += len;
409 				continue;
410 			}
411 			btrfs_page_clear_ordered(fs_info, page, cur, len);
412 		}
413 
414 		/* Now we're fine to update the accounting */
415 		if (unlikely(len > entry->bytes_left)) {
416 			WARN_ON(1);
417 			btrfs_crit(fs_info,
418 "bad ordered extent accounting, root=%llu ino=%llu OE offset=%llu OE len=%llu to_dec=%u left=%llu",
419 				   inode->root->root_key.objectid,
420 				   btrfs_ino(inode),
421 				   entry->file_offset,
422 				   entry->num_bytes,
423 				   len, entry->bytes_left);
424 			entry->bytes_left = 0;
425 		} else {
426 			entry->bytes_left -= len;
427 		}
428 
429 		if (!uptodate)
430 			set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
431 
432 		/*
433 		 * All the IO of the ordered extent is finished, we need to queue
434 		 * the finish_func to be executed.
435 		 */
436 		if (entry->bytes_left == 0) {
437 			set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
438 			cond_wake_up(&entry->wait);
439 			refcount_inc(&entry->refs);
440 			trace_btrfs_ordered_extent_mark_finished(inode, entry);
441 			spin_unlock_irqrestore(&tree->lock, flags);
442 			btrfs_init_work(&entry->work, finish_ordered_fn, NULL, NULL);
443 			btrfs_queue_work(wq, &entry->work);
444 			spin_lock_irqsave(&tree->lock, flags);
445 		}
446 		cur += len;
447 	}
448 	spin_unlock_irqrestore(&tree->lock, flags);
449 }
450 
451 /*
452  * Finish IO for one ordered extent across a given range.  The range can only
453  * contain one ordered extent.
454  *
455  * @cached:	 The cached ordered extent. If not NULL, we can skip the tree
456  *               search and use the ordered extent directly.
457  * 		 Will be also used to store the finished ordered extent.
458  * @file_offset: File offset for the finished IO
459  * @io_size:	 Length of the finish IO range
460  *
461  * Return true if the ordered extent is finished in the range, and update
462  * @cached.
463  * Return false otherwise.
464  *
465  * NOTE: The range can NOT cross multiple ordered extents.
466  * Thus caller should ensure the range doesn't cross ordered extents.
467  */
468 bool btrfs_dec_test_ordered_pending(struct btrfs_inode *inode,
469 				    struct btrfs_ordered_extent **cached,
470 				    u64 file_offset, u64 io_size)
471 {
472 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
473 	struct rb_node *node;
474 	struct btrfs_ordered_extent *entry = NULL;
475 	unsigned long flags;
476 	bool finished = false;
477 
478 	spin_lock_irqsave(&tree->lock, flags);
479 	if (cached && *cached) {
480 		entry = *cached;
481 		goto have_entry;
482 	}
483 
484 	node = tree_search(tree, file_offset);
485 	if (!node)
486 		goto out;
487 
488 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
489 have_entry:
490 	if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
491 		goto out;
492 
493 	if (io_size > entry->bytes_left)
494 		btrfs_crit(inode->root->fs_info,
495 			   "bad ordered accounting left %llu size %llu",
496 		       entry->bytes_left, io_size);
497 
498 	entry->bytes_left -= io_size;
499 
500 	if (entry->bytes_left == 0) {
501 		/*
502 		 * Ensure only one caller can set the flag and finished_ret
503 		 * accordingly
504 		 */
505 		finished = !test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
506 		/* test_and_set_bit implies a barrier */
507 		cond_wake_up_nomb(&entry->wait);
508 	}
509 out:
510 	if (finished && cached && entry) {
511 		*cached = entry;
512 		refcount_inc(&entry->refs);
513 		trace_btrfs_ordered_extent_dec_test_pending(inode, entry);
514 	}
515 	spin_unlock_irqrestore(&tree->lock, flags);
516 	return finished;
517 }
518 
519 /*
520  * used to drop a reference on an ordered extent.  This will free
521  * the extent if the last reference is dropped
522  */
523 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
524 {
525 	struct list_head *cur;
526 	struct btrfs_ordered_sum *sum;
527 
528 	trace_btrfs_ordered_extent_put(BTRFS_I(entry->inode), entry);
529 
530 	if (refcount_dec_and_test(&entry->refs)) {
531 		ASSERT(list_empty(&entry->root_extent_list));
532 		ASSERT(list_empty(&entry->log_list));
533 		ASSERT(RB_EMPTY_NODE(&entry->rb_node));
534 		if (entry->inode)
535 			btrfs_add_delayed_iput(BTRFS_I(entry->inode));
536 		while (!list_empty(&entry->list)) {
537 			cur = entry->list.next;
538 			sum = list_entry(cur, struct btrfs_ordered_sum, list);
539 			list_del(&sum->list);
540 			kvfree(sum);
541 		}
542 		kmem_cache_free(btrfs_ordered_extent_cache, entry);
543 	}
544 }
545 
546 /*
547  * remove an ordered extent from the tree.  No references are dropped
548  * and waiters are woken up.
549  */
550 void btrfs_remove_ordered_extent(struct btrfs_inode *btrfs_inode,
551 				 struct btrfs_ordered_extent *entry)
552 {
553 	struct btrfs_ordered_inode_tree *tree;
554 	struct btrfs_root *root = btrfs_inode->root;
555 	struct btrfs_fs_info *fs_info = root->fs_info;
556 	struct rb_node *node;
557 	bool pending;
558 	bool freespace_inode;
559 
560 	/*
561 	 * If this is a free space inode the thread has not acquired the ordered
562 	 * extents lockdep map.
563 	 */
564 	freespace_inode = btrfs_is_free_space_inode(btrfs_inode);
565 
566 	btrfs_lockdep_acquire(fs_info, btrfs_trans_pending_ordered);
567 	/* This is paired with btrfs_add_ordered_extent. */
568 	spin_lock(&btrfs_inode->lock);
569 	btrfs_mod_outstanding_extents(btrfs_inode, -1);
570 	spin_unlock(&btrfs_inode->lock);
571 	if (root != fs_info->tree_root) {
572 		u64 release;
573 
574 		if (test_bit(BTRFS_ORDERED_ENCODED, &entry->flags))
575 			release = entry->disk_num_bytes;
576 		else
577 			release = entry->num_bytes;
578 		btrfs_delalloc_release_metadata(btrfs_inode, release, false);
579 	}
580 
581 	percpu_counter_add_batch(&fs_info->ordered_bytes, -entry->num_bytes,
582 				 fs_info->delalloc_batch);
583 
584 	tree = &btrfs_inode->ordered_tree;
585 	spin_lock_irq(&tree->lock);
586 	node = &entry->rb_node;
587 	rb_erase(node, &tree->tree);
588 	RB_CLEAR_NODE(node);
589 	if (tree->last == node)
590 		tree->last = NULL;
591 	set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
592 	pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
593 	spin_unlock_irq(&tree->lock);
594 
595 	/*
596 	 * The current running transaction is waiting on us, we need to let it
597 	 * know that we're complete and wake it up.
598 	 */
599 	if (pending) {
600 		struct btrfs_transaction *trans;
601 
602 		/*
603 		 * The checks for trans are just a formality, it should be set,
604 		 * but if it isn't we don't want to deref/assert under the spin
605 		 * lock, so be nice and check if trans is set, but ASSERT() so
606 		 * if it isn't set a developer will notice.
607 		 */
608 		spin_lock(&fs_info->trans_lock);
609 		trans = fs_info->running_transaction;
610 		if (trans)
611 			refcount_inc(&trans->use_count);
612 		spin_unlock(&fs_info->trans_lock);
613 
614 		ASSERT(trans);
615 		if (trans) {
616 			if (atomic_dec_and_test(&trans->pending_ordered))
617 				wake_up(&trans->pending_wait);
618 			btrfs_put_transaction(trans);
619 		}
620 	}
621 
622 	btrfs_lockdep_release(fs_info, btrfs_trans_pending_ordered);
623 
624 	spin_lock(&root->ordered_extent_lock);
625 	list_del_init(&entry->root_extent_list);
626 	root->nr_ordered_extents--;
627 
628 	trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
629 
630 	if (!root->nr_ordered_extents) {
631 		spin_lock(&fs_info->ordered_root_lock);
632 		BUG_ON(list_empty(&root->ordered_root));
633 		list_del_init(&root->ordered_root);
634 		spin_unlock(&fs_info->ordered_root_lock);
635 	}
636 	spin_unlock(&root->ordered_extent_lock);
637 	wake_up(&entry->wait);
638 	if (!freespace_inode)
639 		btrfs_lockdep_release(fs_info, btrfs_ordered_extent);
640 }
641 
642 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
643 {
644 	struct btrfs_ordered_extent *ordered;
645 
646 	ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
647 	btrfs_start_ordered_extent(ordered);
648 	complete(&ordered->completion);
649 }
650 
651 /*
652  * wait for all the ordered extents in a root.  This is done when balancing
653  * space between drives.
654  */
655 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
656 			       const u64 range_start, const u64 range_len)
657 {
658 	struct btrfs_fs_info *fs_info = root->fs_info;
659 	LIST_HEAD(splice);
660 	LIST_HEAD(skipped);
661 	LIST_HEAD(works);
662 	struct btrfs_ordered_extent *ordered, *next;
663 	u64 count = 0;
664 	const u64 range_end = range_start + range_len;
665 
666 	mutex_lock(&root->ordered_extent_mutex);
667 	spin_lock(&root->ordered_extent_lock);
668 	list_splice_init(&root->ordered_extents, &splice);
669 	while (!list_empty(&splice) && nr) {
670 		ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
671 					   root_extent_list);
672 
673 		if (range_end <= ordered->disk_bytenr ||
674 		    ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
675 			list_move_tail(&ordered->root_extent_list, &skipped);
676 			cond_resched_lock(&root->ordered_extent_lock);
677 			continue;
678 		}
679 
680 		list_move_tail(&ordered->root_extent_list,
681 			       &root->ordered_extents);
682 		refcount_inc(&ordered->refs);
683 		spin_unlock(&root->ordered_extent_lock);
684 
685 		btrfs_init_work(&ordered->flush_work,
686 				btrfs_run_ordered_extent_work, NULL, NULL);
687 		list_add_tail(&ordered->work_list, &works);
688 		btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
689 
690 		cond_resched();
691 		spin_lock(&root->ordered_extent_lock);
692 		if (nr != U64_MAX)
693 			nr--;
694 		count++;
695 	}
696 	list_splice_tail(&skipped, &root->ordered_extents);
697 	list_splice_tail(&splice, &root->ordered_extents);
698 	spin_unlock(&root->ordered_extent_lock);
699 
700 	list_for_each_entry_safe(ordered, next, &works, work_list) {
701 		list_del_init(&ordered->work_list);
702 		wait_for_completion(&ordered->completion);
703 		btrfs_put_ordered_extent(ordered);
704 		cond_resched();
705 	}
706 	mutex_unlock(&root->ordered_extent_mutex);
707 
708 	return count;
709 }
710 
711 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
712 			     const u64 range_start, const u64 range_len)
713 {
714 	struct btrfs_root *root;
715 	struct list_head splice;
716 	u64 done;
717 
718 	INIT_LIST_HEAD(&splice);
719 
720 	mutex_lock(&fs_info->ordered_operations_mutex);
721 	spin_lock(&fs_info->ordered_root_lock);
722 	list_splice_init(&fs_info->ordered_roots, &splice);
723 	while (!list_empty(&splice) && nr) {
724 		root = list_first_entry(&splice, struct btrfs_root,
725 					ordered_root);
726 		root = btrfs_grab_root(root);
727 		BUG_ON(!root);
728 		list_move_tail(&root->ordered_root,
729 			       &fs_info->ordered_roots);
730 		spin_unlock(&fs_info->ordered_root_lock);
731 
732 		done = btrfs_wait_ordered_extents(root, nr,
733 						  range_start, range_len);
734 		btrfs_put_root(root);
735 
736 		spin_lock(&fs_info->ordered_root_lock);
737 		if (nr != U64_MAX) {
738 			nr -= done;
739 		}
740 	}
741 	list_splice_tail(&splice, &fs_info->ordered_roots);
742 	spin_unlock(&fs_info->ordered_root_lock);
743 	mutex_unlock(&fs_info->ordered_operations_mutex);
744 }
745 
746 /*
747  * Start IO and wait for a given ordered extent to finish.
748  *
749  * Wait on page writeback for all the pages in the extent and the IO completion
750  * code to insert metadata into the btree corresponding to the extent.
751  */
752 void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry)
753 {
754 	u64 start = entry->file_offset;
755 	u64 end = start + entry->num_bytes - 1;
756 	struct btrfs_inode *inode = BTRFS_I(entry->inode);
757 	bool freespace_inode;
758 
759 	trace_btrfs_ordered_extent_start(inode, entry);
760 
761 	/*
762 	 * If this is a free space inode do not take the ordered extents lockdep
763 	 * map.
764 	 */
765 	freespace_inode = btrfs_is_free_space_inode(inode);
766 
767 	/*
768 	 * pages in the range can be dirty, clean or writeback.  We
769 	 * start IO on any dirty ones so the wait doesn't stall waiting
770 	 * for the flusher thread to find them
771 	 */
772 	if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
773 		filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
774 
775 	if (!freespace_inode)
776 		btrfs_might_wait_for_event(inode->root->fs_info, btrfs_ordered_extent);
777 	wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE, &entry->flags));
778 }
779 
780 /*
781  * Used to wait on ordered extents across a large range of bytes.
782  */
783 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
784 {
785 	int ret = 0;
786 	int ret_wb = 0;
787 	u64 end;
788 	u64 orig_end;
789 	struct btrfs_ordered_extent *ordered;
790 
791 	if (start + len < start) {
792 		orig_end = OFFSET_MAX;
793 	} else {
794 		orig_end = start + len - 1;
795 		if (orig_end > OFFSET_MAX)
796 			orig_end = OFFSET_MAX;
797 	}
798 
799 	/* start IO across the range first to instantiate any delalloc
800 	 * extents
801 	 */
802 	ret = btrfs_fdatawrite_range(inode, start, orig_end);
803 	if (ret)
804 		return ret;
805 
806 	/*
807 	 * If we have a writeback error don't return immediately. Wait first
808 	 * for any ordered extents that haven't completed yet. This is to make
809 	 * sure no one can dirty the same page ranges and call writepages()
810 	 * before the ordered extents complete - to avoid failures (-EEXIST)
811 	 * when adding the new ordered extents to the ordered tree.
812 	 */
813 	ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
814 
815 	end = orig_end;
816 	while (1) {
817 		ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode), end);
818 		if (!ordered)
819 			break;
820 		if (ordered->file_offset > orig_end) {
821 			btrfs_put_ordered_extent(ordered);
822 			break;
823 		}
824 		if (ordered->file_offset + ordered->num_bytes <= start) {
825 			btrfs_put_ordered_extent(ordered);
826 			break;
827 		}
828 		btrfs_start_ordered_extent(ordered);
829 		end = ordered->file_offset;
830 		/*
831 		 * If the ordered extent had an error save the error but don't
832 		 * exit without waiting first for all other ordered extents in
833 		 * the range to complete.
834 		 */
835 		if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
836 			ret = -EIO;
837 		btrfs_put_ordered_extent(ordered);
838 		if (end == 0 || end == start)
839 			break;
840 		end--;
841 	}
842 	return ret_wb ? ret_wb : ret;
843 }
844 
845 /*
846  * find an ordered extent corresponding to file_offset.  return NULL if
847  * nothing is found, otherwise take a reference on the extent and return it
848  */
849 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
850 							 u64 file_offset)
851 {
852 	struct btrfs_ordered_inode_tree *tree;
853 	struct rb_node *node;
854 	struct btrfs_ordered_extent *entry = NULL;
855 	unsigned long flags;
856 
857 	tree = &inode->ordered_tree;
858 	spin_lock_irqsave(&tree->lock, flags);
859 	node = tree_search(tree, file_offset);
860 	if (!node)
861 		goto out;
862 
863 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
864 	if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
865 		entry = NULL;
866 	if (entry) {
867 		refcount_inc(&entry->refs);
868 		trace_btrfs_ordered_extent_lookup(inode, entry);
869 	}
870 out:
871 	spin_unlock_irqrestore(&tree->lock, flags);
872 	return entry;
873 }
874 
875 /* Since the DIO code tries to lock a wide area we need to look for any ordered
876  * extents that exist in the range, rather than just the start of the range.
877  */
878 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
879 		struct btrfs_inode *inode, u64 file_offset, u64 len)
880 {
881 	struct btrfs_ordered_inode_tree *tree;
882 	struct rb_node *node;
883 	struct btrfs_ordered_extent *entry = NULL;
884 
885 	tree = &inode->ordered_tree;
886 	spin_lock_irq(&tree->lock);
887 	node = tree_search(tree, file_offset);
888 	if (!node) {
889 		node = tree_search(tree, file_offset + len);
890 		if (!node)
891 			goto out;
892 	}
893 
894 	while (1) {
895 		entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
896 		if (range_overlaps(entry, file_offset, len))
897 			break;
898 
899 		if (entry->file_offset >= file_offset + len) {
900 			entry = NULL;
901 			break;
902 		}
903 		entry = NULL;
904 		node = rb_next(node);
905 		if (!node)
906 			break;
907 	}
908 out:
909 	if (entry) {
910 		refcount_inc(&entry->refs);
911 		trace_btrfs_ordered_extent_lookup_range(inode, entry);
912 	}
913 	spin_unlock_irq(&tree->lock);
914 	return entry;
915 }
916 
917 /*
918  * Adds all ordered extents to the given list. The list ends up sorted by the
919  * file_offset of the ordered extents.
920  */
921 void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
922 					   struct list_head *list)
923 {
924 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
925 	struct rb_node *n;
926 
927 	ASSERT(inode_is_locked(&inode->vfs_inode));
928 
929 	spin_lock_irq(&tree->lock);
930 	for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
931 		struct btrfs_ordered_extent *ordered;
932 
933 		ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
934 
935 		if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
936 			continue;
937 
938 		ASSERT(list_empty(&ordered->log_list));
939 		list_add_tail(&ordered->log_list, list);
940 		refcount_inc(&ordered->refs);
941 		trace_btrfs_ordered_extent_lookup_for_logging(inode, ordered);
942 	}
943 	spin_unlock_irq(&tree->lock);
944 }
945 
946 /*
947  * lookup and return any extent before 'file_offset'.  NULL is returned
948  * if none is found
949  */
950 struct btrfs_ordered_extent *
951 btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
952 {
953 	struct btrfs_ordered_inode_tree *tree;
954 	struct rb_node *node;
955 	struct btrfs_ordered_extent *entry = NULL;
956 
957 	tree = &inode->ordered_tree;
958 	spin_lock_irq(&tree->lock);
959 	node = tree_search(tree, file_offset);
960 	if (!node)
961 		goto out;
962 
963 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
964 	refcount_inc(&entry->refs);
965 	trace_btrfs_ordered_extent_lookup_first(inode, entry);
966 out:
967 	spin_unlock_irq(&tree->lock);
968 	return entry;
969 }
970 
971 /*
972  * Lookup the first ordered extent that overlaps the range
973  * [@file_offset, @file_offset + @len).
974  *
975  * The difference between this and btrfs_lookup_first_ordered_extent() is
976  * that this one won't return any ordered extent that does not overlap the range.
977  * And the difference against btrfs_lookup_ordered_extent() is, this function
978  * ensures the first ordered extent gets returned.
979  */
980 struct btrfs_ordered_extent *btrfs_lookup_first_ordered_range(
981 			struct btrfs_inode *inode, u64 file_offset, u64 len)
982 {
983 	struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
984 	struct rb_node *node;
985 	struct rb_node *cur;
986 	struct rb_node *prev;
987 	struct rb_node *next;
988 	struct btrfs_ordered_extent *entry = NULL;
989 
990 	spin_lock_irq(&tree->lock);
991 	node = tree->tree.rb_node;
992 	/*
993 	 * Here we don't want to use tree_search() which will use tree->last
994 	 * and screw up the search order.
995 	 * And __tree_search() can't return the adjacent ordered extents
996 	 * either, thus here we do our own search.
997 	 */
998 	while (node) {
999 		entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1000 
1001 		if (file_offset < entry->file_offset) {
1002 			node = node->rb_left;
1003 		} else if (file_offset >= entry_end(entry)) {
1004 			node = node->rb_right;
1005 		} else {
1006 			/*
1007 			 * Direct hit, got an ordered extent that starts at
1008 			 * @file_offset
1009 			 */
1010 			goto out;
1011 		}
1012 	}
1013 	if (!entry) {
1014 		/* Empty tree */
1015 		goto out;
1016 	}
1017 
1018 	cur = &entry->rb_node;
1019 	/* We got an entry around @file_offset, check adjacent entries */
1020 	if (entry->file_offset < file_offset) {
1021 		prev = cur;
1022 		next = rb_next(cur);
1023 	} else {
1024 		prev = rb_prev(cur);
1025 		next = cur;
1026 	}
1027 	if (prev) {
1028 		entry = rb_entry(prev, struct btrfs_ordered_extent, rb_node);
1029 		if (range_overlaps(entry, file_offset, len))
1030 			goto out;
1031 	}
1032 	if (next) {
1033 		entry = rb_entry(next, struct btrfs_ordered_extent, rb_node);
1034 		if (range_overlaps(entry, file_offset, len))
1035 			goto out;
1036 	}
1037 	/* No ordered extent in the range */
1038 	entry = NULL;
1039 out:
1040 	if (entry) {
1041 		refcount_inc(&entry->refs);
1042 		trace_btrfs_ordered_extent_lookup_first_range(inode, entry);
1043 	}
1044 
1045 	spin_unlock_irq(&tree->lock);
1046 	return entry;
1047 }
1048 
1049 /*
1050  * Lock the passed range and ensures all pending ordered extents in it are run
1051  * to completion.
1052  *
1053  * @inode:        Inode whose ordered tree is to be searched
1054  * @start:        Beginning of range to flush
1055  * @end:          Last byte of range to lock
1056  * @cached_state: If passed, will return the extent state responsible for the
1057  *                locked range. It's the caller's responsibility to free the
1058  *                cached state.
1059  *
1060  * Always return with the given range locked, ensuring after it's called no
1061  * order extent can be pending.
1062  */
1063 void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
1064 					u64 end,
1065 					struct extent_state **cached_state)
1066 {
1067 	struct btrfs_ordered_extent *ordered;
1068 	struct extent_state *cache = NULL;
1069 	struct extent_state **cachedp = &cache;
1070 
1071 	if (cached_state)
1072 		cachedp = cached_state;
1073 
1074 	while (1) {
1075 		lock_extent(&inode->io_tree, start, end, cachedp);
1076 		ordered = btrfs_lookup_ordered_range(inode, start,
1077 						     end - start + 1);
1078 		if (!ordered) {
1079 			/*
1080 			 * If no external cached_state has been passed then
1081 			 * decrement the extra ref taken for cachedp since we
1082 			 * aren't exposing it outside of this function
1083 			 */
1084 			if (!cached_state)
1085 				refcount_dec(&cache->refs);
1086 			break;
1087 		}
1088 		unlock_extent(&inode->io_tree, start, end, cachedp);
1089 		btrfs_start_ordered_extent(ordered);
1090 		btrfs_put_ordered_extent(ordered);
1091 	}
1092 }
1093 
1094 /*
1095  * Lock the passed range and ensure all pending ordered extents in it are run
1096  * to completion in nowait mode.
1097  *
1098  * Return true if btrfs_lock_ordered_range does not return any extents,
1099  * otherwise false.
1100  */
1101 bool btrfs_try_lock_ordered_range(struct btrfs_inode *inode, u64 start, u64 end,
1102 				  struct extent_state **cached_state)
1103 {
1104 	struct btrfs_ordered_extent *ordered;
1105 
1106 	if (!try_lock_extent(&inode->io_tree, start, end, cached_state))
1107 		return false;
1108 
1109 	ordered = btrfs_lookup_ordered_range(inode, start, end - start + 1);
1110 	if (!ordered)
1111 		return true;
1112 
1113 	btrfs_put_ordered_extent(ordered);
1114 	unlock_extent(&inode->io_tree, start, end, cached_state);
1115 
1116 	return false;
1117 }
1118 
1119 /* Split out a new ordered extent for this first @len bytes of @ordered. */
1120 int btrfs_split_ordered_extent(struct btrfs_ordered_extent *ordered, u64 len)
1121 {
1122 	struct inode *inode = ordered->inode;
1123 	struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1124 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1125 	u64 file_offset = ordered->file_offset;
1126 	u64 disk_bytenr = ordered->disk_bytenr;
1127 	unsigned long flags = ordered->flags & BTRFS_ORDERED_TYPE_FLAGS;
1128 	struct rb_node *node;
1129 
1130 	trace_btrfs_ordered_extent_split(BTRFS_I(inode), ordered);
1131 
1132 	ASSERT(!(flags & (1U << BTRFS_ORDERED_COMPRESSED)));
1133 
1134 	/*
1135 	 * The entire bio must be covered by the ordered extent, but we can't
1136 	 * reduce the original extent to a zero length either.
1137 	 */
1138 	if (WARN_ON_ONCE(len >= ordered->num_bytes))
1139 		return -EINVAL;
1140 	/* We cannot split once ordered extent is past end_bio. */
1141 	if (WARN_ON_ONCE(ordered->bytes_left != ordered->disk_num_bytes))
1142 		return -EINVAL;
1143 	/* We cannot split a compressed ordered extent. */
1144 	if (WARN_ON_ONCE(ordered->disk_num_bytes != ordered->num_bytes))
1145 		return -EINVAL;
1146 	/* Checksum list should be empty. */
1147 	if (WARN_ON_ONCE(!list_empty(&ordered->list)))
1148 		return -EINVAL;
1149 
1150 	spin_lock_irq(&tree->lock);
1151 	/* Remove from tree once */
1152 	node = &ordered->rb_node;
1153 	rb_erase(node, &tree->tree);
1154 	RB_CLEAR_NODE(node);
1155 	if (tree->last == node)
1156 		tree->last = NULL;
1157 
1158 	ordered->file_offset += len;
1159 	ordered->disk_bytenr += len;
1160 	ordered->num_bytes -= len;
1161 	ordered->disk_num_bytes -= len;
1162 	ordered->bytes_left -= len;
1163 
1164 	/* Re-insert the node */
1165 	node = tree_insert(&tree->tree, ordered->file_offset, &ordered->rb_node);
1166 	if (node)
1167 		btrfs_panic(fs_info, -EEXIST,
1168 			"zoned: inconsistency in ordered tree at offset %llu",
1169 			    ordered->file_offset);
1170 
1171 	spin_unlock_irq(&tree->lock);
1172 
1173 	/*
1174 	 * The splitting extent is already counted and will be added again in
1175 	 * btrfs_add_ordered_extent(). Subtract len to avoid double counting.
1176 	 */
1177 	percpu_counter_add_batch(&fs_info->ordered_bytes, -len, fs_info->delalloc_batch);
1178 
1179 	return btrfs_add_ordered_extent(BTRFS_I(inode), file_offset, len, len,
1180 					disk_bytenr, len, 0, flags,
1181 					ordered->compress_type);
1182 }
1183 
1184 int __init ordered_data_init(void)
1185 {
1186 	btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1187 				     sizeof(struct btrfs_ordered_extent), 0,
1188 				     SLAB_MEM_SPREAD,
1189 				     NULL);
1190 	if (!btrfs_ordered_extent_cache)
1191 		return -ENOMEM;
1192 
1193 	return 0;
1194 }
1195 
1196 void __cold ordered_data_exit(void)
1197 {
1198 	kmem_cache_destroy(btrfs_ordered_extent_cache);
1199 }
1200