xref: /openbmc/linux/fs/btrfs/ordered-data.c (revision 95b384f9)
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 
19 #include <linux/slab.h>
20 #include <linux/blkdev.h>
21 #include <linux/writeback.h>
22 #include <linux/pagevec.h>
23 #include "ctree.h"
24 #include "transaction.h"
25 #include "btrfs_inode.h"
26 #include "extent_io.h"
27 #include "disk-io.h"
28 #include "compression.h"
29 
30 static struct kmem_cache *btrfs_ordered_extent_cache;
31 
32 static u64 entry_end(struct btrfs_ordered_extent *entry)
33 {
34 	if (entry->file_offset + entry->len < entry->file_offset)
35 		return (u64)-1;
36 	return entry->file_offset + entry->len;
37 }
38 
39 /* returns NULL if the insertion worked, or it returns the node it did find
40  * in the tree
41  */
42 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
43 				   struct rb_node *node)
44 {
45 	struct rb_node **p = &root->rb_node;
46 	struct rb_node *parent = NULL;
47 	struct btrfs_ordered_extent *entry;
48 
49 	while (*p) {
50 		parent = *p;
51 		entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
52 
53 		if (file_offset < entry->file_offset)
54 			p = &(*p)->rb_left;
55 		else if (file_offset >= entry_end(entry))
56 			p = &(*p)->rb_right;
57 		else
58 			return parent;
59 	}
60 
61 	rb_link_node(node, parent, p);
62 	rb_insert_color(node, root);
63 	return NULL;
64 }
65 
66 static void ordered_data_tree_panic(struct inode *inode, int errno,
67 					       u64 offset)
68 {
69 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
70 	btrfs_panic(fs_info, errno, "Inconsistency in ordered tree at offset "
71 		    "%llu", offset);
72 }
73 
74 /*
75  * look for a given offset in the tree, and if it can't be found return the
76  * first lesser offset
77  */
78 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
79 				     struct rb_node **prev_ret)
80 {
81 	struct rb_node *n = root->rb_node;
82 	struct rb_node *prev = NULL;
83 	struct rb_node *test;
84 	struct btrfs_ordered_extent *entry;
85 	struct btrfs_ordered_extent *prev_entry = NULL;
86 
87 	while (n) {
88 		entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
89 		prev = n;
90 		prev_entry = entry;
91 
92 		if (file_offset < entry->file_offset)
93 			n = n->rb_left;
94 		else if (file_offset >= entry_end(entry))
95 			n = n->rb_right;
96 		else
97 			return n;
98 	}
99 	if (!prev_ret)
100 		return NULL;
101 
102 	while (prev && file_offset >= entry_end(prev_entry)) {
103 		test = rb_next(prev);
104 		if (!test)
105 			break;
106 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
107 				      rb_node);
108 		if (file_offset < entry_end(prev_entry))
109 			break;
110 
111 		prev = test;
112 	}
113 	if (prev)
114 		prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
115 				      rb_node);
116 	while (prev && file_offset < entry_end(prev_entry)) {
117 		test = rb_prev(prev);
118 		if (!test)
119 			break;
120 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
121 				      rb_node);
122 		prev = test;
123 	}
124 	*prev_ret = prev;
125 	return NULL;
126 }
127 
128 /*
129  * helper to check if a given offset is inside a given entry
130  */
131 static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
132 {
133 	if (file_offset < entry->file_offset ||
134 	    entry->file_offset + entry->len <= file_offset)
135 		return 0;
136 	return 1;
137 }
138 
139 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
140 			  u64 len)
141 {
142 	if (file_offset + len <= entry->file_offset ||
143 	    entry->file_offset + entry->len <= file_offset)
144 		return 0;
145 	return 1;
146 }
147 
148 /*
149  * look find the first ordered struct that has this offset, otherwise
150  * the first one less than this offset
151  */
152 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
153 					  u64 file_offset)
154 {
155 	struct rb_root *root = &tree->tree;
156 	struct rb_node *prev = NULL;
157 	struct rb_node *ret;
158 	struct btrfs_ordered_extent *entry;
159 
160 	if (tree->last) {
161 		entry = rb_entry(tree->last, struct btrfs_ordered_extent,
162 				 rb_node);
163 		if (offset_in_entry(entry, file_offset))
164 			return tree->last;
165 	}
166 	ret = __tree_search(root, file_offset, &prev);
167 	if (!ret)
168 		ret = prev;
169 	if (ret)
170 		tree->last = ret;
171 	return ret;
172 }
173 
174 /* allocate and add a new ordered_extent into the per-inode tree.
175  * file_offset is the logical offset in the file
176  *
177  * start is the disk block number of an extent already reserved in the
178  * extent allocation tree
179  *
180  * len is the length of the extent
181  *
182  * The tree is given a single reference on the ordered extent that was
183  * inserted.
184  */
185 static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
186 				      u64 start, u64 len, u64 disk_len,
187 				      int type, int dio, int compress_type)
188 {
189 	struct btrfs_root *root = BTRFS_I(inode)->root;
190 	struct btrfs_ordered_inode_tree *tree;
191 	struct rb_node *node;
192 	struct btrfs_ordered_extent *entry;
193 
194 	tree = &BTRFS_I(inode)->ordered_tree;
195 	entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
196 	if (!entry)
197 		return -ENOMEM;
198 
199 	entry->file_offset = file_offset;
200 	entry->start = start;
201 	entry->len = len;
202 	entry->disk_len = disk_len;
203 	entry->bytes_left = len;
204 	entry->inode = igrab(inode);
205 	entry->compress_type = compress_type;
206 	entry->truncated_len = (u64)-1;
207 	if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
208 		set_bit(type, &entry->flags);
209 
210 	if (dio)
211 		set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
212 
213 	/* one ref for the tree */
214 	atomic_set(&entry->refs, 1);
215 	init_waitqueue_head(&entry->wait);
216 	INIT_LIST_HEAD(&entry->list);
217 	INIT_LIST_HEAD(&entry->root_extent_list);
218 	INIT_LIST_HEAD(&entry->work_list);
219 	init_completion(&entry->completion);
220 	INIT_LIST_HEAD(&entry->log_list);
221 	INIT_LIST_HEAD(&entry->trans_list);
222 
223 	trace_btrfs_ordered_extent_add(inode, entry);
224 
225 	spin_lock_irq(&tree->lock);
226 	node = tree_insert(&tree->tree, file_offset,
227 			   &entry->rb_node);
228 	if (node)
229 		ordered_data_tree_panic(inode, -EEXIST, file_offset);
230 	spin_unlock_irq(&tree->lock);
231 
232 	spin_lock(&root->ordered_extent_lock);
233 	list_add_tail(&entry->root_extent_list,
234 		      &root->ordered_extents);
235 	root->nr_ordered_extents++;
236 	if (root->nr_ordered_extents == 1) {
237 		spin_lock(&root->fs_info->ordered_root_lock);
238 		BUG_ON(!list_empty(&root->ordered_root));
239 		list_add_tail(&root->ordered_root,
240 			      &root->fs_info->ordered_roots);
241 		spin_unlock(&root->fs_info->ordered_root_lock);
242 	}
243 	spin_unlock(&root->ordered_extent_lock);
244 
245 	return 0;
246 }
247 
248 int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
249 			     u64 start, u64 len, u64 disk_len, int type)
250 {
251 	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
252 					  disk_len, type, 0,
253 					  BTRFS_COMPRESS_NONE);
254 }
255 
256 int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
257 				 u64 start, u64 len, u64 disk_len, int type)
258 {
259 	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
260 					  disk_len, type, 1,
261 					  BTRFS_COMPRESS_NONE);
262 }
263 
264 int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
265 				      u64 start, u64 len, u64 disk_len,
266 				      int type, int compress_type)
267 {
268 	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
269 					  disk_len, type, 0,
270 					  compress_type);
271 }
272 
273 /*
274  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
275  * when an ordered extent is finished.  If the list covers more than one
276  * ordered extent, it is split across multiples.
277  */
278 void btrfs_add_ordered_sum(struct inode *inode,
279 			   struct btrfs_ordered_extent *entry,
280 			   struct btrfs_ordered_sum *sum)
281 {
282 	struct btrfs_ordered_inode_tree *tree;
283 
284 	tree = &BTRFS_I(inode)->ordered_tree;
285 	spin_lock_irq(&tree->lock);
286 	list_add_tail(&sum->list, &entry->list);
287 	spin_unlock_irq(&tree->lock);
288 }
289 
290 /*
291  * this is used to account for finished IO across a given range
292  * of the file.  The IO may span ordered extents.  If
293  * a given ordered_extent is completely done, 1 is returned, otherwise
294  * 0.
295  *
296  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
297  * to make sure this function only returns 1 once for a given ordered extent.
298  *
299  * file_offset is updated to one byte past the range that is recorded as
300  * complete.  This allows you to walk forward in the file.
301  */
302 int btrfs_dec_test_first_ordered_pending(struct inode *inode,
303 				   struct btrfs_ordered_extent **cached,
304 				   u64 *file_offset, u64 io_size, int uptodate)
305 {
306 	struct btrfs_ordered_inode_tree *tree;
307 	struct rb_node *node;
308 	struct btrfs_ordered_extent *entry = NULL;
309 	int ret;
310 	unsigned long flags;
311 	u64 dec_end;
312 	u64 dec_start;
313 	u64 to_dec;
314 
315 	tree = &BTRFS_I(inode)->ordered_tree;
316 	spin_lock_irqsave(&tree->lock, flags);
317 	node = tree_search(tree, *file_offset);
318 	if (!node) {
319 		ret = 1;
320 		goto out;
321 	}
322 
323 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
324 	if (!offset_in_entry(entry, *file_offset)) {
325 		ret = 1;
326 		goto out;
327 	}
328 
329 	dec_start = max(*file_offset, entry->file_offset);
330 	dec_end = min(*file_offset + io_size, entry->file_offset +
331 		      entry->len);
332 	*file_offset = dec_end;
333 	if (dec_start > dec_end) {
334 		btrfs_crit(BTRFS_I(inode)->root->fs_info,
335 			"bad ordering dec_start %llu end %llu", dec_start, dec_end);
336 	}
337 	to_dec = dec_end - dec_start;
338 	if (to_dec > entry->bytes_left) {
339 		btrfs_crit(BTRFS_I(inode)->root->fs_info,
340 			"bad ordered accounting left %llu size %llu",
341 			entry->bytes_left, to_dec);
342 	}
343 	entry->bytes_left -= to_dec;
344 	if (!uptodate)
345 		set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
346 
347 	if (entry->bytes_left == 0) {
348 		ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
349 		/*
350 		 * Implicit memory barrier after test_and_set_bit
351 		 */
352 		if (waitqueue_active(&entry->wait))
353 			wake_up(&entry->wait);
354 	} else {
355 		ret = 1;
356 	}
357 out:
358 	if (!ret && cached && entry) {
359 		*cached = entry;
360 		atomic_inc(&entry->refs);
361 	}
362 	spin_unlock_irqrestore(&tree->lock, flags);
363 	return ret == 0;
364 }
365 
366 /*
367  * this is used to account for finished IO across a given range
368  * of the file.  The IO should not span ordered extents.  If
369  * a given ordered_extent is completely done, 1 is returned, otherwise
370  * 0.
371  *
372  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
373  * to make sure this function only returns 1 once for a given ordered extent.
374  */
375 int btrfs_dec_test_ordered_pending(struct inode *inode,
376 				   struct btrfs_ordered_extent **cached,
377 				   u64 file_offset, u64 io_size, int uptodate)
378 {
379 	struct btrfs_ordered_inode_tree *tree;
380 	struct rb_node *node;
381 	struct btrfs_ordered_extent *entry = NULL;
382 	unsigned long flags;
383 	int ret;
384 
385 	tree = &BTRFS_I(inode)->ordered_tree;
386 	spin_lock_irqsave(&tree->lock, flags);
387 	if (cached && *cached) {
388 		entry = *cached;
389 		goto have_entry;
390 	}
391 
392 	node = tree_search(tree, file_offset);
393 	if (!node) {
394 		ret = 1;
395 		goto out;
396 	}
397 
398 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
399 have_entry:
400 	if (!offset_in_entry(entry, file_offset)) {
401 		ret = 1;
402 		goto out;
403 	}
404 
405 	if (io_size > entry->bytes_left) {
406 		btrfs_crit(BTRFS_I(inode)->root->fs_info,
407 			   "bad ordered accounting left %llu size %llu",
408 		       entry->bytes_left, io_size);
409 	}
410 	entry->bytes_left -= io_size;
411 	if (!uptodate)
412 		set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
413 
414 	if (entry->bytes_left == 0) {
415 		ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
416 		/*
417 		 * Implicit memory barrier after test_and_set_bit
418 		 */
419 		if (waitqueue_active(&entry->wait))
420 			wake_up(&entry->wait);
421 	} else {
422 		ret = 1;
423 	}
424 out:
425 	if (!ret && cached && entry) {
426 		*cached = entry;
427 		atomic_inc(&entry->refs);
428 	}
429 	spin_unlock_irqrestore(&tree->lock, flags);
430 	return ret == 0;
431 }
432 
433 /* Needs to either be called under a log transaction or the log_mutex */
434 void btrfs_get_logged_extents(struct inode *inode,
435 			      struct list_head *logged_list,
436 			      const loff_t start,
437 			      const loff_t end)
438 {
439 	struct btrfs_ordered_inode_tree *tree;
440 	struct btrfs_ordered_extent *ordered;
441 	struct rb_node *n;
442 	struct rb_node *prev;
443 
444 	tree = &BTRFS_I(inode)->ordered_tree;
445 	spin_lock_irq(&tree->lock);
446 	n = __tree_search(&tree->tree, end, &prev);
447 	if (!n)
448 		n = prev;
449 	for (; n; n = rb_prev(n)) {
450 		ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
451 		if (ordered->file_offset > end)
452 			continue;
453 		if (entry_end(ordered) <= start)
454 			break;
455 		if (test_and_set_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
456 			continue;
457 		list_add(&ordered->log_list, logged_list);
458 		atomic_inc(&ordered->refs);
459 	}
460 	spin_unlock_irq(&tree->lock);
461 }
462 
463 void btrfs_put_logged_extents(struct list_head *logged_list)
464 {
465 	struct btrfs_ordered_extent *ordered;
466 
467 	while (!list_empty(logged_list)) {
468 		ordered = list_first_entry(logged_list,
469 					   struct btrfs_ordered_extent,
470 					   log_list);
471 		list_del_init(&ordered->log_list);
472 		btrfs_put_ordered_extent(ordered);
473 	}
474 }
475 
476 void btrfs_submit_logged_extents(struct list_head *logged_list,
477 				 struct btrfs_root *log)
478 {
479 	int index = log->log_transid % 2;
480 
481 	spin_lock_irq(&log->log_extents_lock[index]);
482 	list_splice_tail(logged_list, &log->logged_list[index]);
483 	spin_unlock_irq(&log->log_extents_lock[index]);
484 }
485 
486 void btrfs_wait_logged_extents(struct btrfs_trans_handle *trans,
487 			       struct btrfs_root *log, u64 transid)
488 {
489 	struct btrfs_ordered_extent *ordered;
490 	int index = transid % 2;
491 
492 	spin_lock_irq(&log->log_extents_lock[index]);
493 	while (!list_empty(&log->logged_list[index])) {
494 		struct inode *inode;
495 		ordered = list_first_entry(&log->logged_list[index],
496 					   struct btrfs_ordered_extent,
497 					   log_list);
498 		list_del_init(&ordered->log_list);
499 		inode = ordered->inode;
500 		spin_unlock_irq(&log->log_extents_lock[index]);
501 
502 		if (!test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) &&
503 		    !test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
504 			u64 start = ordered->file_offset;
505 			u64 end = ordered->file_offset + ordered->len - 1;
506 
507 			WARN_ON(!inode);
508 			filemap_fdatawrite_range(inode->i_mapping, start, end);
509 		}
510 		wait_event(ordered->wait, test_bit(BTRFS_ORDERED_IO_DONE,
511 						   &ordered->flags));
512 
513 		/*
514 		 * In order to keep us from losing our ordered extent
515 		 * information when committing the transaction we have to make
516 		 * sure that any logged extents are completed when we go to
517 		 * commit the transaction.  To do this we simply increase the
518 		 * current transactions pending_ordered counter and decrement it
519 		 * when the ordered extent completes.
520 		 */
521 		if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
522 			struct btrfs_ordered_inode_tree *tree;
523 
524 			tree = &BTRFS_I(inode)->ordered_tree;
525 			spin_lock_irq(&tree->lock);
526 			if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
527 				set_bit(BTRFS_ORDERED_PENDING, &ordered->flags);
528 				atomic_inc(&trans->transaction->pending_ordered);
529 			}
530 			spin_unlock_irq(&tree->lock);
531 		}
532 		btrfs_put_ordered_extent(ordered);
533 		spin_lock_irq(&log->log_extents_lock[index]);
534 	}
535 	spin_unlock_irq(&log->log_extents_lock[index]);
536 }
537 
538 void btrfs_free_logged_extents(struct btrfs_root *log, u64 transid)
539 {
540 	struct btrfs_ordered_extent *ordered;
541 	int index = transid % 2;
542 
543 	spin_lock_irq(&log->log_extents_lock[index]);
544 	while (!list_empty(&log->logged_list[index])) {
545 		ordered = list_first_entry(&log->logged_list[index],
546 					   struct btrfs_ordered_extent,
547 					   log_list);
548 		list_del_init(&ordered->log_list);
549 		spin_unlock_irq(&log->log_extents_lock[index]);
550 		btrfs_put_ordered_extent(ordered);
551 		spin_lock_irq(&log->log_extents_lock[index]);
552 	}
553 	spin_unlock_irq(&log->log_extents_lock[index]);
554 }
555 
556 /*
557  * used to drop a reference on an ordered extent.  This will free
558  * the extent if the last reference is dropped
559  */
560 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
561 {
562 	struct list_head *cur;
563 	struct btrfs_ordered_sum *sum;
564 
565 	trace_btrfs_ordered_extent_put(entry->inode, entry);
566 
567 	if (atomic_dec_and_test(&entry->refs)) {
568 		ASSERT(list_empty(&entry->log_list));
569 		ASSERT(list_empty(&entry->trans_list));
570 		ASSERT(list_empty(&entry->root_extent_list));
571 		ASSERT(RB_EMPTY_NODE(&entry->rb_node));
572 		if (entry->inode)
573 			btrfs_add_delayed_iput(entry->inode);
574 		while (!list_empty(&entry->list)) {
575 			cur = entry->list.next;
576 			sum = list_entry(cur, struct btrfs_ordered_sum, list);
577 			list_del(&sum->list);
578 			kfree(sum);
579 		}
580 		kmem_cache_free(btrfs_ordered_extent_cache, entry);
581 	}
582 }
583 
584 /*
585  * remove an ordered extent from the tree.  No references are dropped
586  * and waiters are woken up.
587  */
588 void btrfs_remove_ordered_extent(struct inode *inode,
589 				 struct btrfs_ordered_extent *entry)
590 {
591 	struct btrfs_ordered_inode_tree *tree;
592 	struct btrfs_root *root = BTRFS_I(inode)->root;
593 	struct rb_node *node;
594 	bool dec_pending_ordered = false;
595 
596 	tree = &BTRFS_I(inode)->ordered_tree;
597 	spin_lock_irq(&tree->lock);
598 	node = &entry->rb_node;
599 	rb_erase(node, &tree->tree);
600 	RB_CLEAR_NODE(node);
601 	if (tree->last == node)
602 		tree->last = NULL;
603 	set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
604 	if (test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags))
605 		dec_pending_ordered = true;
606 	spin_unlock_irq(&tree->lock);
607 
608 	/*
609 	 * The current running transaction is waiting on us, we need to let it
610 	 * know that we're complete and wake it up.
611 	 */
612 	if (dec_pending_ordered) {
613 		struct btrfs_transaction *trans;
614 
615 		/*
616 		 * The checks for trans are just a formality, it should be set,
617 		 * but if it isn't we don't want to deref/assert under the spin
618 		 * lock, so be nice and check if trans is set, but ASSERT() so
619 		 * if it isn't set a developer will notice.
620 		 */
621 		spin_lock(&root->fs_info->trans_lock);
622 		trans = root->fs_info->running_transaction;
623 		if (trans)
624 			atomic_inc(&trans->use_count);
625 		spin_unlock(&root->fs_info->trans_lock);
626 
627 		ASSERT(trans);
628 		if (trans) {
629 			if (atomic_dec_and_test(&trans->pending_ordered))
630 				wake_up(&trans->pending_wait);
631 			btrfs_put_transaction(trans);
632 		}
633 	}
634 
635 	spin_lock(&root->ordered_extent_lock);
636 	list_del_init(&entry->root_extent_list);
637 	root->nr_ordered_extents--;
638 
639 	trace_btrfs_ordered_extent_remove(inode, entry);
640 
641 	if (!root->nr_ordered_extents) {
642 		spin_lock(&root->fs_info->ordered_root_lock);
643 		BUG_ON(list_empty(&root->ordered_root));
644 		list_del_init(&root->ordered_root);
645 		spin_unlock(&root->fs_info->ordered_root_lock);
646 	}
647 	spin_unlock(&root->ordered_extent_lock);
648 	wake_up(&entry->wait);
649 }
650 
651 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
652 {
653 	struct btrfs_ordered_extent *ordered;
654 
655 	ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
656 	btrfs_start_ordered_extent(ordered->inode, ordered, 1);
657 	complete(&ordered->completion);
658 }
659 
660 /*
661  * wait for all the ordered extents in a root.  This is done when balancing
662  * space between drives.
663  */
664 int btrfs_wait_ordered_extents(struct btrfs_root *root, int nr,
665 			       const u64 range_start, const u64 range_len)
666 {
667 	LIST_HEAD(splice);
668 	LIST_HEAD(skipped);
669 	LIST_HEAD(works);
670 	struct btrfs_ordered_extent *ordered, *next;
671 	int count = 0;
672 	const u64 range_end = range_start + range_len;
673 
674 	mutex_lock(&root->ordered_extent_mutex);
675 	spin_lock(&root->ordered_extent_lock);
676 	list_splice_init(&root->ordered_extents, &splice);
677 	while (!list_empty(&splice) && nr) {
678 		ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
679 					   root_extent_list);
680 
681 		if (range_end <= ordered->start ||
682 		    ordered->start + ordered->disk_len <= range_start) {
683 			list_move_tail(&ordered->root_extent_list, &skipped);
684 			cond_resched_lock(&root->ordered_extent_lock);
685 			continue;
686 		}
687 
688 		list_move_tail(&ordered->root_extent_list,
689 			       &root->ordered_extents);
690 		atomic_inc(&ordered->refs);
691 		spin_unlock(&root->ordered_extent_lock);
692 
693 		btrfs_init_work(&ordered->flush_work,
694 				btrfs_flush_delalloc_helper,
695 				btrfs_run_ordered_extent_work, NULL, NULL);
696 		list_add_tail(&ordered->work_list, &works);
697 		btrfs_queue_work(root->fs_info->flush_workers,
698 				 &ordered->flush_work);
699 
700 		cond_resched();
701 		spin_lock(&root->ordered_extent_lock);
702 		if (nr != -1)
703 			nr--;
704 		count++;
705 	}
706 	list_splice_tail(&skipped, &root->ordered_extents);
707 	list_splice_tail(&splice, &root->ordered_extents);
708 	spin_unlock(&root->ordered_extent_lock);
709 
710 	list_for_each_entry_safe(ordered, next, &works, work_list) {
711 		list_del_init(&ordered->work_list);
712 		wait_for_completion(&ordered->completion);
713 		btrfs_put_ordered_extent(ordered);
714 		cond_resched();
715 	}
716 	mutex_unlock(&root->ordered_extent_mutex);
717 
718 	return count;
719 }
720 
721 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, int nr,
722 			      const u64 range_start, const u64 range_len)
723 {
724 	struct btrfs_root *root;
725 	struct list_head splice;
726 	int done;
727 
728 	INIT_LIST_HEAD(&splice);
729 
730 	mutex_lock(&fs_info->ordered_operations_mutex);
731 	spin_lock(&fs_info->ordered_root_lock);
732 	list_splice_init(&fs_info->ordered_roots, &splice);
733 	while (!list_empty(&splice) && nr) {
734 		root = list_first_entry(&splice, struct btrfs_root,
735 					ordered_root);
736 		root = btrfs_grab_fs_root(root);
737 		BUG_ON(!root);
738 		list_move_tail(&root->ordered_root,
739 			       &fs_info->ordered_roots);
740 		spin_unlock(&fs_info->ordered_root_lock);
741 
742 		done = btrfs_wait_ordered_extents(root, nr,
743 						  range_start, range_len);
744 		btrfs_put_fs_root(root);
745 
746 		spin_lock(&fs_info->ordered_root_lock);
747 		if (nr != -1) {
748 			nr -= done;
749 			WARN_ON(nr < 0);
750 		}
751 	}
752 	list_splice_tail(&splice, &fs_info->ordered_roots);
753 	spin_unlock(&fs_info->ordered_root_lock);
754 	mutex_unlock(&fs_info->ordered_operations_mutex);
755 }
756 
757 /*
758  * Used to start IO or wait for a given ordered extent to finish.
759  *
760  * If wait is one, this effectively waits on page writeback for all the pages
761  * in the extent, and it waits on the io completion code to insert
762  * metadata into the btree corresponding to the extent
763  */
764 void btrfs_start_ordered_extent(struct inode *inode,
765 				       struct btrfs_ordered_extent *entry,
766 				       int wait)
767 {
768 	u64 start = entry->file_offset;
769 	u64 end = start + entry->len - 1;
770 
771 	trace_btrfs_ordered_extent_start(inode, entry);
772 
773 	/*
774 	 * pages in the range can be dirty, clean or writeback.  We
775 	 * start IO on any dirty ones so the wait doesn't stall waiting
776 	 * for the flusher thread to find them
777 	 */
778 	if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
779 		filemap_fdatawrite_range(inode->i_mapping, start, end);
780 	if (wait) {
781 		wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
782 						 &entry->flags));
783 	}
784 }
785 
786 /*
787  * Used to wait on ordered extents across a large range of bytes.
788  */
789 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
790 {
791 	int ret = 0;
792 	int ret_wb = 0;
793 	u64 end;
794 	u64 orig_end;
795 	struct btrfs_ordered_extent *ordered;
796 
797 	if (start + len < start) {
798 		orig_end = INT_LIMIT(loff_t);
799 	} else {
800 		orig_end = start + len - 1;
801 		if (orig_end > INT_LIMIT(loff_t))
802 			orig_end = INT_LIMIT(loff_t);
803 	}
804 
805 	/* start IO across the range first to instantiate any delalloc
806 	 * extents
807 	 */
808 	ret = btrfs_fdatawrite_range(inode, start, orig_end);
809 	if (ret)
810 		return ret;
811 
812 	/*
813 	 * If we have a writeback error don't return immediately. Wait first
814 	 * for any ordered extents that haven't completed yet. This is to make
815 	 * sure no one can dirty the same page ranges and call writepages()
816 	 * before the ordered extents complete - to avoid failures (-EEXIST)
817 	 * when adding the new ordered extents to the ordered tree.
818 	 */
819 	ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
820 
821 	end = orig_end;
822 	while (1) {
823 		ordered = btrfs_lookup_first_ordered_extent(inode, end);
824 		if (!ordered)
825 			break;
826 		if (ordered->file_offset > orig_end) {
827 			btrfs_put_ordered_extent(ordered);
828 			break;
829 		}
830 		if (ordered->file_offset + ordered->len <= start) {
831 			btrfs_put_ordered_extent(ordered);
832 			break;
833 		}
834 		btrfs_start_ordered_extent(inode, ordered, 1);
835 		end = ordered->file_offset;
836 		if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
837 			ret = -EIO;
838 		btrfs_put_ordered_extent(ordered);
839 		if (ret || end == 0 || end == start)
840 			break;
841 		end--;
842 	}
843 	return ret_wb ? ret_wb : ret;
844 }
845 
846 /*
847  * find an ordered extent corresponding to file_offset.  return NULL if
848  * nothing is found, otherwise take a reference on the extent and return it
849  */
850 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
851 							 u64 file_offset)
852 {
853 	struct btrfs_ordered_inode_tree *tree;
854 	struct rb_node *node;
855 	struct btrfs_ordered_extent *entry = NULL;
856 
857 	tree = &BTRFS_I(inode)->ordered_tree;
858 	spin_lock_irq(&tree->lock);
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 (!offset_in_entry(entry, file_offset))
865 		entry = NULL;
866 	if (entry)
867 		atomic_inc(&entry->refs);
868 out:
869 	spin_unlock_irq(&tree->lock);
870 	return entry;
871 }
872 
873 /* Since the DIO code tries to lock a wide area we need to look for any ordered
874  * extents that exist in the range, rather than just the start of the range.
875  */
876 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
877 							u64 file_offset,
878 							u64 len)
879 {
880 	struct btrfs_ordered_inode_tree *tree;
881 	struct rb_node *node;
882 	struct btrfs_ordered_extent *entry = NULL;
883 
884 	tree = &BTRFS_I(inode)->ordered_tree;
885 	spin_lock_irq(&tree->lock);
886 	node = tree_search(tree, file_offset);
887 	if (!node) {
888 		node = tree_search(tree, file_offset + len);
889 		if (!node)
890 			goto out;
891 	}
892 
893 	while (1) {
894 		entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
895 		if (range_overlaps(entry, file_offset, len))
896 			break;
897 
898 		if (entry->file_offset >= file_offset + len) {
899 			entry = NULL;
900 			break;
901 		}
902 		entry = NULL;
903 		node = rb_next(node);
904 		if (!node)
905 			break;
906 	}
907 out:
908 	if (entry)
909 		atomic_inc(&entry->refs);
910 	spin_unlock_irq(&tree->lock);
911 	return entry;
912 }
913 
914 bool btrfs_have_ordered_extents_in_range(struct inode *inode,
915 					 u64 file_offset,
916 					 u64 len)
917 {
918 	struct btrfs_ordered_extent *oe;
919 
920 	oe = btrfs_lookup_ordered_range(inode, file_offset, len);
921 	if (oe) {
922 		btrfs_put_ordered_extent(oe);
923 		return true;
924 	}
925 	return false;
926 }
927 
928 /*
929  * lookup and return any extent before 'file_offset'.  NULL is returned
930  * if none is found
931  */
932 struct btrfs_ordered_extent *
933 btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
934 {
935 	struct btrfs_ordered_inode_tree *tree;
936 	struct rb_node *node;
937 	struct btrfs_ordered_extent *entry = NULL;
938 
939 	tree = &BTRFS_I(inode)->ordered_tree;
940 	spin_lock_irq(&tree->lock);
941 	node = tree_search(tree, file_offset);
942 	if (!node)
943 		goto out;
944 
945 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
946 	atomic_inc(&entry->refs);
947 out:
948 	spin_unlock_irq(&tree->lock);
949 	return entry;
950 }
951 
952 /*
953  * After an extent is done, call this to conditionally update the on disk
954  * i_size.  i_size is updated to cover any fully written part of the file.
955  */
956 int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
957 				struct btrfs_ordered_extent *ordered)
958 {
959 	struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
960 	u64 disk_i_size;
961 	u64 new_i_size;
962 	u64 i_size = i_size_read(inode);
963 	struct rb_node *node;
964 	struct rb_node *prev = NULL;
965 	struct btrfs_ordered_extent *test;
966 	int ret = 1;
967 
968 	spin_lock_irq(&tree->lock);
969 	if (ordered) {
970 		offset = entry_end(ordered);
971 		if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags))
972 			offset = min(offset,
973 				     ordered->file_offset +
974 				     ordered->truncated_len);
975 	} else {
976 		offset = ALIGN(offset, BTRFS_I(inode)->root->sectorsize);
977 	}
978 	disk_i_size = BTRFS_I(inode)->disk_i_size;
979 
980 	/* truncate file */
981 	if (disk_i_size > i_size) {
982 		BTRFS_I(inode)->disk_i_size = i_size;
983 		ret = 0;
984 		goto out;
985 	}
986 
987 	/*
988 	 * if the disk i_size is already at the inode->i_size, or
989 	 * this ordered extent is inside the disk i_size, we're done
990 	 */
991 	if (disk_i_size == i_size)
992 		goto out;
993 
994 	/*
995 	 * We still need to update disk_i_size if outstanding_isize is greater
996 	 * than disk_i_size.
997 	 */
998 	if (offset <= disk_i_size &&
999 	    (!ordered || ordered->outstanding_isize <= disk_i_size))
1000 		goto out;
1001 
1002 	/*
1003 	 * walk backward from this ordered extent to disk_i_size.
1004 	 * if we find an ordered extent then we can't update disk i_size
1005 	 * yet
1006 	 */
1007 	if (ordered) {
1008 		node = rb_prev(&ordered->rb_node);
1009 	} else {
1010 		prev = tree_search(tree, offset);
1011 		/*
1012 		 * we insert file extents without involving ordered struct,
1013 		 * so there should be no ordered struct cover this offset
1014 		 */
1015 		if (prev) {
1016 			test = rb_entry(prev, struct btrfs_ordered_extent,
1017 					rb_node);
1018 			BUG_ON(offset_in_entry(test, offset));
1019 		}
1020 		node = prev;
1021 	}
1022 	for (; node; node = rb_prev(node)) {
1023 		test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1024 
1025 		/* We treat this entry as if it doesn't exist */
1026 		if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
1027 			continue;
1028 		if (test->file_offset + test->len <= disk_i_size)
1029 			break;
1030 		if (test->file_offset >= i_size)
1031 			break;
1032 		if (entry_end(test) > disk_i_size) {
1033 			/*
1034 			 * we don't update disk_i_size now, so record this
1035 			 * undealt i_size. Or we will not know the real
1036 			 * i_size.
1037 			 */
1038 			if (test->outstanding_isize < offset)
1039 				test->outstanding_isize = offset;
1040 			if (ordered &&
1041 			    ordered->outstanding_isize >
1042 			    test->outstanding_isize)
1043 				test->outstanding_isize =
1044 						ordered->outstanding_isize;
1045 			goto out;
1046 		}
1047 	}
1048 	new_i_size = min_t(u64, offset, i_size);
1049 
1050 	/*
1051 	 * Some ordered extents may completed before the current one, and
1052 	 * we hold the real i_size in ->outstanding_isize.
1053 	 */
1054 	if (ordered && ordered->outstanding_isize > new_i_size)
1055 		new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
1056 	BTRFS_I(inode)->disk_i_size = new_i_size;
1057 	ret = 0;
1058 out:
1059 	/*
1060 	 * We need to do this because we can't remove ordered extents until
1061 	 * after the i_disk_size has been updated and then the inode has been
1062 	 * updated to reflect the change, so we need to tell anybody who finds
1063 	 * this ordered extent that we've already done all the real work, we
1064 	 * just haven't completed all the other work.
1065 	 */
1066 	if (ordered)
1067 		set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
1068 	spin_unlock_irq(&tree->lock);
1069 	return ret;
1070 }
1071 
1072 /*
1073  * search the ordered extents for one corresponding to 'offset' and
1074  * try to find a checksum.  This is used because we allow pages to
1075  * be reclaimed before their checksum is actually put into the btree
1076  */
1077 int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
1078 			   u32 *sum, int len)
1079 {
1080 	struct btrfs_ordered_sum *ordered_sum;
1081 	struct btrfs_ordered_extent *ordered;
1082 	struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1083 	unsigned long num_sectors;
1084 	unsigned long i;
1085 	u32 sectorsize = BTRFS_I(inode)->root->sectorsize;
1086 	int index = 0;
1087 
1088 	ordered = btrfs_lookup_ordered_extent(inode, offset);
1089 	if (!ordered)
1090 		return 0;
1091 
1092 	spin_lock_irq(&tree->lock);
1093 	list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
1094 		if (disk_bytenr >= ordered_sum->bytenr &&
1095 		    disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
1096 			i = (disk_bytenr - ordered_sum->bytenr) >>
1097 			    inode->i_sb->s_blocksize_bits;
1098 			num_sectors = ordered_sum->len >>
1099 				      inode->i_sb->s_blocksize_bits;
1100 			num_sectors = min_t(int, len - index, num_sectors - i);
1101 			memcpy(sum + index, ordered_sum->sums + i,
1102 			       num_sectors);
1103 
1104 			index += (int)num_sectors;
1105 			if (index == len)
1106 				goto out;
1107 			disk_bytenr += num_sectors * sectorsize;
1108 		}
1109 	}
1110 out:
1111 	spin_unlock_irq(&tree->lock);
1112 	btrfs_put_ordered_extent(ordered);
1113 	return index;
1114 }
1115 
1116 int __init ordered_data_init(void)
1117 {
1118 	btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1119 				     sizeof(struct btrfs_ordered_extent), 0,
1120 				     SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
1121 				     NULL);
1122 	if (!btrfs_ordered_extent_cache)
1123 		return -ENOMEM;
1124 
1125 	return 0;
1126 }
1127 
1128 void ordered_data_exit(void)
1129 {
1130 	kmem_cache_destroy(btrfs_ordered_extent_cache);
1131 }
1132