xref: /openbmc/linux/fs/btrfs/ordered-data.c (revision 62e7ca52)
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 
29 static struct kmem_cache *btrfs_ordered_extent_cache;
30 
31 static u64 entry_end(struct btrfs_ordered_extent *entry)
32 {
33 	if (entry->file_offset + entry->len < entry->file_offset)
34 		return (u64)-1;
35 	return entry->file_offset + entry->len;
36 }
37 
38 /* returns NULL if the insertion worked, or it returns the node it did find
39  * in the tree
40  */
41 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
42 				   struct rb_node *node)
43 {
44 	struct rb_node **p = &root->rb_node;
45 	struct rb_node *parent = NULL;
46 	struct btrfs_ordered_extent *entry;
47 
48 	while (*p) {
49 		parent = *p;
50 		entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
51 
52 		if (file_offset < entry->file_offset)
53 			p = &(*p)->rb_left;
54 		else if (file_offset >= entry_end(entry))
55 			p = &(*p)->rb_right;
56 		else
57 			return parent;
58 	}
59 
60 	rb_link_node(node, parent, p);
61 	rb_insert_color(node, root);
62 	return NULL;
63 }
64 
65 static void ordered_data_tree_panic(struct inode *inode, int errno,
66 					       u64 offset)
67 {
68 	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
69 	btrfs_panic(fs_info, errno, "Inconsistency in ordered tree at offset "
70 		    "%llu", offset);
71 }
72 
73 /*
74  * look for a given offset in the tree, and if it can't be found return the
75  * first lesser offset
76  */
77 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
78 				     struct rb_node **prev_ret)
79 {
80 	struct rb_node *n = root->rb_node;
81 	struct rb_node *prev = NULL;
82 	struct rb_node *test;
83 	struct btrfs_ordered_extent *entry;
84 	struct btrfs_ordered_extent *prev_entry = NULL;
85 
86 	while (n) {
87 		entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
88 		prev = n;
89 		prev_entry = entry;
90 
91 		if (file_offset < entry->file_offset)
92 			n = n->rb_left;
93 		else if (file_offset >= entry_end(entry))
94 			n = n->rb_right;
95 		else
96 			return n;
97 	}
98 	if (!prev_ret)
99 		return NULL;
100 
101 	while (prev && file_offset >= entry_end(prev_entry)) {
102 		test = rb_next(prev);
103 		if (!test)
104 			break;
105 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
106 				      rb_node);
107 		if (file_offset < entry_end(prev_entry))
108 			break;
109 
110 		prev = test;
111 	}
112 	if (prev)
113 		prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
114 				      rb_node);
115 	while (prev && file_offset < entry_end(prev_entry)) {
116 		test = rb_prev(prev);
117 		if (!test)
118 			break;
119 		prev_entry = rb_entry(test, struct btrfs_ordered_extent,
120 				      rb_node);
121 		prev = test;
122 	}
123 	*prev_ret = prev;
124 	return NULL;
125 }
126 
127 /*
128  * helper to check if a given offset is inside a given entry
129  */
130 static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
131 {
132 	if (file_offset < entry->file_offset ||
133 	    entry->file_offset + entry->len <= file_offset)
134 		return 0;
135 	return 1;
136 }
137 
138 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
139 			  u64 len)
140 {
141 	if (file_offset + len <= entry->file_offset ||
142 	    entry->file_offset + entry->len <= file_offset)
143 		return 0;
144 	return 1;
145 }
146 
147 /*
148  * look find the first ordered struct that has this offset, otherwise
149  * the first one less than this offset
150  */
151 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
152 					  u64 file_offset)
153 {
154 	struct rb_root *root = &tree->tree;
155 	struct rb_node *prev = NULL;
156 	struct rb_node *ret;
157 	struct btrfs_ordered_extent *entry;
158 
159 	if (tree->last) {
160 		entry = rb_entry(tree->last, struct btrfs_ordered_extent,
161 				 rb_node);
162 		if (offset_in_entry(entry, file_offset))
163 			return tree->last;
164 	}
165 	ret = __tree_search(root, file_offset, &prev);
166 	if (!ret)
167 		ret = prev;
168 	if (ret)
169 		tree->last = ret;
170 	return ret;
171 }
172 
173 /* allocate and add a new ordered_extent into the per-inode tree.
174  * file_offset is the logical offset in the file
175  *
176  * start is the disk block number of an extent already reserved in the
177  * extent allocation tree
178  *
179  * len is the length of the extent
180  *
181  * The tree is given a single reference on the ordered extent that was
182  * inserted.
183  */
184 static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
185 				      u64 start, u64 len, u64 disk_len,
186 				      int type, int dio, int compress_type)
187 {
188 	struct btrfs_root *root = BTRFS_I(inode)->root;
189 	struct btrfs_ordered_inode_tree *tree;
190 	struct rb_node *node;
191 	struct btrfs_ordered_extent *entry;
192 
193 	tree = &BTRFS_I(inode)->ordered_tree;
194 	entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
195 	if (!entry)
196 		return -ENOMEM;
197 
198 	entry->file_offset = file_offset;
199 	entry->start = start;
200 	entry->len = len;
201 	if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM) &&
202 	    !(type == BTRFS_ORDERED_NOCOW))
203 		entry->csum_bytes_left = disk_len;
204 	entry->disk_len = disk_len;
205 	entry->bytes_left = len;
206 	entry->inode = igrab(inode);
207 	entry->compress_type = compress_type;
208 	entry->truncated_len = (u64)-1;
209 	if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
210 		set_bit(type, &entry->flags);
211 
212 	if (dio)
213 		set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
214 
215 	/* one ref for the tree */
216 	atomic_set(&entry->refs, 1);
217 	init_waitqueue_head(&entry->wait);
218 	INIT_LIST_HEAD(&entry->list);
219 	INIT_LIST_HEAD(&entry->root_extent_list);
220 	INIT_LIST_HEAD(&entry->work_list);
221 	init_completion(&entry->completion);
222 	INIT_LIST_HEAD(&entry->log_list);
223 
224 	trace_btrfs_ordered_extent_add(inode, entry);
225 
226 	spin_lock_irq(&tree->lock);
227 	node = tree_insert(&tree->tree, file_offset,
228 			   &entry->rb_node);
229 	if (node)
230 		ordered_data_tree_panic(inode, -EEXIST, file_offset);
231 	spin_unlock_irq(&tree->lock);
232 
233 	spin_lock(&root->ordered_extent_lock);
234 	list_add_tail(&entry->root_extent_list,
235 		      &root->ordered_extents);
236 	root->nr_ordered_extents++;
237 	if (root->nr_ordered_extents == 1) {
238 		spin_lock(&root->fs_info->ordered_root_lock);
239 		BUG_ON(!list_empty(&root->ordered_root));
240 		list_add_tail(&root->ordered_root,
241 			      &root->fs_info->ordered_roots);
242 		spin_unlock(&root->fs_info->ordered_root_lock);
243 	}
244 	spin_unlock(&root->ordered_extent_lock);
245 
246 	return 0;
247 }
248 
249 int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
250 			     u64 start, u64 len, u64 disk_len, int type)
251 {
252 	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
253 					  disk_len, type, 0,
254 					  BTRFS_COMPRESS_NONE);
255 }
256 
257 int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
258 				 u64 start, u64 len, u64 disk_len, int type)
259 {
260 	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
261 					  disk_len, type, 1,
262 					  BTRFS_COMPRESS_NONE);
263 }
264 
265 int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
266 				      u64 start, u64 len, u64 disk_len,
267 				      int type, int compress_type)
268 {
269 	return __btrfs_add_ordered_extent(inode, file_offset, start, len,
270 					  disk_len, type, 0,
271 					  compress_type);
272 }
273 
274 /*
275  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
276  * when an ordered extent is finished.  If the list covers more than one
277  * ordered extent, it is split across multiples.
278  */
279 void btrfs_add_ordered_sum(struct inode *inode,
280 			   struct btrfs_ordered_extent *entry,
281 			   struct btrfs_ordered_sum *sum)
282 {
283 	struct btrfs_ordered_inode_tree *tree;
284 
285 	tree = &BTRFS_I(inode)->ordered_tree;
286 	spin_lock_irq(&tree->lock);
287 	list_add_tail(&sum->list, &entry->list);
288 	WARN_ON(entry->csum_bytes_left < sum->len);
289 	entry->csum_bytes_left -= sum->len;
290 	if (entry->csum_bytes_left == 0)
291 		wake_up(&entry->wait);
292 	spin_unlock_irq(&tree->lock);
293 }
294 
295 /*
296  * this is used to account for finished IO across a given range
297  * of the file.  The IO may span ordered extents.  If
298  * a given ordered_extent is completely done, 1 is returned, otherwise
299  * 0.
300  *
301  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
302  * to make sure this function only returns 1 once for a given ordered extent.
303  *
304  * file_offset is updated to one byte past the range that is recorded as
305  * complete.  This allows you to walk forward in the file.
306  */
307 int btrfs_dec_test_first_ordered_pending(struct inode *inode,
308 				   struct btrfs_ordered_extent **cached,
309 				   u64 *file_offset, u64 io_size, int uptodate)
310 {
311 	struct btrfs_ordered_inode_tree *tree;
312 	struct rb_node *node;
313 	struct btrfs_ordered_extent *entry = NULL;
314 	int ret;
315 	unsigned long flags;
316 	u64 dec_end;
317 	u64 dec_start;
318 	u64 to_dec;
319 
320 	tree = &BTRFS_I(inode)->ordered_tree;
321 	spin_lock_irqsave(&tree->lock, flags);
322 	node = tree_search(tree, *file_offset);
323 	if (!node) {
324 		ret = 1;
325 		goto out;
326 	}
327 
328 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
329 	if (!offset_in_entry(entry, *file_offset)) {
330 		ret = 1;
331 		goto out;
332 	}
333 
334 	dec_start = max(*file_offset, entry->file_offset);
335 	dec_end = min(*file_offset + io_size, entry->file_offset +
336 		      entry->len);
337 	*file_offset = dec_end;
338 	if (dec_start > dec_end) {
339 		btrfs_crit(BTRFS_I(inode)->root->fs_info,
340 			"bad ordering dec_start %llu end %llu", dec_start, dec_end);
341 	}
342 	to_dec = dec_end - dec_start;
343 	if (to_dec > entry->bytes_left) {
344 		btrfs_crit(BTRFS_I(inode)->root->fs_info,
345 			"bad ordered accounting left %llu size %llu",
346 			entry->bytes_left, to_dec);
347 	}
348 	entry->bytes_left -= to_dec;
349 	if (!uptodate)
350 		set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
351 
352 	if (entry->bytes_left == 0) {
353 		ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
354 		if (waitqueue_active(&entry->wait))
355 			wake_up(&entry->wait);
356 	} else {
357 		ret = 1;
358 	}
359 out:
360 	if (!ret && cached && entry) {
361 		*cached = entry;
362 		atomic_inc(&entry->refs);
363 	}
364 	spin_unlock_irqrestore(&tree->lock, flags);
365 	return ret == 0;
366 }
367 
368 /*
369  * this is used to account for finished IO across a given range
370  * of the file.  The IO should not span ordered extents.  If
371  * a given ordered_extent is completely done, 1 is returned, otherwise
372  * 0.
373  *
374  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
375  * to make sure this function only returns 1 once for a given ordered extent.
376  */
377 int btrfs_dec_test_ordered_pending(struct inode *inode,
378 				   struct btrfs_ordered_extent **cached,
379 				   u64 file_offset, u64 io_size, int uptodate)
380 {
381 	struct btrfs_ordered_inode_tree *tree;
382 	struct rb_node *node;
383 	struct btrfs_ordered_extent *entry = NULL;
384 	unsigned long flags;
385 	int ret;
386 
387 	tree = &BTRFS_I(inode)->ordered_tree;
388 	spin_lock_irqsave(&tree->lock, flags);
389 	if (cached && *cached) {
390 		entry = *cached;
391 		goto have_entry;
392 	}
393 
394 	node = tree_search(tree, file_offset);
395 	if (!node) {
396 		ret = 1;
397 		goto out;
398 	}
399 
400 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
401 have_entry:
402 	if (!offset_in_entry(entry, file_offset)) {
403 		ret = 1;
404 		goto out;
405 	}
406 
407 	if (io_size > entry->bytes_left) {
408 		btrfs_crit(BTRFS_I(inode)->root->fs_info,
409 			   "bad ordered accounting left %llu size %llu",
410 		       entry->bytes_left, io_size);
411 	}
412 	entry->bytes_left -= io_size;
413 	if (!uptodate)
414 		set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
415 
416 	if (entry->bytes_left == 0) {
417 		ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
418 		if (waitqueue_active(&entry->wait))
419 			wake_up(&entry->wait);
420 	} else {
421 		ret = 1;
422 	}
423 out:
424 	if (!ret && cached && entry) {
425 		*cached = entry;
426 		atomic_inc(&entry->refs);
427 	}
428 	spin_unlock_irqrestore(&tree->lock, flags);
429 	return ret == 0;
430 }
431 
432 /* Needs to either be called under a log transaction or the log_mutex */
433 void btrfs_get_logged_extents(struct inode *inode,
434 			      struct list_head *logged_list)
435 {
436 	struct btrfs_ordered_inode_tree *tree;
437 	struct btrfs_ordered_extent *ordered;
438 	struct rb_node *n;
439 
440 	tree = &BTRFS_I(inode)->ordered_tree;
441 	spin_lock_irq(&tree->lock);
442 	for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
443 		ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
444 		if (!list_empty(&ordered->log_list))
445 			continue;
446 		list_add_tail(&ordered->log_list, logged_list);
447 		atomic_inc(&ordered->refs);
448 	}
449 	spin_unlock_irq(&tree->lock);
450 }
451 
452 void btrfs_put_logged_extents(struct list_head *logged_list)
453 {
454 	struct btrfs_ordered_extent *ordered;
455 
456 	while (!list_empty(logged_list)) {
457 		ordered = list_first_entry(logged_list,
458 					   struct btrfs_ordered_extent,
459 					   log_list);
460 		list_del_init(&ordered->log_list);
461 		btrfs_put_ordered_extent(ordered);
462 	}
463 }
464 
465 void btrfs_submit_logged_extents(struct list_head *logged_list,
466 				 struct btrfs_root *log)
467 {
468 	int index = log->log_transid % 2;
469 
470 	spin_lock_irq(&log->log_extents_lock[index]);
471 	list_splice_tail(logged_list, &log->logged_list[index]);
472 	spin_unlock_irq(&log->log_extents_lock[index]);
473 }
474 
475 void btrfs_wait_logged_extents(struct btrfs_root *log, u64 transid)
476 {
477 	struct btrfs_ordered_extent *ordered;
478 	int index = transid % 2;
479 
480 	spin_lock_irq(&log->log_extents_lock[index]);
481 	while (!list_empty(&log->logged_list[index])) {
482 		ordered = list_first_entry(&log->logged_list[index],
483 					   struct btrfs_ordered_extent,
484 					   log_list);
485 		list_del_init(&ordered->log_list);
486 		spin_unlock_irq(&log->log_extents_lock[index]);
487 
488 		if (!test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) &&
489 		    !test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
490 			struct inode *inode = ordered->inode;
491 			u64 start = ordered->file_offset;
492 			u64 end = ordered->file_offset + ordered->len - 1;
493 
494 			WARN_ON(!inode);
495 			filemap_fdatawrite_range(inode->i_mapping, start, end);
496 		}
497 		wait_event(ordered->wait, test_bit(BTRFS_ORDERED_IO_DONE,
498 						   &ordered->flags));
499 
500 		btrfs_put_ordered_extent(ordered);
501 		spin_lock_irq(&log->log_extents_lock[index]);
502 	}
503 	spin_unlock_irq(&log->log_extents_lock[index]);
504 }
505 
506 void btrfs_free_logged_extents(struct btrfs_root *log, u64 transid)
507 {
508 	struct btrfs_ordered_extent *ordered;
509 	int index = transid % 2;
510 
511 	spin_lock_irq(&log->log_extents_lock[index]);
512 	while (!list_empty(&log->logged_list[index])) {
513 		ordered = list_first_entry(&log->logged_list[index],
514 					   struct btrfs_ordered_extent,
515 					   log_list);
516 		list_del_init(&ordered->log_list);
517 		spin_unlock_irq(&log->log_extents_lock[index]);
518 		btrfs_put_ordered_extent(ordered);
519 		spin_lock_irq(&log->log_extents_lock[index]);
520 	}
521 	spin_unlock_irq(&log->log_extents_lock[index]);
522 }
523 
524 /*
525  * used to drop a reference on an ordered extent.  This will free
526  * the extent if the last reference is dropped
527  */
528 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
529 {
530 	struct list_head *cur;
531 	struct btrfs_ordered_sum *sum;
532 
533 	trace_btrfs_ordered_extent_put(entry->inode, entry);
534 
535 	if (atomic_dec_and_test(&entry->refs)) {
536 		if (entry->inode)
537 			btrfs_add_delayed_iput(entry->inode);
538 		while (!list_empty(&entry->list)) {
539 			cur = entry->list.next;
540 			sum = list_entry(cur, struct btrfs_ordered_sum, list);
541 			list_del(&sum->list);
542 			kfree(sum);
543 		}
544 		kmem_cache_free(btrfs_ordered_extent_cache, entry);
545 	}
546 }
547 
548 /*
549  * remove an ordered extent from the tree.  No references are dropped
550  * and waiters are woken up.
551  */
552 void btrfs_remove_ordered_extent(struct inode *inode,
553 				 struct btrfs_ordered_extent *entry)
554 {
555 	struct btrfs_ordered_inode_tree *tree;
556 	struct btrfs_root *root = BTRFS_I(inode)->root;
557 	struct rb_node *node;
558 
559 	tree = &BTRFS_I(inode)->ordered_tree;
560 	spin_lock_irq(&tree->lock);
561 	node = &entry->rb_node;
562 	rb_erase(node, &tree->tree);
563 	if (tree->last == node)
564 		tree->last = NULL;
565 	set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
566 	spin_unlock_irq(&tree->lock);
567 
568 	spin_lock(&root->ordered_extent_lock);
569 	list_del_init(&entry->root_extent_list);
570 	root->nr_ordered_extents--;
571 
572 	trace_btrfs_ordered_extent_remove(inode, entry);
573 
574 	/*
575 	 * we have no more ordered extents for this inode and
576 	 * no dirty pages.  We can safely remove it from the
577 	 * list of ordered extents
578 	 */
579 	if (RB_EMPTY_ROOT(&tree->tree) &&
580 	    !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
581 		spin_lock(&root->fs_info->ordered_root_lock);
582 		list_del_init(&BTRFS_I(inode)->ordered_operations);
583 		spin_unlock(&root->fs_info->ordered_root_lock);
584 	}
585 
586 	if (!root->nr_ordered_extents) {
587 		spin_lock(&root->fs_info->ordered_root_lock);
588 		BUG_ON(list_empty(&root->ordered_root));
589 		list_del_init(&root->ordered_root);
590 		spin_unlock(&root->fs_info->ordered_root_lock);
591 	}
592 	spin_unlock(&root->ordered_extent_lock);
593 	wake_up(&entry->wait);
594 }
595 
596 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
597 {
598 	struct btrfs_ordered_extent *ordered;
599 
600 	ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
601 	btrfs_start_ordered_extent(ordered->inode, ordered, 1);
602 	complete(&ordered->completion);
603 }
604 
605 /*
606  * wait for all the ordered extents in a root.  This is done when balancing
607  * space between drives.
608  */
609 int btrfs_wait_ordered_extents(struct btrfs_root *root, int nr)
610 {
611 	struct list_head splice, works;
612 	struct btrfs_ordered_extent *ordered, *next;
613 	int count = 0;
614 
615 	INIT_LIST_HEAD(&splice);
616 	INIT_LIST_HEAD(&works);
617 
618 	mutex_lock(&root->ordered_extent_mutex);
619 	spin_lock(&root->ordered_extent_lock);
620 	list_splice_init(&root->ordered_extents, &splice);
621 	while (!list_empty(&splice) && nr) {
622 		ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
623 					   root_extent_list);
624 		list_move_tail(&ordered->root_extent_list,
625 			       &root->ordered_extents);
626 		atomic_inc(&ordered->refs);
627 		spin_unlock(&root->ordered_extent_lock);
628 
629 		btrfs_init_work(&ordered->flush_work,
630 				btrfs_run_ordered_extent_work, NULL, NULL);
631 		list_add_tail(&ordered->work_list, &works);
632 		btrfs_queue_work(root->fs_info->flush_workers,
633 				 &ordered->flush_work);
634 
635 		cond_resched();
636 		spin_lock(&root->ordered_extent_lock);
637 		if (nr != -1)
638 			nr--;
639 		count++;
640 	}
641 	list_splice_tail(&splice, &root->ordered_extents);
642 	spin_unlock(&root->ordered_extent_lock);
643 
644 	list_for_each_entry_safe(ordered, next, &works, work_list) {
645 		list_del_init(&ordered->work_list);
646 		wait_for_completion(&ordered->completion);
647 		btrfs_put_ordered_extent(ordered);
648 		cond_resched();
649 	}
650 	mutex_unlock(&root->ordered_extent_mutex);
651 
652 	return count;
653 }
654 
655 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, int nr)
656 {
657 	struct btrfs_root *root;
658 	struct list_head splice;
659 	int done;
660 
661 	INIT_LIST_HEAD(&splice);
662 
663 	mutex_lock(&fs_info->ordered_operations_mutex);
664 	spin_lock(&fs_info->ordered_root_lock);
665 	list_splice_init(&fs_info->ordered_roots, &splice);
666 	while (!list_empty(&splice) && nr) {
667 		root = list_first_entry(&splice, struct btrfs_root,
668 					ordered_root);
669 		root = btrfs_grab_fs_root(root);
670 		BUG_ON(!root);
671 		list_move_tail(&root->ordered_root,
672 			       &fs_info->ordered_roots);
673 		spin_unlock(&fs_info->ordered_root_lock);
674 
675 		done = btrfs_wait_ordered_extents(root, nr);
676 		btrfs_put_fs_root(root);
677 
678 		spin_lock(&fs_info->ordered_root_lock);
679 		if (nr != -1) {
680 			nr -= done;
681 			WARN_ON(nr < 0);
682 		}
683 	}
684 	list_splice_tail(&splice, &fs_info->ordered_roots);
685 	spin_unlock(&fs_info->ordered_root_lock);
686 	mutex_unlock(&fs_info->ordered_operations_mutex);
687 }
688 
689 /*
690  * this is used during transaction commit to write all the inodes
691  * added to the ordered operation list.  These files must be fully on
692  * disk before the transaction commits.
693  *
694  * we have two modes here, one is to just start the IO via filemap_flush
695  * and the other is to wait for all the io.  When we wait, we have an
696  * extra check to make sure the ordered operation list really is empty
697  * before we return
698  */
699 int btrfs_run_ordered_operations(struct btrfs_trans_handle *trans,
700 				 struct btrfs_root *root, int wait)
701 {
702 	struct btrfs_inode *btrfs_inode;
703 	struct inode *inode;
704 	struct btrfs_transaction *cur_trans = trans->transaction;
705 	struct list_head splice;
706 	struct list_head works;
707 	struct btrfs_delalloc_work *work, *next;
708 	int ret = 0;
709 
710 	INIT_LIST_HEAD(&splice);
711 	INIT_LIST_HEAD(&works);
712 
713 	mutex_lock(&root->fs_info->ordered_extent_flush_mutex);
714 	spin_lock(&root->fs_info->ordered_root_lock);
715 	list_splice_init(&cur_trans->ordered_operations, &splice);
716 	while (!list_empty(&splice)) {
717 		btrfs_inode = list_entry(splice.next, struct btrfs_inode,
718 				   ordered_operations);
719 		inode = &btrfs_inode->vfs_inode;
720 
721 		list_del_init(&btrfs_inode->ordered_operations);
722 
723 		/*
724 		 * the inode may be getting freed (in sys_unlink path).
725 		 */
726 		inode = igrab(inode);
727 		if (!inode)
728 			continue;
729 
730 		if (!wait)
731 			list_add_tail(&BTRFS_I(inode)->ordered_operations,
732 				      &cur_trans->ordered_operations);
733 		spin_unlock(&root->fs_info->ordered_root_lock);
734 
735 		work = btrfs_alloc_delalloc_work(inode, wait, 1);
736 		if (!work) {
737 			spin_lock(&root->fs_info->ordered_root_lock);
738 			if (list_empty(&BTRFS_I(inode)->ordered_operations))
739 				list_add_tail(&btrfs_inode->ordered_operations,
740 					      &splice);
741 			list_splice_tail(&splice,
742 					 &cur_trans->ordered_operations);
743 			spin_unlock(&root->fs_info->ordered_root_lock);
744 			ret = -ENOMEM;
745 			goto out;
746 		}
747 		list_add_tail(&work->list, &works);
748 		btrfs_queue_work(root->fs_info->flush_workers,
749 				 &work->work);
750 
751 		cond_resched();
752 		spin_lock(&root->fs_info->ordered_root_lock);
753 	}
754 	spin_unlock(&root->fs_info->ordered_root_lock);
755 out:
756 	list_for_each_entry_safe(work, next, &works, list) {
757 		list_del_init(&work->list);
758 		btrfs_wait_and_free_delalloc_work(work);
759 	}
760 	mutex_unlock(&root->fs_info->ordered_extent_flush_mutex);
761 	return ret;
762 }
763 
764 /*
765  * Used to start IO or wait for a given ordered extent to finish.
766  *
767  * If wait is one, this effectively waits on page writeback for all the pages
768  * in the extent, and it waits on the io completion code to insert
769  * metadata into the btree corresponding to the extent
770  */
771 void btrfs_start_ordered_extent(struct inode *inode,
772 				       struct btrfs_ordered_extent *entry,
773 				       int wait)
774 {
775 	u64 start = entry->file_offset;
776 	u64 end = start + entry->len - 1;
777 
778 	trace_btrfs_ordered_extent_start(inode, entry);
779 
780 	/*
781 	 * pages in the range can be dirty, clean or writeback.  We
782 	 * start IO on any dirty ones so the wait doesn't stall waiting
783 	 * for the flusher thread to find them
784 	 */
785 	if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
786 		filemap_fdatawrite_range(inode->i_mapping, start, end);
787 	if (wait) {
788 		wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
789 						 &entry->flags));
790 	}
791 }
792 
793 /*
794  * Used to wait on ordered extents across a large range of bytes.
795  */
796 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
797 {
798 	int ret = 0;
799 	u64 end;
800 	u64 orig_end;
801 	struct btrfs_ordered_extent *ordered;
802 
803 	if (start + len < start) {
804 		orig_end = INT_LIMIT(loff_t);
805 	} else {
806 		orig_end = start + len - 1;
807 		if (orig_end > INT_LIMIT(loff_t))
808 			orig_end = INT_LIMIT(loff_t);
809 	}
810 
811 	/* start IO across the range first to instantiate any delalloc
812 	 * extents
813 	 */
814 	ret = filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
815 	if (ret)
816 		return ret;
817 	/*
818 	 * So with compression we will find and lock a dirty page and clear the
819 	 * first one as dirty, setup an async extent, and immediately return
820 	 * with the entire range locked but with nobody actually marked with
821 	 * writeback.  So we can't just filemap_write_and_wait_range() and
822 	 * expect it to work since it will just kick off a thread to do the
823 	 * actual work.  So we need to call filemap_fdatawrite_range _again_
824 	 * since it will wait on the page lock, which won't be unlocked until
825 	 * after the pages have been marked as writeback and so we're good to go
826 	 * from there.  We have to do this otherwise we'll miss the ordered
827 	 * extents and that results in badness.  Please Josef, do not think you
828 	 * know better and pull this out at some point in the future, it is
829 	 * right and you are wrong.
830 	 */
831 	if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
832 		     &BTRFS_I(inode)->runtime_flags)) {
833 		ret = filemap_fdatawrite_range(inode->i_mapping, start,
834 					       orig_end);
835 		if (ret)
836 			return ret;
837 	}
838 	ret = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
839 	if (ret)
840 		return ret;
841 
842 	end = orig_end;
843 	while (1) {
844 		ordered = btrfs_lookup_first_ordered_extent(inode, end);
845 		if (!ordered)
846 			break;
847 		if (ordered->file_offset > orig_end) {
848 			btrfs_put_ordered_extent(ordered);
849 			break;
850 		}
851 		if (ordered->file_offset + ordered->len <= start) {
852 			btrfs_put_ordered_extent(ordered);
853 			break;
854 		}
855 		btrfs_start_ordered_extent(inode, ordered, 1);
856 		end = ordered->file_offset;
857 		if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
858 			ret = -EIO;
859 		btrfs_put_ordered_extent(ordered);
860 		if (ret || end == 0 || end == start)
861 			break;
862 		end--;
863 	}
864 	return ret;
865 }
866 
867 /*
868  * find an ordered extent corresponding to file_offset.  return NULL if
869  * nothing is found, otherwise take a reference on the extent and return it
870  */
871 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
872 							 u64 file_offset)
873 {
874 	struct btrfs_ordered_inode_tree *tree;
875 	struct rb_node *node;
876 	struct btrfs_ordered_extent *entry = NULL;
877 
878 	tree = &BTRFS_I(inode)->ordered_tree;
879 	spin_lock_irq(&tree->lock);
880 	node = tree_search(tree, file_offset);
881 	if (!node)
882 		goto out;
883 
884 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
885 	if (!offset_in_entry(entry, file_offset))
886 		entry = NULL;
887 	if (entry)
888 		atomic_inc(&entry->refs);
889 out:
890 	spin_unlock_irq(&tree->lock);
891 	return entry;
892 }
893 
894 /* Since the DIO code tries to lock a wide area we need to look for any ordered
895  * extents that exist in the range, rather than just the start of the range.
896  */
897 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
898 							u64 file_offset,
899 							u64 len)
900 {
901 	struct btrfs_ordered_inode_tree *tree;
902 	struct rb_node *node;
903 	struct btrfs_ordered_extent *entry = NULL;
904 
905 	tree = &BTRFS_I(inode)->ordered_tree;
906 	spin_lock_irq(&tree->lock);
907 	node = tree_search(tree, file_offset);
908 	if (!node) {
909 		node = tree_search(tree, file_offset + len);
910 		if (!node)
911 			goto out;
912 	}
913 
914 	while (1) {
915 		entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
916 		if (range_overlaps(entry, file_offset, len))
917 			break;
918 
919 		if (entry->file_offset >= file_offset + len) {
920 			entry = NULL;
921 			break;
922 		}
923 		entry = NULL;
924 		node = rb_next(node);
925 		if (!node)
926 			break;
927 	}
928 out:
929 	if (entry)
930 		atomic_inc(&entry->refs);
931 	spin_unlock_irq(&tree->lock);
932 	return entry;
933 }
934 
935 /*
936  * lookup and return any extent before 'file_offset'.  NULL is returned
937  * if none is found
938  */
939 struct btrfs_ordered_extent *
940 btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
941 {
942 	struct btrfs_ordered_inode_tree *tree;
943 	struct rb_node *node;
944 	struct btrfs_ordered_extent *entry = NULL;
945 
946 	tree = &BTRFS_I(inode)->ordered_tree;
947 	spin_lock_irq(&tree->lock);
948 	node = tree_search(tree, file_offset);
949 	if (!node)
950 		goto out;
951 
952 	entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
953 	atomic_inc(&entry->refs);
954 out:
955 	spin_unlock_irq(&tree->lock);
956 	return entry;
957 }
958 
959 /*
960  * After an extent is done, call this to conditionally update the on disk
961  * i_size.  i_size is updated to cover any fully written part of the file.
962  */
963 int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
964 				struct btrfs_ordered_extent *ordered)
965 {
966 	struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
967 	u64 disk_i_size;
968 	u64 new_i_size;
969 	u64 i_size = i_size_read(inode);
970 	struct rb_node *node;
971 	struct rb_node *prev = NULL;
972 	struct btrfs_ordered_extent *test;
973 	int ret = 1;
974 
975 	spin_lock_irq(&tree->lock);
976 	if (ordered) {
977 		offset = entry_end(ordered);
978 		if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags))
979 			offset = min(offset,
980 				     ordered->file_offset +
981 				     ordered->truncated_len);
982 	} else {
983 		offset = ALIGN(offset, BTRFS_I(inode)->root->sectorsize);
984 	}
985 	disk_i_size = BTRFS_I(inode)->disk_i_size;
986 
987 	/* truncate file */
988 	if (disk_i_size > i_size) {
989 		BTRFS_I(inode)->disk_i_size = i_size;
990 		ret = 0;
991 		goto out;
992 	}
993 
994 	/*
995 	 * if the disk i_size is already at the inode->i_size, or
996 	 * this ordered extent is inside the disk i_size, we're done
997 	 */
998 	if (disk_i_size == i_size)
999 		goto out;
1000 
1001 	/*
1002 	 * We still need to update disk_i_size if outstanding_isize is greater
1003 	 * than disk_i_size.
1004 	 */
1005 	if (offset <= disk_i_size &&
1006 	    (!ordered || ordered->outstanding_isize <= disk_i_size))
1007 		goto out;
1008 
1009 	/*
1010 	 * walk backward from this ordered extent to disk_i_size.
1011 	 * if we find an ordered extent then we can't update disk i_size
1012 	 * yet
1013 	 */
1014 	if (ordered) {
1015 		node = rb_prev(&ordered->rb_node);
1016 	} else {
1017 		prev = tree_search(tree, offset);
1018 		/*
1019 		 * we insert file extents without involving ordered struct,
1020 		 * so there should be no ordered struct cover this offset
1021 		 */
1022 		if (prev) {
1023 			test = rb_entry(prev, struct btrfs_ordered_extent,
1024 					rb_node);
1025 			BUG_ON(offset_in_entry(test, offset));
1026 		}
1027 		node = prev;
1028 	}
1029 	for (; node; node = rb_prev(node)) {
1030 		test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1031 
1032 		/* We treat this entry as if it doesnt exist */
1033 		if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
1034 			continue;
1035 		if (test->file_offset + test->len <= disk_i_size)
1036 			break;
1037 		if (test->file_offset >= i_size)
1038 			break;
1039 		if (entry_end(test) > disk_i_size) {
1040 			/*
1041 			 * we don't update disk_i_size now, so record this
1042 			 * undealt i_size. Or we will not know the real
1043 			 * i_size.
1044 			 */
1045 			if (test->outstanding_isize < offset)
1046 				test->outstanding_isize = offset;
1047 			if (ordered &&
1048 			    ordered->outstanding_isize >
1049 			    test->outstanding_isize)
1050 				test->outstanding_isize =
1051 						ordered->outstanding_isize;
1052 			goto out;
1053 		}
1054 	}
1055 	new_i_size = min_t(u64, offset, i_size);
1056 
1057 	/*
1058 	 * Some ordered extents may completed before the current one, and
1059 	 * we hold the real i_size in ->outstanding_isize.
1060 	 */
1061 	if (ordered && ordered->outstanding_isize > new_i_size)
1062 		new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
1063 	BTRFS_I(inode)->disk_i_size = new_i_size;
1064 	ret = 0;
1065 out:
1066 	/*
1067 	 * We need to do this because we can't remove ordered extents until
1068 	 * after the i_disk_size has been updated and then the inode has been
1069 	 * updated to reflect the change, so we need to tell anybody who finds
1070 	 * this ordered extent that we've already done all the real work, we
1071 	 * just haven't completed all the other work.
1072 	 */
1073 	if (ordered)
1074 		set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
1075 	spin_unlock_irq(&tree->lock);
1076 	return ret;
1077 }
1078 
1079 /*
1080  * search the ordered extents for one corresponding to 'offset' and
1081  * try to find a checksum.  This is used because we allow pages to
1082  * be reclaimed before their checksum is actually put into the btree
1083  */
1084 int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
1085 			   u32 *sum, int len)
1086 {
1087 	struct btrfs_ordered_sum *ordered_sum;
1088 	struct btrfs_ordered_extent *ordered;
1089 	struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1090 	unsigned long num_sectors;
1091 	unsigned long i;
1092 	u32 sectorsize = BTRFS_I(inode)->root->sectorsize;
1093 	int index = 0;
1094 
1095 	ordered = btrfs_lookup_ordered_extent(inode, offset);
1096 	if (!ordered)
1097 		return 0;
1098 
1099 	spin_lock_irq(&tree->lock);
1100 	list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
1101 		if (disk_bytenr >= ordered_sum->bytenr &&
1102 		    disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
1103 			i = (disk_bytenr - ordered_sum->bytenr) >>
1104 			    inode->i_sb->s_blocksize_bits;
1105 			num_sectors = ordered_sum->len >>
1106 				      inode->i_sb->s_blocksize_bits;
1107 			num_sectors = min_t(int, len - index, num_sectors - i);
1108 			memcpy(sum + index, ordered_sum->sums + i,
1109 			       num_sectors);
1110 
1111 			index += (int)num_sectors;
1112 			if (index == len)
1113 				goto out;
1114 			disk_bytenr += num_sectors * sectorsize;
1115 		}
1116 	}
1117 out:
1118 	spin_unlock_irq(&tree->lock);
1119 	btrfs_put_ordered_extent(ordered);
1120 	return index;
1121 }
1122 
1123 
1124 /*
1125  * add a given inode to the list of inodes that must be fully on
1126  * disk before a transaction commit finishes.
1127  *
1128  * This basically gives us the ext3 style data=ordered mode, and it is mostly
1129  * used to make sure renamed files are fully on disk.
1130  *
1131  * It is a noop if the inode is already fully on disk.
1132  *
1133  * If trans is not null, we'll do a friendly check for a transaction that
1134  * is already flushing things and force the IO down ourselves.
1135  */
1136 void btrfs_add_ordered_operation(struct btrfs_trans_handle *trans,
1137 				 struct btrfs_root *root, struct inode *inode)
1138 {
1139 	struct btrfs_transaction *cur_trans = trans->transaction;
1140 	u64 last_mod;
1141 
1142 	last_mod = max(BTRFS_I(inode)->generation, BTRFS_I(inode)->last_trans);
1143 
1144 	/*
1145 	 * if this file hasn't been changed since the last transaction
1146 	 * commit, we can safely return without doing anything
1147 	 */
1148 	if (last_mod <= root->fs_info->last_trans_committed)
1149 		return;
1150 
1151 	spin_lock(&root->fs_info->ordered_root_lock);
1152 	if (list_empty(&BTRFS_I(inode)->ordered_operations)) {
1153 		list_add_tail(&BTRFS_I(inode)->ordered_operations,
1154 			      &cur_trans->ordered_operations);
1155 	}
1156 	spin_unlock(&root->fs_info->ordered_root_lock);
1157 }
1158 
1159 int __init ordered_data_init(void)
1160 {
1161 	btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1162 				     sizeof(struct btrfs_ordered_extent), 0,
1163 				     SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
1164 				     NULL);
1165 	if (!btrfs_ordered_extent_cache)
1166 		return -ENOMEM;
1167 
1168 	return 0;
1169 }
1170 
1171 void ordered_data_exit(void)
1172 {
1173 	if (btrfs_ordered_extent_cache)
1174 		kmem_cache_destroy(btrfs_ordered_extent_cache);
1175 }
1176