Loading...
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
28static u64 entry_end(struct btrfs_ordered_extent *entry)
29{
30 if (entry->file_offset + entry->len < entry->file_offset)
31 return (u64)-1;
32 return entry->file_offset + entry->len;
33}
34
35/* returns NULL if the insertion worked, or it returns the node it did find
36 * in the tree
37 */
38static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
39 struct rb_node *node)
40{
41 struct rb_node **p = &root->rb_node;
42 struct rb_node *parent = NULL;
43 struct btrfs_ordered_extent *entry;
44
45 while (*p) {
46 parent = *p;
47 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
48
49 if (file_offset < entry->file_offset)
50 p = &(*p)->rb_left;
51 else if (file_offset >= entry_end(entry))
52 p = &(*p)->rb_right;
53 else
54 return parent;
55 }
56
57 rb_link_node(node, parent, p);
58 rb_insert_color(node, root);
59 return NULL;
60}
61
62/*
63 * look for a given offset in the tree, and if it can't be found return the
64 * first lesser offset
65 */
66static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
67 struct rb_node **prev_ret)
68{
69 struct rb_node *n = root->rb_node;
70 struct rb_node *prev = NULL;
71 struct rb_node *test;
72 struct btrfs_ordered_extent *entry;
73 struct btrfs_ordered_extent *prev_entry = NULL;
74
75 while (n) {
76 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
77 prev = n;
78 prev_entry = entry;
79
80 if (file_offset < entry->file_offset)
81 n = n->rb_left;
82 else if (file_offset >= entry_end(entry))
83 n = n->rb_right;
84 else
85 return n;
86 }
87 if (!prev_ret)
88 return NULL;
89
90 while (prev && file_offset >= entry_end(prev_entry)) {
91 test = rb_next(prev);
92 if (!test)
93 break;
94 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
95 rb_node);
96 if (file_offset < entry_end(prev_entry))
97 break;
98
99 prev = test;
100 }
101 if (prev)
102 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
103 rb_node);
104 while (prev && file_offset < entry_end(prev_entry)) {
105 test = rb_prev(prev);
106 if (!test)
107 break;
108 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
109 rb_node);
110 prev = test;
111 }
112 *prev_ret = prev;
113 return NULL;
114}
115
116/*
117 * helper to check if a given offset is inside a given entry
118 */
119static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
120{
121 if (file_offset < entry->file_offset ||
122 entry->file_offset + entry->len <= file_offset)
123 return 0;
124 return 1;
125}
126
127static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
128 u64 len)
129{
130 if (file_offset + len <= entry->file_offset ||
131 entry->file_offset + entry->len <= file_offset)
132 return 0;
133 return 1;
134}
135
136/*
137 * look find the first ordered struct that has this offset, otherwise
138 * the first one less than this offset
139 */
140static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
141 u64 file_offset)
142{
143 struct rb_root *root = &tree->tree;
144 struct rb_node *prev = NULL;
145 struct rb_node *ret;
146 struct btrfs_ordered_extent *entry;
147
148 if (tree->last) {
149 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
150 rb_node);
151 if (offset_in_entry(entry, file_offset))
152 return tree->last;
153 }
154 ret = __tree_search(root, file_offset, &prev);
155 if (!ret)
156 ret = prev;
157 if (ret)
158 tree->last = ret;
159 return ret;
160}
161
162/* allocate and add a new ordered_extent into the per-inode tree.
163 * file_offset is the logical offset in the file
164 *
165 * start is the disk block number of an extent already reserved in the
166 * extent allocation tree
167 *
168 * len is the length of the extent
169 *
170 * The tree is given a single reference on the ordered extent that was
171 * inserted.
172 */
173static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
174 u64 start, u64 len, u64 disk_len,
175 int type, int dio, int compress_type)
176{
177 struct btrfs_ordered_inode_tree *tree;
178 struct rb_node *node;
179 struct btrfs_ordered_extent *entry;
180
181 tree = &BTRFS_I(inode)->ordered_tree;
182 entry = kzalloc(sizeof(*entry), GFP_NOFS);
183 if (!entry)
184 return -ENOMEM;
185
186 entry->file_offset = file_offset;
187 entry->start = start;
188 entry->len = len;
189 entry->disk_len = disk_len;
190 entry->bytes_left = len;
191 entry->inode = inode;
192 entry->compress_type = compress_type;
193 if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
194 set_bit(type, &entry->flags);
195
196 if (dio)
197 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
198
199 /* one ref for the tree */
200 atomic_set(&entry->refs, 1);
201 init_waitqueue_head(&entry->wait);
202 INIT_LIST_HEAD(&entry->list);
203 INIT_LIST_HEAD(&entry->root_extent_list);
204
205 trace_btrfs_ordered_extent_add(inode, entry);
206
207 spin_lock(&tree->lock);
208 node = tree_insert(&tree->tree, file_offset,
209 &entry->rb_node);
210 BUG_ON(node);
211 spin_unlock(&tree->lock);
212
213 spin_lock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
214 list_add_tail(&entry->root_extent_list,
215 &BTRFS_I(inode)->root->fs_info->ordered_extents);
216 spin_unlock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
217
218 BUG_ON(node);
219 return 0;
220}
221
222int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
223 u64 start, u64 len, u64 disk_len, int type)
224{
225 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
226 disk_len, type, 0,
227 BTRFS_COMPRESS_NONE);
228}
229
230int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
231 u64 start, u64 len, u64 disk_len, int type)
232{
233 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
234 disk_len, type, 1,
235 BTRFS_COMPRESS_NONE);
236}
237
238int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
239 u64 start, u64 len, u64 disk_len,
240 int type, int compress_type)
241{
242 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
243 disk_len, type, 0,
244 compress_type);
245}
246
247/*
248 * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
249 * when an ordered extent is finished. If the list covers more than one
250 * ordered extent, it is split across multiples.
251 */
252int btrfs_add_ordered_sum(struct inode *inode,
253 struct btrfs_ordered_extent *entry,
254 struct btrfs_ordered_sum *sum)
255{
256 struct btrfs_ordered_inode_tree *tree;
257
258 tree = &BTRFS_I(inode)->ordered_tree;
259 spin_lock(&tree->lock);
260 list_add_tail(&sum->list, &entry->list);
261 spin_unlock(&tree->lock);
262 return 0;
263}
264
265/*
266 * this is used to account for finished IO across a given range
267 * of the file. The IO may span ordered extents. If
268 * a given ordered_extent is completely done, 1 is returned, otherwise
269 * 0.
270 *
271 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
272 * to make sure this function only returns 1 once for a given ordered extent.
273 *
274 * file_offset is updated to one byte past the range that is recorded as
275 * complete. This allows you to walk forward in the file.
276 */
277int btrfs_dec_test_first_ordered_pending(struct inode *inode,
278 struct btrfs_ordered_extent **cached,
279 u64 *file_offset, u64 io_size)
280{
281 struct btrfs_ordered_inode_tree *tree;
282 struct rb_node *node;
283 struct btrfs_ordered_extent *entry = NULL;
284 int ret;
285 u64 dec_end;
286 u64 dec_start;
287 u64 to_dec;
288
289 tree = &BTRFS_I(inode)->ordered_tree;
290 spin_lock(&tree->lock);
291 node = tree_search(tree, *file_offset);
292 if (!node) {
293 ret = 1;
294 goto out;
295 }
296
297 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
298 if (!offset_in_entry(entry, *file_offset)) {
299 ret = 1;
300 goto out;
301 }
302
303 dec_start = max(*file_offset, entry->file_offset);
304 dec_end = min(*file_offset + io_size, entry->file_offset +
305 entry->len);
306 *file_offset = dec_end;
307 if (dec_start > dec_end) {
308 printk(KERN_CRIT "bad ordering dec_start %llu end %llu\n",
309 (unsigned long long)dec_start,
310 (unsigned long long)dec_end);
311 }
312 to_dec = dec_end - dec_start;
313 if (to_dec > entry->bytes_left) {
314 printk(KERN_CRIT "bad ordered accounting left %llu size %llu\n",
315 (unsigned long long)entry->bytes_left,
316 (unsigned long long)to_dec);
317 }
318 entry->bytes_left -= to_dec;
319 if (entry->bytes_left == 0)
320 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
321 else
322 ret = 1;
323out:
324 if (!ret && cached && entry) {
325 *cached = entry;
326 atomic_inc(&entry->refs);
327 }
328 spin_unlock(&tree->lock);
329 return ret == 0;
330}
331
332/*
333 * this is used to account for finished IO across a given range
334 * of the file. The IO should not span ordered extents. If
335 * a given ordered_extent is completely done, 1 is returned, otherwise
336 * 0.
337 *
338 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
339 * to make sure this function only returns 1 once for a given ordered extent.
340 */
341int btrfs_dec_test_ordered_pending(struct inode *inode,
342 struct btrfs_ordered_extent **cached,
343 u64 file_offset, u64 io_size)
344{
345 struct btrfs_ordered_inode_tree *tree;
346 struct rb_node *node;
347 struct btrfs_ordered_extent *entry = NULL;
348 int ret;
349
350 tree = &BTRFS_I(inode)->ordered_tree;
351 spin_lock(&tree->lock);
352 node = tree_search(tree, file_offset);
353 if (!node) {
354 ret = 1;
355 goto out;
356 }
357
358 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
359 if (!offset_in_entry(entry, file_offset)) {
360 ret = 1;
361 goto out;
362 }
363
364 if (io_size > entry->bytes_left) {
365 printk(KERN_CRIT "bad ordered accounting left %llu size %llu\n",
366 (unsigned long long)entry->bytes_left,
367 (unsigned long long)io_size);
368 }
369 entry->bytes_left -= io_size;
370 if (entry->bytes_left == 0)
371 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
372 else
373 ret = 1;
374out:
375 if (!ret && cached && entry) {
376 *cached = entry;
377 atomic_inc(&entry->refs);
378 }
379 spin_unlock(&tree->lock);
380 return ret == 0;
381}
382
383/*
384 * used to drop a reference on an ordered extent. This will free
385 * the extent if the last reference is dropped
386 */
387int btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
388{
389 struct list_head *cur;
390 struct btrfs_ordered_sum *sum;
391
392 trace_btrfs_ordered_extent_put(entry->inode, entry);
393
394 if (atomic_dec_and_test(&entry->refs)) {
395 while (!list_empty(&entry->list)) {
396 cur = entry->list.next;
397 sum = list_entry(cur, struct btrfs_ordered_sum, list);
398 list_del(&sum->list);
399 kfree(sum);
400 }
401 kfree(entry);
402 }
403 return 0;
404}
405
406/*
407 * remove an ordered extent from the tree. No references are dropped
408 * and you must wake_up entry->wait. You must hold the tree lock
409 * while you call this function.
410 */
411static int __btrfs_remove_ordered_extent(struct inode *inode,
412 struct btrfs_ordered_extent *entry)
413{
414 struct btrfs_ordered_inode_tree *tree;
415 struct btrfs_root *root = BTRFS_I(inode)->root;
416 struct rb_node *node;
417
418 tree = &BTRFS_I(inode)->ordered_tree;
419 node = &entry->rb_node;
420 rb_erase(node, &tree->tree);
421 tree->last = NULL;
422 set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
423
424 spin_lock(&root->fs_info->ordered_extent_lock);
425 list_del_init(&entry->root_extent_list);
426
427 trace_btrfs_ordered_extent_remove(inode, entry);
428
429 /*
430 * we have no more ordered extents for this inode and
431 * no dirty pages. We can safely remove it from the
432 * list of ordered extents
433 */
434 if (RB_EMPTY_ROOT(&tree->tree) &&
435 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
436 list_del_init(&BTRFS_I(inode)->ordered_operations);
437 }
438 spin_unlock(&root->fs_info->ordered_extent_lock);
439
440 return 0;
441}
442
443/*
444 * remove an ordered extent from the tree. No references are dropped
445 * but any waiters are woken.
446 */
447int btrfs_remove_ordered_extent(struct inode *inode,
448 struct btrfs_ordered_extent *entry)
449{
450 struct btrfs_ordered_inode_tree *tree;
451 int ret;
452
453 tree = &BTRFS_I(inode)->ordered_tree;
454 spin_lock(&tree->lock);
455 ret = __btrfs_remove_ordered_extent(inode, entry);
456 spin_unlock(&tree->lock);
457 wake_up(&entry->wait);
458
459 return ret;
460}
461
462/*
463 * wait for all the ordered extents in a root. This is done when balancing
464 * space between drives.
465 */
466int btrfs_wait_ordered_extents(struct btrfs_root *root,
467 int nocow_only, int delay_iput)
468{
469 struct list_head splice;
470 struct list_head *cur;
471 struct btrfs_ordered_extent *ordered;
472 struct inode *inode;
473
474 INIT_LIST_HEAD(&splice);
475
476 spin_lock(&root->fs_info->ordered_extent_lock);
477 list_splice_init(&root->fs_info->ordered_extents, &splice);
478 while (!list_empty(&splice)) {
479 cur = splice.next;
480 ordered = list_entry(cur, struct btrfs_ordered_extent,
481 root_extent_list);
482 if (nocow_only &&
483 !test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags) &&
484 !test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags)) {
485 list_move(&ordered->root_extent_list,
486 &root->fs_info->ordered_extents);
487 cond_resched_lock(&root->fs_info->ordered_extent_lock);
488 continue;
489 }
490
491 list_del_init(&ordered->root_extent_list);
492 atomic_inc(&ordered->refs);
493
494 /*
495 * the inode may be getting freed (in sys_unlink path).
496 */
497 inode = igrab(ordered->inode);
498
499 spin_unlock(&root->fs_info->ordered_extent_lock);
500
501 if (inode) {
502 btrfs_start_ordered_extent(inode, ordered, 1);
503 btrfs_put_ordered_extent(ordered);
504 if (delay_iput)
505 btrfs_add_delayed_iput(inode);
506 else
507 iput(inode);
508 } else {
509 btrfs_put_ordered_extent(ordered);
510 }
511
512 spin_lock(&root->fs_info->ordered_extent_lock);
513 }
514 spin_unlock(&root->fs_info->ordered_extent_lock);
515 return 0;
516}
517
518/*
519 * this is used during transaction commit to write all the inodes
520 * added to the ordered operation list. These files must be fully on
521 * disk before the transaction commits.
522 *
523 * we have two modes here, one is to just start the IO via filemap_flush
524 * and the other is to wait for all the io. When we wait, we have an
525 * extra check to make sure the ordered operation list really is empty
526 * before we return
527 */
528int btrfs_run_ordered_operations(struct btrfs_root *root, int wait)
529{
530 struct btrfs_inode *btrfs_inode;
531 struct inode *inode;
532 struct list_head splice;
533
534 INIT_LIST_HEAD(&splice);
535
536 mutex_lock(&root->fs_info->ordered_operations_mutex);
537 spin_lock(&root->fs_info->ordered_extent_lock);
538again:
539 list_splice_init(&root->fs_info->ordered_operations, &splice);
540
541 while (!list_empty(&splice)) {
542 btrfs_inode = list_entry(splice.next, struct btrfs_inode,
543 ordered_operations);
544
545 inode = &btrfs_inode->vfs_inode;
546
547 list_del_init(&btrfs_inode->ordered_operations);
548
549 /*
550 * the inode may be getting freed (in sys_unlink path).
551 */
552 inode = igrab(inode);
553
554 if (!wait && inode) {
555 list_add_tail(&BTRFS_I(inode)->ordered_operations,
556 &root->fs_info->ordered_operations);
557 }
558 spin_unlock(&root->fs_info->ordered_extent_lock);
559
560 if (inode) {
561 if (wait)
562 btrfs_wait_ordered_range(inode, 0, (u64)-1);
563 else
564 filemap_flush(inode->i_mapping);
565 btrfs_add_delayed_iput(inode);
566 }
567
568 cond_resched();
569 spin_lock(&root->fs_info->ordered_extent_lock);
570 }
571 if (wait && !list_empty(&root->fs_info->ordered_operations))
572 goto again;
573
574 spin_unlock(&root->fs_info->ordered_extent_lock);
575 mutex_unlock(&root->fs_info->ordered_operations_mutex);
576
577 return 0;
578}
579
580/*
581 * Used to start IO or wait for a given ordered extent to finish.
582 *
583 * If wait is one, this effectively waits on page writeback for all the pages
584 * in the extent, and it waits on the io completion code to insert
585 * metadata into the btree corresponding to the extent
586 */
587void btrfs_start_ordered_extent(struct inode *inode,
588 struct btrfs_ordered_extent *entry,
589 int wait)
590{
591 u64 start = entry->file_offset;
592 u64 end = start + entry->len - 1;
593
594 trace_btrfs_ordered_extent_start(inode, entry);
595
596 /*
597 * pages in the range can be dirty, clean or writeback. We
598 * start IO on any dirty ones so the wait doesn't stall waiting
599 * for pdflush to find them
600 */
601 if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
602 filemap_fdatawrite_range(inode->i_mapping, start, end);
603 if (wait) {
604 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
605 &entry->flags));
606 }
607}
608
609/*
610 * Used to wait on ordered extents across a large range of bytes.
611 */
612int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
613{
614 u64 end;
615 u64 orig_end;
616 struct btrfs_ordered_extent *ordered;
617 int found;
618
619 if (start + len < start) {
620 orig_end = INT_LIMIT(loff_t);
621 } else {
622 orig_end = start + len - 1;
623 if (orig_end > INT_LIMIT(loff_t))
624 orig_end = INT_LIMIT(loff_t);
625 }
626again:
627 /* start IO across the range first to instantiate any delalloc
628 * extents
629 */
630 filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
631
632 /* The compression code will leave pages locked but return from
633 * writepage without setting the page writeback. Starting again
634 * with WB_SYNC_ALL will end up waiting for the IO to actually start.
635 */
636 filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
637
638 filemap_fdatawait_range(inode->i_mapping, start, orig_end);
639
640 end = orig_end;
641 found = 0;
642 while (1) {
643 ordered = btrfs_lookup_first_ordered_extent(inode, end);
644 if (!ordered)
645 break;
646 if (ordered->file_offset > orig_end) {
647 btrfs_put_ordered_extent(ordered);
648 break;
649 }
650 if (ordered->file_offset + ordered->len < start) {
651 btrfs_put_ordered_extent(ordered);
652 break;
653 }
654 found++;
655 btrfs_start_ordered_extent(inode, ordered, 1);
656 end = ordered->file_offset;
657 btrfs_put_ordered_extent(ordered);
658 if (end == 0 || end == start)
659 break;
660 end--;
661 }
662 if (found || test_range_bit(&BTRFS_I(inode)->io_tree, start, orig_end,
663 EXTENT_DELALLOC, 0, NULL)) {
664 schedule_timeout(1);
665 goto again;
666 }
667 return 0;
668}
669
670/*
671 * find an ordered extent corresponding to file_offset. return NULL if
672 * nothing is found, otherwise take a reference on the extent and return it
673 */
674struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
675 u64 file_offset)
676{
677 struct btrfs_ordered_inode_tree *tree;
678 struct rb_node *node;
679 struct btrfs_ordered_extent *entry = NULL;
680
681 tree = &BTRFS_I(inode)->ordered_tree;
682 spin_lock(&tree->lock);
683 node = tree_search(tree, file_offset);
684 if (!node)
685 goto out;
686
687 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
688 if (!offset_in_entry(entry, file_offset))
689 entry = NULL;
690 if (entry)
691 atomic_inc(&entry->refs);
692out:
693 spin_unlock(&tree->lock);
694 return entry;
695}
696
697/* Since the DIO code tries to lock a wide area we need to look for any ordered
698 * extents that exist in the range, rather than just the start of the range.
699 */
700struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
701 u64 file_offset,
702 u64 len)
703{
704 struct btrfs_ordered_inode_tree *tree;
705 struct rb_node *node;
706 struct btrfs_ordered_extent *entry = NULL;
707
708 tree = &BTRFS_I(inode)->ordered_tree;
709 spin_lock(&tree->lock);
710 node = tree_search(tree, file_offset);
711 if (!node) {
712 node = tree_search(tree, file_offset + len);
713 if (!node)
714 goto out;
715 }
716
717 while (1) {
718 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
719 if (range_overlaps(entry, file_offset, len))
720 break;
721
722 if (entry->file_offset >= file_offset + len) {
723 entry = NULL;
724 break;
725 }
726 entry = NULL;
727 node = rb_next(node);
728 if (!node)
729 break;
730 }
731out:
732 if (entry)
733 atomic_inc(&entry->refs);
734 spin_unlock(&tree->lock);
735 return entry;
736}
737
738/*
739 * lookup and return any extent before 'file_offset'. NULL is returned
740 * if none is found
741 */
742struct btrfs_ordered_extent *
743btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
744{
745 struct btrfs_ordered_inode_tree *tree;
746 struct rb_node *node;
747 struct btrfs_ordered_extent *entry = NULL;
748
749 tree = &BTRFS_I(inode)->ordered_tree;
750 spin_lock(&tree->lock);
751 node = tree_search(tree, file_offset);
752 if (!node)
753 goto out;
754
755 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
756 atomic_inc(&entry->refs);
757out:
758 spin_unlock(&tree->lock);
759 return entry;
760}
761
762/*
763 * After an extent is done, call this to conditionally update the on disk
764 * i_size. i_size is updated to cover any fully written part of the file.
765 */
766int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
767 struct btrfs_ordered_extent *ordered)
768{
769 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
770 struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
771 u64 disk_i_size;
772 u64 new_i_size;
773 u64 i_size_test;
774 u64 i_size = i_size_read(inode);
775 struct rb_node *node;
776 struct rb_node *prev = NULL;
777 struct btrfs_ordered_extent *test;
778 int ret = 1;
779
780 if (ordered)
781 offset = entry_end(ordered);
782 else
783 offset = ALIGN(offset, BTRFS_I(inode)->root->sectorsize);
784
785 spin_lock(&tree->lock);
786 disk_i_size = BTRFS_I(inode)->disk_i_size;
787
788 /* truncate file */
789 if (disk_i_size > i_size) {
790 BTRFS_I(inode)->disk_i_size = i_size;
791 ret = 0;
792 goto out;
793 }
794
795 /*
796 * if the disk i_size is already at the inode->i_size, or
797 * this ordered extent is inside the disk i_size, we're done
798 */
799 if (disk_i_size == i_size || offset <= disk_i_size) {
800 goto out;
801 }
802
803 /*
804 * we can't update the disk_isize if there are delalloc bytes
805 * between disk_i_size and this ordered extent
806 */
807 if (test_range_bit(io_tree, disk_i_size, offset - 1,
808 EXTENT_DELALLOC, 0, NULL)) {
809 goto out;
810 }
811 /*
812 * walk backward from this ordered extent to disk_i_size.
813 * if we find an ordered extent then we can't update disk i_size
814 * yet
815 */
816 if (ordered) {
817 node = rb_prev(&ordered->rb_node);
818 } else {
819 prev = tree_search(tree, offset);
820 /*
821 * we insert file extents without involving ordered struct,
822 * so there should be no ordered struct cover this offset
823 */
824 if (prev) {
825 test = rb_entry(prev, struct btrfs_ordered_extent,
826 rb_node);
827 BUG_ON(offset_in_entry(test, offset));
828 }
829 node = prev;
830 }
831 while (node) {
832 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
833 if (test->file_offset + test->len <= disk_i_size)
834 break;
835 if (test->file_offset >= i_size)
836 break;
837 if (test->file_offset >= disk_i_size)
838 goto out;
839 node = rb_prev(node);
840 }
841 new_i_size = min_t(u64, offset, i_size);
842
843 /*
844 * at this point, we know we can safely update i_size to at least
845 * the offset from this ordered extent. But, we need to
846 * walk forward and see if ios from higher up in the file have
847 * finished.
848 */
849 if (ordered) {
850 node = rb_next(&ordered->rb_node);
851 } else {
852 if (prev)
853 node = rb_next(prev);
854 else
855 node = rb_first(&tree->tree);
856 }
857 i_size_test = 0;
858 if (node) {
859 /*
860 * do we have an area where IO might have finished
861 * between our ordered extent and the next one.
862 */
863 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
864 if (test->file_offset > offset)
865 i_size_test = test->file_offset;
866 } else {
867 i_size_test = i_size;
868 }
869
870 /*
871 * i_size_test is the end of a region after this ordered
872 * extent where there are no ordered extents. As long as there
873 * are no delalloc bytes in this area, it is safe to update
874 * disk_i_size to the end of the region.
875 */
876 if (i_size_test > offset &&
877 !test_range_bit(io_tree, offset, i_size_test - 1,
878 EXTENT_DELALLOC, 0, NULL)) {
879 new_i_size = min_t(u64, i_size_test, i_size);
880 }
881 BTRFS_I(inode)->disk_i_size = new_i_size;
882 ret = 0;
883out:
884 /*
885 * we need to remove the ordered extent with the tree lock held
886 * so that other people calling this function don't find our fully
887 * processed ordered entry and skip updating the i_size
888 */
889 if (ordered)
890 __btrfs_remove_ordered_extent(inode, ordered);
891 spin_unlock(&tree->lock);
892 if (ordered)
893 wake_up(&ordered->wait);
894 return ret;
895}
896
897/*
898 * search the ordered extents for one corresponding to 'offset' and
899 * try to find a checksum. This is used because we allow pages to
900 * be reclaimed before their checksum is actually put into the btree
901 */
902int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
903 u32 *sum)
904{
905 struct btrfs_ordered_sum *ordered_sum;
906 struct btrfs_sector_sum *sector_sums;
907 struct btrfs_ordered_extent *ordered;
908 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
909 unsigned long num_sectors;
910 unsigned long i;
911 u32 sectorsize = BTRFS_I(inode)->root->sectorsize;
912 int ret = 1;
913
914 ordered = btrfs_lookup_ordered_extent(inode, offset);
915 if (!ordered)
916 return 1;
917
918 spin_lock(&tree->lock);
919 list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
920 if (disk_bytenr >= ordered_sum->bytenr) {
921 num_sectors = ordered_sum->len / sectorsize;
922 sector_sums = ordered_sum->sums;
923 for (i = 0; i < num_sectors; i++) {
924 if (sector_sums[i].bytenr == disk_bytenr) {
925 *sum = sector_sums[i].sum;
926 ret = 0;
927 goto out;
928 }
929 }
930 }
931 }
932out:
933 spin_unlock(&tree->lock);
934 btrfs_put_ordered_extent(ordered);
935 return ret;
936}
937
938
939/*
940 * add a given inode to the list of inodes that must be fully on
941 * disk before a transaction commit finishes.
942 *
943 * This basically gives us the ext3 style data=ordered mode, and it is mostly
944 * used to make sure renamed files are fully on disk.
945 *
946 * It is a noop if the inode is already fully on disk.
947 *
948 * If trans is not null, we'll do a friendly check for a transaction that
949 * is already flushing things and force the IO down ourselves.
950 */
951int btrfs_add_ordered_operation(struct btrfs_trans_handle *trans,
952 struct btrfs_root *root,
953 struct inode *inode)
954{
955 u64 last_mod;
956
957 last_mod = max(BTRFS_I(inode)->generation, BTRFS_I(inode)->last_trans);
958
959 /*
960 * if this file hasn't been changed since the last transaction
961 * commit, we can safely return without doing anything
962 */
963 if (last_mod < root->fs_info->last_trans_committed)
964 return 0;
965
966 /*
967 * the transaction is already committing. Just start the IO and
968 * don't bother with all of this list nonsense
969 */
970 if (trans && root->fs_info->running_transaction->blocked) {
971 btrfs_wait_ordered_range(inode, 0, (u64)-1);
972 return 0;
973 }
974
975 spin_lock(&root->fs_info->ordered_extent_lock);
976 if (list_empty(&BTRFS_I(inode)->ordered_operations)) {
977 list_add_tail(&BTRFS_I(inode)->ordered_operations,
978 &root->fs_info->ordered_operations);
979 }
980 spin_unlock(&root->fs_info->ordered_extent_lock);
981
982 return 0;
983}
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
30static struct kmem_cache *btrfs_ordered_extent_cache;
31
32static 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 */
42static 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
66static 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,
71 "Inconsistency in ordered tree at offset %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 */
78static 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 */
131static 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
139static 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 */
152static 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 */
185static 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_fs_info *fs_info = btrfs_sb(inode->i_sb);
190 struct btrfs_root *root = BTRFS_I(inode)->root;
191 struct btrfs_ordered_inode_tree *tree;
192 struct rb_node *node;
193 struct btrfs_ordered_extent *entry;
194
195 tree = &BTRFS_I(inode)->ordered_tree;
196 entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
197 if (!entry)
198 return -ENOMEM;
199
200 entry->file_offset = file_offset;
201 entry->start = start;
202 entry->len = len;
203 entry->disk_len = disk_len;
204 entry->bytes_left = len;
205 entry->inode = igrab(inode);
206 entry->compress_type = compress_type;
207 entry->truncated_len = (u64)-1;
208 if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
209 set_bit(type, &entry->flags);
210
211 if (dio)
212 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
213
214 /* one ref for the tree */
215 atomic_set(&entry->refs, 1);
216 init_waitqueue_head(&entry->wait);
217 INIT_LIST_HEAD(&entry->list);
218 INIT_LIST_HEAD(&entry->root_extent_list);
219 INIT_LIST_HEAD(&entry->work_list);
220 init_completion(&entry->completion);
221 INIT_LIST_HEAD(&entry->log_list);
222 INIT_LIST_HEAD(&entry->trans_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(&fs_info->ordered_root_lock);
239 BUG_ON(!list_empty(&root->ordered_root));
240 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
241 spin_unlock(&fs_info->ordered_root_lock);
242 }
243 spin_unlock(&root->ordered_extent_lock);
244
245 return 0;
246}
247
248int 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
256int 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
264int 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 */
278void 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 */
302int 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_fs_info *fs_info = btrfs_sb(inode->i_sb);
307 struct btrfs_ordered_inode_tree *tree;
308 struct rb_node *node;
309 struct btrfs_ordered_extent *entry = NULL;
310 int ret;
311 unsigned long flags;
312 u64 dec_end;
313 u64 dec_start;
314 u64 to_dec;
315
316 tree = &BTRFS_I(inode)->ordered_tree;
317 spin_lock_irqsave(&tree->lock, flags);
318 node = tree_search(tree, *file_offset);
319 if (!node) {
320 ret = 1;
321 goto out;
322 }
323
324 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
325 if (!offset_in_entry(entry, *file_offset)) {
326 ret = 1;
327 goto out;
328 }
329
330 dec_start = max(*file_offset, entry->file_offset);
331 dec_end = min(*file_offset + io_size, entry->file_offset +
332 entry->len);
333 *file_offset = dec_end;
334 if (dec_start > dec_end) {
335 btrfs_crit(fs_info, "bad ordering dec_start %llu end %llu",
336 dec_start, dec_end);
337 }
338 to_dec = dec_end - dec_start;
339 if (to_dec > entry->bytes_left) {
340 btrfs_crit(fs_info,
341 "bad ordered accounting left %llu size %llu",
342 entry->bytes_left, to_dec);
343 }
344 entry->bytes_left -= to_dec;
345 if (!uptodate)
346 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
347
348 if (entry->bytes_left == 0) {
349 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
350 /*
351 * Implicit memory barrier after test_and_set_bit
352 */
353 if (waitqueue_active(&entry->wait))
354 wake_up(&entry->wait);
355 } else {
356 ret = 1;
357 }
358out:
359 if (!ret && cached && entry) {
360 *cached = entry;
361 atomic_inc(&entry->refs);
362 }
363 spin_unlock_irqrestore(&tree->lock, flags);
364 return ret == 0;
365}
366
367/*
368 * this is used to account for finished IO across a given range
369 * of the file. The IO should not span ordered extents. If
370 * a given ordered_extent is completely done, 1 is returned, otherwise
371 * 0.
372 *
373 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
374 * to make sure this function only returns 1 once for a given ordered extent.
375 */
376int btrfs_dec_test_ordered_pending(struct inode *inode,
377 struct btrfs_ordered_extent **cached,
378 u64 file_offset, u64 io_size, int uptodate)
379{
380 struct btrfs_ordered_inode_tree *tree;
381 struct rb_node *node;
382 struct btrfs_ordered_extent *entry = NULL;
383 unsigned long flags;
384 int ret;
385
386 tree = &BTRFS_I(inode)->ordered_tree;
387 spin_lock_irqsave(&tree->lock, flags);
388 if (cached && *cached) {
389 entry = *cached;
390 goto have_entry;
391 }
392
393 node = tree_search(tree, file_offset);
394 if (!node) {
395 ret = 1;
396 goto out;
397 }
398
399 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
400have_entry:
401 if (!offset_in_entry(entry, file_offset)) {
402 ret = 1;
403 goto out;
404 }
405
406 if (io_size > entry->bytes_left) {
407 btrfs_crit(BTRFS_I(inode)->root->fs_info,
408 "bad ordered accounting left %llu size %llu",
409 entry->bytes_left, io_size);
410 }
411 entry->bytes_left -= io_size;
412 if (!uptodate)
413 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
414
415 if (entry->bytes_left == 0) {
416 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
417 /*
418 * Implicit memory barrier after test_and_set_bit
419 */
420 if (waitqueue_active(&entry->wait))
421 wake_up(&entry->wait);
422 } else {
423 ret = 1;
424 }
425out:
426 if (!ret && cached && entry) {
427 *cached = entry;
428 atomic_inc(&entry->refs);
429 }
430 spin_unlock_irqrestore(&tree->lock, flags);
431 return ret == 0;
432}
433
434/* Needs to either be called under a log transaction or the log_mutex */
435void btrfs_get_logged_extents(struct inode *inode,
436 struct list_head *logged_list,
437 const loff_t start,
438 const loff_t end)
439{
440 struct btrfs_ordered_inode_tree *tree;
441 struct btrfs_ordered_extent *ordered;
442 struct rb_node *n;
443 struct rb_node *prev;
444
445 tree = &BTRFS_I(inode)->ordered_tree;
446 spin_lock_irq(&tree->lock);
447 n = __tree_search(&tree->tree, end, &prev);
448 if (!n)
449 n = prev;
450 for (; n; n = rb_prev(n)) {
451 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
452 if (ordered->file_offset > end)
453 continue;
454 if (entry_end(ordered) <= start)
455 break;
456 if (test_and_set_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
457 continue;
458 list_add(&ordered->log_list, logged_list);
459 atomic_inc(&ordered->refs);
460 }
461 spin_unlock_irq(&tree->lock);
462}
463
464void btrfs_put_logged_extents(struct list_head *logged_list)
465{
466 struct btrfs_ordered_extent *ordered;
467
468 while (!list_empty(logged_list)) {
469 ordered = list_first_entry(logged_list,
470 struct btrfs_ordered_extent,
471 log_list);
472 list_del_init(&ordered->log_list);
473 btrfs_put_ordered_extent(ordered);
474 }
475}
476
477void btrfs_submit_logged_extents(struct list_head *logged_list,
478 struct btrfs_root *log)
479{
480 int index = log->log_transid % 2;
481
482 spin_lock_irq(&log->log_extents_lock[index]);
483 list_splice_tail(logged_list, &log->logged_list[index]);
484 spin_unlock_irq(&log->log_extents_lock[index]);
485}
486
487void btrfs_wait_logged_extents(struct btrfs_trans_handle *trans,
488 struct btrfs_root *log, u64 transid)
489{
490 struct btrfs_ordered_extent *ordered;
491 int index = transid % 2;
492
493 spin_lock_irq(&log->log_extents_lock[index]);
494 while (!list_empty(&log->logged_list[index])) {
495 struct inode *inode;
496 ordered = list_first_entry(&log->logged_list[index],
497 struct btrfs_ordered_extent,
498 log_list);
499 list_del_init(&ordered->log_list);
500 inode = ordered->inode;
501 spin_unlock_irq(&log->log_extents_lock[index]);
502
503 if (!test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) &&
504 !test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
505 u64 start = ordered->file_offset;
506 u64 end = ordered->file_offset + ordered->len - 1;
507
508 WARN_ON(!inode);
509 filemap_fdatawrite_range(inode->i_mapping, start, end);
510 }
511 wait_event(ordered->wait, test_bit(BTRFS_ORDERED_IO_DONE,
512 &ordered->flags));
513
514 /*
515 * In order to keep us from losing our ordered extent
516 * information when committing the transaction we have to make
517 * sure that any logged extents are completed when we go to
518 * commit the transaction. To do this we simply increase the
519 * current transactions pending_ordered counter and decrement it
520 * when the ordered extent completes.
521 */
522 if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
523 struct btrfs_ordered_inode_tree *tree;
524
525 tree = &BTRFS_I(inode)->ordered_tree;
526 spin_lock_irq(&tree->lock);
527 if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
528 set_bit(BTRFS_ORDERED_PENDING, &ordered->flags);
529 atomic_inc(&trans->transaction->pending_ordered);
530 }
531 spin_unlock_irq(&tree->lock);
532 }
533 btrfs_put_ordered_extent(ordered);
534 spin_lock_irq(&log->log_extents_lock[index]);
535 }
536 spin_unlock_irq(&log->log_extents_lock[index]);
537}
538
539void btrfs_free_logged_extents(struct btrfs_root *log, u64 transid)
540{
541 struct btrfs_ordered_extent *ordered;
542 int index = transid % 2;
543
544 spin_lock_irq(&log->log_extents_lock[index]);
545 while (!list_empty(&log->logged_list[index])) {
546 ordered = list_first_entry(&log->logged_list[index],
547 struct btrfs_ordered_extent,
548 log_list);
549 list_del_init(&ordered->log_list);
550 spin_unlock_irq(&log->log_extents_lock[index]);
551 btrfs_put_ordered_extent(ordered);
552 spin_lock_irq(&log->log_extents_lock[index]);
553 }
554 spin_unlock_irq(&log->log_extents_lock[index]);
555}
556
557/*
558 * used to drop a reference on an ordered extent. This will free
559 * the extent if the last reference is dropped
560 */
561void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
562{
563 struct list_head *cur;
564 struct btrfs_ordered_sum *sum;
565
566 trace_btrfs_ordered_extent_put(entry->inode, entry);
567
568 if (atomic_dec_and_test(&entry->refs)) {
569 ASSERT(list_empty(&entry->log_list));
570 ASSERT(list_empty(&entry->trans_list));
571 ASSERT(list_empty(&entry->root_extent_list));
572 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
573 if (entry->inode)
574 btrfs_add_delayed_iput(entry->inode);
575 while (!list_empty(&entry->list)) {
576 cur = entry->list.next;
577 sum = list_entry(cur, struct btrfs_ordered_sum, list);
578 list_del(&sum->list);
579 kfree(sum);
580 }
581 kmem_cache_free(btrfs_ordered_extent_cache, entry);
582 }
583}
584
585/*
586 * remove an ordered extent from the tree. No references are dropped
587 * and waiters are woken up.
588 */
589void btrfs_remove_ordered_extent(struct inode *inode,
590 struct btrfs_ordered_extent *entry)
591{
592 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
593 struct btrfs_ordered_inode_tree *tree;
594 struct btrfs_root *root = BTRFS_I(inode)->root;
595 struct rb_node *node;
596 bool dec_pending_ordered = false;
597
598 tree = &BTRFS_I(inode)->ordered_tree;
599 spin_lock_irq(&tree->lock);
600 node = &entry->rb_node;
601 rb_erase(node, &tree->tree);
602 RB_CLEAR_NODE(node);
603 if (tree->last == node)
604 tree->last = NULL;
605 set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
606 if (test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags))
607 dec_pending_ordered = true;
608 spin_unlock_irq(&tree->lock);
609
610 /*
611 * The current running transaction is waiting on us, we need to let it
612 * know that we're complete and wake it up.
613 */
614 if (dec_pending_ordered) {
615 struct btrfs_transaction *trans;
616
617 /*
618 * The checks for trans are just a formality, it should be set,
619 * but if it isn't we don't want to deref/assert under the spin
620 * lock, so be nice and check if trans is set, but ASSERT() so
621 * if it isn't set a developer will notice.
622 */
623 spin_lock(&fs_info->trans_lock);
624 trans = fs_info->running_transaction;
625 if (trans)
626 atomic_inc(&trans->use_count);
627 spin_unlock(&fs_info->trans_lock);
628
629 ASSERT(trans);
630 if (trans) {
631 if (atomic_dec_and_test(&trans->pending_ordered))
632 wake_up(&trans->pending_wait);
633 btrfs_put_transaction(trans);
634 }
635 }
636
637 spin_lock(&root->ordered_extent_lock);
638 list_del_init(&entry->root_extent_list);
639 root->nr_ordered_extents--;
640
641 trace_btrfs_ordered_extent_remove(inode, entry);
642
643 if (!root->nr_ordered_extents) {
644 spin_lock(&fs_info->ordered_root_lock);
645 BUG_ON(list_empty(&root->ordered_root));
646 list_del_init(&root->ordered_root);
647 spin_unlock(&fs_info->ordered_root_lock);
648 }
649 spin_unlock(&root->ordered_extent_lock);
650 wake_up(&entry->wait);
651}
652
653static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
654{
655 struct btrfs_ordered_extent *ordered;
656
657 ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
658 btrfs_start_ordered_extent(ordered->inode, ordered, 1);
659 complete(&ordered->completion);
660}
661
662/*
663 * wait for all the ordered extents in a root. This is done when balancing
664 * space between drives.
665 */
666int btrfs_wait_ordered_extents(struct btrfs_root *root, int nr,
667 const u64 range_start, const u64 range_len)
668{
669 struct btrfs_fs_info *fs_info = root->fs_info;
670 LIST_HEAD(splice);
671 LIST_HEAD(skipped);
672 LIST_HEAD(works);
673 struct btrfs_ordered_extent *ordered, *next;
674 int count = 0;
675 const u64 range_end = range_start + range_len;
676
677 mutex_lock(&root->ordered_extent_mutex);
678 spin_lock(&root->ordered_extent_lock);
679 list_splice_init(&root->ordered_extents, &splice);
680 while (!list_empty(&splice) && nr) {
681 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
682 root_extent_list);
683
684 if (range_end <= ordered->start ||
685 ordered->start + ordered->disk_len <= range_start) {
686 list_move_tail(&ordered->root_extent_list, &skipped);
687 cond_resched_lock(&root->ordered_extent_lock);
688 continue;
689 }
690
691 list_move_tail(&ordered->root_extent_list,
692 &root->ordered_extents);
693 atomic_inc(&ordered->refs);
694 spin_unlock(&root->ordered_extent_lock);
695
696 btrfs_init_work(&ordered->flush_work,
697 btrfs_flush_delalloc_helper,
698 btrfs_run_ordered_extent_work, NULL, NULL);
699 list_add_tail(&ordered->work_list, &works);
700 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
701
702 cond_resched();
703 spin_lock(&root->ordered_extent_lock);
704 if (nr != -1)
705 nr--;
706 count++;
707 }
708 list_splice_tail(&skipped, &root->ordered_extents);
709 list_splice_tail(&splice, &root->ordered_extents);
710 spin_unlock(&root->ordered_extent_lock);
711
712 list_for_each_entry_safe(ordered, next, &works, work_list) {
713 list_del_init(&ordered->work_list);
714 wait_for_completion(&ordered->completion);
715 btrfs_put_ordered_extent(ordered);
716 cond_resched();
717 }
718 mutex_unlock(&root->ordered_extent_mutex);
719
720 return count;
721}
722
723int btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, int nr,
724 const u64 range_start, const u64 range_len)
725{
726 struct btrfs_root *root;
727 struct list_head splice;
728 int done;
729 int total_done = 0;
730
731 INIT_LIST_HEAD(&splice);
732
733 mutex_lock(&fs_info->ordered_operations_mutex);
734 spin_lock(&fs_info->ordered_root_lock);
735 list_splice_init(&fs_info->ordered_roots, &splice);
736 while (!list_empty(&splice) && nr) {
737 root = list_first_entry(&splice, struct btrfs_root,
738 ordered_root);
739 root = btrfs_grab_fs_root(root);
740 BUG_ON(!root);
741 list_move_tail(&root->ordered_root,
742 &fs_info->ordered_roots);
743 spin_unlock(&fs_info->ordered_root_lock);
744
745 done = btrfs_wait_ordered_extents(root, nr,
746 range_start, range_len);
747 btrfs_put_fs_root(root);
748 total_done += done;
749
750 spin_lock(&fs_info->ordered_root_lock);
751 if (nr != -1) {
752 nr -= done;
753 WARN_ON(nr < 0);
754 }
755 }
756 list_splice_tail(&splice, &fs_info->ordered_roots);
757 spin_unlock(&fs_info->ordered_root_lock);
758 mutex_unlock(&fs_info->ordered_operations_mutex);
759
760 return total_done;
761}
762
763/*
764 * Used to start IO or wait for a given ordered extent to finish.
765 *
766 * If wait is one, this effectively waits on page writeback for all the pages
767 * in the extent, and it waits on the io completion code to insert
768 * metadata into the btree corresponding to the extent
769 */
770void btrfs_start_ordered_extent(struct inode *inode,
771 struct btrfs_ordered_extent *entry,
772 int wait)
773{
774 u64 start = entry->file_offset;
775 u64 end = start + entry->len - 1;
776
777 trace_btrfs_ordered_extent_start(inode, entry);
778
779 /*
780 * pages in the range can be dirty, clean or writeback. We
781 * start IO on any dirty ones so the wait doesn't stall waiting
782 * for the flusher thread to find them
783 */
784 if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
785 filemap_fdatawrite_range(inode->i_mapping, start, end);
786 if (wait) {
787 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
788 &entry->flags));
789 }
790}
791
792/*
793 * Used to wait on ordered extents across a large range of bytes.
794 */
795int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
796{
797 int ret = 0;
798 int ret_wb = 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 = btrfs_fdatawrite_range(inode, start, orig_end);
815 if (ret)
816 return ret;
817
818 /*
819 * If we have a writeback error don't return immediately. Wait first
820 * for any ordered extents that haven't completed yet. This is to make
821 * sure no one can dirty the same page ranges and call writepages()
822 * before the ordered extents complete - to avoid failures (-EEXIST)
823 * when adding the new ordered extents to the ordered tree.
824 */
825 ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
826
827 end = orig_end;
828 while (1) {
829 ordered = btrfs_lookup_first_ordered_extent(inode, end);
830 if (!ordered)
831 break;
832 if (ordered->file_offset > orig_end) {
833 btrfs_put_ordered_extent(ordered);
834 break;
835 }
836 if (ordered->file_offset + ordered->len <= start) {
837 btrfs_put_ordered_extent(ordered);
838 break;
839 }
840 btrfs_start_ordered_extent(inode, ordered, 1);
841 end = ordered->file_offset;
842 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
843 ret = -EIO;
844 btrfs_put_ordered_extent(ordered);
845 if (ret || end == 0 || end == start)
846 break;
847 end--;
848 }
849 return ret_wb ? ret_wb : ret;
850}
851
852/*
853 * find an ordered extent corresponding to file_offset. return NULL if
854 * nothing is found, otherwise take a reference on the extent and return it
855 */
856struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
857 u64 file_offset)
858{
859 struct btrfs_ordered_inode_tree *tree;
860 struct rb_node *node;
861 struct btrfs_ordered_extent *entry = NULL;
862
863 tree = &BTRFS_I(inode)->ordered_tree;
864 spin_lock_irq(&tree->lock);
865 node = tree_search(tree, file_offset);
866 if (!node)
867 goto out;
868
869 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
870 if (!offset_in_entry(entry, file_offset))
871 entry = NULL;
872 if (entry)
873 atomic_inc(&entry->refs);
874out:
875 spin_unlock_irq(&tree->lock);
876 return entry;
877}
878
879/* Since the DIO code tries to lock a wide area we need to look for any ordered
880 * extents that exist in the range, rather than just the start of the range.
881 */
882struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
883 u64 file_offset,
884 u64 len)
885{
886 struct btrfs_ordered_inode_tree *tree;
887 struct rb_node *node;
888 struct btrfs_ordered_extent *entry = NULL;
889
890 tree = &BTRFS_I(inode)->ordered_tree;
891 spin_lock_irq(&tree->lock);
892 node = tree_search(tree, file_offset);
893 if (!node) {
894 node = tree_search(tree, file_offset + len);
895 if (!node)
896 goto out;
897 }
898
899 while (1) {
900 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
901 if (range_overlaps(entry, file_offset, len))
902 break;
903
904 if (entry->file_offset >= file_offset + len) {
905 entry = NULL;
906 break;
907 }
908 entry = NULL;
909 node = rb_next(node);
910 if (!node)
911 break;
912 }
913out:
914 if (entry)
915 atomic_inc(&entry->refs);
916 spin_unlock_irq(&tree->lock);
917 return entry;
918}
919
920bool btrfs_have_ordered_extents_in_range(struct inode *inode,
921 u64 file_offset,
922 u64 len)
923{
924 struct btrfs_ordered_extent *oe;
925
926 oe = btrfs_lookup_ordered_range(inode, file_offset, len);
927 if (oe) {
928 btrfs_put_ordered_extent(oe);
929 return true;
930 }
931 return false;
932}
933
934/*
935 * lookup and return any extent before 'file_offset'. NULL is returned
936 * if none is found
937 */
938struct btrfs_ordered_extent *
939btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
940{
941 struct btrfs_ordered_inode_tree *tree;
942 struct rb_node *node;
943 struct btrfs_ordered_extent *entry = NULL;
944
945 tree = &BTRFS_I(inode)->ordered_tree;
946 spin_lock_irq(&tree->lock);
947 node = tree_search(tree, file_offset);
948 if (!node)
949 goto out;
950
951 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
952 atomic_inc(&entry->refs);
953out:
954 spin_unlock_irq(&tree->lock);
955 return entry;
956}
957
958/*
959 * After an extent is done, call this to conditionally update the on disk
960 * i_size. i_size is updated to cover any fully written part of the file.
961 */
962int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
963 struct btrfs_ordered_extent *ordered)
964{
965 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
966 u64 disk_i_size;
967 u64 new_i_size;
968 u64 i_size = i_size_read(inode);
969 struct rb_node *node;
970 struct rb_node *prev = NULL;
971 struct btrfs_ordered_extent *test;
972 int ret = 1;
973 u64 orig_offset = offset;
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_inode_sectorsize(inode));
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 = orig_offset;
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 doesn't 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;
1065out:
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 */
1084int 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_inode_sectorsize(inode);
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 }
1117out:
1118 spin_unlock_irq(&tree->lock);
1119 btrfs_put_ordered_extent(ordered);
1120 return index;
1121}
1122
1123int __init ordered_data_init(void)
1124{
1125 btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1126 sizeof(struct btrfs_ordered_extent), 0,
1127 SLAB_MEM_SPREAD,
1128 NULL);
1129 if (!btrfs_ordered_extent_cache)
1130 return -ENOMEM;
1131
1132 return 0;
1133}
1134
1135void ordered_data_exit(void)
1136{
1137 kmem_cache_destroy(btrfs_ordered_extent_cache);
1138}