Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6#include <linux/slab.h>
7#include <linux/blkdev.h>
8#include <linux/writeback.h>
9#include <linux/sched/mm.h>
10#include "misc.h"
11#include "ctree.h"
12#include "transaction.h"
13#include "btrfs_inode.h"
14#include "extent_io.h"
15#include "disk-io.h"
16#include "compression.h"
17#include "delalloc-space.h"
18
19static struct kmem_cache *btrfs_ordered_extent_cache;
20
21static u64 entry_end(struct btrfs_ordered_extent *entry)
22{
23 if (entry->file_offset + entry->len < entry->file_offset)
24 return (u64)-1;
25 return entry->file_offset + entry->len;
26}
27
28/* returns NULL if the insertion worked, or it returns the node it did find
29 * in the tree
30 */
31static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
32 struct rb_node *node)
33{
34 struct rb_node **p = &root->rb_node;
35 struct rb_node *parent = NULL;
36 struct btrfs_ordered_extent *entry;
37
38 while (*p) {
39 parent = *p;
40 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
41
42 if (file_offset < entry->file_offset)
43 p = &(*p)->rb_left;
44 else if (file_offset >= entry_end(entry))
45 p = &(*p)->rb_right;
46 else
47 return parent;
48 }
49
50 rb_link_node(node, parent, p);
51 rb_insert_color(node, root);
52 return NULL;
53}
54
55static void ordered_data_tree_panic(struct inode *inode, int errno,
56 u64 offset)
57{
58 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
59 btrfs_panic(fs_info, errno,
60 "Inconsistency in ordered tree at offset %llu", offset);
61}
62
63/*
64 * look for a given offset in the tree, and if it can't be found return the
65 * first lesser offset
66 */
67static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
68 struct rb_node **prev_ret)
69{
70 struct rb_node *n = root->rb_node;
71 struct rb_node *prev = NULL;
72 struct rb_node *test;
73 struct btrfs_ordered_extent *entry;
74 struct btrfs_ordered_extent *prev_entry = NULL;
75
76 while (n) {
77 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
78 prev = n;
79 prev_entry = entry;
80
81 if (file_offset < entry->file_offset)
82 n = n->rb_left;
83 else if (file_offset >= entry_end(entry))
84 n = n->rb_right;
85 else
86 return n;
87 }
88 if (!prev_ret)
89 return NULL;
90
91 while (prev && file_offset >= entry_end(prev_entry)) {
92 test = rb_next(prev);
93 if (!test)
94 break;
95 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
96 rb_node);
97 if (file_offset < entry_end(prev_entry))
98 break;
99
100 prev = test;
101 }
102 if (prev)
103 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
104 rb_node);
105 while (prev && file_offset < entry_end(prev_entry)) {
106 test = rb_prev(prev);
107 if (!test)
108 break;
109 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
110 rb_node);
111 prev = test;
112 }
113 *prev_ret = prev;
114 return NULL;
115}
116
117/*
118 * helper to check if a given offset is inside a given entry
119 */
120static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
121{
122 if (file_offset < entry->file_offset ||
123 entry->file_offset + entry->len <= file_offset)
124 return 0;
125 return 1;
126}
127
128static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
129 u64 len)
130{
131 if (file_offset + len <= entry->file_offset ||
132 entry->file_offset + entry->len <= file_offset)
133 return 0;
134 return 1;
135}
136
137/*
138 * look find the first ordered struct that has this offset, otherwise
139 * the first one less than this offset
140 */
141static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
142 u64 file_offset)
143{
144 struct rb_root *root = &tree->tree;
145 struct rb_node *prev = NULL;
146 struct rb_node *ret;
147 struct btrfs_ordered_extent *entry;
148
149 if (tree->last) {
150 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
151 rb_node);
152 if (offset_in_entry(entry, file_offset))
153 return tree->last;
154 }
155 ret = __tree_search(root, file_offset, &prev);
156 if (!ret)
157 ret = prev;
158 if (ret)
159 tree->last = ret;
160 return ret;
161}
162
163/* allocate and add a new ordered_extent into the per-inode tree.
164 * file_offset is the logical offset in the file
165 *
166 * start is the disk block number of an extent already reserved in the
167 * extent allocation tree
168 *
169 * len is the length of the extent
170 *
171 * The tree is given a single reference on the ordered extent that was
172 * inserted.
173 */
174static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
175 u64 start, u64 len, u64 disk_len,
176 int type, int dio, int compress_type)
177{
178 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
179 struct btrfs_root *root = BTRFS_I(inode)->root;
180 struct btrfs_ordered_inode_tree *tree;
181 struct rb_node *node;
182 struct btrfs_ordered_extent *entry;
183
184 tree = &BTRFS_I(inode)->ordered_tree;
185 entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
186 if (!entry)
187 return -ENOMEM;
188
189 entry->file_offset = file_offset;
190 entry->start = start;
191 entry->len = len;
192 entry->disk_len = disk_len;
193 entry->bytes_left = len;
194 entry->inode = igrab(inode);
195 entry->compress_type = compress_type;
196 entry->truncated_len = (u64)-1;
197 if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
198 set_bit(type, &entry->flags);
199
200 if (dio) {
201 percpu_counter_add_batch(&fs_info->dio_bytes, len,
202 fs_info->delalloc_batch);
203 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
204 }
205
206 /* one ref for the tree */
207 refcount_set(&entry->refs, 1);
208 init_waitqueue_head(&entry->wait);
209 INIT_LIST_HEAD(&entry->list);
210 INIT_LIST_HEAD(&entry->root_extent_list);
211 INIT_LIST_HEAD(&entry->work_list);
212 init_completion(&entry->completion);
213 INIT_LIST_HEAD(&entry->log_list);
214 INIT_LIST_HEAD(&entry->trans_list);
215
216 trace_btrfs_ordered_extent_add(inode, entry);
217
218 spin_lock_irq(&tree->lock);
219 node = tree_insert(&tree->tree, file_offset,
220 &entry->rb_node);
221 if (node)
222 ordered_data_tree_panic(inode, -EEXIST, file_offset);
223 spin_unlock_irq(&tree->lock);
224
225 spin_lock(&root->ordered_extent_lock);
226 list_add_tail(&entry->root_extent_list,
227 &root->ordered_extents);
228 root->nr_ordered_extents++;
229 if (root->nr_ordered_extents == 1) {
230 spin_lock(&fs_info->ordered_root_lock);
231 BUG_ON(!list_empty(&root->ordered_root));
232 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
233 spin_unlock(&fs_info->ordered_root_lock);
234 }
235 spin_unlock(&root->ordered_extent_lock);
236
237 /*
238 * We don't need the count_max_extents here, we can assume that all of
239 * that work has been done at higher layers, so this is truly the
240 * smallest the extent is going to get.
241 */
242 spin_lock(&BTRFS_I(inode)->lock);
243 btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
244 spin_unlock(&BTRFS_I(inode)->lock);
245
246 return 0;
247}
248
249int 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
257int 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
265int 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 */
279void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
280 struct btrfs_ordered_sum *sum)
281{
282 struct btrfs_ordered_inode_tree *tree;
283
284 tree = &BTRFS_I(entry->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 /* test_and_set_bit implies a barrier */
351 cond_wake_up_nomb(&entry->wait);
352 } else {
353 ret = 1;
354 }
355out:
356 if (!ret && cached && entry) {
357 *cached = entry;
358 refcount_inc(&entry->refs);
359 }
360 spin_unlock_irqrestore(&tree->lock, flags);
361 return ret == 0;
362}
363
364/*
365 * this is used to account for finished IO across a given range
366 * of the file. The IO should not span ordered extents. If
367 * a given ordered_extent is completely done, 1 is returned, otherwise
368 * 0.
369 *
370 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
371 * to make sure this function only returns 1 once for a given ordered extent.
372 */
373int btrfs_dec_test_ordered_pending(struct inode *inode,
374 struct btrfs_ordered_extent **cached,
375 u64 file_offset, u64 io_size, int uptodate)
376{
377 struct btrfs_ordered_inode_tree *tree;
378 struct rb_node *node;
379 struct btrfs_ordered_extent *entry = NULL;
380 unsigned long flags;
381 int ret;
382
383 tree = &BTRFS_I(inode)->ordered_tree;
384 spin_lock_irqsave(&tree->lock, flags);
385 if (cached && *cached) {
386 entry = *cached;
387 goto have_entry;
388 }
389
390 node = tree_search(tree, file_offset);
391 if (!node) {
392 ret = 1;
393 goto out;
394 }
395
396 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
397have_entry:
398 if (!offset_in_entry(entry, file_offset)) {
399 ret = 1;
400 goto out;
401 }
402
403 if (io_size > entry->bytes_left) {
404 btrfs_crit(BTRFS_I(inode)->root->fs_info,
405 "bad ordered accounting left %llu size %llu",
406 entry->bytes_left, io_size);
407 }
408 entry->bytes_left -= io_size;
409 if (!uptodate)
410 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
411
412 if (entry->bytes_left == 0) {
413 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
414 /* test_and_set_bit implies a barrier */
415 cond_wake_up_nomb(&entry->wait);
416 } else {
417 ret = 1;
418 }
419out:
420 if (!ret && cached && entry) {
421 *cached = entry;
422 refcount_inc(&entry->refs);
423 }
424 spin_unlock_irqrestore(&tree->lock, flags);
425 return ret == 0;
426}
427
428/*
429 * used to drop a reference on an ordered extent. This will free
430 * the extent if the last reference is dropped
431 */
432void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
433{
434 struct list_head *cur;
435 struct btrfs_ordered_sum *sum;
436
437 trace_btrfs_ordered_extent_put(entry->inode, entry);
438
439 if (refcount_dec_and_test(&entry->refs)) {
440 ASSERT(list_empty(&entry->log_list));
441 ASSERT(list_empty(&entry->trans_list));
442 ASSERT(list_empty(&entry->root_extent_list));
443 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
444 if (entry->inode)
445 btrfs_add_delayed_iput(entry->inode);
446 while (!list_empty(&entry->list)) {
447 cur = entry->list.next;
448 sum = list_entry(cur, struct btrfs_ordered_sum, list);
449 list_del(&sum->list);
450 kvfree(sum);
451 }
452 kmem_cache_free(btrfs_ordered_extent_cache, entry);
453 }
454}
455
456/*
457 * remove an ordered extent from the tree. No references are dropped
458 * and waiters are woken up.
459 */
460void btrfs_remove_ordered_extent(struct inode *inode,
461 struct btrfs_ordered_extent *entry)
462{
463 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
464 struct btrfs_ordered_inode_tree *tree;
465 struct btrfs_inode *btrfs_inode = BTRFS_I(inode);
466 struct btrfs_root *root = btrfs_inode->root;
467 struct rb_node *node;
468
469 /* This is paired with btrfs_add_ordered_extent. */
470 spin_lock(&btrfs_inode->lock);
471 btrfs_mod_outstanding_extents(btrfs_inode, -1);
472 spin_unlock(&btrfs_inode->lock);
473 if (root != fs_info->tree_root)
474 btrfs_delalloc_release_metadata(btrfs_inode, entry->len, false);
475
476 if (test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
477 percpu_counter_add_batch(&fs_info->dio_bytes, -entry->len,
478 fs_info->delalloc_batch);
479
480 tree = &btrfs_inode->ordered_tree;
481 spin_lock_irq(&tree->lock);
482 node = &entry->rb_node;
483 rb_erase(node, &tree->tree);
484 RB_CLEAR_NODE(node);
485 if (tree->last == node)
486 tree->last = NULL;
487 set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
488 spin_unlock_irq(&tree->lock);
489
490 spin_lock(&root->ordered_extent_lock);
491 list_del_init(&entry->root_extent_list);
492 root->nr_ordered_extents--;
493
494 trace_btrfs_ordered_extent_remove(inode, entry);
495
496 if (!root->nr_ordered_extents) {
497 spin_lock(&fs_info->ordered_root_lock);
498 BUG_ON(list_empty(&root->ordered_root));
499 list_del_init(&root->ordered_root);
500 spin_unlock(&fs_info->ordered_root_lock);
501 }
502 spin_unlock(&root->ordered_extent_lock);
503 wake_up(&entry->wait);
504}
505
506static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
507{
508 struct btrfs_ordered_extent *ordered;
509
510 ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
511 btrfs_start_ordered_extent(ordered->inode, ordered, 1);
512 complete(&ordered->completion);
513}
514
515/*
516 * wait for all the ordered extents in a root. This is done when balancing
517 * space between drives.
518 */
519u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
520 const u64 range_start, const u64 range_len)
521{
522 struct btrfs_fs_info *fs_info = root->fs_info;
523 LIST_HEAD(splice);
524 LIST_HEAD(skipped);
525 LIST_HEAD(works);
526 struct btrfs_ordered_extent *ordered, *next;
527 u64 count = 0;
528 const u64 range_end = range_start + range_len;
529
530 mutex_lock(&root->ordered_extent_mutex);
531 spin_lock(&root->ordered_extent_lock);
532 list_splice_init(&root->ordered_extents, &splice);
533 while (!list_empty(&splice) && nr) {
534 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
535 root_extent_list);
536
537 if (range_end <= ordered->start ||
538 ordered->start + ordered->disk_len <= range_start) {
539 list_move_tail(&ordered->root_extent_list, &skipped);
540 cond_resched_lock(&root->ordered_extent_lock);
541 continue;
542 }
543
544 list_move_tail(&ordered->root_extent_list,
545 &root->ordered_extents);
546 refcount_inc(&ordered->refs);
547 spin_unlock(&root->ordered_extent_lock);
548
549 btrfs_init_work(&ordered->flush_work,
550 btrfs_flush_delalloc_helper,
551 btrfs_run_ordered_extent_work, NULL, NULL);
552 list_add_tail(&ordered->work_list, &works);
553 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
554
555 cond_resched();
556 spin_lock(&root->ordered_extent_lock);
557 if (nr != U64_MAX)
558 nr--;
559 count++;
560 }
561 list_splice_tail(&skipped, &root->ordered_extents);
562 list_splice_tail(&splice, &root->ordered_extents);
563 spin_unlock(&root->ordered_extent_lock);
564
565 list_for_each_entry_safe(ordered, next, &works, work_list) {
566 list_del_init(&ordered->work_list);
567 wait_for_completion(&ordered->completion);
568 btrfs_put_ordered_extent(ordered);
569 cond_resched();
570 }
571 mutex_unlock(&root->ordered_extent_mutex);
572
573 return count;
574}
575
576u64 btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
577 const u64 range_start, const u64 range_len)
578{
579 struct btrfs_root *root;
580 struct list_head splice;
581 u64 total_done = 0;
582 u64 done;
583
584 INIT_LIST_HEAD(&splice);
585
586 mutex_lock(&fs_info->ordered_operations_mutex);
587 spin_lock(&fs_info->ordered_root_lock);
588 list_splice_init(&fs_info->ordered_roots, &splice);
589 while (!list_empty(&splice) && nr) {
590 root = list_first_entry(&splice, struct btrfs_root,
591 ordered_root);
592 root = btrfs_grab_fs_root(root);
593 BUG_ON(!root);
594 list_move_tail(&root->ordered_root,
595 &fs_info->ordered_roots);
596 spin_unlock(&fs_info->ordered_root_lock);
597
598 done = btrfs_wait_ordered_extents(root, nr,
599 range_start, range_len);
600 btrfs_put_fs_root(root);
601 total_done += done;
602
603 spin_lock(&fs_info->ordered_root_lock);
604 if (nr != U64_MAX) {
605 nr -= done;
606 }
607 }
608 list_splice_tail(&splice, &fs_info->ordered_roots);
609 spin_unlock(&fs_info->ordered_root_lock);
610 mutex_unlock(&fs_info->ordered_operations_mutex);
611
612 return total_done;
613}
614
615/*
616 * Used to start IO or wait for a given ordered extent to finish.
617 *
618 * If wait is one, this effectively waits on page writeback for all the pages
619 * in the extent, and it waits on the io completion code to insert
620 * metadata into the btree corresponding to the extent
621 */
622void btrfs_start_ordered_extent(struct inode *inode,
623 struct btrfs_ordered_extent *entry,
624 int wait)
625{
626 u64 start = entry->file_offset;
627 u64 end = start + entry->len - 1;
628
629 trace_btrfs_ordered_extent_start(inode, entry);
630
631 /*
632 * pages in the range can be dirty, clean or writeback. We
633 * start IO on any dirty ones so the wait doesn't stall waiting
634 * for the flusher thread to find them
635 */
636 if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
637 filemap_fdatawrite_range(inode->i_mapping, start, end);
638 if (wait) {
639 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
640 &entry->flags));
641 }
642}
643
644/*
645 * Used to wait on ordered extents across a large range of bytes.
646 */
647int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
648{
649 int ret = 0;
650 int ret_wb = 0;
651 u64 end;
652 u64 orig_end;
653 struct btrfs_ordered_extent *ordered;
654
655 if (start + len < start) {
656 orig_end = INT_LIMIT(loff_t);
657 } else {
658 orig_end = start + len - 1;
659 if (orig_end > INT_LIMIT(loff_t))
660 orig_end = INT_LIMIT(loff_t);
661 }
662
663 /* start IO across the range first to instantiate any delalloc
664 * extents
665 */
666 ret = btrfs_fdatawrite_range(inode, start, orig_end);
667 if (ret)
668 return ret;
669
670 /*
671 * If we have a writeback error don't return immediately. Wait first
672 * for any ordered extents that haven't completed yet. This is to make
673 * sure no one can dirty the same page ranges and call writepages()
674 * before the ordered extents complete - to avoid failures (-EEXIST)
675 * when adding the new ordered extents to the ordered tree.
676 */
677 ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
678
679 end = orig_end;
680 while (1) {
681 ordered = btrfs_lookup_first_ordered_extent(inode, end);
682 if (!ordered)
683 break;
684 if (ordered->file_offset > orig_end) {
685 btrfs_put_ordered_extent(ordered);
686 break;
687 }
688 if (ordered->file_offset + ordered->len <= start) {
689 btrfs_put_ordered_extent(ordered);
690 break;
691 }
692 btrfs_start_ordered_extent(inode, ordered, 1);
693 end = ordered->file_offset;
694 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
695 ret = -EIO;
696 btrfs_put_ordered_extent(ordered);
697 if (ret || end == 0 || end == start)
698 break;
699 end--;
700 }
701 return ret_wb ? ret_wb : ret;
702}
703
704/*
705 * find an ordered extent corresponding to file_offset. return NULL if
706 * nothing is found, otherwise take a reference on the extent and return it
707 */
708struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
709 u64 file_offset)
710{
711 struct btrfs_ordered_inode_tree *tree;
712 struct rb_node *node;
713 struct btrfs_ordered_extent *entry = NULL;
714
715 tree = &BTRFS_I(inode)->ordered_tree;
716 spin_lock_irq(&tree->lock);
717 node = tree_search(tree, file_offset);
718 if (!node)
719 goto out;
720
721 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
722 if (!offset_in_entry(entry, file_offset))
723 entry = NULL;
724 if (entry)
725 refcount_inc(&entry->refs);
726out:
727 spin_unlock_irq(&tree->lock);
728 return entry;
729}
730
731/* Since the DIO code tries to lock a wide area we need to look for any ordered
732 * extents that exist in the range, rather than just the start of the range.
733 */
734struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
735 struct btrfs_inode *inode, u64 file_offset, u64 len)
736{
737 struct btrfs_ordered_inode_tree *tree;
738 struct rb_node *node;
739 struct btrfs_ordered_extent *entry = NULL;
740
741 tree = &inode->ordered_tree;
742 spin_lock_irq(&tree->lock);
743 node = tree_search(tree, file_offset);
744 if (!node) {
745 node = tree_search(tree, file_offset + len);
746 if (!node)
747 goto out;
748 }
749
750 while (1) {
751 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
752 if (range_overlaps(entry, file_offset, len))
753 break;
754
755 if (entry->file_offset >= file_offset + len) {
756 entry = NULL;
757 break;
758 }
759 entry = NULL;
760 node = rb_next(node);
761 if (!node)
762 break;
763 }
764out:
765 if (entry)
766 refcount_inc(&entry->refs);
767 spin_unlock_irq(&tree->lock);
768 return entry;
769}
770
771/*
772 * lookup and return any extent before 'file_offset'. NULL is returned
773 * if none is found
774 */
775struct btrfs_ordered_extent *
776btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
777{
778 struct btrfs_ordered_inode_tree *tree;
779 struct rb_node *node;
780 struct btrfs_ordered_extent *entry = NULL;
781
782 tree = &BTRFS_I(inode)->ordered_tree;
783 spin_lock_irq(&tree->lock);
784 node = tree_search(tree, file_offset);
785 if (!node)
786 goto out;
787
788 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
789 refcount_inc(&entry->refs);
790out:
791 spin_unlock_irq(&tree->lock);
792 return entry;
793}
794
795/*
796 * After an extent is done, call this to conditionally update the on disk
797 * i_size. i_size is updated to cover any fully written part of the file.
798 */
799int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
800 struct btrfs_ordered_extent *ordered)
801{
802 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
803 u64 disk_i_size;
804 u64 new_i_size;
805 u64 i_size = i_size_read(inode);
806 struct rb_node *node;
807 struct rb_node *prev = NULL;
808 struct btrfs_ordered_extent *test;
809 int ret = 1;
810 u64 orig_offset = offset;
811
812 spin_lock_irq(&tree->lock);
813 if (ordered) {
814 offset = entry_end(ordered);
815 if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags))
816 offset = min(offset,
817 ordered->file_offset +
818 ordered->truncated_len);
819 } else {
820 offset = ALIGN(offset, btrfs_inode_sectorsize(inode));
821 }
822 disk_i_size = BTRFS_I(inode)->disk_i_size;
823
824 /*
825 * truncate file.
826 * If ordered is not NULL, then this is called from endio and
827 * disk_i_size will be updated by either truncate itself or any
828 * in-flight IOs which are inside the disk_i_size.
829 *
830 * Because btrfs_setsize() may set i_size with disk_i_size if truncate
831 * fails somehow, we need to make sure we have a precise disk_i_size by
832 * updating it as usual.
833 *
834 */
835 if (!ordered && disk_i_size > i_size) {
836 BTRFS_I(inode)->disk_i_size = orig_offset;
837 ret = 0;
838 goto out;
839 }
840
841 /*
842 * if the disk i_size is already at the inode->i_size, or
843 * this ordered extent is inside the disk i_size, we're done
844 */
845 if (disk_i_size == i_size)
846 goto out;
847
848 /*
849 * We still need to update disk_i_size if outstanding_isize is greater
850 * than disk_i_size.
851 */
852 if (offset <= disk_i_size &&
853 (!ordered || ordered->outstanding_isize <= disk_i_size))
854 goto out;
855
856 /*
857 * walk backward from this ordered extent to disk_i_size.
858 * if we find an ordered extent then we can't update disk i_size
859 * yet
860 */
861 if (ordered) {
862 node = rb_prev(&ordered->rb_node);
863 } else {
864 prev = tree_search(tree, offset);
865 /*
866 * we insert file extents without involving ordered struct,
867 * so there should be no ordered struct cover this offset
868 */
869 if (prev) {
870 test = rb_entry(prev, struct btrfs_ordered_extent,
871 rb_node);
872 BUG_ON(offset_in_entry(test, offset));
873 }
874 node = prev;
875 }
876 for (; node; node = rb_prev(node)) {
877 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
878
879 /* We treat this entry as if it doesn't exist */
880 if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
881 continue;
882
883 if (entry_end(test) <= disk_i_size)
884 break;
885 if (test->file_offset >= i_size)
886 break;
887
888 /*
889 * We don't update disk_i_size now, so record this undealt
890 * i_size. Or we will not know the real i_size.
891 */
892 if (test->outstanding_isize < offset)
893 test->outstanding_isize = offset;
894 if (ordered &&
895 ordered->outstanding_isize > test->outstanding_isize)
896 test->outstanding_isize = ordered->outstanding_isize;
897 goto out;
898 }
899 new_i_size = min_t(u64, offset, i_size);
900
901 /*
902 * Some ordered extents may completed before the current one, and
903 * we hold the real i_size in ->outstanding_isize.
904 */
905 if (ordered && ordered->outstanding_isize > new_i_size)
906 new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
907 BTRFS_I(inode)->disk_i_size = new_i_size;
908 ret = 0;
909out:
910 /*
911 * We need to do this because we can't remove ordered extents until
912 * after the i_disk_size has been updated and then the inode has been
913 * updated to reflect the change, so we need to tell anybody who finds
914 * this ordered extent that we've already done all the real work, we
915 * just haven't completed all the other work.
916 */
917 if (ordered)
918 set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
919 spin_unlock_irq(&tree->lock);
920 return ret;
921}
922
923/*
924 * search the ordered extents for one corresponding to 'offset' and
925 * try to find a checksum. This is used because we allow pages to
926 * be reclaimed before their checksum is actually put into the btree
927 */
928int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
929 u8 *sum, int len)
930{
931 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
932 struct btrfs_ordered_sum *ordered_sum;
933 struct btrfs_ordered_extent *ordered;
934 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
935 unsigned long num_sectors;
936 unsigned long i;
937 u32 sectorsize = btrfs_inode_sectorsize(inode);
938 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
939 int index = 0;
940
941 ordered = btrfs_lookup_ordered_extent(inode, offset);
942 if (!ordered)
943 return 0;
944
945 spin_lock_irq(&tree->lock);
946 list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
947 if (disk_bytenr >= ordered_sum->bytenr &&
948 disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
949 i = (disk_bytenr - ordered_sum->bytenr) >>
950 inode->i_sb->s_blocksize_bits;
951 num_sectors = ordered_sum->len >>
952 inode->i_sb->s_blocksize_bits;
953 num_sectors = min_t(int, len - index, num_sectors - i);
954 memcpy(sum + index, ordered_sum->sums + i * csum_size,
955 num_sectors * csum_size);
956
957 index += (int)num_sectors * csum_size;
958 if (index == len)
959 goto out;
960 disk_bytenr += num_sectors * sectorsize;
961 }
962 }
963out:
964 spin_unlock_irq(&tree->lock);
965 btrfs_put_ordered_extent(ordered);
966 return index;
967}
968
969/*
970 * btrfs_flush_ordered_range - Lock the passed range and ensures all pending
971 * ordered extents in it are run to completion.
972 *
973 * @tree: IO tree used for locking out other users of the range
974 * @inode: Inode whose ordered tree is to be searched
975 * @start: Beginning of range to flush
976 * @end: Last byte of range to lock
977 * @cached_state: If passed, will return the extent state responsible for the
978 * locked range. It's the caller's responsibility to free the cached state.
979 *
980 * This function always returns with the given range locked, ensuring after it's
981 * called no order extent can be pending.
982 */
983void btrfs_lock_and_flush_ordered_range(struct extent_io_tree *tree,
984 struct btrfs_inode *inode, u64 start,
985 u64 end,
986 struct extent_state **cached_state)
987{
988 struct btrfs_ordered_extent *ordered;
989 struct extent_state *cache = NULL;
990 struct extent_state **cachedp = &cache;
991
992 if (cached_state)
993 cachedp = cached_state;
994
995 while (1) {
996 lock_extent_bits(tree, start, end, cachedp);
997 ordered = btrfs_lookup_ordered_range(inode, start,
998 end - start + 1);
999 if (!ordered) {
1000 /*
1001 * If no external cached_state has been passed then
1002 * decrement the extra ref taken for cachedp since we
1003 * aren't exposing it outside of this function
1004 */
1005 if (!cached_state)
1006 refcount_dec(&cache->refs);
1007 break;
1008 }
1009 unlock_extent_cached(tree, start, end, cachedp);
1010 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1011 btrfs_put_ordered_extent(ordered);
1012 }
1013}
1014
1015int __init ordered_data_init(void)
1016{
1017 btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1018 sizeof(struct btrfs_ordered_extent), 0,
1019 SLAB_MEM_SPREAD,
1020 NULL);
1021 if (!btrfs_ordered_extent_cache)
1022 return -ENOMEM;
1023
1024 return 0;
1025}
1026
1027void __cold ordered_data_exit(void)
1028{
1029 kmem_cache_destroy(btrfs_ordered_extent_cache);
1030}
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
62static void ordered_data_tree_panic(struct inode *inode, int errno,
63 u64 offset)
64{
65 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
66 btrfs_panic(fs_info, errno, "Inconsistency in ordered tree at offset "
67 "%llu\n", (unsigned long long)offset);
68}
69
70/*
71 * look for a given offset in the tree, and if it can't be found return the
72 * first lesser offset
73 */
74static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
75 struct rb_node **prev_ret)
76{
77 struct rb_node *n = root->rb_node;
78 struct rb_node *prev = NULL;
79 struct rb_node *test;
80 struct btrfs_ordered_extent *entry;
81 struct btrfs_ordered_extent *prev_entry = NULL;
82
83 while (n) {
84 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
85 prev = n;
86 prev_entry = entry;
87
88 if (file_offset < entry->file_offset)
89 n = n->rb_left;
90 else if (file_offset >= entry_end(entry))
91 n = n->rb_right;
92 else
93 return n;
94 }
95 if (!prev_ret)
96 return NULL;
97
98 while (prev && file_offset >= entry_end(prev_entry)) {
99 test = rb_next(prev);
100 if (!test)
101 break;
102 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
103 rb_node);
104 if (file_offset < entry_end(prev_entry))
105 break;
106
107 prev = test;
108 }
109 if (prev)
110 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
111 rb_node);
112 while (prev && file_offset < entry_end(prev_entry)) {
113 test = rb_prev(prev);
114 if (!test)
115 break;
116 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
117 rb_node);
118 prev = test;
119 }
120 *prev_ret = prev;
121 return NULL;
122}
123
124/*
125 * helper to check if a given offset is inside a given entry
126 */
127static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
128{
129 if (file_offset < entry->file_offset ||
130 entry->file_offset + entry->len <= file_offset)
131 return 0;
132 return 1;
133}
134
135static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
136 u64 len)
137{
138 if (file_offset + len <= entry->file_offset ||
139 entry->file_offset + entry->len <= file_offset)
140 return 0;
141 return 1;
142}
143
144/*
145 * look find the first ordered struct that has this offset, otherwise
146 * the first one less than this offset
147 */
148static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
149 u64 file_offset)
150{
151 struct rb_root *root = &tree->tree;
152 struct rb_node *prev = NULL;
153 struct rb_node *ret;
154 struct btrfs_ordered_extent *entry;
155
156 if (tree->last) {
157 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
158 rb_node);
159 if (offset_in_entry(entry, file_offset))
160 return tree->last;
161 }
162 ret = __tree_search(root, file_offset, &prev);
163 if (!ret)
164 ret = prev;
165 if (ret)
166 tree->last = ret;
167 return ret;
168}
169
170/* allocate and add a new ordered_extent into the per-inode tree.
171 * file_offset is the logical offset in the file
172 *
173 * start is the disk block number of an extent already reserved in the
174 * extent allocation tree
175 *
176 * len is the length of the extent
177 *
178 * The tree is given a single reference on the ordered extent that was
179 * inserted.
180 */
181static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
182 u64 start, u64 len, u64 disk_len,
183 int type, int dio, int compress_type)
184{
185 struct btrfs_ordered_inode_tree *tree;
186 struct rb_node *node;
187 struct btrfs_ordered_extent *entry;
188
189 tree = &BTRFS_I(inode)->ordered_tree;
190 entry = kzalloc(sizeof(*entry), GFP_NOFS);
191 if (!entry)
192 return -ENOMEM;
193
194 entry->file_offset = file_offset;
195 entry->start = start;
196 entry->len = len;
197 entry->disk_len = disk_len;
198 entry->bytes_left = len;
199 entry->inode = igrab(inode);
200 entry->compress_type = compress_type;
201 if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
202 set_bit(type, &entry->flags);
203
204 if (dio)
205 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
206
207 /* one ref for the tree */
208 atomic_set(&entry->refs, 1);
209 init_waitqueue_head(&entry->wait);
210 INIT_LIST_HEAD(&entry->list);
211 INIT_LIST_HEAD(&entry->root_extent_list);
212
213 trace_btrfs_ordered_extent_add(inode, entry);
214
215 spin_lock_irq(&tree->lock);
216 node = tree_insert(&tree->tree, file_offset,
217 &entry->rb_node);
218 if (node)
219 ordered_data_tree_panic(inode, -EEXIST, file_offset);
220 spin_unlock_irq(&tree->lock);
221
222 spin_lock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
223 list_add_tail(&entry->root_extent_list,
224 &BTRFS_I(inode)->root->fs_info->ordered_extents);
225 spin_unlock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
226
227 return 0;
228}
229
230int btrfs_add_ordered_extent(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, 0,
235 BTRFS_COMPRESS_NONE);
236}
237
238int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
239 u64 start, u64 len, u64 disk_len, int type)
240{
241 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
242 disk_len, type, 1,
243 BTRFS_COMPRESS_NONE);
244}
245
246int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
247 u64 start, u64 len, u64 disk_len,
248 int type, int compress_type)
249{
250 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
251 disk_len, type, 0,
252 compress_type);
253}
254
255/*
256 * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
257 * when an ordered extent is finished. If the list covers more than one
258 * ordered extent, it is split across multiples.
259 */
260void btrfs_add_ordered_sum(struct inode *inode,
261 struct btrfs_ordered_extent *entry,
262 struct btrfs_ordered_sum *sum)
263{
264 struct btrfs_ordered_inode_tree *tree;
265
266 tree = &BTRFS_I(inode)->ordered_tree;
267 spin_lock_irq(&tree->lock);
268 list_add_tail(&sum->list, &entry->list);
269 spin_unlock_irq(&tree->lock);
270}
271
272/*
273 * this is used to account for finished IO across a given range
274 * of the file. The IO may span ordered extents. If
275 * a given ordered_extent is completely done, 1 is returned, otherwise
276 * 0.
277 *
278 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
279 * to make sure this function only returns 1 once for a given ordered extent.
280 *
281 * file_offset is updated to one byte past the range that is recorded as
282 * complete. This allows you to walk forward in the file.
283 */
284int btrfs_dec_test_first_ordered_pending(struct inode *inode,
285 struct btrfs_ordered_extent **cached,
286 u64 *file_offset, u64 io_size, int uptodate)
287{
288 struct btrfs_ordered_inode_tree *tree;
289 struct rb_node *node;
290 struct btrfs_ordered_extent *entry = NULL;
291 int ret;
292 unsigned long flags;
293 u64 dec_end;
294 u64 dec_start;
295 u64 to_dec;
296
297 tree = &BTRFS_I(inode)->ordered_tree;
298 spin_lock_irqsave(&tree->lock, flags);
299 node = tree_search(tree, *file_offset);
300 if (!node) {
301 ret = 1;
302 goto out;
303 }
304
305 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
306 if (!offset_in_entry(entry, *file_offset)) {
307 ret = 1;
308 goto out;
309 }
310
311 dec_start = max(*file_offset, entry->file_offset);
312 dec_end = min(*file_offset + io_size, entry->file_offset +
313 entry->len);
314 *file_offset = dec_end;
315 if (dec_start > dec_end) {
316 printk(KERN_CRIT "bad ordering dec_start %llu end %llu\n",
317 (unsigned long long)dec_start,
318 (unsigned long long)dec_end);
319 }
320 to_dec = dec_end - dec_start;
321 if (to_dec > entry->bytes_left) {
322 printk(KERN_CRIT "bad ordered accounting left %llu size %llu\n",
323 (unsigned long long)entry->bytes_left,
324 (unsigned long long)to_dec);
325 }
326 entry->bytes_left -= to_dec;
327 if (!uptodate)
328 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
329
330 if (entry->bytes_left == 0)
331 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
332 else
333 ret = 1;
334out:
335 if (!ret && cached && entry) {
336 *cached = entry;
337 atomic_inc(&entry->refs);
338 }
339 spin_unlock_irqrestore(&tree->lock, flags);
340 return ret == 0;
341}
342
343/*
344 * this is used to account for finished IO across a given range
345 * of the file. The IO should not span ordered extents. If
346 * a given ordered_extent is completely done, 1 is returned, otherwise
347 * 0.
348 *
349 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
350 * to make sure this function only returns 1 once for a given ordered extent.
351 */
352int btrfs_dec_test_ordered_pending(struct inode *inode,
353 struct btrfs_ordered_extent **cached,
354 u64 file_offset, u64 io_size, int uptodate)
355{
356 struct btrfs_ordered_inode_tree *tree;
357 struct rb_node *node;
358 struct btrfs_ordered_extent *entry = NULL;
359 unsigned long flags;
360 int ret;
361
362 tree = &BTRFS_I(inode)->ordered_tree;
363 spin_lock_irqsave(&tree->lock, flags);
364 if (cached && *cached) {
365 entry = *cached;
366 goto have_entry;
367 }
368
369 node = tree_search(tree, file_offset);
370 if (!node) {
371 ret = 1;
372 goto out;
373 }
374
375 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
376have_entry:
377 if (!offset_in_entry(entry, file_offset)) {
378 ret = 1;
379 goto out;
380 }
381
382 if (io_size > entry->bytes_left) {
383 printk(KERN_CRIT "bad ordered accounting left %llu size %llu\n",
384 (unsigned long long)entry->bytes_left,
385 (unsigned long long)io_size);
386 }
387 entry->bytes_left -= io_size;
388 if (!uptodate)
389 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
390
391 if (entry->bytes_left == 0)
392 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
393 else
394 ret = 1;
395out:
396 if (!ret && cached && entry) {
397 *cached = entry;
398 atomic_inc(&entry->refs);
399 }
400 spin_unlock_irqrestore(&tree->lock, flags);
401 return ret == 0;
402}
403
404/*
405 * used to drop a reference on an ordered extent. This will free
406 * the extent if the last reference is dropped
407 */
408void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
409{
410 struct list_head *cur;
411 struct btrfs_ordered_sum *sum;
412
413 trace_btrfs_ordered_extent_put(entry->inode, entry);
414
415 if (atomic_dec_and_test(&entry->refs)) {
416 if (entry->inode)
417 btrfs_add_delayed_iput(entry->inode);
418 while (!list_empty(&entry->list)) {
419 cur = entry->list.next;
420 sum = list_entry(cur, struct btrfs_ordered_sum, list);
421 list_del(&sum->list);
422 kfree(sum);
423 }
424 kfree(entry);
425 }
426}
427
428/*
429 * remove an ordered extent from the tree. No references are dropped
430 * and waiters are woken up.
431 */
432void btrfs_remove_ordered_extent(struct inode *inode,
433 struct btrfs_ordered_extent *entry)
434{
435 struct btrfs_ordered_inode_tree *tree;
436 struct btrfs_root *root = BTRFS_I(inode)->root;
437 struct rb_node *node;
438
439 tree = &BTRFS_I(inode)->ordered_tree;
440 spin_lock_irq(&tree->lock);
441 node = &entry->rb_node;
442 rb_erase(node, &tree->tree);
443 tree->last = NULL;
444 set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
445 spin_unlock_irq(&tree->lock);
446
447 spin_lock(&root->fs_info->ordered_extent_lock);
448 list_del_init(&entry->root_extent_list);
449
450 trace_btrfs_ordered_extent_remove(inode, entry);
451
452 /*
453 * we have no more ordered extents for this inode and
454 * no dirty pages. We can safely remove it from the
455 * list of ordered extents
456 */
457 if (RB_EMPTY_ROOT(&tree->tree) &&
458 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
459 list_del_init(&BTRFS_I(inode)->ordered_operations);
460 }
461 spin_unlock(&root->fs_info->ordered_extent_lock);
462 wake_up(&entry->wait);
463}
464
465/*
466 * wait for all the ordered extents in a root. This is done when balancing
467 * space between drives.
468 */
469void btrfs_wait_ordered_extents(struct btrfs_root *root,
470 int nocow_only, int delay_iput)
471{
472 struct list_head splice;
473 struct list_head *cur;
474 struct btrfs_ordered_extent *ordered;
475 struct inode *inode;
476
477 INIT_LIST_HEAD(&splice);
478
479 spin_lock(&root->fs_info->ordered_extent_lock);
480 list_splice_init(&root->fs_info->ordered_extents, &splice);
481 while (!list_empty(&splice)) {
482 cur = splice.next;
483 ordered = list_entry(cur, struct btrfs_ordered_extent,
484 root_extent_list);
485 if (nocow_only &&
486 !test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags) &&
487 !test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags)) {
488 list_move(&ordered->root_extent_list,
489 &root->fs_info->ordered_extents);
490 cond_resched_lock(&root->fs_info->ordered_extent_lock);
491 continue;
492 }
493
494 list_del_init(&ordered->root_extent_list);
495 atomic_inc(&ordered->refs);
496
497 /*
498 * the inode may be getting freed (in sys_unlink path).
499 */
500 inode = igrab(ordered->inode);
501
502 spin_unlock(&root->fs_info->ordered_extent_lock);
503
504 if (inode) {
505 btrfs_start_ordered_extent(inode, ordered, 1);
506 btrfs_put_ordered_extent(ordered);
507 if (delay_iput)
508 btrfs_add_delayed_iput(inode);
509 else
510 iput(inode);
511 } else {
512 btrfs_put_ordered_extent(ordered);
513 }
514
515 spin_lock(&root->fs_info->ordered_extent_lock);
516 }
517 spin_unlock(&root->fs_info->ordered_extent_lock);
518}
519
520/*
521 * this is used during transaction commit to write all the inodes
522 * added to the ordered operation list. These files must be fully on
523 * disk before the transaction commits.
524 *
525 * we have two modes here, one is to just start the IO via filemap_flush
526 * and the other is to wait for all the io. When we wait, we have an
527 * extra check to make sure the ordered operation list really is empty
528 * before we return
529 */
530void btrfs_run_ordered_operations(struct btrfs_root *root, int wait)
531{
532 struct btrfs_inode *btrfs_inode;
533 struct inode *inode;
534 struct list_head splice;
535
536 INIT_LIST_HEAD(&splice);
537
538 mutex_lock(&root->fs_info->ordered_operations_mutex);
539 spin_lock(&root->fs_info->ordered_extent_lock);
540again:
541 list_splice_init(&root->fs_info->ordered_operations, &splice);
542
543 while (!list_empty(&splice)) {
544 btrfs_inode = list_entry(splice.next, struct btrfs_inode,
545 ordered_operations);
546
547 inode = &btrfs_inode->vfs_inode;
548
549 list_del_init(&btrfs_inode->ordered_operations);
550
551 /*
552 * the inode may be getting freed (in sys_unlink path).
553 */
554 inode = igrab(inode);
555
556 if (!wait && inode) {
557 list_add_tail(&BTRFS_I(inode)->ordered_operations,
558 &root->fs_info->ordered_operations);
559 }
560 spin_unlock(&root->fs_info->ordered_extent_lock);
561
562 if (inode) {
563 if (wait)
564 btrfs_wait_ordered_range(inode, 0, (u64)-1);
565 else
566 filemap_flush(inode->i_mapping);
567 btrfs_add_delayed_iput(inode);
568 }
569
570 cond_resched();
571 spin_lock(&root->fs_info->ordered_extent_lock);
572 }
573 if (wait && !list_empty(&root->fs_info->ordered_operations))
574 goto again;
575
576 spin_unlock(&root->fs_info->ordered_extent_lock);
577 mutex_unlock(&root->fs_info->ordered_operations_mutex);
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 */
612void 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 }
626
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 /*
633 * So with compression we will find and lock a dirty page and clear the
634 * first one as dirty, setup an async extent, and immediately return
635 * with the entire range locked but with nobody actually marked with
636 * writeback. So we can't just filemap_write_and_wait_range() and
637 * expect it to work since it will just kick off a thread to do the
638 * actual work. So we need to call filemap_fdatawrite_range _again_
639 * since it will wait on the page lock, which won't be unlocked until
640 * after the pages have been marked as writeback and so we're good to go
641 * from there. We have to do this otherwise we'll miss the ordered
642 * extents and that results in badness. Please Josef, do not think you
643 * know better and pull this out at some point in the future, it is
644 * right and you are wrong.
645 */
646 if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
647 &BTRFS_I(inode)->runtime_flags))
648 filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
649
650 filemap_fdatawait_range(inode->i_mapping, start, orig_end);
651
652 end = orig_end;
653 found = 0;
654 while (1) {
655 ordered = btrfs_lookup_first_ordered_extent(inode, end);
656 if (!ordered)
657 break;
658 if (ordered->file_offset > orig_end) {
659 btrfs_put_ordered_extent(ordered);
660 break;
661 }
662 if (ordered->file_offset + ordered->len < start) {
663 btrfs_put_ordered_extent(ordered);
664 break;
665 }
666 found++;
667 btrfs_start_ordered_extent(inode, ordered, 1);
668 end = ordered->file_offset;
669 btrfs_put_ordered_extent(ordered);
670 if (end == 0 || end == start)
671 break;
672 end--;
673 }
674}
675
676/*
677 * find an ordered extent corresponding to file_offset. return NULL if
678 * nothing is found, otherwise take a reference on the extent and return it
679 */
680struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
681 u64 file_offset)
682{
683 struct btrfs_ordered_inode_tree *tree;
684 struct rb_node *node;
685 struct btrfs_ordered_extent *entry = NULL;
686
687 tree = &BTRFS_I(inode)->ordered_tree;
688 spin_lock_irq(&tree->lock);
689 node = tree_search(tree, file_offset);
690 if (!node)
691 goto out;
692
693 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
694 if (!offset_in_entry(entry, file_offset))
695 entry = NULL;
696 if (entry)
697 atomic_inc(&entry->refs);
698out:
699 spin_unlock_irq(&tree->lock);
700 return entry;
701}
702
703/* Since the DIO code tries to lock a wide area we need to look for any ordered
704 * extents that exist in the range, rather than just the start of the range.
705 */
706struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
707 u64 file_offset,
708 u64 len)
709{
710 struct btrfs_ordered_inode_tree *tree;
711 struct rb_node *node;
712 struct btrfs_ordered_extent *entry = NULL;
713
714 tree = &BTRFS_I(inode)->ordered_tree;
715 spin_lock_irq(&tree->lock);
716 node = tree_search(tree, file_offset);
717 if (!node) {
718 node = tree_search(tree, file_offset + len);
719 if (!node)
720 goto out;
721 }
722
723 while (1) {
724 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
725 if (range_overlaps(entry, file_offset, len))
726 break;
727
728 if (entry->file_offset >= file_offset + len) {
729 entry = NULL;
730 break;
731 }
732 entry = NULL;
733 node = rb_next(node);
734 if (!node)
735 break;
736 }
737out:
738 if (entry)
739 atomic_inc(&entry->refs);
740 spin_unlock_irq(&tree->lock);
741 return entry;
742}
743
744/*
745 * lookup and return any extent before 'file_offset'. NULL is returned
746 * if none is found
747 */
748struct btrfs_ordered_extent *
749btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
750{
751 struct btrfs_ordered_inode_tree *tree;
752 struct rb_node *node;
753 struct btrfs_ordered_extent *entry = NULL;
754
755 tree = &BTRFS_I(inode)->ordered_tree;
756 spin_lock_irq(&tree->lock);
757 node = tree_search(tree, file_offset);
758 if (!node)
759 goto out;
760
761 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
762 atomic_inc(&entry->refs);
763out:
764 spin_unlock_irq(&tree->lock);
765 return entry;
766}
767
768/*
769 * After an extent is done, call this to conditionally update the on disk
770 * i_size. i_size is updated to cover any fully written part of the file.
771 */
772int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
773 struct btrfs_ordered_extent *ordered)
774{
775 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
776 u64 disk_i_size;
777 u64 new_i_size;
778 u64 i_size_test;
779 u64 i_size = i_size_read(inode);
780 struct rb_node *node;
781 struct rb_node *prev = NULL;
782 struct btrfs_ordered_extent *test;
783 int ret = 1;
784
785 if (ordered)
786 offset = entry_end(ordered);
787 else
788 offset = ALIGN(offset, BTRFS_I(inode)->root->sectorsize);
789
790 spin_lock_irq(&tree->lock);
791 disk_i_size = BTRFS_I(inode)->disk_i_size;
792
793 /* truncate file */
794 if (disk_i_size > i_size) {
795 BTRFS_I(inode)->disk_i_size = i_size;
796 ret = 0;
797 goto out;
798 }
799
800 /*
801 * if the disk i_size is already at the inode->i_size, or
802 * this ordered extent is inside the disk i_size, we're done
803 */
804 if (disk_i_size == i_size || offset <= disk_i_size) {
805 goto out;
806 }
807
808 /*
809 * walk backward from this ordered extent to disk_i_size.
810 * if we find an ordered extent then we can't update disk i_size
811 * yet
812 */
813 if (ordered) {
814 node = rb_prev(&ordered->rb_node);
815 } else {
816 prev = tree_search(tree, offset);
817 /*
818 * we insert file extents without involving ordered struct,
819 * so there should be no ordered struct cover this offset
820 */
821 if (prev) {
822 test = rb_entry(prev, struct btrfs_ordered_extent,
823 rb_node);
824 BUG_ON(offset_in_entry(test, offset));
825 }
826 node = prev;
827 }
828 for (; node; node = rb_prev(node)) {
829 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
830
831 /* We treat this entry as if it doesnt exist */
832 if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
833 continue;
834 if (test->file_offset + test->len <= disk_i_size)
835 break;
836 if (test->file_offset >= i_size)
837 break;
838 if (test->file_offset >= disk_i_size)
839 goto out;
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
858 /*
859 * We are looking for an area between our current extent and the next
860 * ordered extent to update the i_size to. There are 3 cases here
861 *
862 * 1) We don't actually have anything and we can update to i_size.
863 * 2) We have stuff but they already did their i_size update so again we
864 * can just update to i_size.
865 * 3) We have an outstanding ordered extent so the most we can update
866 * our disk_i_size to is the start of the next offset.
867 */
868 i_size_test = i_size;
869 for (; node; node = rb_next(node)) {
870 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
871
872 if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
873 continue;
874 if (test->file_offset > offset) {
875 i_size_test = test->file_offset;
876 break;
877 }
878 }
879
880 /*
881 * i_size_test is the end of a region after this ordered
882 * extent where there are no ordered extents, we can safely set
883 * disk_i_size to this.
884 */
885 if (i_size_test > offset)
886 new_i_size = min_t(u64, i_size_test, i_size);
887 BTRFS_I(inode)->disk_i_size = new_i_size;
888 ret = 0;
889out:
890 /*
891 * We need to do this because we can't remove ordered extents until
892 * after the i_disk_size has been updated and then the inode has been
893 * updated to reflect the change, so we need to tell anybody who finds
894 * this ordered extent that we've already done all the real work, we
895 * just haven't completed all the other work.
896 */
897 if (ordered)
898 set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
899 spin_unlock_irq(&tree->lock);
900 return ret;
901}
902
903/*
904 * search the ordered extents for one corresponding to 'offset' and
905 * try to find a checksum. This is used because we allow pages to
906 * be reclaimed before their checksum is actually put into the btree
907 */
908int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
909 u32 *sum)
910{
911 struct btrfs_ordered_sum *ordered_sum;
912 struct btrfs_sector_sum *sector_sums;
913 struct btrfs_ordered_extent *ordered;
914 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
915 unsigned long num_sectors;
916 unsigned long i;
917 u32 sectorsize = BTRFS_I(inode)->root->sectorsize;
918 int ret = 1;
919
920 ordered = btrfs_lookup_ordered_extent(inode, offset);
921 if (!ordered)
922 return 1;
923
924 spin_lock_irq(&tree->lock);
925 list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
926 if (disk_bytenr >= ordered_sum->bytenr) {
927 num_sectors = ordered_sum->len / sectorsize;
928 sector_sums = ordered_sum->sums;
929 for (i = 0; i < num_sectors; i++) {
930 if (sector_sums[i].bytenr == disk_bytenr) {
931 *sum = sector_sums[i].sum;
932 ret = 0;
933 goto out;
934 }
935 }
936 }
937 }
938out:
939 spin_unlock_irq(&tree->lock);
940 btrfs_put_ordered_extent(ordered);
941 return ret;
942}
943
944
945/*
946 * add a given inode to the list of inodes that must be fully on
947 * disk before a transaction commit finishes.
948 *
949 * This basically gives us the ext3 style data=ordered mode, and it is mostly
950 * used to make sure renamed files are fully on disk.
951 *
952 * It is a noop if the inode is already fully on disk.
953 *
954 * If trans is not null, we'll do a friendly check for a transaction that
955 * is already flushing things and force the IO down ourselves.
956 */
957void btrfs_add_ordered_operation(struct btrfs_trans_handle *trans,
958 struct btrfs_root *root, struct inode *inode)
959{
960 u64 last_mod;
961
962 last_mod = max(BTRFS_I(inode)->generation, BTRFS_I(inode)->last_trans);
963
964 /*
965 * if this file hasn't been changed since the last transaction
966 * commit, we can safely return without doing anything
967 */
968 if (last_mod < root->fs_info->last_trans_committed)
969 return;
970
971 /*
972 * the transaction is already committing. Just start the IO and
973 * don't bother with all of this list nonsense
974 */
975 if (trans && root->fs_info->running_transaction->blocked) {
976 btrfs_wait_ordered_range(inode, 0, (u64)-1);
977 return;
978 }
979
980 spin_lock(&root->fs_info->ordered_extent_lock);
981 if (list_empty(&BTRFS_I(inode)->ordered_operations)) {
982 list_add_tail(&BTRFS_I(inode)->ordered_operations,
983 &root->fs_info->ordered_operations);
984 }
985 spin_unlock(&root->fs_info->ordered_extent_lock);
986}