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