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#include "disk-io.h"
28#include "compression.h"
29
30static struct kmem_cache *btrfs_ordered_extent_cache;
31
32static u64 entry_end(struct btrfs_ordered_extent *entry)
33{
34 if (entry->file_offset + entry->len < entry->file_offset)
35 return (u64)-1;
36 return entry->file_offset + entry->len;
37}
38
39/* returns NULL if the insertion worked, or it returns the node it did find
40 * in the tree
41 */
42static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
43 struct rb_node *node)
44{
45 struct rb_node **p = &root->rb_node;
46 struct rb_node *parent = NULL;
47 struct btrfs_ordered_extent *entry;
48
49 while (*p) {
50 parent = *p;
51 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
52
53 if (file_offset < entry->file_offset)
54 p = &(*p)->rb_left;
55 else if (file_offset >= entry_end(entry))
56 p = &(*p)->rb_right;
57 else
58 return parent;
59 }
60
61 rb_link_node(node, parent, p);
62 rb_insert_color(node, root);
63 return NULL;
64}
65
66static void ordered_data_tree_panic(struct inode *inode, int errno,
67 u64 offset)
68{
69 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
70 btrfs_panic(fs_info, errno, "Inconsistency in ordered tree at offset "
71 "%llu", offset);
72}
73
74/*
75 * look for a given offset in the tree, and if it can't be found return the
76 * first lesser offset
77 */
78static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
79 struct rb_node **prev_ret)
80{
81 struct rb_node *n = root->rb_node;
82 struct rb_node *prev = NULL;
83 struct rb_node *test;
84 struct btrfs_ordered_extent *entry;
85 struct btrfs_ordered_extent *prev_entry = NULL;
86
87 while (n) {
88 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
89 prev = n;
90 prev_entry = entry;
91
92 if (file_offset < entry->file_offset)
93 n = n->rb_left;
94 else if (file_offset >= entry_end(entry))
95 n = n->rb_right;
96 else
97 return n;
98 }
99 if (!prev_ret)
100 return NULL;
101
102 while (prev && file_offset >= entry_end(prev_entry)) {
103 test = rb_next(prev);
104 if (!test)
105 break;
106 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
107 rb_node);
108 if (file_offset < entry_end(prev_entry))
109 break;
110
111 prev = test;
112 }
113 if (prev)
114 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
115 rb_node);
116 while (prev && file_offset < entry_end(prev_entry)) {
117 test = rb_prev(prev);
118 if (!test)
119 break;
120 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
121 rb_node);
122 prev = test;
123 }
124 *prev_ret = prev;
125 return NULL;
126}
127
128/*
129 * helper to check if a given offset is inside a given entry
130 */
131static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
132{
133 if (file_offset < entry->file_offset ||
134 entry->file_offset + entry->len <= file_offset)
135 return 0;
136 return 1;
137}
138
139static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
140 u64 len)
141{
142 if (file_offset + len <= entry->file_offset ||
143 entry->file_offset + entry->len <= file_offset)
144 return 0;
145 return 1;
146}
147
148/*
149 * look find the first ordered struct that has this offset, otherwise
150 * the first one less than this offset
151 */
152static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
153 u64 file_offset)
154{
155 struct rb_root *root = &tree->tree;
156 struct rb_node *prev = NULL;
157 struct rb_node *ret;
158 struct btrfs_ordered_extent *entry;
159
160 if (tree->last) {
161 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
162 rb_node);
163 if (offset_in_entry(entry, file_offset))
164 return tree->last;
165 }
166 ret = __tree_search(root, file_offset, &prev);
167 if (!ret)
168 ret = prev;
169 if (ret)
170 tree->last = ret;
171 return ret;
172}
173
174/* allocate and add a new ordered_extent into the per-inode tree.
175 * file_offset is the logical offset in the file
176 *
177 * start is the disk block number of an extent already reserved in the
178 * extent allocation tree
179 *
180 * len is the length of the extent
181 *
182 * The tree is given a single reference on the ordered extent that was
183 * inserted.
184 */
185static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
186 u64 start, u64 len, u64 disk_len,
187 int type, int dio, int compress_type)
188{
189 struct btrfs_root *root = BTRFS_I(inode)->root;
190 struct btrfs_ordered_inode_tree *tree;
191 struct rb_node *node;
192 struct btrfs_ordered_extent *entry;
193
194 tree = &BTRFS_I(inode)->ordered_tree;
195 entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
196 if (!entry)
197 return -ENOMEM;
198
199 entry->file_offset = file_offset;
200 entry->start = start;
201 entry->len = len;
202 entry->disk_len = disk_len;
203 entry->bytes_left = len;
204 entry->inode = igrab(inode);
205 entry->compress_type = compress_type;
206 entry->truncated_len = (u64)-1;
207 if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
208 set_bit(type, &entry->flags);
209
210 if (dio)
211 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
212
213 /* one ref for the tree */
214 atomic_set(&entry->refs, 1);
215 init_waitqueue_head(&entry->wait);
216 INIT_LIST_HEAD(&entry->list);
217 INIT_LIST_HEAD(&entry->root_extent_list);
218 INIT_LIST_HEAD(&entry->work_list);
219 init_completion(&entry->completion);
220 INIT_LIST_HEAD(&entry->log_list);
221 INIT_LIST_HEAD(&entry->trans_list);
222
223 trace_btrfs_ordered_extent_add(inode, entry);
224
225 spin_lock_irq(&tree->lock);
226 node = tree_insert(&tree->tree, file_offset,
227 &entry->rb_node);
228 if (node)
229 ordered_data_tree_panic(inode, -EEXIST, file_offset);
230 spin_unlock_irq(&tree->lock);
231
232 spin_lock(&root->ordered_extent_lock);
233 list_add_tail(&entry->root_extent_list,
234 &root->ordered_extents);
235 root->nr_ordered_extents++;
236 if (root->nr_ordered_extents == 1) {
237 spin_lock(&root->fs_info->ordered_root_lock);
238 BUG_ON(!list_empty(&root->ordered_root));
239 list_add_tail(&root->ordered_root,
240 &root->fs_info->ordered_roots);
241 spin_unlock(&root->fs_info->ordered_root_lock);
242 }
243 spin_unlock(&root->ordered_extent_lock);
244
245 return 0;
246}
247
248int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
249 u64 start, u64 len, u64 disk_len, int type)
250{
251 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
252 disk_len, type, 0,
253 BTRFS_COMPRESS_NONE);
254}
255
256int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
257 u64 start, u64 len, u64 disk_len, int type)
258{
259 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
260 disk_len, type, 1,
261 BTRFS_COMPRESS_NONE);
262}
263
264int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
265 u64 start, u64 len, u64 disk_len,
266 int type, int compress_type)
267{
268 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
269 disk_len, type, 0,
270 compress_type);
271}
272
273/*
274 * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
275 * when an ordered extent is finished. If the list covers more than one
276 * ordered extent, it is split across multiples.
277 */
278void btrfs_add_ordered_sum(struct inode *inode,
279 struct btrfs_ordered_extent *entry,
280 struct btrfs_ordered_sum *sum)
281{
282 struct btrfs_ordered_inode_tree *tree;
283
284 tree = &BTRFS_I(inode)->ordered_tree;
285 spin_lock_irq(&tree->lock);
286 list_add_tail(&sum->list, &entry->list);
287 spin_unlock_irq(&tree->lock);
288}
289
290/*
291 * this is used to account for finished IO across a given range
292 * of the file. The IO may span ordered extents. If
293 * a given ordered_extent is completely done, 1 is returned, otherwise
294 * 0.
295 *
296 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
297 * to make sure this function only returns 1 once for a given ordered extent.
298 *
299 * file_offset is updated to one byte past the range that is recorded as
300 * complete. This allows you to walk forward in the file.
301 */
302int btrfs_dec_test_first_ordered_pending(struct inode *inode,
303 struct btrfs_ordered_extent **cached,
304 u64 *file_offset, u64 io_size, int uptodate)
305{
306 struct btrfs_ordered_inode_tree *tree;
307 struct rb_node *node;
308 struct btrfs_ordered_extent *entry = NULL;
309 int ret;
310 unsigned long flags;
311 u64 dec_end;
312 u64 dec_start;
313 u64 to_dec;
314
315 tree = &BTRFS_I(inode)->ordered_tree;
316 spin_lock_irqsave(&tree->lock, flags);
317 node = tree_search(tree, *file_offset);
318 if (!node) {
319 ret = 1;
320 goto out;
321 }
322
323 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
324 if (!offset_in_entry(entry, *file_offset)) {
325 ret = 1;
326 goto out;
327 }
328
329 dec_start = max(*file_offset, entry->file_offset);
330 dec_end = min(*file_offset + io_size, entry->file_offset +
331 entry->len);
332 *file_offset = dec_end;
333 if (dec_start > dec_end) {
334 btrfs_crit(BTRFS_I(inode)->root->fs_info,
335 "bad ordering dec_start %llu end %llu", dec_start, dec_end);
336 }
337 to_dec = dec_end - dec_start;
338 if (to_dec > entry->bytes_left) {
339 btrfs_crit(BTRFS_I(inode)->root->fs_info,
340 "bad ordered accounting left %llu size %llu",
341 entry->bytes_left, to_dec);
342 }
343 entry->bytes_left -= to_dec;
344 if (!uptodate)
345 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
346
347 if (entry->bytes_left == 0) {
348 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
349 /*
350 * Implicit memory barrier after test_and_set_bit
351 */
352 if (waitqueue_active(&entry->wait))
353 wake_up(&entry->wait);
354 } else {
355 ret = 1;
356 }
357out:
358 if (!ret && cached && entry) {
359 *cached = entry;
360 atomic_inc(&entry->refs);
361 }
362 spin_unlock_irqrestore(&tree->lock, flags);
363 return ret == 0;
364}
365
366/*
367 * this is used to account for finished IO across a given range
368 * of the file. The IO should not span ordered extents. If
369 * a given ordered_extent is completely done, 1 is returned, otherwise
370 * 0.
371 *
372 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
373 * to make sure this function only returns 1 once for a given ordered extent.
374 */
375int btrfs_dec_test_ordered_pending(struct inode *inode,
376 struct btrfs_ordered_extent **cached,
377 u64 file_offset, u64 io_size, int uptodate)
378{
379 struct btrfs_ordered_inode_tree *tree;
380 struct rb_node *node;
381 struct btrfs_ordered_extent *entry = NULL;
382 unsigned long flags;
383 int ret;
384
385 tree = &BTRFS_I(inode)->ordered_tree;
386 spin_lock_irqsave(&tree->lock, flags);
387 if (cached && *cached) {
388 entry = *cached;
389 goto have_entry;
390 }
391
392 node = tree_search(tree, file_offset);
393 if (!node) {
394 ret = 1;
395 goto out;
396 }
397
398 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
399have_entry:
400 if (!offset_in_entry(entry, file_offset)) {
401 ret = 1;
402 goto out;
403 }
404
405 if (io_size > entry->bytes_left) {
406 btrfs_crit(BTRFS_I(inode)->root->fs_info,
407 "bad ordered accounting left %llu size %llu",
408 entry->bytes_left, io_size);
409 }
410 entry->bytes_left -= io_size;
411 if (!uptodate)
412 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
413
414 if (entry->bytes_left == 0) {
415 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
416 /*
417 * Implicit memory barrier after test_and_set_bit
418 */
419 if (waitqueue_active(&entry->wait))
420 wake_up(&entry->wait);
421 } else {
422 ret = 1;
423 }
424out:
425 if (!ret && cached && entry) {
426 *cached = entry;
427 atomic_inc(&entry->refs);
428 }
429 spin_unlock_irqrestore(&tree->lock, flags);
430 return ret == 0;
431}
432
433/* Needs to either be called under a log transaction or the log_mutex */
434void btrfs_get_logged_extents(struct inode *inode,
435 struct list_head *logged_list,
436 const loff_t start,
437 const loff_t end)
438{
439 struct btrfs_ordered_inode_tree *tree;
440 struct btrfs_ordered_extent *ordered;
441 struct rb_node *n;
442 struct rb_node *prev;
443
444 tree = &BTRFS_I(inode)->ordered_tree;
445 spin_lock_irq(&tree->lock);
446 n = __tree_search(&tree->tree, end, &prev);
447 if (!n)
448 n = prev;
449 for (; n; n = rb_prev(n)) {
450 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
451 if (ordered->file_offset > end)
452 continue;
453 if (entry_end(ordered) <= start)
454 break;
455 if (test_and_set_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
456 continue;
457 list_add(&ordered->log_list, logged_list);
458 atomic_inc(&ordered->refs);
459 }
460 spin_unlock_irq(&tree->lock);
461}
462
463void btrfs_put_logged_extents(struct list_head *logged_list)
464{
465 struct btrfs_ordered_extent *ordered;
466
467 while (!list_empty(logged_list)) {
468 ordered = list_first_entry(logged_list,
469 struct btrfs_ordered_extent,
470 log_list);
471 list_del_init(&ordered->log_list);
472 btrfs_put_ordered_extent(ordered);
473 }
474}
475
476void btrfs_submit_logged_extents(struct list_head *logged_list,
477 struct btrfs_root *log)
478{
479 int index = log->log_transid % 2;
480
481 spin_lock_irq(&log->log_extents_lock[index]);
482 list_splice_tail(logged_list, &log->logged_list[index]);
483 spin_unlock_irq(&log->log_extents_lock[index]);
484}
485
486void btrfs_wait_logged_extents(struct btrfs_trans_handle *trans,
487 struct btrfs_root *log, u64 transid)
488{
489 struct btrfs_ordered_extent *ordered;
490 int index = transid % 2;
491
492 spin_lock_irq(&log->log_extents_lock[index]);
493 while (!list_empty(&log->logged_list[index])) {
494 struct inode *inode;
495 ordered = list_first_entry(&log->logged_list[index],
496 struct btrfs_ordered_extent,
497 log_list);
498 list_del_init(&ordered->log_list);
499 inode = ordered->inode;
500 spin_unlock_irq(&log->log_extents_lock[index]);
501
502 if (!test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags) &&
503 !test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags)) {
504 u64 start = ordered->file_offset;
505 u64 end = ordered->file_offset + ordered->len - 1;
506
507 WARN_ON(!inode);
508 filemap_fdatawrite_range(inode->i_mapping, start, end);
509 }
510 wait_event(ordered->wait, test_bit(BTRFS_ORDERED_IO_DONE,
511 &ordered->flags));
512
513 /*
514 * In order to keep us from losing our ordered extent
515 * information when committing the transaction we have to make
516 * sure that any logged extents are completed when we go to
517 * commit the transaction. To do this we simply increase the
518 * current transactions pending_ordered counter and decrement it
519 * when the ordered extent completes.
520 */
521 if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
522 struct btrfs_ordered_inode_tree *tree;
523
524 tree = &BTRFS_I(inode)->ordered_tree;
525 spin_lock_irq(&tree->lock);
526 if (!test_bit(BTRFS_ORDERED_COMPLETE, &ordered->flags)) {
527 set_bit(BTRFS_ORDERED_PENDING, &ordered->flags);
528 atomic_inc(&trans->transaction->pending_ordered);
529 }
530 spin_unlock_irq(&tree->lock);
531 }
532 btrfs_put_ordered_extent(ordered);
533 spin_lock_irq(&log->log_extents_lock[index]);
534 }
535 spin_unlock_irq(&log->log_extents_lock[index]);
536}
537
538void btrfs_free_logged_extents(struct btrfs_root *log, u64 transid)
539{
540 struct btrfs_ordered_extent *ordered;
541 int index = transid % 2;
542
543 spin_lock_irq(&log->log_extents_lock[index]);
544 while (!list_empty(&log->logged_list[index])) {
545 ordered = list_first_entry(&log->logged_list[index],
546 struct btrfs_ordered_extent,
547 log_list);
548 list_del_init(&ordered->log_list);
549 spin_unlock_irq(&log->log_extents_lock[index]);
550 btrfs_put_ordered_extent(ordered);
551 spin_lock_irq(&log->log_extents_lock[index]);
552 }
553 spin_unlock_irq(&log->log_extents_lock[index]);
554}
555
556/*
557 * used to drop a reference on an ordered extent. This will free
558 * the extent if the last reference is dropped
559 */
560void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
561{
562 struct list_head *cur;
563 struct btrfs_ordered_sum *sum;
564
565 trace_btrfs_ordered_extent_put(entry->inode, entry);
566
567 if (atomic_dec_and_test(&entry->refs)) {
568 ASSERT(list_empty(&entry->log_list));
569 ASSERT(list_empty(&entry->trans_list));
570 ASSERT(list_empty(&entry->root_extent_list));
571 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
572 if (entry->inode)
573 btrfs_add_delayed_iput(entry->inode);
574 while (!list_empty(&entry->list)) {
575 cur = entry->list.next;
576 sum = list_entry(cur, struct btrfs_ordered_sum, list);
577 list_del(&sum->list);
578 kfree(sum);
579 }
580 kmem_cache_free(btrfs_ordered_extent_cache, entry);
581 }
582}
583
584/*
585 * remove an ordered extent from the tree. No references are dropped
586 * and waiters are woken up.
587 */
588void btrfs_remove_ordered_extent(struct inode *inode,
589 struct btrfs_ordered_extent *entry)
590{
591 struct btrfs_ordered_inode_tree *tree;
592 struct btrfs_root *root = BTRFS_I(inode)->root;
593 struct rb_node *node;
594 bool dec_pending_ordered = false;
595
596 tree = &BTRFS_I(inode)->ordered_tree;
597 spin_lock_irq(&tree->lock);
598 node = &entry->rb_node;
599 rb_erase(node, &tree->tree);
600 RB_CLEAR_NODE(node);
601 if (tree->last == node)
602 tree->last = NULL;
603 set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
604 if (test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags))
605 dec_pending_ordered = true;
606 spin_unlock_irq(&tree->lock);
607
608 /*
609 * The current running transaction is waiting on us, we need to let it
610 * know that we're complete and wake it up.
611 */
612 if (dec_pending_ordered) {
613 struct btrfs_transaction *trans;
614
615 /*
616 * The checks for trans are just a formality, it should be set,
617 * but if it isn't we don't want to deref/assert under the spin
618 * lock, so be nice and check if trans is set, but ASSERT() so
619 * if it isn't set a developer will notice.
620 */
621 spin_lock(&root->fs_info->trans_lock);
622 trans = root->fs_info->running_transaction;
623 if (trans)
624 atomic_inc(&trans->use_count);
625 spin_unlock(&root->fs_info->trans_lock);
626
627 ASSERT(trans);
628 if (trans) {
629 if (atomic_dec_and_test(&trans->pending_ordered))
630 wake_up(&trans->pending_wait);
631 btrfs_put_transaction(trans);
632 }
633 }
634
635 spin_lock(&root->ordered_extent_lock);
636 list_del_init(&entry->root_extent_list);
637 root->nr_ordered_extents--;
638
639 trace_btrfs_ordered_extent_remove(inode, entry);
640
641 if (!root->nr_ordered_extents) {
642 spin_lock(&root->fs_info->ordered_root_lock);
643 BUG_ON(list_empty(&root->ordered_root));
644 list_del_init(&root->ordered_root);
645 spin_unlock(&root->fs_info->ordered_root_lock);
646 }
647 spin_unlock(&root->ordered_extent_lock);
648 wake_up(&entry->wait);
649}
650
651static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
652{
653 struct btrfs_ordered_extent *ordered;
654
655 ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
656 btrfs_start_ordered_extent(ordered->inode, ordered, 1);
657 complete(&ordered->completion);
658}
659
660/*
661 * wait for all the ordered extents in a root. This is done when balancing
662 * space between drives.
663 */
664int btrfs_wait_ordered_extents(struct btrfs_root *root, int nr)
665{
666 struct list_head splice, works;
667 struct btrfs_ordered_extent *ordered, *next;
668 int count = 0;
669
670 INIT_LIST_HEAD(&splice);
671 INIT_LIST_HEAD(&works);
672
673 mutex_lock(&root->ordered_extent_mutex);
674 spin_lock(&root->ordered_extent_lock);
675 list_splice_init(&root->ordered_extents, &splice);
676 while (!list_empty(&splice) && nr) {
677 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
678 root_extent_list);
679 list_move_tail(&ordered->root_extent_list,
680 &root->ordered_extents);
681 atomic_inc(&ordered->refs);
682 spin_unlock(&root->ordered_extent_lock);
683
684 btrfs_init_work(&ordered->flush_work,
685 btrfs_flush_delalloc_helper,
686 btrfs_run_ordered_extent_work, NULL, NULL);
687 list_add_tail(&ordered->work_list, &works);
688 btrfs_queue_work(root->fs_info->flush_workers,
689 &ordered->flush_work);
690
691 cond_resched();
692 spin_lock(&root->ordered_extent_lock);
693 if (nr != -1)
694 nr--;
695 count++;
696 }
697 list_splice_tail(&splice, &root->ordered_extents);
698 spin_unlock(&root->ordered_extent_lock);
699
700 list_for_each_entry_safe(ordered, next, &works, work_list) {
701 list_del_init(&ordered->work_list);
702 wait_for_completion(&ordered->completion);
703 btrfs_put_ordered_extent(ordered);
704 cond_resched();
705 }
706 mutex_unlock(&root->ordered_extent_mutex);
707
708 return count;
709}
710
711void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, int nr)
712{
713 struct btrfs_root *root;
714 struct list_head splice;
715 int done;
716
717 INIT_LIST_HEAD(&splice);
718
719 mutex_lock(&fs_info->ordered_operations_mutex);
720 spin_lock(&fs_info->ordered_root_lock);
721 list_splice_init(&fs_info->ordered_roots, &splice);
722 while (!list_empty(&splice) && nr) {
723 root = list_first_entry(&splice, struct btrfs_root,
724 ordered_root);
725 root = btrfs_grab_fs_root(root);
726 BUG_ON(!root);
727 list_move_tail(&root->ordered_root,
728 &fs_info->ordered_roots);
729 spin_unlock(&fs_info->ordered_root_lock);
730
731 done = btrfs_wait_ordered_extents(root, nr);
732 btrfs_put_fs_root(root);
733
734 spin_lock(&fs_info->ordered_root_lock);
735 if (nr != -1) {
736 nr -= done;
737 WARN_ON(nr < 0);
738 }
739 }
740 list_splice_tail(&splice, &fs_info->ordered_roots);
741 spin_unlock(&fs_info->ordered_root_lock);
742 mutex_unlock(&fs_info->ordered_operations_mutex);
743}
744
745/*
746 * Used to start IO or wait for a given ordered extent to finish.
747 *
748 * If wait is one, this effectively waits on page writeback for all the pages
749 * in the extent, and it waits on the io completion code to insert
750 * metadata into the btree corresponding to the extent
751 */
752void btrfs_start_ordered_extent(struct inode *inode,
753 struct btrfs_ordered_extent *entry,
754 int wait)
755{
756 u64 start = entry->file_offset;
757 u64 end = start + entry->len - 1;
758
759 trace_btrfs_ordered_extent_start(inode, entry);
760
761 /*
762 * pages in the range can be dirty, clean or writeback. We
763 * start IO on any dirty ones so the wait doesn't stall waiting
764 * for the flusher thread to find them
765 */
766 if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
767 filemap_fdatawrite_range(inode->i_mapping, start, end);
768 if (wait) {
769 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
770 &entry->flags));
771 }
772}
773
774/*
775 * Used to wait on ordered extents across a large range of bytes.
776 */
777int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
778{
779 int ret = 0;
780 int ret_wb = 0;
781 u64 end;
782 u64 orig_end;
783 struct btrfs_ordered_extent *ordered;
784
785 if (start + len < start) {
786 orig_end = INT_LIMIT(loff_t);
787 } else {
788 orig_end = start + len - 1;
789 if (orig_end > INT_LIMIT(loff_t))
790 orig_end = INT_LIMIT(loff_t);
791 }
792
793 /* start IO across the range first to instantiate any delalloc
794 * extents
795 */
796 ret = btrfs_fdatawrite_range(inode, start, orig_end);
797 if (ret)
798 return ret;
799
800 /*
801 * If we have a writeback error don't return immediately. Wait first
802 * for any ordered extents that haven't completed yet. This is to make
803 * sure no one can dirty the same page ranges and call writepages()
804 * before the ordered extents complete - to avoid failures (-EEXIST)
805 * when adding the new ordered extents to the ordered tree.
806 */
807 ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
808
809 end = orig_end;
810 while (1) {
811 ordered = btrfs_lookup_first_ordered_extent(inode, end);
812 if (!ordered)
813 break;
814 if (ordered->file_offset > orig_end) {
815 btrfs_put_ordered_extent(ordered);
816 break;
817 }
818 if (ordered->file_offset + ordered->len <= start) {
819 btrfs_put_ordered_extent(ordered);
820 break;
821 }
822 btrfs_start_ordered_extent(inode, ordered, 1);
823 end = ordered->file_offset;
824 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
825 ret = -EIO;
826 btrfs_put_ordered_extent(ordered);
827 if (ret || end == 0 || end == start)
828 break;
829 end--;
830 }
831 return ret_wb ? ret_wb : ret;
832}
833
834/*
835 * find an ordered extent corresponding to file_offset. return NULL if
836 * nothing is found, otherwise take a reference on the extent and return it
837 */
838struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
839 u64 file_offset)
840{
841 struct btrfs_ordered_inode_tree *tree;
842 struct rb_node *node;
843 struct btrfs_ordered_extent *entry = NULL;
844
845 tree = &BTRFS_I(inode)->ordered_tree;
846 spin_lock_irq(&tree->lock);
847 node = tree_search(tree, file_offset);
848 if (!node)
849 goto out;
850
851 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
852 if (!offset_in_entry(entry, file_offset))
853 entry = NULL;
854 if (entry)
855 atomic_inc(&entry->refs);
856out:
857 spin_unlock_irq(&tree->lock);
858 return entry;
859}
860
861/* Since the DIO code tries to lock a wide area we need to look for any ordered
862 * extents that exist in the range, rather than just the start of the range.
863 */
864struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
865 u64 file_offset,
866 u64 len)
867{
868 struct btrfs_ordered_inode_tree *tree;
869 struct rb_node *node;
870 struct btrfs_ordered_extent *entry = NULL;
871
872 tree = &BTRFS_I(inode)->ordered_tree;
873 spin_lock_irq(&tree->lock);
874 node = tree_search(tree, file_offset);
875 if (!node) {
876 node = tree_search(tree, file_offset + len);
877 if (!node)
878 goto out;
879 }
880
881 while (1) {
882 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
883 if (range_overlaps(entry, file_offset, len))
884 break;
885
886 if (entry->file_offset >= file_offset + len) {
887 entry = NULL;
888 break;
889 }
890 entry = NULL;
891 node = rb_next(node);
892 if (!node)
893 break;
894 }
895out:
896 if (entry)
897 atomic_inc(&entry->refs);
898 spin_unlock_irq(&tree->lock);
899 return entry;
900}
901
902bool btrfs_have_ordered_extents_in_range(struct inode *inode,
903 u64 file_offset,
904 u64 len)
905{
906 struct btrfs_ordered_extent *oe;
907
908 oe = btrfs_lookup_ordered_range(inode, file_offset, len);
909 if (oe) {
910 btrfs_put_ordered_extent(oe);
911 return true;
912 }
913 return false;
914}
915
916/*
917 * lookup and return any extent before 'file_offset'. NULL is returned
918 * if none is found
919 */
920struct btrfs_ordered_extent *
921btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
922{
923 struct btrfs_ordered_inode_tree *tree;
924 struct rb_node *node;
925 struct btrfs_ordered_extent *entry = NULL;
926
927 tree = &BTRFS_I(inode)->ordered_tree;
928 spin_lock_irq(&tree->lock);
929 node = tree_search(tree, file_offset);
930 if (!node)
931 goto out;
932
933 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
934 atomic_inc(&entry->refs);
935out:
936 spin_unlock_irq(&tree->lock);
937 return entry;
938}
939
940/*
941 * After an extent is done, call this to conditionally update the on disk
942 * i_size. i_size is updated to cover any fully written part of the file.
943 */
944int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
945 struct btrfs_ordered_extent *ordered)
946{
947 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
948 u64 disk_i_size;
949 u64 new_i_size;
950 u64 i_size = i_size_read(inode);
951 struct rb_node *node;
952 struct rb_node *prev = NULL;
953 struct btrfs_ordered_extent *test;
954 int ret = 1;
955
956 spin_lock_irq(&tree->lock);
957 if (ordered) {
958 offset = entry_end(ordered);
959 if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags))
960 offset = min(offset,
961 ordered->file_offset +
962 ordered->truncated_len);
963 } else {
964 offset = ALIGN(offset, BTRFS_I(inode)->root->sectorsize);
965 }
966 disk_i_size = BTRFS_I(inode)->disk_i_size;
967
968 /* truncate file */
969 if (disk_i_size > i_size) {
970 BTRFS_I(inode)->disk_i_size = i_size;
971 ret = 0;
972 goto out;
973 }
974
975 /*
976 * if the disk i_size is already at the inode->i_size, or
977 * this ordered extent is inside the disk i_size, we're done
978 */
979 if (disk_i_size == i_size)
980 goto out;
981
982 /*
983 * We still need to update disk_i_size if outstanding_isize is greater
984 * than disk_i_size.
985 */
986 if (offset <= disk_i_size &&
987 (!ordered || ordered->outstanding_isize <= disk_i_size))
988 goto out;
989
990 /*
991 * walk backward from this ordered extent to disk_i_size.
992 * if we find an ordered extent then we can't update disk i_size
993 * yet
994 */
995 if (ordered) {
996 node = rb_prev(&ordered->rb_node);
997 } else {
998 prev = tree_search(tree, offset);
999 /*
1000 * we insert file extents without involving ordered struct,
1001 * so there should be no ordered struct cover this offset
1002 */
1003 if (prev) {
1004 test = rb_entry(prev, struct btrfs_ordered_extent,
1005 rb_node);
1006 BUG_ON(offset_in_entry(test, offset));
1007 }
1008 node = prev;
1009 }
1010 for (; node; node = rb_prev(node)) {
1011 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1012
1013 /* We treat this entry as if it doesn't exist */
1014 if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
1015 continue;
1016 if (test->file_offset + test->len <= disk_i_size)
1017 break;
1018 if (test->file_offset >= i_size)
1019 break;
1020 if (entry_end(test) > disk_i_size) {
1021 /*
1022 * we don't update disk_i_size now, so record this
1023 * undealt i_size. Or we will not know the real
1024 * i_size.
1025 */
1026 if (test->outstanding_isize < offset)
1027 test->outstanding_isize = offset;
1028 if (ordered &&
1029 ordered->outstanding_isize >
1030 test->outstanding_isize)
1031 test->outstanding_isize =
1032 ordered->outstanding_isize;
1033 goto out;
1034 }
1035 }
1036 new_i_size = min_t(u64, offset, i_size);
1037
1038 /*
1039 * Some ordered extents may completed before the current one, and
1040 * we hold the real i_size in ->outstanding_isize.
1041 */
1042 if (ordered && ordered->outstanding_isize > new_i_size)
1043 new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
1044 BTRFS_I(inode)->disk_i_size = new_i_size;
1045 ret = 0;
1046out:
1047 /*
1048 * We need to do this because we can't remove ordered extents until
1049 * after the i_disk_size has been updated and then the inode has been
1050 * updated to reflect the change, so we need to tell anybody who finds
1051 * this ordered extent that we've already done all the real work, we
1052 * just haven't completed all the other work.
1053 */
1054 if (ordered)
1055 set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
1056 spin_unlock_irq(&tree->lock);
1057 return ret;
1058}
1059
1060/*
1061 * search the ordered extents for one corresponding to 'offset' and
1062 * try to find a checksum. This is used because we allow pages to
1063 * be reclaimed before their checksum is actually put into the btree
1064 */
1065int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
1066 u32 *sum, int len)
1067{
1068 struct btrfs_ordered_sum *ordered_sum;
1069 struct btrfs_ordered_extent *ordered;
1070 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1071 unsigned long num_sectors;
1072 unsigned long i;
1073 u32 sectorsize = BTRFS_I(inode)->root->sectorsize;
1074 int index = 0;
1075
1076 ordered = btrfs_lookup_ordered_extent(inode, offset);
1077 if (!ordered)
1078 return 0;
1079
1080 spin_lock_irq(&tree->lock);
1081 list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
1082 if (disk_bytenr >= ordered_sum->bytenr &&
1083 disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
1084 i = (disk_bytenr - ordered_sum->bytenr) >>
1085 inode->i_sb->s_blocksize_bits;
1086 num_sectors = ordered_sum->len >>
1087 inode->i_sb->s_blocksize_bits;
1088 num_sectors = min_t(int, len - index, num_sectors - i);
1089 memcpy(sum + index, ordered_sum->sums + i,
1090 num_sectors);
1091
1092 index += (int)num_sectors;
1093 if (index == len)
1094 goto out;
1095 disk_bytenr += num_sectors * sectorsize;
1096 }
1097 }
1098out:
1099 spin_unlock_irq(&tree->lock);
1100 btrfs_put_ordered_extent(ordered);
1101 return index;
1102}
1103
1104int __init ordered_data_init(void)
1105{
1106 btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1107 sizeof(struct btrfs_ordered_extent), 0,
1108 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
1109 NULL);
1110 if (!btrfs_ordered_extent_cache)
1111 return -ENOMEM;
1112
1113 return 0;
1114}
1115
1116void ordered_data_exit(void)
1117{
1118 kmem_cache_destroy(btrfs_ordered_extent_cache);
1119}
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
20static struct kmem_cache *btrfs_ordered_extent_cache;
21
22static u64 entry_end(struct btrfs_ordered_extent *entry)
23{
24 if (entry->file_offset + entry->num_bytes < entry->file_offset)
25 return (u64)-1;
26 return entry->file_offset + entry->num_bytes;
27}
28
29/* returns NULL if the insertion worked, or it returns the node it did find
30 * in the tree
31 */
32static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
33 struct rb_node *node)
34{
35 struct rb_node **p = &root->rb_node;
36 struct rb_node *parent = NULL;
37 struct btrfs_ordered_extent *entry;
38
39 while (*p) {
40 parent = *p;
41 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
42
43 if (file_offset < entry->file_offset)
44 p = &(*p)->rb_left;
45 else if (file_offset >= entry_end(entry))
46 p = &(*p)->rb_right;
47 else
48 return parent;
49 }
50
51 rb_link_node(node, parent, p);
52 rb_insert_color(node, root);
53 return NULL;
54}
55
56/*
57 * look for a given offset in the tree, and if it can't be found return the
58 * first lesser offset
59 */
60static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
61 struct rb_node **prev_ret)
62{
63 struct rb_node *n = root->rb_node;
64 struct rb_node *prev = NULL;
65 struct rb_node *test;
66 struct btrfs_ordered_extent *entry;
67 struct btrfs_ordered_extent *prev_entry = NULL;
68
69 while (n) {
70 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
71 prev = n;
72 prev_entry = entry;
73
74 if (file_offset < entry->file_offset)
75 n = n->rb_left;
76 else if (file_offset >= entry_end(entry))
77 n = n->rb_right;
78 else
79 return n;
80 }
81 if (!prev_ret)
82 return NULL;
83
84 while (prev && file_offset >= entry_end(prev_entry)) {
85 test = rb_next(prev);
86 if (!test)
87 break;
88 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
89 rb_node);
90 if (file_offset < entry_end(prev_entry))
91 break;
92
93 prev = test;
94 }
95 if (prev)
96 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
97 rb_node);
98 while (prev && file_offset < entry_end(prev_entry)) {
99 test = rb_prev(prev);
100 if (!test)
101 break;
102 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
103 rb_node);
104 prev = test;
105 }
106 *prev_ret = prev;
107 return NULL;
108}
109
110/*
111 * helper to check if a given offset is inside a given entry
112 */
113static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
114{
115 if (file_offset < entry->file_offset ||
116 entry->file_offset + entry->num_bytes <= file_offset)
117 return 0;
118 return 1;
119}
120
121static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
122 u64 len)
123{
124 if (file_offset + len <= entry->file_offset ||
125 entry->file_offset + entry->num_bytes <= file_offset)
126 return 0;
127 return 1;
128}
129
130/*
131 * look find the first ordered struct that has this offset, otherwise
132 * the first one less than this offset
133 */
134static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
135 u64 file_offset)
136{
137 struct rb_root *root = &tree->tree;
138 struct rb_node *prev = NULL;
139 struct rb_node *ret;
140 struct btrfs_ordered_extent *entry;
141
142 if (tree->last) {
143 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
144 rb_node);
145 if (offset_in_entry(entry, file_offset))
146 return tree->last;
147 }
148 ret = __tree_search(root, file_offset, &prev);
149 if (!ret)
150 ret = prev;
151 if (ret)
152 tree->last = ret;
153 return ret;
154}
155
156/*
157 * Allocate and add a new ordered_extent into the per-inode tree.
158 *
159 * The tree is given a single reference on the ordered extent that was
160 * inserted.
161 */
162static int __btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
163 u64 disk_bytenr, u64 num_bytes,
164 u64 disk_num_bytes, int type, int dio,
165 int compress_type)
166{
167 struct btrfs_root *root = inode->root;
168 struct btrfs_fs_info *fs_info = root->fs_info;
169 struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
170 struct rb_node *node;
171 struct btrfs_ordered_extent *entry;
172 int ret;
173
174 if (type == BTRFS_ORDERED_NOCOW || type == BTRFS_ORDERED_PREALLOC) {
175 /* For nocow write, we can release the qgroup rsv right now */
176 ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes);
177 if (ret < 0)
178 return ret;
179 ret = 0;
180 } else {
181 /*
182 * The ordered extent has reserved qgroup space, release now
183 * and pass the reserved number for qgroup_record to free.
184 */
185 ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes);
186 if (ret < 0)
187 return ret;
188 }
189 entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
190 if (!entry)
191 return -ENOMEM;
192
193 entry->file_offset = file_offset;
194 entry->disk_bytenr = disk_bytenr;
195 entry->num_bytes = num_bytes;
196 entry->disk_num_bytes = disk_num_bytes;
197 entry->bytes_left = num_bytes;
198 entry->inode = igrab(&inode->vfs_inode);
199 entry->compress_type = compress_type;
200 entry->truncated_len = (u64)-1;
201 entry->qgroup_rsv = ret;
202 if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
203 set_bit(type, &entry->flags);
204
205 if (dio) {
206 percpu_counter_add_batch(&fs_info->dio_bytes, num_bytes,
207 fs_info->delalloc_batch);
208 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
209 }
210
211 /* one ref for the tree */
212 refcount_set(&entry->refs, 1);
213 init_waitqueue_head(&entry->wait);
214 INIT_LIST_HEAD(&entry->list);
215 INIT_LIST_HEAD(&entry->root_extent_list);
216 INIT_LIST_HEAD(&entry->work_list);
217 init_completion(&entry->completion);
218
219 trace_btrfs_ordered_extent_add(&inode->vfs_inode, entry);
220
221 spin_lock_irq(&tree->lock);
222 node = tree_insert(&tree->tree, file_offset,
223 &entry->rb_node);
224 if (node)
225 btrfs_panic(fs_info, -EEXIST,
226 "inconsistency in ordered tree at offset %llu",
227 file_offset);
228 spin_unlock_irq(&tree->lock);
229
230 spin_lock(&root->ordered_extent_lock);
231 list_add_tail(&entry->root_extent_list,
232 &root->ordered_extents);
233 root->nr_ordered_extents++;
234 if (root->nr_ordered_extents == 1) {
235 spin_lock(&fs_info->ordered_root_lock);
236 BUG_ON(!list_empty(&root->ordered_root));
237 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
238 spin_unlock(&fs_info->ordered_root_lock);
239 }
240 spin_unlock(&root->ordered_extent_lock);
241
242 /*
243 * We don't need the count_max_extents here, we can assume that all of
244 * that work has been done at higher layers, so this is truly the
245 * smallest the extent is going to get.
246 */
247 spin_lock(&inode->lock);
248 btrfs_mod_outstanding_extents(inode, 1);
249 spin_unlock(&inode->lock);
250
251 return 0;
252}
253
254int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
255 u64 disk_bytenr, u64 num_bytes, u64 disk_num_bytes,
256 int type)
257{
258 return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
259 num_bytes, disk_num_bytes, type, 0,
260 BTRFS_COMPRESS_NONE);
261}
262
263int btrfs_add_ordered_extent_dio(struct btrfs_inode *inode, u64 file_offset,
264 u64 disk_bytenr, u64 num_bytes,
265 u64 disk_num_bytes, int type)
266{
267 return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
268 num_bytes, disk_num_bytes, type, 1,
269 BTRFS_COMPRESS_NONE);
270}
271
272int btrfs_add_ordered_extent_compress(struct btrfs_inode *inode, u64 file_offset,
273 u64 disk_bytenr, u64 num_bytes,
274 u64 disk_num_bytes, int type,
275 int compress_type)
276{
277 return __btrfs_add_ordered_extent(inode, file_offset, disk_bytenr,
278 num_bytes, disk_num_bytes, type, 0,
279 compress_type);
280}
281
282/*
283 * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
284 * when an ordered extent is finished. If the list covers more than one
285 * ordered extent, it is split across multiples.
286 */
287void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
288 struct btrfs_ordered_sum *sum)
289{
290 struct btrfs_ordered_inode_tree *tree;
291
292 tree = &BTRFS_I(entry->inode)->ordered_tree;
293 spin_lock_irq(&tree->lock);
294 list_add_tail(&sum->list, &entry->list);
295 spin_unlock_irq(&tree->lock);
296}
297
298/*
299 * this is used to account for finished IO across a given range
300 * of the file. The IO may span ordered extents. If
301 * a given ordered_extent is completely done, 1 is returned, otherwise
302 * 0.
303 *
304 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
305 * to make sure this function only returns 1 once for a given ordered extent.
306 *
307 * file_offset is updated to one byte past the range that is recorded as
308 * complete. This allows you to walk forward in the file.
309 */
310int btrfs_dec_test_first_ordered_pending(struct btrfs_inode *inode,
311 struct btrfs_ordered_extent **cached,
312 u64 *file_offset, u64 io_size, int uptodate)
313{
314 struct btrfs_fs_info *fs_info = inode->root->fs_info;
315 struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
316 struct rb_node *node;
317 struct btrfs_ordered_extent *entry = NULL;
318 int ret;
319 unsigned long flags;
320 u64 dec_end;
321 u64 dec_start;
322 u64 to_dec;
323
324 spin_lock_irqsave(&tree->lock, flags);
325 node = tree_search(tree, *file_offset);
326 if (!node) {
327 ret = 1;
328 goto out;
329 }
330
331 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
332 if (!offset_in_entry(entry, *file_offset)) {
333 ret = 1;
334 goto out;
335 }
336
337 dec_start = max(*file_offset, entry->file_offset);
338 dec_end = min(*file_offset + io_size,
339 entry->file_offset + entry->num_bytes);
340 *file_offset = dec_end;
341 if (dec_start > dec_end) {
342 btrfs_crit(fs_info, "bad ordering dec_start %llu end %llu",
343 dec_start, dec_end);
344 }
345 to_dec = dec_end - dec_start;
346 if (to_dec > entry->bytes_left) {
347 btrfs_crit(fs_info,
348 "bad ordered accounting left %llu size %llu",
349 entry->bytes_left, to_dec);
350 }
351 entry->bytes_left -= to_dec;
352 if (!uptodate)
353 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
354
355 if (entry->bytes_left == 0) {
356 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
357 /* test_and_set_bit implies a barrier */
358 cond_wake_up_nomb(&entry->wait);
359 } else {
360 ret = 1;
361 }
362out:
363 if (!ret && cached && entry) {
364 *cached = entry;
365 refcount_inc(&entry->refs);
366 }
367 spin_unlock_irqrestore(&tree->lock, flags);
368 return ret == 0;
369}
370
371/*
372 * this is used to account for finished IO across a given range
373 * of the file. The IO should not span ordered extents. If
374 * a given ordered_extent is completely done, 1 is returned, otherwise
375 * 0.
376 *
377 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
378 * to make sure this function only returns 1 once for a given ordered extent.
379 */
380int btrfs_dec_test_ordered_pending(struct inode *inode,
381 struct btrfs_ordered_extent **cached,
382 u64 file_offset, u64 io_size, int uptodate)
383{
384 struct btrfs_ordered_inode_tree *tree;
385 struct rb_node *node;
386 struct btrfs_ordered_extent *entry = NULL;
387 unsigned long flags;
388 int ret;
389
390 tree = &BTRFS_I(inode)->ordered_tree;
391 spin_lock_irqsave(&tree->lock, flags);
392 if (cached && *cached) {
393 entry = *cached;
394 goto have_entry;
395 }
396
397 node = tree_search(tree, file_offset);
398 if (!node) {
399 ret = 1;
400 goto out;
401 }
402
403 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
404have_entry:
405 if (!offset_in_entry(entry, file_offset)) {
406 ret = 1;
407 goto out;
408 }
409
410 if (io_size > entry->bytes_left) {
411 btrfs_crit(BTRFS_I(inode)->root->fs_info,
412 "bad ordered accounting left %llu size %llu",
413 entry->bytes_left, io_size);
414 }
415 entry->bytes_left -= io_size;
416 if (!uptodate)
417 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
418
419 if (entry->bytes_left == 0) {
420 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
421 /* test_and_set_bit implies a barrier */
422 cond_wake_up_nomb(&entry->wait);
423 } else {
424 ret = 1;
425 }
426out:
427 if (!ret && cached && entry) {
428 *cached = entry;
429 refcount_inc(&entry->refs);
430 }
431 spin_unlock_irqrestore(&tree->lock, flags);
432 return ret == 0;
433}
434
435/*
436 * used to drop a reference on an ordered extent. This will free
437 * the extent if the last reference is dropped
438 */
439void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
440{
441 struct list_head *cur;
442 struct btrfs_ordered_sum *sum;
443
444 trace_btrfs_ordered_extent_put(entry->inode, entry);
445
446 if (refcount_dec_and_test(&entry->refs)) {
447 ASSERT(list_empty(&entry->root_extent_list));
448 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
449 if (entry->inode)
450 btrfs_add_delayed_iput(entry->inode);
451 while (!list_empty(&entry->list)) {
452 cur = entry->list.next;
453 sum = list_entry(cur, struct btrfs_ordered_sum, list);
454 list_del(&sum->list);
455 kvfree(sum);
456 }
457 kmem_cache_free(btrfs_ordered_extent_cache, entry);
458 }
459}
460
461/*
462 * remove an ordered extent from the tree. No references are dropped
463 * and waiters are woken up.
464 */
465void btrfs_remove_ordered_extent(struct inode *inode,
466 struct btrfs_ordered_extent *entry)
467{
468 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
469 struct btrfs_ordered_inode_tree *tree;
470 struct btrfs_inode *btrfs_inode = BTRFS_I(inode);
471 struct btrfs_root *root = btrfs_inode->root;
472 struct rb_node *node;
473
474 /* This is paired with btrfs_add_ordered_extent. */
475 spin_lock(&btrfs_inode->lock);
476 btrfs_mod_outstanding_extents(btrfs_inode, -1);
477 spin_unlock(&btrfs_inode->lock);
478 if (root != fs_info->tree_root)
479 btrfs_delalloc_release_metadata(btrfs_inode, entry->num_bytes,
480 false);
481
482 if (test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
483 percpu_counter_add_batch(&fs_info->dio_bytes, -entry->num_bytes,
484 fs_info->delalloc_batch);
485
486 tree = &btrfs_inode->ordered_tree;
487 spin_lock_irq(&tree->lock);
488 node = &entry->rb_node;
489 rb_erase(node, &tree->tree);
490 RB_CLEAR_NODE(node);
491 if (tree->last == node)
492 tree->last = NULL;
493 set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
494 spin_unlock_irq(&tree->lock);
495
496 spin_lock(&root->ordered_extent_lock);
497 list_del_init(&entry->root_extent_list);
498 root->nr_ordered_extents--;
499
500 trace_btrfs_ordered_extent_remove(inode, entry);
501
502 if (!root->nr_ordered_extents) {
503 spin_lock(&fs_info->ordered_root_lock);
504 BUG_ON(list_empty(&root->ordered_root));
505 list_del_init(&root->ordered_root);
506 spin_unlock(&fs_info->ordered_root_lock);
507 }
508 spin_unlock(&root->ordered_extent_lock);
509 wake_up(&entry->wait);
510}
511
512static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
513{
514 struct btrfs_ordered_extent *ordered;
515
516 ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
517 btrfs_start_ordered_extent(ordered->inode, ordered, 1);
518 complete(&ordered->completion);
519}
520
521/*
522 * wait for all the ordered extents in a root. This is done when balancing
523 * space between drives.
524 */
525u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
526 const u64 range_start, const u64 range_len)
527{
528 struct btrfs_fs_info *fs_info = root->fs_info;
529 LIST_HEAD(splice);
530 LIST_HEAD(skipped);
531 LIST_HEAD(works);
532 struct btrfs_ordered_extent *ordered, *next;
533 u64 count = 0;
534 const u64 range_end = range_start + range_len;
535
536 mutex_lock(&root->ordered_extent_mutex);
537 spin_lock(&root->ordered_extent_lock);
538 list_splice_init(&root->ordered_extents, &splice);
539 while (!list_empty(&splice) && nr) {
540 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
541 root_extent_list);
542
543 if (range_end <= ordered->disk_bytenr ||
544 ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
545 list_move_tail(&ordered->root_extent_list, &skipped);
546 cond_resched_lock(&root->ordered_extent_lock);
547 continue;
548 }
549
550 list_move_tail(&ordered->root_extent_list,
551 &root->ordered_extents);
552 refcount_inc(&ordered->refs);
553 spin_unlock(&root->ordered_extent_lock);
554
555 btrfs_init_work(&ordered->flush_work,
556 btrfs_run_ordered_extent_work, NULL, NULL);
557 list_add_tail(&ordered->work_list, &works);
558 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
559
560 cond_resched();
561 spin_lock(&root->ordered_extent_lock);
562 if (nr != U64_MAX)
563 nr--;
564 count++;
565 }
566 list_splice_tail(&skipped, &root->ordered_extents);
567 list_splice_tail(&splice, &root->ordered_extents);
568 spin_unlock(&root->ordered_extent_lock);
569
570 list_for_each_entry_safe(ordered, next, &works, work_list) {
571 list_del_init(&ordered->work_list);
572 wait_for_completion(&ordered->completion);
573 btrfs_put_ordered_extent(ordered);
574 cond_resched();
575 }
576 mutex_unlock(&root->ordered_extent_mutex);
577
578 return count;
579}
580
581void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
582 const u64 range_start, const u64 range_len)
583{
584 struct btrfs_root *root;
585 struct list_head splice;
586 u64 done;
587
588 INIT_LIST_HEAD(&splice);
589
590 mutex_lock(&fs_info->ordered_operations_mutex);
591 spin_lock(&fs_info->ordered_root_lock);
592 list_splice_init(&fs_info->ordered_roots, &splice);
593 while (!list_empty(&splice) && nr) {
594 root = list_first_entry(&splice, struct btrfs_root,
595 ordered_root);
596 root = btrfs_grab_root(root);
597 BUG_ON(!root);
598 list_move_tail(&root->ordered_root,
599 &fs_info->ordered_roots);
600 spin_unlock(&fs_info->ordered_root_lock);
601
602 done = btrfs_wait_ordered_extents(root, nr,
603 range_start, range_len);
604 btrfs_put_root(root);
605
606 spin_lock(&fs_info->ordered_root_lock);
607 if (nr != U64_MAX) {
608 nr -= done;
609 }
610 }
611 list_splice_tail(&splice, &fs_info->ordered_roots);
612 spin_unlock(&fs_info->ordered_root_lock);
613 mutex_unlock(&fs_info->ordered_operations_mutex);
614}
615
616/*
617 * Used to start IO or wait for a given ordered extent to finish.
618 *
619 * If wait is one, this effectively waits on page writeback for all the pages
620 * in the extent, and it waits on the io completion code to insert
621 * metadata into the btree corresponding to the extent
622 */
623void btrfs_start_ordered_extent(struct inode *inode,
624 struct btrfs_ordered_extent *entry,
625 int wait)
626{
627 u64 start = entry->file_offset;
628 u64 end = start + entry->num_bytes - 1;
629
630 trace_btrfs_ordered_extent_start(inode, entry);
631
632 /*
633 * pages in the range can be dirty, clean or writeback. We
634 * start IO on any dirty ones so the wait doesn't stall waiting
635 * for the flusher thread to find them
636 */
637 if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
638 filemap_fdatawrite_range(inode->i_mapping, start, end);
639 if (wait) {
640 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
641 &entry->flags));
642 }
643}
644
645/*
646 * Used to wait on ordered extents across a large range of bytes.
647 */
648int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
649{
650 int ret = 0;
651 int ret_wb = 0;
652 u64 end;
653 u64 orig_end;
654 struct btrfs_ordered_extent *ordered;
655
656 if (start + len < start) {
657 orig_end = INT_LIMIT(loff_t);
658 } else {
659 orig_end = start + len - 1;
660 if (orig_end > INT_LIMIT(loff_t))
661 orig_end = INT_LIMIT(loff_t);
662 }
663
664 /* start IO across the range first to instantiate any delalloc
665 * extents
666 */
667 ret = btrfs_fdatawrite_range(inode, start, orig_end);
668 if (ret)
669 return ret;
670
671 /*
672 * If we have a writeback error don't return immediately. Wait first
673 * for any ordered extents that haven't completed yet. This is to make
674 * sure no one can dirty the same page ranges and call writepages()
675 * before the ordered extents complete - to avoid failures (-EEXIST)
676 * when adding the new ordered extents to the ordered tree.
677 */
678 ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
679
680 end = orig_end;
681 while (1) {
682 ordered = btrfs_lookup_first_ordered_extent(inode, end);
683 if (!ordered)
684 break;
685 if (ordered->file_offset > orig_end) {
686 btrfs_put_ordered_extent(ordered);
687 break;
688 }
689 if (ordered->file_offset + ordered->num_bytes <= start) {
690 btrfs_put_ordered_extent(ordered);
691 break;
692 }
693 btrfs_start_ordered_extent(inode, ordered, 1);
694 end = ordered->file_offset;
695 /*
696 * If the ordered extent had an error save the error but don't
697 * exit without waiting first for all other ordered extents in
698 * the range to complete.
699 */
700 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
701 ret = -EIO;
702 btrfs_put_ordered_extent(ordered);
703 if (end == 0 || end == start)
704 break;
705 end--;
706 }
707 return ret_wb ? ret_wb : ret;
708}
709
710/*
711 * find an ordered extent corresponding to file_offset. return NULL if
712 * nothing is found, otherwise take a reference on the extent and return it
713 */
714struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
715 u64 file_offset)
716{
717 struct btrfs_ordered_inode_tree *tree;
718 struct rb_node *node;
719 struct btrfs_ordered_extent *entry = NULL;
720
721 tree = &inode->ordered_tree;
722 spin_lock_irq(&tree->lock);
723 node = tree_search(tree, file_offset);
724 if (!node)
725 goto out;
726
727 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
728 if (!offset_in_entry(entry, file_offset))
729 entry = NULL;
730 if (entry)
731 refcount_inc(&entry->refs);
732out:
733 spin_unlock_irq(&tree->lock);
734 return entry;
735}
736
737/* Since the DIO code tries to lock a wide area we need to look for any ordered
738 * extents that exist in the range, rather than just the start of the range.
739 */
740struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
741 struct btrfs_inode *inode, u64 file_offset, u64 len)
742{
743 struct btrfs_ordered_inode_tree *tree;
744 struct rb_node *node;
745 struct btrfs_ordered_extent *entry = NULL;
746
747 tree = &inode->ordered_tree;
748 spin_lock_irq(&tree->lock);
749 node = tree_search(tree, file_offset);
750 if (!node) {
751 node = tree_search(tree, file_offset + len);
752 if (!node)
753 goto out;
754 }
755
756 while (1) {
757 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
758 if (range_overlaps(entry, file_offset, len))
759 break;
760
761 if (entry->file_offset >= file_offset + len) {
762 entry = NULL;
763 break;
764 }
765 entry = NULL;
766 node = rb_next(node);
767 if (!node)
768 break;
769 }
770out:
771 if (entry)
772 refcount_inc(&entry->refs);
773 spin_unlock_irq(&tree->lock);
774 return entry;
775}
776
777/*
778 * lookup and return any extent before 'file_offset'. NULL is returned
779 * if none is found
780 */
781struct btrfs_ordered_extent *
782btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
783{
784 struct btrfs_ordered_inode_tree *tree;
785 struct rb_node *node;
786 struct btrfs_ordered_extent *entry = NULL;
787
788 tree = &BTRFS_I(inode)->ordered_tree;
789 spin_lock_irq(&tree->lock);
790 node = tree_search(tree, file_offset);
791 if (!node)
792 goto out;
793
794 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
795 refcount_inc(&entry->refs);
796out:
797 spin_unlock_irq(&tree->lock);
798 return entry;
799}
800
801/*
802 * search the ordered extents for one corresponding to 'offset' and
803 * try to find a checksum. This is used because we allow pages to
804 * be reclaimed before their checksum is actually put into the btree
805 */
806int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
807 u8 *sum, int len)
808{
809 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
810 struct btrfs_ordered_sum *ordered_sum;
811 struct btrfs_ordered_extent *ordered;
812 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
813 unsigned long num_sectors;
814 unsigned long i;
815 u32 sectorsize = btrfs_inode_sectorsize(inode);
816 const u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
817 int index = 0;
818
819 ordered = btrfs_lookup_ordered_extent(BTRFS_I(inode), offset);
820 if (!ordered)
821 return 0;
822
823 spin_lock_irq(&tree->lock);
824 list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
825 if (disk_bytenr >= ordered_sum->bytenr &&
826 disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
827 i = (disk_bytenr - ordered_sum->bytenr) >>
828 inode->i_sb->s_blocksize_bits;
829 num_sectors = ordered_sum->len >>
830 inode->i_sb->s_blocksize_bits;
831 num_sectors = min_t(int, len - index, num_sectors - i);
832 memcpy(sum + index, ordered_sum->sums + i * csum_size,
833 num_sectors * csum_size);
834
835 index += (int)num_sectors * csum_size;
836 if (index == len)
837 goto out;
838 disk_bytenr += num_sectors * sectorsize;
839 }
840 }
841out:
842 spin_unlock_irq(&tree->lock);
843 btrfs_put_ordered_extent(ordered);
844 return index;
845}
846
847/*
848 * btrfs_flush_ordered_range - Lock the passed range and ensures all pending
849 * ordered extents in it are run to completion.
850 *
851 * @inode: Inode whose ordered tree is to be searched
852 * @start: Beginning of range to flush
853 * @end: Last byte of range to lock
854 * @cached_state: If passed, will return the extent state responsible for the
855 * locked range. It's the caller's responsibility to free the cached state.
856 *
857 * This function always returns with the given range locked, ensuring after it's
858 * called no order extent can be pending.
859 */
860void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
861 u64 end,
862 struct extent_state **cached_state)
863{
864 struct btrfs_ordered_extent *ordered;
865 struct extent_state *cache = NULL;
866 struct extent_state **cachedp = &cache;
867
868 if (cached_state)
869 cachedp = cached_state;
870
871 while (1) {
872 lock_extent_bits(&inode->io_tree, start, end, cachedp);
873 ordered = btrfs_lookup_ordered_range(inode, start,
874 end - start + 1);
875 if (!ordered) {
876 /*
877 * If no external cached_state has been passed then
878 * decrement the extra ref taken for cachedp since we
879 * aren't exposing it outside of this function
880 */
881 if (!cached_state)
882 refcount_dec(&cache->refs);
883 break;
884 }
885 unlock_extent_cached(&inode->io_tree, start, end, cachedp);
886 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
887 btrfs_put_ordered_extent(ordered);
888 }
889}
890
891int __init ordered_data_init(void)
892{
893 btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
894 sizeof(struct btrfs_ordered_extent), 0,
895 SLAB_MEM_SPREAD,
896 NULL);
897 if (!btrfs_ordered_extent_cache)
898 return -ENOMEM;
899
900 return 0;
901}
902
903void __cold ordered_data_exit(void)
904{
905 kmem_cache_destroy(btrfs_ordered_extent_cache);
906}