Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6#include <linux/slab.h>
7#include <linux/blkdev.h>
8#include <linux/writeback.h>
9#include <linux/sched/mm.h>
10#include "messages.h"
11#include "misc.h"
12#include "ctree.h"
13#include "transaction.h"
14#include "btrfs_inode.h"
15#include "extent_io.h"
16#include "disk-io.h"
17#include "compression.h"
18#include "delalloc-space.h"
19#include "qgroup.h"
20#include "subpage.h"
21#include "file.h"
22#include "super.h"
23
24static struct kmem_cache *btrfs_ordered_extent_cache;
25
26static u64 entry_end(struct btrfs_ordered_extent *entry)
27{
28 if (entry->file_offset + entry->num_bytes < entry->file_offset)
29 return (u64)-1;
30 return entry->file_offset + entry->num_bytes;
31}
32
33/* returns NULL if the insertion worked, or it returns the node it did find
34 * in the tree
35 */
36static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
37 struct rb_node *node)
38{
39 struct rb_node **p = &root->rb_node;
40 struct rb_node *parent = NULL;
41 struct btrfs_ordered_extent *entry;
42
43 while (*p) {
44 parent = *p;
45 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
46
47 if (file_offset < entry->file_offset)
48 p = &(*p)->rb_left;
49 else if (file_offset >= entry_end(entry))
50 p = &(*p)->rb_right;
51 else
52 return parent;
53 }
54
55 rb_link_node(node, parent, p);
56 rb_insert_color(node, root);
57 return NULL;
58}
59
60/*
61 * look for a given offset in the tree, and if it can't be found return the
62 * first lesser offset
63 */
64static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
65 struct rb_node **prev_ret)
66{
67 struct rb_node *n = root->rb_node;
68 struct rb_node *prev = NULL;
69 struct rb_node *test;
70 struct btrfs_ordered_extent *entry;
71 struct btrfs_ordered_extent *prev_entry = NULL;
72
73 while (n) {
74 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
75 prev = n;
76 prev_entry = entry;
77
78 if (file_offset < entry->file_offset)
79 n = n->rb_left;
80 else if (file_offset >= entry_end(entry))
81 n = n->rb_right;
82 else
83 return n;
84 }
85 if (!prev_ret)
86 return NULL;
87
88 while (prev && file_offset >= entry_end(prev_entry)) {
89 test = rb_next(prev);
90 if (!test)
91 break;
92 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
93 rb_node);
94 if (file_offset < entry_end(prev_entry))
95 break;
96
97 prev = test;
98 }
99 if (prev)
100 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
101 rb_node);
102 while (prev && file_offset < entry_end(prev_entry)) {
103 test = rb_prev(prev);
104 if (!test)
105 break;
106 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
107 rb_node);
108 prev = test;
109 }
110 *prev_ret = prev;
111 return NULL;
112}
113
114static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
115 u64 len)
116{
117 if (file_offset + len <= entry->file_offset ||
118 entry->file_offset + entry->num_bytes <= file_offset)
119 return 0;
120 return 1;
121}
122
123/*
124 * look find the first ordered struct that has this offset, otherwise
125 * the first one less than this offset
126 */
127static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
128 u64 file_offset)
129{
130 struct rb_root *root = &tree->tree;
131 struct rb_node *prev = NULL;
132 struct rb_node *ret;
133 struct btrfs_ordered_extent *entry;
134
135 if (tree->last) {
136 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
137 rb_node);
138 if (in_range(file_offset, entry->file_offset, entry->num_bytes))
139 return tree->last;
140 }
141 ret = __tree_search(root, file_offset, &prev);
142 if (!ret)
143 ret = prev;
144 if (ret)
145 tree->last = ret;
146 return ret;
147}
148
149/*
150 * Add an ordered extent to the per-inode tree.
151 *
152 * @inode: Inode that this extent is for.
153 * @file_offset: Logical offset in file where the extent starts.
154 * @num_bytes: Logical length of extent in file.
155 * @ram_bytes: Full length of unencoded data.
156 * @disk_bytenr: Offset of extent on disk.
157 * @disk_num_bytes: Size of extent on disk.
158 * @offset: Offset into unencoded data where file data starts.
159 * @flags: Flags specifying type of extent (1 << BTRFS_ORDERED_*).
160 * @compress_type: Compression algorithm used for data.
161 *
162 * Most of these parameters correspond to &struct btrfs_file_extent_item. The
163 * tree is given a single reference on the ordered extent that was inserted.
164 *
165 * Return: 0 or -ENOMEM.
166 */
167int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
168 u64 num_bytes, u64 ram_bytes, u64 disk_bytenr,
169 u64 disk_num_bytes, u64 offset, unsigned flags,
170 int compress_type)
171{
172 struct btrfs_root *root = inode->root;
173 struct btrfs_fs_info *fs_info = root->fs_info;
174 struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
175 struct rb_node *node;
176 struct btrfs_ordered_extent *entry;
177 int ret;
178
179 if (flags &
180 ((1 << BTRFS_ORDERED_NOCOW) | (1 << BTRFS_ORDERED_PREALLOC))) {
181 /* For nocow write, we can release the qgroup rsv right now */
182 ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes);
183 if (ret < 0)
184 return ret;
185 ret = 0;
186 } else {
187 /*
188 * The ordered extent has reserved qgroup space, release now
189 * and pass the reserved number for qgroup_record to free.
190 */
191 ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes);
192 if (ret < 0)
193 return ret;
194 }
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->num_bytes = num_bytes;
201 entry->ram_bytes = ram_bytes;
202 entry->disk_bytenr = disk_bytenr;
203 entry->disk_num_bytes = disk_num_bytes;
204 entry->offset = offset;
205 entry->bytes_left = num_bytes;
206 entry->inode = igrab(&inode->vfs_inode);
207 entry->compress_type = compress_type;
208 entry->truncated_len = (u64)-1;
209 entry->qgroup_rsv = ret;
210 entry->physical = (u64)-1;
211
212 ASSERT((flags & ~BTRFS_ORDERED_TYPE_FLAGS) == 0);
213 entry->flags = flags;
214
215 percpu_counter_add_batch(&fs_info->ordered_bytes, num_bytes,
216 fs_info->delalloc_batch);
217
218 /* one ref for the tree */
219 refcount_set(&entry->refs, 1);
220 init_waitqueue_head(&entry->wait);
221 INIT_LIST_HEAD(&entry->list);
222 INIT_LIST_HEAD(&entry->log_list);
223 INIT_LIST_HEAD(&entry->root_extent_list);
224 INIT_LIST_HEAD(&entry->work_list);
225 init_completion(&entry->completion);
226
227 trace_btrfs_ordered_extent_add(inode, entry);
228
229 spin_lock_irq(&tree->lock);
230 node = tree_insert(&tree->tree, file_offset,
231 &entry->rb_node);
232 if (node)
233 btrfs_panic(fs_info, -EEXIST,
234 "inconsistency in ordered tree at offset %llu",
235 file_offset);
236 spin_unlock_irq(&tree->lock);
237
238 spin_lock(&root->ordered_extent_lock);
239 list_add_tail(&entry->root_extent_list,
240 &root->ordered_extents);
241 root->nr_ordered_extents++;
242 if (root->nr_ordered_extents == 1) {
243 spin_lock(&fs_info->ordered_root_lock);
244 BUG_ON(!list_empty(&root->ordered_root));
245 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
246 spin_unlock(&fs_info->ordered_root_lock);
247 }
248 spin_unlock(&root->ordered_extent_lock);
249
250 /*
251 * We don't need the count_max_extents here, we can assume that all of
252 * that work has been done at higher layers, so this is truly the
253 * smallest the extent is going to get.
254 */
255 spin_lock(&inode->lock);
256 btrfs_mod_outstanding_extents(inode, 1);
257 spin_unlock(&inode->lock);
258
259 return 0;
260}
261
262/*
263 * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
264 * when an ordered extent is finished. If the list covers more than one
265 * ordered extent, it is split across multiples.
266 */
267void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
268 struct btrfs_ordered_sum *sum)
269{
270 struct btrfs_ordered_inode_tree *tree;
271
272 tree = &BTRFS_I(entry->inode)->ordered_tree;
273 spin_lock_irq(&tree->lock);
274 list_add_tail(&sum->list, &entry->list);
275 spin_unlock_irq(&tree->lock);
276}
277
278static void finish_ordered_fn(struct btrfs_work *work)
279{
280 struct btrfs_ordered_extent *ordered_extent;
281
282 ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
283 btrfs_finish_ordered_io(ordered_extent);
284}
285
286/*
287 * Mark all ordered extents io inside the specified range finished.
288 *
289 * @page: The involved page for the operation.
290 * For uncompressed buffered IO, the page status also needs to be
291 * updated to indicate whether the pending ordered io is finished.
292 * Can be NULL for direct IO and compressed write.
293 * For these cases, callers are ensured they won't execute the
294 * endio function twice.
295 *
296 * This function is called for endio, thus the range must have ordered
297 * extent(s) covering it.
298 */
299void btrfs_mark_ordered_io_finished(struct btrfs_inode *inode,
300 struct page *page, u64 file_offset,
301 u64 num_bytes, bool uptodate)
302{
303 struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
304 struct btrfs_fs_info *fs_info = inode->root->fs_info;
305 struct btrfs_workqueue *wq;
306 struct rb_node *node;
307 struct btrfs_ordered_extent *entry = NULL;
308 unsigned long flags;
309 u64 cur = file_offset;
310
311 if (btrfs_is_free_space_inode(inode))
312 wq = fs_info->endio_freespace_worker;
313 else
314 wq = fs_info->endio_write_workers;
315
316 if (page)
317 ASSERT(page->mapping && page_offset(page) <= file_offset &&
318 file_offset + num_bytes <= page_offset(page) + PAGE_SIZE);
319
320 spin_lock_irqsave(&tree->lock, flags);
321 while (cur < file_offset + num_bytes) {
322 u64 entry_end;
323 u64 end;
324 u32 len;
325
326 node = tree_search(tree, cur);
327 /* No ordered extents at all */
328 if (!node)
329 break;
330
331 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
332 entry_end = entry->file_offset + entry->num_bytes;
333 /*
334 * |<-- OE --->| |
335 * cur
336 * Go to next OE.
337 */
338 if (cur >= entry_end) {
339 node = rb_next(node);
340 /* No more ordered extents, exit */
341 if (!node)
342 break;
343 entry = rb_entry(node, struct btrfs_ordered_extent,
344 rb_node);
345
346 /* Go to next ordered extent and continue */
347 cur = entry->file_offset;
348 continue;
349 }
350 /*
351 * | |<--- OE --->|
352 * cur
353 * Go to the start of OE.
354 */
355 if (cur < entry->file_offset) {
356 cur = entry->file_offset;
357 continue;
358 }
359
360 /*
361 * Now we are definitely inside one ordered extent.
362 *
363 * |<--- OE --->|
364 * |
365 * cur
366 */
367 end = min(entry->file_offset + entry->num_bytes,
368 file_offset + num_bytes) - 1;
369 ASSERT(end + 1 - cur < U32_MAX);
370 len = end + 1 - cur;
371
372 if (page) {
373 /*
374 * Ordered (Private2) bit indicates whether we still
375 * have pending io unfinished for the ordered extent.
376 *
377 * If there's no such bit, we need to skip to next range.
378 */
379 if (!btrfs_page_test_ordered(fs_info, page, cur, len)) {
380 cur += len;
381 continue;
382 }
383 btrfs_page_clear_ordered(fs_info, page, cur, len);
384 }
385
386 /* Now we're fine to update the accounting */
387 if (unlikely(len > entry->bytes_left)) {
388 WARN_ON(1);
389 btrfs_crit(fs_info,
390"bad ordered extent accounting, root=%llu ino=%llu OE offset=%llu OE len=%llu to_dec=%u left=%llu",
391 inode->root->root_key.objectid,
392 btrfs_ino(inode),
393 entry->file_offset,
394 entry->num_bytes,
395 len, entry->bytes_left);
396 entry->bytes_left = 0;
397 } else {
398 entry->bytes_left -= len;
399 }
400
401 if (!uptodate)
402 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
403
404 /*
405 * All the IO of the ordered extent is finished, we need to queue
406 * the finish_func to be executed.
407 */
408 if (entry->bytes_left == 0) {
409 set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
410 cond_wake_up(&entry->wait);
411 refcount_inc(&entry->refs);
412 trace_btrfs_ordered_extent_mark_finished(inode, entry);
413 spin_unlock_irqrestore(&tree->lock, flags);
414 btrfs_init_work(&entry->work, finish_ordered_fn, NULL, NULL);
415 btrfs_queue_work(wq, &entry->work);
416 spin_lock_irqsave(&tree->lock, flags);
417 }
418 cur += len;
419 }
420 spin_unlock_irqrestore(&tree->lock, flags);
421}
422
423/*
424 * Finish IO for one ordered extent across a given range. The range can only
425 * contain one ordered extent.
426 *
427 * @cached: The cached ordered extent. If not NULL, we can skip the tree
428 * search and use the ordered extent directly.
429 * Will be also used to store the finished ordered extent.
430 * @file_offset: File offset for the finished IO
431 * @io_size: Length of the finish IO range
432 *
433 * Return true if the ordered extent is finished in the range, and update
434 * @cached.
435 * Return false otherwise.
436 *
437 * NOTE: The range can NOT cross multiple ordered extents.
438 * Thus caller should ensure the range doesn't cross ordered extents.
439 */
440bool btrfs_dec_test_ordered_pending(struct btrfs_inode *inode,
441 struct btrfs_ordered_extent **cached,
442 u64 file_offset, u64 io_size)
443{
444 struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
445 struct rb_node *node;
446 struct btrfs_ordered_extent *entry = NULL;
447 unsigned long flags;
448 bool finished = false;
449
450 spin_lock_irqsave(&tree->lock, flags);
451 if (cached && *cached) {
452 entry = *cached;
453 goto have_entry;
454 }
455
456 node = tree_search(tree, file_offset);
457 if (!node)
458 goto out;
459
460 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
461have_entry:
462 if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
463 goto out;
464
465 if (io_size > entry->bytes_left)
466 btrfs_crit(inode->root->fs_info,
467 "bad ordered accounting left %llu size %llu",
468 entry->bytes_left, io_size);
469
470 entry->bytes_left -= io_size;
471
472 if (entry->bytes_left == 0) {
473 /*
474 * Ensure only one caller can set the flag and finished_ret
475 * accordingly
476 */
477 finished = !test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
478 /* test_and_set_bit implies a barrier */
479 cond_wake_up_nomb(&entry->wait);
480 }
481out:
482 if (finished && cached && entry) {
483 *cached = entry;
484 refcount_inc(&entry->refs);
485 trace_btrfs_ordered_extent_dec_test_pending(inode, entry);
486 }
487 spin_unlock_irqrestore(&tree->lock, flags);
488 return finished;
489}
490
491/*
492 * used to drop a reference on an ordered extent. This will free
493 * the extent if the last reference is dropped
494 */
495void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
496{
497 struct list_head *cur;
498 struct btrfs_ordered_sum *sum;
499
500 trace_btrfs_ordered_extent_put(BTRFS_I(entry->inode), entry);
501
502 if (refcount_dec_and_test(&entry->refs)) {
503 ASSERT(list_empty(&entry->root_extent_list));
504 ASSERT(list_empty(&entry->log_list));
505 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
506 if (entry->inode)
507 btrfs_add_delayed_iput(BTRFS_I(entry->inode));
508 while (!list_empty(&entry->list)) {
509 cur = entry->list.next;
510 sum = list_entry(cur, struct btrfs_ordered_sum, list);
511 list_del(&sum->list);
512 kvfree(sum);
513 }
514 kmem_cache_free(btrfs_ordered_extent_cache, entry);
515 }
516}
517
518/*
519 * remove an ordered extent from the tree. No references are dropped
520 * and waiters are woken up.
521 */
522void btrfs_remove_ordered_extent(struct btrfs_inode *btrfs_inode,
523 struct btrfs_ordered_extent *entry)
524{
525 struct btrfs_ordered_inode_tree *tree;
526 struct btrfs_root *root = btrfs_inode->root;
527 struct btrfs_fs_info *fs_info = root->fs_info;
528 struct rb_node *node;
529 bool pending;
530 bool freespace_inode;
531
532 /*
533 * If this is a free space inode the thread has not acquired the ordered
534 * extents lockdep map.
535 */
536 freespace_inode = btrfs_is_free_space_inode(btrfs_inode);
537
538 btrfs_lockdep_acquire(fs_info, btrfs_trans_pending_ordered);
539 /* This is paired with btrfs_add_ordered_extent. */
540 spin_lock(&btrfs_inode->lock);
541 btrfs_mod_outstanding_extents(btrfs_inode, -1);
542 spin_unlock(&btrfs_inode->lock);
543 if (root != fs_info->tree_root) {
544 u64 release;
545
546 if (test_bit(BTRFS_ORDERED_ENCODED, &entry->flags))
547 release = entry->disk_num_bytes;
548 else
549 release = entry->num_bytes;
550 btrfs_delalloc_release_metadata(btrfs_inode, release, false);
551 }
552
553 percpu_counter_add_batch(&fs_info->ordered_bytes, -entry->num_bytes,
554 fs_info->delalloc_batch);
555
556 tree = &btrfs_inode->ordered_tree;
557 spin_lock_irq(&tree->lock);
558 node = &entry->rb_node;
559 rb_erase(node, &tree->tree);
560 RB_CLEAR_NODE(node);
561 if (tree->last == node)
562 tree->last = NULL;
563 set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
564 pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
565 spin_unlock_irq(&tree->lock);
566
567 /*
568 * The current running transaction is waiting on us, we need to let it
569 * know that we're complete and wake it up.
570 */
571 if (pending) {
572 struct btrfs_transaction *trans;
573
574 /*
575 * The checks for trans are just a formality, it should be set,
576 * but if it isn't we don't want to deref/assert under the spin
577 * lock, so be nice and check if trans is set, but ASSERT() so
578 * if it isn't set a developer will notice.
579 */
580 spin_lock(&fs_info->trans_lock);
581 trans = fs_info->running_transaction;
582 if (trans)
583 refcount_inc(&trans->use_count);
584 spin_unlock(&fs_info->trans_lock);
585
586 ASSERT(trans);
587 if (trans) {
588 if (atomic_dec_and_test(&trans->pending_ordered))
589 wake_up(&trans->pending_wait);
590 btrfs_put_transaction(trans);
591 }
592 }
593
594 btrfs_lockdep_release(fs_info, btrfs_trans_pending_ordered);
595
596 spin_lock(&root->ordered_extent_lock);
597 list_del_init(&entry->root_extent_list);
598 root->nr_ordered_extents--;
599
600 trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
601
602 if (!root->nr_ordered_extents) {
603 spin_lock(&fs_info->ordered_root_lock);
604 BUG_ON(list_empty(&root->ordered_root));
605 list_del_init(&root->ordered_root);
606 spin_unlock(&fs_info->ordered_root_lock);
607 }
608 spin_unlock(&root->ordered_extent_lock);
609 wake_up(&entry->wait);
610 if (!freespace_inode)
611 btrfs_lockdep_release(fs_info, btrfs_ordered_extent);
612}
613
614static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
615{
616 struct btrfs_ordered_extent *ordered;
617
618 ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
619 btrfs_start_ordered_extent(ordered, 1);
620 complete(&ordered->completion);
621}
622
623/*
624 * wait for all the ordered extents in a root. This is done when balancing
625 * space between drives.
626 */
627u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
628 const u64 range_start, const u64 range_len)
629{
630 struct btrfs_fs_info *fs_info = root->fs_info;
631 LIST_HEAD(splice);
632 LIST_HEAD(skipped);
633 LIST_HEAD(works);
634 struct btrfs_ordered_extent *ordered, *next;
635 u64 count = 0;
636 const u64 range_end = range_start + range_len;
637
638 mutex_lock(&root->ordered_extent_mutex);
639 spin_lock(&root->ordered_extent_lock);
640 list_splice_init(&root->ordered_extents, &splice);
641 while (!list_empty(&splice) && nr) {
642 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
643 root_extent_list);
644
645 if (range_end <= ordered->disk_bytenr ||
646 ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
647 list_move_tail(&ordered->root_extent_list, &skipped);
648 cond_resched_lock(&root->ordered_extent_lock);
649 continue;
650 }
651
652 list_move_tail(&ordered->root_extent_list,
653 &root->ordered_extents);
654 refcount_inc(&ordered->refs);
655 spin_unlock(&root->ordered_extent_lock);
656
657 btrfs_init_work(&ordered->flush_work,
658 btrfs_run_ordered_extent_work, NULL, NULL);
659 list_add_tail(&ordered->work_list, &works);
660 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
661
662 cond_resched();
663 spin_lock(&root->ordered_extent_lock);
664 if (nr != U64_MAX)
665 nr--;
666 count++;
667 }
668 list_splice_tail(&skipped, &root->ordered_extents);
669 list_splice_tail(&splice, &root->ordered_extents);
670 spin_unlock(&root->ordered_extent_lock);
671
672 list_for_each_entry_safe(ordered, next, &works, work_list) {
673 list_del_init(&ordered->work_list);
674 wait_for_completion(&ordered->completion);
675 btrfs_put_ordered_extent(ordered);
676 cond_resched();
677 }
678 mutex_unlock(&root->ordered_extent_mutex);
679
680 return count;
681}
682
683void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
684 const u64 range_start, const u64 range_len)
685{
686 struct btrfs_root *root;
687 struct list_head splice;
688 u64 done;
689
690 INIT_LIST_HEAD(&splice);
691
692 mutex_lock(&fs_info->ordered_operations_mutex);
693 spin_lock(&fs_info->ordered_root_lock);
694 list_splice_init(&fs_info->ordered_roots, &splice);
695 while (!list_empty(&splice) && nr) {
696 root = list_first_entry(&splice, struct btrfs_root,
697 ordered_root);
698 root = btrfs_grab_root(root);
699 BUG_ON(!root);
700 list_move_tail(&root->ordered_root,
701 &fs_info->ordered_roots);
702 spin_unlock(&fs_info->ordered_root_lock);
703
704 done = btrfs_wait_ordered_extents(root, nr,
705 range_start, range_len);
706 btrfs_put_root(root);
707
708 spin_lock(&fs_info->ordered_root_lock);
709 if (nr != U64_MAX) {
710 nr -= done;
711 }
712 }
713 list_splice_tail(&splice, &fs_info->ordered_roots);
714 spin_unlock(&fs_info->ordered_root_lock);
715 mutex_unlock(&fs_info->ordered_operations_mutex);
716}
717
718/*
719 * Used to start IO or wait for a given ordered extent to finish.
720 *
721 * If wait is one, this effectively waits on page writeback for all the pages
722 * in the extent, and it waits on the io completion code to insert
723 * metadata into the btree corresponding to the extent
724 */
725void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry, int wait)
726{
727 u64 start = entry->file_offset;
728 u64 end = start + entry->num_bytes - 1;
729 struct btrfs_inode *inode = BTRFS_I(entry->inode);
730 bool freespace_inode;
731
732 trace_btrfs_ordered_extent_start(inode, entry);
733
734 /*
735 * If this is a free space inode do not take the ordered extents lockdep
736 * map.
737 */
738 freespace_inode = btrfs_is_free_space_inode(inode);
739
740 /*
741 * pages in the range can be dirty, clean or writeback. We
742 * start IO on any dirty ones so the wait doesn't stall waiting
743 * for the flusher thread to find them
744 */
745 if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
746 filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
747 if (wait) {
748 if (!freespace_inode)
749 btrfs_might_wait_for_event(inode->root->fs_info, btrfs_ordered_extent);
750 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
751 &entry->flags));
752 }
753}
754
755/*
756 * Used to wait on ordered extents across a large range of bytes.
757 */
758int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
759{
760 int ret = 0;
761 int ret_wb = 0;
762 u64 end;
763 u64 orig_end;
764 struct btrfs_ordered_extent *ordered;
765
766 if (start + len < start) {
767 orig_end = OFFSET_MAX;
768 } else {
769 orig_end = start + len - 1;
770 if (orig_end > OFFSET_MAX)
771 orig_end = OFFSET_MAX;
772 }
773
774 /* start IO across the range first to instantiate any delalloc
775 * extents
776 */
777 ret = btrfs_fdatawrite_range(inode, start, orig_end);
778 if (ret)
779 return ret;
780
781 /*
782 * If we have a writeback error don't return immediately. Wait first
783 * for any ordered extents that haven't completed yet. This is to make
784 * sure no one can dirty the same page ranges and call writepages()
785 * before the ordered extents complete - to avoid failures (-EEXIST)
786 * when adding the new ordered extents to the ordered tree.
787 */
788 ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
789
790 end = orig_end;
791 while (1) {
792 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode), end);
793 if (!ordered)
794 break;
795 if (ordered->file_offset > orig_end) {
796 btrfs_put_ordered_extent(ordered);
797 break;
798 }
799 if (ordered->file_offset + ordered->num_bytes <= start) {
800 btrfs_put_ordered_extent(ordered);
801 break;
802 }
803 btrfs_start_ordered_extent(ordered, 1);
804 end = ordered->file_offset;
805 /*
806 * If the ordered extent had an error save the error but don't
807 * exit without waiting first for all other ordered extents in
808 * the range to complete.
809 */
810 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
811 ret = -EIO;
812 btrfs_put_ordered_extent(ordered);
813 if (end == 0 || end == start)
814 break;
815 end--;
816 }
817 return ret_wb ? ret_wb : ret;
818}
819
820/*
821 * find an ordered extent corresponding to file_offset. return NULL if
822 * nothing is found, otherwise take a reference on the extent and return it
823 */
824struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
825 u64 file_offset)
826{
827 struct btrfs_ordered_inode_tree *tree;
828 struct rb_node *node;
829 struct btrfs_ordered_extent *entry = NULL;
830 unsigned long flags;
831
832 tree = &inode->ordered_tree;
833 spin_lock_irqsave(&tree->lock, flags);
834 node = tree_search(tree, file_offset);
835 if (!node)
836 goto out;
837
838 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
839 if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
840 entry = NULL;
841 if (entry) {
842 refcount_inc(&entry->refs);
843 trace_btrfs_ordered_extent_lookup(inode, entry);
844 }
845out:
846 spin_unlock_irqrestore(&tree->lock, flags);
847 return entry;
848}
849
850/* Since the DIO code tries to lock a wide area we need to look for any ordered
851 * extents that exist in the range, rather than just the start of the range.
852 */
853struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
854 struct btrfs_inode *inode, u64 file_offset, u64 len)
855{
856 struct btrfs_ordered_inode_tree *tree;
857 struct rb_node *node;
858 struct btrfs_ordered_extent *entry = NULL;
859
860 tree = &inode->ordered_tree;
861 spin_lock_irq(&tree->lock);
862 node = tree_search(tree, file_offset);
863 if (!node) {
864 node = tree_search(tree, file_offset + len);
865 if (!node)
866 goto out;
867 }
868
869 while (1) {
870 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
871 if (range_overlaps(entry, file_offset, len))
872 break;
873
874 if (entry->file_offset >= file_offset + len) {
875 entry = NULL;
876 break;
877 }
878 entry = NULL;
879 node = rb_next(node);
880 if (!node)
881 break;
882 }
883out:
884 if (entry) {
885 refcount_inc(&entry->refs);
886 trace_btrfs_ordered_extent_lookup_range(inode, entry);
887 }
888 spin_unlock_irq(&tree->lock);
889 return entry;
890}
891
892/*
893 * Adds all ordered extents to the given list. The list ends up sorted by the
894 * file_offset of the ordered extents.
895 */
896void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
897 struct list_head *list)
898{
899 struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
900 struct rb_node *n;
901
902 ASSERT(inode_is_locked(&inode->vfs_inode));
903
904 spin_lock_irq(&tree->lock);
905 for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
906 struct btrfs_ordered_extent *ordered;
907
908 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
909
910 if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
911 continue;
912
913 ASSERT(list_empty(&ordered->log_list));
914 list_add_tail(&ordered->log_list, list);
915 refcount_inc(&ordered->refs);
916 trace_btrfs_ordered_extent_lookup_for_logging(inode, ordered);
917 }
918 spin_unlock_irq(&tree->lock);
919}
920
921/*
922 * lookup and return any extent before 'file_offset'. NULL is returned
923 * if none is found
924 */
925struct btrfs_ordered_extent *
926btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
927{
928 struct btrfs_ordered_inode_tree *tree;
929 struct rb_node *node;
930 struct btrfs_ordered_extent *entry = NULL;
931
932 tree = &inode->ordered_tree;
933 spin_lock_irq(&tree->lock);
934 node = tree_search(tree, file_offset);
935 if (!node)
936 goto out;
937
938 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
939 refcount_inc(&entry->refs);
940 trace_btrfs_ordered_extent_lookup_first(inode, entry);
941out:
942 spin_unlock_irq(&tree->lock);
943 return entry;
944}
945
946/*
947 * Lookup the first ordered extent that overlaps the range
948 * [@file_offset, @file_offset + @len).
949 *
950 * The difference between this and btrfs_lookup_first_ordered_extent() is
951 * that this one won't return any ordered extent that does not overlap the range.
952 * And the difference against btrfs_lookup_ordered_extent() is, this function
953 * ensures the first ordered extent gets returned.
954 */
955struct btrfs_ordered_extent *btrfs_lookup_first_ordered_range(
956 struct btrfs_inode *inode, u64 file_offset, u64 len)
957{
958 struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
959 struct rb_node *node;
960 struct rb_node *cur;
961 struct rb_node *prev;
962 struct rb_node *next;
963 struct btrfs_ordered_extent *entry = NULL;
964
965 spin_lock_irq(&tree->lock);
966 node = tree->tree.rb_node;
967 /*
968 * Here we don't want to use tree_search() which will use tree->last
969 * and screw up the search order.
970 * And __tree_search() can't return the adjacent ordered extents
971 * either, thus here we do our own search.
972 */
973 while (node) {
974 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
975
976 if (file_offset < entry->file_offset) {
977 node = node->rb_left;
978 } else if (file_offset >= entry_end(entry)) {
979 node = node->rb_right;
980 } else {
981 /*
982 * Direct hit, got an ordered extent that starts at
983 * @file_offset
984 */
985 goto out;
986 }
987 }
988 if (!entry) {
989 /* Empty tree */
990 goto out;
991 }
992
993 cur = &entry->rb_node;
994 /* We got an entry around @file_offset, check adjacent entries */
995 if (entry->file_offset < file_offset) {
996 prev = cur;
997 next = rb_next(cur);
998 } else {
999 prev = rb_prev(cur);
1000 next = cur;
1001 }
1002 if (prev) {
1003 entry = rb_entry(prev, struct btrfs_ordered_extent, rb_node);
1004 if (range_overlaps(entry, file_offset, len))
1005 goto out;
1006 }
1007 if (next) {
1008 entry = rb_entry(next, struct btrfs_ordered_extent, rb_node);
1009 if (range_overlaps(entry, file_offset, len))
1010 goto out;
1011 }
1012 /* No ordered extent in the range */
1013 entry = NULL;
1014out:
1015 if (entry) {
1016 refcount_inc(&entry->refs);
1017 trace_btrfs_ordered_extent_lookup_first_range(inode, entry);
1018 }
1019
1020 spin_unlock_irq(&tree->lock);
1021 return entry;
1022}
1023
1024/*
1025 * Lock the passed range and ensures all pending ordered extents in it are run
1026 * to completion.
1027 *
1028 * @inode: Inode whose ordered tree is to be searched
1029 * @start: Beginning of range to flush
1030 * @end: Last byte of range to lock
1031 * @cached_state: If passed, will return the extent state responsible for the
1032 * locked range. It's the caller's responsibility to free the
1033 * cached state.
1034 *
1035 * Always return with the given range locked, ensuring after it's called no
1036 * order extent can be pending.
1037 */
1038void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
1039 u64 end,
1040 struct extent_state **cached_state)
1041{
1042 struct btrfs_ordered_extent *ordered;
1043 struct extent_state *cache = NULL;
1044 struct extent_state **cachedp = &cache;
1045
1046 if (cached_state)
1047 cachedp = cached_state;
1048
1049 while (1) {
1050 lock_extent(&inode->io_tree, start, end, cachedp);
1051 ordered = btrfs_lookup_ordered_range(inode, start,
1052 end - start + 1);
1053 if (!ordered) {
1054 /*
1055 * If no external cached_state has been passed then
1056 * decrement the extra ref taken for cachedp since we
1057 * aren't exposing it outside of this function
1058 */
1059 if (!cached_state)
1060 refcount_dec(&cache->refs);
1061 break;
1062 }
1063 unlock_extent(&inode->io_tree, start, end, cachedp);
1064 btrfs_start_ordered_extent(ordered, 1);
1065 btrfs_put_ordered_extent(ordered);
1066 }
1067}
1068
1069/*
1070 * Lock the passed range and ensure all pending ordered extents in it are run
1071 * to completion in nowait mode.
1072 *
1073 * Return true if btrfs_lock_ordered_range does not return any extents,
1074 * otherwise false.
1075 */
1076bool btrfs_try_lock_ordered_range(struct btrfs_inode *inode, u64 start, u64 end,
1077 struct extent_state **cached_state)
1078{
1079 struct btrfs_ordered_extent *ordered;
1080
1081 if (!try_lock_extent(&inode->io_tree, start, end, cached_state))
1082 return false;
1083
1084 ordered = btrfs_lookup_ordered_range(inode, start, end - start + 1);
1085 if (!ordered)
1086 return true;
1087
1088 btrfs_put_ordered_extent(ordered);
1089 unlock_extent(&inode->io_tree, start, end, cached_state);
1090
1091 return false;
1092}
1093
1094
1095static int clone_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pos,
1096 u64 len)
1097{
1098 struct inode *inode = ordered->inode;
1099 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1100 u64 file_offset = ordered->file_offset + pos;
1101 u64 disk_bytenr = ordered->disk_bytenr + pos;
1102 unsigned long flags = ordered->flags & BTRFS_ORDERED_TYPE_FLAGS;
1103
1104 /*
1105 * The splitting extent is already counted and will be added again in
1106 * btrfs_add_ordered_extent_*(). Subtract len to avoid double counting.
1107 */
1108 percpu_counter_add_batch(&fs_info->ordered_bytes, -len,
1109 fs_info->delalloc_batch);
1110 WARN_ON_ONCE(flags & (1 << BTRFS_ORDERED_COMPRESSED));
1111 return btrfs_add_ordered_extent(BTRFS_I(inode), file_offset, len, len,
1112 disk_bytenr, len, 0, flags,
1113 ordered->compress_type);
1114}
1115
1116int btrfs_split_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pre,
1117 u64 post)
1118{
1119 struct inode *inode = ordered->inode;
1120 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1121 struct rb_node *node;
1122 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1123 int ret = 0;
1124
1125 trace_btrfs_ordered_extent_split(BTRFS_I(inode), ordered);
1126
1127 spin_lock_irq(&tree->lock);
1128 /* Remove from tree once */
1129 node = &ordered->rb_node;
1130 rb_erase(node, &tree->tree);
1131 RB_CLEAR_NODE(node);
1132 if (tree->last == node)
1133 tree->last = NULL;
1134
1135 ordered->file_offset += pre;
1136 ordered->disk_bytenr += pre;
1137 ordered->num_bytes -= (pre + post);
1138 ordered->disk_num_bytes -= (pre + post);
1139 ordered->bytes_left -= (pre + post);
1140
1141 /* Re-insert the node */
1142 node = tree_insert(&tree->tree, ordered->file_offset, &ordered->rb_node);
1143 if (node)
1144 btrfs_panic(fs_info, -EEXIST,
1145 "zoned: inconsistency in ordered tree at offset %llu",
1146 ordered->file_offset);
1147
1148 spin_unlock_irq(&tree->lock);
1149
1150 if (pre)
1151 ret = clone_ordered_extent(ordered, 0, pre);
1152 if (ret == 0 && post)
1153 ret = clone_ordered_extent(ordered, pre + ordered->disk_num_bytes,
1154 post);
1155
1156 return ret;
1157}
1158
1159int __init ordered_data_init(void)
1160{
1161 btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1162 sizeof(struct btrfs_ordered_extent), 0,
1163 SLAB_MEM_SPREAD,
1164 NULL);
1165 if (!btrfs_ordered_extent_cache)
1166 return -ENOMEM;
1167
1168 return 0;
1169}
1170
1171void __cold ordered_data_exit(void)
1172{
1173 kmem_cache_destroy(btrfs_ordered_extent_cache);
1174}
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 "messages.h"
11#include "misc.h"
12#include "ctree.h"
13#include "transaction.h"
14#include "btrfs_inode.h"
15#include "extent_io.h"
16#include "disk-io.h"
17#include "compression.h"
18#include "delalloc-space.h"
19#include "qgroup.h"
20#include "subpage.h"
21#include "file.h"
22#include "block-group.h"
23
24static struct kmem_cache *btrfs_ordered_extent_cache;
25
26static u64 entry_end(struct btrfs_ordered_extent *entry)
27{
28 if (entry->file_offset + entry->num_bytes < entry->file_offset)
29 return (u64)-1;
30 return entry->file_offset + entry->num_bytes;
31}
32
33/* returns NULL if the insertion worked, or it returns the node it did find
34 * in the tree
35 */
36static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
37 struct rb_node *node)
38{
39 struct rb_node **p = &root->rb_node;
40 struct rb_node *parent = NULL;
41 struct btrfs_ordered_extent *entry;
42
43 while (*p) {
44 parent = *p;
45 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
46
47 if (file_offset < entry->file_offset)
48 p = &(*p)->rb_left;
49 else if (file_offset >= entry_end(entry))
50 p = &(*p)->rb_right;
51 else
52 return parent;
53 }
54
55 rb_link_node(node, parent, p);
56 rb_insert_color(node, root);
57 return NULL;
58}
59
60/*
61 * look for a given offset in the tree, and if it can't be found return the
62 * first lesser offset
63 */
64static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
65 struct rb_node **prev_ret)
66{
67 struct rb_node *n = root->rb_node;
68 struct rb_node *prev = NULL;
69 struct rb_node *test;
70 struct btrfs_ordered_extent *entry;
71 struct btrfs_ordered_extent *prev_entry = NULL;
72
73 while (n) {
74 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
75 prev = n;
76 prev_entry = entry;
77
78 if (file_offset < entry->file_offset)
79 n = n->rb_left;
80 else if (file_offset >= entry_end(entry))
81 n = n->rb_right;
82 else
83 return n;
84 }
85 if (!prev_ret)
86 return NULL;
87
88 while (prev && file_offset >= entry_end(prev_entry)) {
89 test = rb_next(prev);
90 if (!test)
91 break;
92 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
93 rb_node);
94 if (file_offset < entry_end(prev_entry))
95 break;
96
97 prev = test;
98 }
99 if (prev)
100 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
101 rb_node);
102 while (prev && file_offset < entry_end(prev_entry)) {
103 test = rb_prev(prev);
104 if (!test)
105 break;
106 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
107 rb_node);
108 prev = test;
109 }
110 *prev_ret = prev;
111 return NULL;
112}
113
114static int btrfs_range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
115 u64 len)
116{
117 if (file_offset + len <= entry->file_offset ||
118 entry->file_offset + entry->num_bytes <= file_offset)
119 return 0;
120 return 1;
121}
122
123/*
124 * look find the first ordered struct that has this offset, otherwise
125 * the first one less than this offset
126 */
127static inline struct rb_node *ordered_tree_search(struct btrfs_inode *inode,
128 u64 file_offset)
129{
130 struct rb_node *prev = NULL;
131 struct rb_node *ret;
132 struct btrfs_ordered_extent *entry;
133
134 if (inode->ordered_tree_last) {
135 entry = rb_entry(inode->ordered_tree_last, struct btrfs_ordered_extent,
136 rb_node);
137 if (in_range(file_offset, entry->file_offset, entry->num_bytes))
138 return inode->ordered_tree_last;
139 }
140 ret = __tree_search(&inode->ordered_tree, file_offset, &prev);
141 if (!ret)
142 ret = prev;
143 if (ret)
144 inode->ordered_tree_last = ret;
145 return ret;
146}
147
148static struct btrfs_ordered_extent *alloc_ordered_extent(
149 struct btrfs_inode *inode, u64 file_offset, u64 num_bytes,
150 u64 ram_bytes, u64 disk_bytenr, u64 disk_num_bytes,
151 u64 offset, unsigned long flags, int compress_type)
152{
153 struct btrfs_ordered_extent *entry;
154 int ret;
155 u64 qgroup_rsv = 0;
156
157 if (flags &
158 ((1 << BTRFS_ORDERED_NOCOW) | (1 << BTRFS_ORDERED_PREALLOC))) {
159 /* For nocow write, we can release the qgroup rsv right now */
160 ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes, &qgroup_rsv);
161 if (ret < 0)
162 return ERR_PTR(ret);
163 } else {
164 /*
165 * The ordered extent has reserved qgroup space, release now
166 * and pass the reserved number for qgroup_record to free.
167 */
168 ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes, &qgroup_rsv);
169 if (ret < 0)
170 return ERR_PTR(ret);
171 }
172 entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
173 if (!entry)
174 return ERR_PTR(-ENOMEM);
175
176 entry->file_offset = file_offset;
177 entry->num_bytes = num_bytes;
178 entry->ram_bytes = ram_bytes;
179 entry->disk_bytenr = disk_bytenr;
180 entry->disk_num_bytes = disk_num_bytes;
181 entry->offset = offset;
182 entry->bytes_left = num_bytes;
183 entry->inode = BTRFS_I(igrab(&inode->vfs_inode));
184 entry->compress_type = compress_type;
185 entry->truncated_len = (u64)-1;
186 entry->qgroup_rsv = qgroup_rsv;
187 entry->flags = flags;
188 refcount_set(&entry->refs, 1);
189 init_waitqueue_head(&entry->wait);
190 INIT_LIST_HEAD(&entry->list);
191 INIT_LIST_HEAD(&entry->log_list);
192 INIT_LIST_HEAD(&entry->root_extent_list);
193 INIT_LIST_HEAD(&entry->work_list);
194 INIT_LIST_HEAD(&entry->bioc_list);
195 init_completion(&entry->completion);
196
197 /*
198 * We don't need the count_max_extents here, we can assume that all of
199 * that work has been done at higher layers, so this is truly the
200 * smallest the extent is going to get.
201 */
202 spin_lock(&inode->lock);
203 btrfs_mod_outstanding_extents(inode, 1);
204 spin_unlock(&inode->lock);
205
206 return entry;
207}
208
209static void insert_ordered_extent(struct btrfs_ordered_extent *entry)
210{
211 struct btrfs_inode *inode = entry->inode;
212 struct btrfs_root *root = inode->root;
213 struct btrfs_fs_info *fs_info = root->fs_info;
214 struct rb_node *node;
215
216 trace_btrfs_ordered_extent_add(inode, entry);
217
218 percpu_counter_add_batch(&fs_info->ordered_bytes, entry->num_bytes,
219 fs_info->delalloc_batch);
220
221 /* One ref for the tree. */
222 refcount_inc(&entry->refs);
223
224 spin_lock_irq(&inode->ordered_tree_lock);
225 node = tree_insert(&inode->ordered_tree, entry->file_offset,
226 &entry->rb_node);
227 if (unlikely(node))
228 btrfs_panic(fs_info, -EEXIST,
229 "inconsistency in ordered tree at offset %llu",
230 entry->file_offset);
231 spin_unlock_irq(&inode->ordered_tree_lock);
232
233 spin_lock(&root->ordered_extent_lock);
234 list_add_tail(&entry->root_extent_list,
235 &root->ordered_extents);
236 root->nr_ordered_extents++;
237 if (root->nr_ordered_extents == 1) {
238 spin_lock(&fs_info->ordered_root_lock);
239 BUG_ON(!list_empty(&root->ordered_root));
240 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
241 spin_unlock(&fs_info->ordered_root_lock);
242 }
243 spin_unlock(&root->ordered_extent_lock);
244}
245
246/*
247 * Add an ordered extent to the per-inode tree.
248 *
249 * @inode: Inode that this extent is for.
250 * @file_offset: Logical offset in file where the extent starts.
251 * @num_bytes: Logical length of extent in file.
252 * @ram_bytes: Full length of unencoded data.
253 * @disk_bytenr: Offset of extent on disk.
254 * @disk_num_bytes: Size of extent on disk.
255 * @offset: Offset into unencoded data where file data starts.
256 * @flags: Flags specifying type of extent (1 << BTRFS_ORDERED_*).
257 * @compress_type: Compression algorithm used for data.
258 *
259 * Most of these parameters correspond to &struct btrfs_file_extent_item. The
260 * tree is given a single reference on the ordered extent that was inserted, and
261 * the returned pointer is given a second reference.
262 *
263 * Return: the new ordered extent or error pointer.
264 */
265struct btrfs_ordered_extent *btrfs_alloc_ordered_extent(
266 struct btrfs_inode *inode, u64 file_offset,
267 const struct btrfs_file_extent *file_extent, unsigned long flags)
268{
269 struct btrfs_ordered_extent *entry;
270
271 ASSERT((flags & ~BTRFS_ORDERED_TYPE_FLAGS) == 0);
272
273 /*
274 * For regular writes, we just use the members in @file_extent.
275 *
276 * For NOCOW, we don't really care about the numbers except @start and
277 * file_extent->num_bytes, as we won't insert a file extent item at all.
278 *
279 * For PREALLOC, we do not use ordered extent members, but
280 * btrfs_mark_extent_written() handles everything.
281 *
282 * So here we always pass 0 as offset for NOCOW/PREALLOC ordered extents,
283 * or btrfs_split_ordered_extent() cannot handle it correctly.
284 */
285 if (flags & ((1U << BTRFS_ORDERED_NOCOW) | (1U << BTRFS_ORDERED_PREALLOC)))
286 entry = alloc_ordered_extent(inode, file_offset,
287 file_extent->num_bytes,
288 file_extent->num_bytes,
289 file_extent->disk_bytenr + file_extent->offset,
290 file_extent->num_bytes, 0, flags,
291 file_extent->compression);
292 else
293 entry = alloc_ordered_extent(inode, file_offset,
294 file_extent->num_bytes,
295 file_extent->ram_bytes,
296 file_extent->disk_bytenr,
297 file_extent->disk_num_bytes,
298 file_extent->offset, flags,
299 file_extent->compression);
300 if (!IS_ERR(entry))
301 insert_ordered_extent(entry);
302 return entry;
303}
304
305/*
306 * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
307 * when an ordered extent is finished. If the list covers more than one
308 * ordered extent, it is split across multiples.
309 */
310void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
311 struct btrfs_ordered_sum *sum)
312{
313 struct btrfs_inode *inode = entry->inode;
314
315 spin_lock_irq(&inode->ordered_tree_lock);
316 list_add_tail(&sum->list, &entry->list);
317 spin_unlock_irq(&inode->ordered_tree_lock);
318}
319
320void btrfs_mark_ordered_extent_error(struct btrfs_ordered_extent *ordered)
321{
322 if (!test_and_set_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
323 mapping_set_error(ordered->inode->vfs_inode.i_mapping, -EIO);
324}
325
326static void finish_ordered_fn(struct btrfs_work *work)
327{
328 struct btrfs_ordered_extent *ordered_extent;
329
330 ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
331 btrfs_finish_ordered_io(ordered_extent);
332}
333
334static bool can_finish_ordered_extent(struct btrfs_ordered_extent *ordered,
335 struct folio *folio, u64 file_offset,
336 u64 len, bool uptodate)
337{
338 struct btrfs_inode *inode = ordered->inode;
339 struct btrfs_fs_info *fs_info = inode->root->fs_info;
340
341 lockdep_assert_held(&inode->ordered_tree_lock);
342
343 if (folio) {
344 ASSERT(folio->mapping);
345 ASSERT(folio_pos(folio) <= file_offset);
346 ASSERT(file_offset + len <= folio_pos(folio) + folio_size(folio));
347
348 /*
349 * Ordered flag indicates whether we still have
350 * pending io unfinished for the ordered extent.
351 *
352 * If it's not set, we need to skip to next range.
353 */
354 if (!btrfs_folio_test_ordered(fs_info, folio, file_offset, len))
355 return false;
356 btrfs_folio_clear_ordered(fs_info, folio, file_offset, len);
357 }
358
359 /* Now we're fine to update the accounting. */
360 if (WARN_ON_ONCE(len > ordered->bytes_left)) {
361 btrfs_crit(fs_info,
362"bad ordered extent accounting, root=%llu ino=%llu OE offset=%llu OE len=%llu to_dec=%llu left=%llu",
363 btrfs_root_id(inode->root), btrfs_ino(inode),
364 ordered->file_offset, ordered->num_bytes,
365 len, ordered->bytes_left);
366 ordered->bytes_left = 0;
367 } else {
368 ordered->bytes_left -= len;
369 }
370
371 if (!uptodate)
372 set_bit(BTRFS_ORDERED_IOERR, &ordered->flags);
373
374 if (ordered->bytes_left)
375 return false;
376
377 /*
378 * All the IO of the ordered extent is finished, we need to queue
379 * the finish_func to be executed.
380 */
381 set_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags);
382 cond_wake_up(&ordered->wait);
383 refcount_inc(&ordered->refs);
384 trace_btrfs_ordered_extent_mark_finished(inode, ordered);
385 return true;
386}
387
388static void btrfs_queue_ordered_fn(struct btrfs_ordered_extent *ordered)
389{
390 struct btrfs_inode *inode = ordered->inode;
391 struct btrfs_fs_info *fs_info = inode->root->fs_info;
392 struct btrfs_workqueue *wq = btrfs_is_free_space_inode(inode) ?
393 fs_info->endio_freespace_worker : fs_info->endio_write_workers;
394
395 btrfs_init_work(&ordered->work, finish_ordered_fn, NULL);
396 btrfs_queue_work(wq, &ordered->work);
397}
398
399void btrfs_finish_ordered_extent(struct btrfs_ordered_extent *ordered,
400 struct folio *folio, u64 file_offset, u64 len,
401 bool uptodate)
402{
403 struct btrfs_inode *inode = ordered->inode;
404 unsigned long flags;
405 bool ret;
406
407 trace_btrfs_finish_ordered_extent(inode, file_offset, len, uptodate);
408
409 spin_lock_irqsave(&inode->ordered_tree_lock, flags);
410 ret = can_finish_ordered_extent(ordered, folio, file_offset, len,
411 uptodate);
412 spin_unlock_irqrestore(&inode->ordered_tree_lock, flags);
413
414 /*
415 * If this is a COW write it means we created new extent maps for the
416 * range and they point to unwritten locations if we got an error either
417 * before submitting a bio or during IO.
418 *
419 * We have marked the ordered extent with BTRFS_ORDERED_IOERR, and we
420 * are queuing its completion below. During completion, at
421 * btrfs_finish_one_ordered(), we will drop the extent maps for the
422 * unwritten extents.
423 *
424 * However because completion runs in a work queue we can end up having
425 * a fast fsync running before that. In the case of direct IO, once we
426 * unlock the inode the fsync might start, and we queue the completion
427 * before unlocking the inode. In the case of buffered IO when writeback
428 * finishes (end_bbio_data_write()) we queue the completion, so if the
429 * writeback was triggered by a fast fsync, the fsync might start
430 * logging before ordered extent completion runs in the work queue.
431 *
432 * The fast fsync will log file extent items based on the extent maps it
433 * finds, so if by the time it collects extent maps the ordered extent
434 * completion didn't happen yet, it will log file extent items that
435 * point to unwritten extents, resulting in a corruption if a crash
436 * happens and the log tree is replayed. Note that a fast fsync does not
437 * wait for completion of ordered extents in order to reduce latency.
438 *
439 * Set a flag in the inode so that the next fast fsync will wait for
440 * ordered extents to complete before starting to log.
441 */
442 if (!uptodate && !test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags))
443 set_bit(BTRFS_INODE_COW_WRITE_ERROR, &inode->runtime_flags);
444
445 if (ret)
446 btrfs_queue_ordered_fn(ordered);
447}
448
449/*
450 * Mark all ordered extents io inside the specified range finished.
451 *
452 * @folio: The involved folio for the operation.
453 * For uncompressed buffered IO, the folio status also needs to be
454 * updated to indicate whether the pending ordered io is finished.
455 * Can be NULL for direct IO and compressed write.
456 * For these cases, callers are ensured they won't execute the
457 * endio function twice.
458 *
459 * This function is called for endio, thus the range must have ordered
460 * extent(s) covering it.
461 */
462void btrfs_mark_ordered_io_finished(struct btrfs_inode *inode,
463 struct folio *folio, u64 file_offset,
464 u64 num_bytes, bool uptodate)
465{
466 struct rb_node *node;
467 struct btrfs_ordered_extent *entry = NULL;
468 unsigned long flags;
469 u64 cur = file_offset;
470
471 trace_btrfs_writepage_end_io_hook(inode, file_offset,
472 file_offset + num_bytes - 1,
473 uptodate);
474
475 spin_lock_irqsave(&inode->ordered_tree_lock, flags);
476 while (cur < file_offset + num_bytes) {
477 u64 entry_end;
478 u64 end;
479 u32 len;
480
481 node = ordered_tree_search(inode, cur);
482 /* No ordered extents at all */
483 if (!node)
484 break;
485
486 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
487 entry_end = entry->file_offset + entry->num_bytes;
488 /*
489 * |<-- OE --->| |
490 * cur
491 * Go to next OE.
492 */
493 if (cur >= entry_end) {
494 node = rb_next(node);
495 /* No more ordered extents, exit */
496 if (!node)
497 break;
498 entry = rb_entry(node, struct btrfs_ordered_extent,
499 rb_node);
500
501 /* Go to next ordered extent and continue */
502 cur = entry->file_offset;
503 continue;
504 }
505 /*
506 * | |<--- OE --->|
507 * cur
508 * Go to the start of OE.
509 */
510 if (cur < entry->file_offset) {
511 cur = entry->file_offset;
512 continue;
513 }
514
515 /*
516 * Now we are definitely inside one ordered extent.
517 *
518 * |<--- OE --->|
519 * |
520 * cur
521 */
522 end = min(entry->file_offset + entry->num_bytes,
523 file_offset + num_bytes) - 1;
524 ASSERT(end + 1 - cur < U32_MAX);
525 len = end + 1 - cur;
526
527 if (can_finish_ordered_extent(entry, folio, cur, len, uptodate)) {
528 spin_unlock_irqrestore(&inode->ordered_tree_lock, flags);
529 btrfs_queue_ordered_fn(entry);
530 spin_lock_irqsave(&inode->ordered_tree_lock, flags);
531 }
532 cur += len;
533 }
534 spin_unlock_irqrestore(&inode->ordered_tree_lock, flags);
535}
536
537/*
538 * Finish IO for one ordered extent across a given range. The range can only
539 * contain one ordered extent.
540 *
541 * @cached: The cached ordered extent. If not NULL, we can skip the tree
542 * search and use the ordered extent directly.
543 * Will be also used to store the finished ordered extent.
544 * @file_offset: File offset for the finished IO
545 * @io_size: Length of the finish IO range
546 *
547 * Return true if the ordered extent is finished in the range, and update
548 * @cached.
549 * Return false otherwise.
550 *
551 * NOTE: The range can NOT cross multiple ordered extents.
552 * Thus caller should ensure the range doesn't cross ordered extents.
553 */
554bool btrfs_dec_test_ordered_pending(struct btrfs_inode *inode,
555 struct btrfs_ordered_extent **cached,
556 u64 file_offset, u64 io_size)
557{
558 struct rb_node *node;
559 struct btrfs_ordered_extent *entry = NULL;
560 unsigned long flags;
561 bool finished = false;
562
563 spin_lock_irqsave(&inode->ordered_tree_lock, flags);
564 if (cached && *cached) {
565 entry = *cached;
566 goto have_entry;
567 }
568
569 node = ordered_tree_search(inode, file_offset);
570 if (!node)
571 goto out;
572
573 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
574have_entry:
575 if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
576 goto out;
577
578 if (io_size > entry->bytes_left)
579 btrfs_crit(inode->root->fs_info,
580 "bad ordered accounting left %llu size %llu",
581 entry->bytes_left, io_size);
582
583 entry->bytes_left -= io_size;
584
585 if (entry->bytes_left == 0) {
586 /*
587 * Ensure only one caller can set the flag and finished_ret
588 * accordingly
589 */
590 finished = !test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
591 /* test_and_set_bit implies a barrier */
592 cond_wake_up_nomb(&entry->wait);
593 }
594out:
595 if (finished && cached && entry) {
596 *cached = entry;
597 refcount_inc(&entry->refs);
598 trace_btrfs_ordered_extent_dec_test_pending(inode, entry);
599 }
600 spin_unlock_irqrestore(&inode->ordered_tree_lock, flags);
601 return finished;
602}
603
604/*
605 * used to drop a reference on an ordered extent. This will free
606 * the extent if the last reference is dropped
607 */
608void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
609{
610 struct list_head *cur;
611 struct btrfs_ordered_sum *sum;
612
613 trace_btrfs_ordered_extent_put(entry->inode, entry);
614
615 if (refcount_dec_and_test(&entry->refs)) {
616 ASSERT(list_empty(&entry->root_extent_list));
617 ASSERT(list_empty(&entry->log_list));
618 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
619 if (entry->inode)
620 btrfs_add_delayed_iput(entry->inode);
621 while (!list_empty(&entry->list)) {
622 cur = entry->list.next;
623 sum = list_entry(cur, struct btrfs_ordered_sum, list);
624 list_del(&sum->list);
625 kvfree(sum);
626 }
627 kmem_cache_free(btrfs_ordered_extent_cache, entry);
628 }
629}
630
631/*
632 * remove an ordered extent from the tree. No references are dropped
633 * and waiters are woken up.
634 */
635void btrfs_remove_ordered_extent(struct btrfs_inode *btrfs_inode,
636 struct btrfs_ordered_extent *entry)
637{
638 struct btrfs_root *root = btrfs_inode->root;
639 struct btrfs_fs_info *fs_info = root->fs_info;
640 struct rb_node *node;
641 bool pending;
642 bool freespace_inode;
643
644 /*
645 * If this is a free space inode the thread has not acquired the ordered
646 * extents lockdep map.
647 */
648 freespace_inode = btrfs_is_free_space_inode(btrfs_inode);
649
650 btrfs_lockdep_acquire(fs_info, btrfs_trans_pending_ordered);
651 /* This is paired with alloc_ordered_extent(). */
652 spin_lock(&btrfs_inode->lock);
653 btrfs_mod_outstanding_extents(btrfs_inode, -1);
654 spin_unlock(&btrfs_inode->lock);
655 if (root != fs_info->tree_root) {
656 u64 release;
657
658 if (test_bit(BTRFS_ORDERED_ENCODED, &entry->flags))
659 release = entry->disk_num_bytes;
660 else
661 release = entry->num_bytes;
662 btrfs_delalloc_release_metadata(btrfs_inode, release,
663 test_bit(BTRFS_ORDERED_IOERR,
664 &entry->flags));
665 }
666
667 percpu_counter_add_batch(&fs_info->ordered_bytes, -entry->num_bytes,
668 fs_info->delalloc_batch);
669
670 spin_lock_irq(&btrfs_inode->ordered_tree_lock);
671 node = &entry->rb_node;
672 rb_erase(node, &btrfs_inode->ordered_tree);
673 RB_CLEAR_NODE(node);
674 if (btrfs_inode->ordered_tree_last == node)
675 btrfs_inode->ordered_tree_last = NULL;
676 set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
677 pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
678 spin_unlock_irq(&btrfs_inode->ordered_tree_lock);
679
680 /*
681 * The current running transaction is waiting on us, we need to let it
682 * know that we're complete and wake it up.
683 */
684 if (pending) {
685 struct btrfs_transaction *trans;
686
687 /*
688 * The checks for trans are just a formality, it should be set,
689 * but if it isn't we don't want to deref/assert under the spin
690 * lock, so be nice and check if trans is set, but ASSERT() so
691 * if it isn't set a developer will notice.
692 */
693 spin_lock(&fs_info->trans_lock);
694 trans = fs_info->running_transaction;
695 if (trans)
696 refcount_inc(&trans->use_count);
697 spin_unlock(&fs_info->trans_lock);
698
699 ASSERT(trans || BTRFS_FS_ERROR(fs_info));
700 if (trans) {
701 if (atomic_dec_and_test(&trans->pending_ordered))
702 wake_up(&trans->pending_wait);
703 btrfs_put_transaction(trans);
704 }
705 }
706
707 btrfs_lockdep_release(fs_info, btrfs_trans_pending_ordered);
708
709 spin_lock(&root->ordered_extent_lock);
710 list_del_init(&entry->root_extent_list);
711 root->nr_ordered_extents--;
712
713 trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
714
715 if (!root->nr_ordered_extents) {
716 spin_lock(&fs_info->ordered_root_lock);
717 BUG_ON(list_empty(&root->ordered_root));
718 list_del_init(&root->ordered_root);
719 spin_unlock(&fs_info->ordered_root_lock);
720 }
721 spin_unlock(&root->ordered_extent_lock);
722 wake_up(&entry->wait);
723 if (!freespace_inode)
724 btrfs_lockdep_release(fs_info, btrfs_ordered_extent);
725}
726
727static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
728{
729 struct btrfs_ordered_extent *ordered;
730
731 ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
732 btrfs_start_ordered_extent(ordered);
733 complete(&ordered->completion);
734}
735
736/*
737 * Wait for all the ordered extents in a root. Use @bg as range or do whole
738 * range if it's NULL.
739 */
740u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
741 const struct btrfs_block_group *bg)
742{
743 struct btrfs_fs_info *fs_info = root->fs_info;
744 LIST_HEAD(splice);
745 LIST_HEAD(skipped);
746 LIST_HEAD(works);
747 struct btrfs_ordered_extent *ordered, *next;
748 u64 count = 0;
749 u64 range_start, range_len;
750 u64 range_end;
751
752 if (bg) {
753 range_start = bg->start;
754 range_len = bg->length;
755 } else {
756 range_start = 0;
757 range_len = U64_MAX;
758 }
759 range_end = range_start + range_len;
760
761 mutex_lock(&root->ordered_extent_mutex);
762 spin_lock(&root->ordered_extent_lock);
763 list_splice_init(&root->ordered_extents, &splice);
764 while (!list_empty(&splice) && nr) {
765 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
766 root_extent_list);
767
768 if (range_end <= ordered->disk_bytenr ||
769 ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
770 list_move_tail(&ordered->root_extent_list, &skipped);
771 cond_resched_lock(&root->ordered_extent_lock);
772 continue;
773 }
774
775 list_move_tail(&ordered->root_extent_list,
776 &root->ordered_extents);
777 refcount_inc(&ordered->refs);
778 spin_unlock(&root->ordered_extent_lock);
779
780 btrfs_init_work(&ordered->flush_work,
781 btrfs_run_ordered_extent_work, NULL);
782 list_add_tail(&ordered->work_list, &works);
783 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
784
785 cond_resched();
786 if (nr != U64_MAX)
787 nr--;
788 count++;
789 spin_lock(&root->ordered_extent_lock);
790 }
791 list_splice_tail(&skipped, &root->ordered_extents);
792 list_splice_tail(&splice, &root->ordered_extents);
793 spin_unlock(&root->ordered_extent_lock);
794
795 list_for_each_entry_safe(ordered, next, &works, work_list) {
796 list_del_init(&ordered->work_list);
797 wait_for_completion(&ordered->completion);
798 btrfs_put_ordered_extent(ordered);
799 cond_resched();
800 }
801 mutex_unlock(&root->ordered_extent_mutex);
802
803 return count;
804}
805
806/*
807 * Wait for @nr ordered extents that intersect the @bg, or the whole range of
808 * the filesystem if @bg is NULL.
809 */
810void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
811 const struct btrfs_block_group *bg)
812{
813 struct btrfs_root *root;
814 LIST_HEAD(splice);
815 u64 done;
816
817 mutex_lock(&fs_info->ordered_operations_mutex);
818 spin_lock(&fs_info->ordered_root_lock);
819 list_splice_init(&fs_info->ordered_roots, &splice);
820 while (!list_empty(&splice) && nr) {
821 root = list_first_entry(&splice, struct btrfs_root,
822 ordered_root);
823 root = btrfs_grab_root(root);
824 BUG_ON(!root);
825 list_move_tail(&root->ordered_root,
826 &fs_info->ordered_roots);
827 spin_unlock(&fs_info->ordered_root_lock);
828
829 done = btrfs_wait_ordered_extents(root, nr, bg);
830 btrfs_put_root(root);
831
832 if (nr != U64_MAX)
833 nr -= done;
834
835 spin_lock(&fs_info->ordered_root_lock);
836 }
837 list_splice_tail(&splice, &fs_info->ordered_roots);
838 spin_unlock(&fs_info->ordered_root_lock);
839 mutex_unlock(&fs_info->ordered_operations_mutex);
840}
841
842/*
843 * Start IO and wait for a given ordered extent to finish.
844 *
845 * Wait on page writeback for all the pages in the extent and the IO completion
846 * code to insert metadata into the btree corresponding to the extent.
847 */
848void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry)
849{
850 u64 start = entry->file_offset;
851 u64 end = start + entry->num_bytes - 1;
852 struct btrfs_inode *inode = entry->inode;
853 bool freespace_inode;
854
855 trace_btrfs_ordered_extent_start(inode, entry);
856
857 /*
858 * If this is a free space inode do not take the ordered extents lockdep
859 * map.
860 */
861 freespace_inode = btrfs_is_free_space_inode(inode);
862
863 /*
864 * pages in the range can be dirty, clean or writeback. We
865 * start IO on any dirty ones so the wait doesn't stall waiting
866 * for the flusher thread to find them
867 */
868 if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
869 filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
870
871 if (!freespace_inode)
872 btrfs_might_wait_for_event(inode->root->fs_info, btrfs_ordered_extent);
873 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE, &entry->flags));
874}
875
876/*
877 * Used to wait on ordered extents across a large range of bytes.
878 */
879int btrfs_wait_ordered_range(struct btrfs_inode *inode, u64 start, u64 len)
880{
881 int ret = 0;
882 int ret_wb = 0;
883 u64 end;
884 u64 orig_end;
885 struct btrfs_ordered_extent *ordered;
886
887 if (start + len < start) {
888 orig_end = OFFSET_MAX;
889 } else {
890 orig_end = start + len - 1;
891 if (orig_end > OFFSET_MAX)
892 orig_end = OFFSET_MAX;
893 }
894
895 /* start IO across the range first to instantiate any delalloc
896 * extents
897 */
898 ret = btrfs_fdatawrite_range(inode, start, orig_end);
899 if (ret)
900 return ret;
901
902 /*
903 * If we have a writeback error don't return immediately. Wait first
904 * for any ordered extents that haven't completed yet. This is to make
905 * sure no one can dirty the same page ranges and call writepages()
906 * before the ordered extents complete - to avoid failures (-EEXIST)
907 * when adding the new ordered extents to the ordered tree.
908 */
909 ret_wb = filemap_fdatawait_range(inode->vfs_inode.i_mapping, start, orig_end);
910
911 end = orig_end;
912 while (1) {
913 ordered = btrfs_lookup_first_ordered_extent(inode, end);
914 if (!ordered)
915 break;
916 if (ordered->file_offset > orig_end) {
917 btrfs_put_ordered_extent(ordered);
918 break;
919 }
920 if (ordered->file_offset + ordered->num_bytes <= start) {
921 btrfs_put_ordered_extent(ordered);
922 break;
923 }
924 btrfs_start_ordered_extent(ordered);
925 end = ordered->file_offset;
926 /*
927 * If the ordered extent had an error save the error but don't
928 * exit without waiting first for all other ordered extents in
929 * the range to complete.
930 */
931 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
932 ret = -EIO;
933 btrfs_put_ordered_extent(ordered);
934 if (end == 0 || end == start)
935 break;
936 end--;
937 }
938 return ret_wb ? ret_wb : ret;
939}
940
941/*
942 * find an ordered extent corresponding to file_offset. return NULL if
943 * nothing is found, otherwise take a reference on the extent and return it
944 */
945struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
946 u64 file_offset)
947{
948 struct rb_node *node;
949 struct btrfs_ordered_extent *entry = NULL;
950 unsigned long flags;
951
952 spin_lock_irqsave(&inode->ordered_tree_lock, flags);
953 node = ordered_tree_search(inode, file_offset);
954 if (!node)
955 goto out;
956
957 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
958 if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
959 entry = NULL;
960 if (entry) {
961 refcount_inc(&entry->refs);
962 trace_btrfs_ordered_extent_lookup(inode, entry);
963 }
964out:
965 spin_unlock_irqrestore(&inode->ordered_tree_lock, flags);
966 return entry;
967}
968
969/* Since the DIO code tries to lock a wide area we need to look for any ordered
970 * extents that exist in the range, rather than just the start of the range.
971 */
972struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
973 struct btrfs_inode *inode, u64 file_offset, u64 len)
974{
975 struct rb_node *node;
976 struct btrfs_ordered_extent *entry = NULL;
977
978 spin_lock_irq(&inode->ordered_tree_lock);
979 node = ordered_tree_search(inode, file_offset);
980 if (!node) {
981 node = ordered_tree_search(inode, file_offset + len);
982 if (!node)
983 goto out;
984 }
985
986 while (1) {
987 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
988 if (btrfs_range_overlaps(entry, file_offset, len))
989 break;
990
991 if (entry->file_offset >= file_offset + len) {
992 entry = NULL;
993 break;
994 }
995 entry = NULL;
996 node = rb_next(node);
997 if (!node)
998 break;
999 }
1000out:
1001 if (entry) {
1002 refcount_inc(&entry->refs);
1003 trace_btrfs_ordered_extent_lookup_range(inode, entry);
1004 }
1005 spin_unlock_irq(&inode->ordered_tree_lock);
1006 return entry;
1007}
1008
1009/*
1010 * Adds all ordered extents to the given list. The list ends up sorted by the
1011 * file_offset of the ordered extents.
1012 */
1013void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
1014 struct list_head *list)
1015{
1016 struct rb_node *n;
1017
1018 btrfs_assert_inode_locked(inode);
1019
1020 spin_lock_irq(&inode->ordered_tree_lock);
1021 for (n = rb_first(&inode->ordered_tree); n; n = rb_next(n)) {
1022 struct btrfs_ordered_extent *ordered;
1023
1024 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
1025
1026 if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
1027 continue;
1028
1029 ASSERT(list_empty(&ordered->log_list));
1030 list_add_tail(&ordered->log_list, list);
1031 refcount_inc(&ordered->refs);
1032 trace_btrfs_ordered_extent_lookup_for_logging(inode, ordered);
1033 }
1034 spin_unlock_irq(&inode->ordered_tree_lock);
1035}
1036
1037/*
1038 * lookup and return any extent before 'file_offset'. NULL is returned
1039 * if none is found
1040 */
1041struct btrfs_ordered_extent *
1042btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
1043{
1044 struct rb_node *node;
1045 struct btrfs_ordered_extent *entry = NULL;
1046
1047 spin_lock_irq(&inode->ordered_tree_lock);
1048 node = ordered_tree_search(inode, file_offset);
1049 if (!node)
1050 goto out;
1051
1052 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1053 refcount_inc(&entry->refs);
1054 trace_btrfs_ordered_extent_lookup_first(inode, entry);
1055out:
1056 spin_unlock_irq(&inode->ordered_tree_lock);
1057 return entry;
1058}
1059
1060/*
1061 * Lookup the first ordered extent that overlaps the range
1062 * [@file_offset, @file_offset + @len).
1063 *
1064 * The difference between this and btrfs_lookup_first_ordered_extent() is
1065 * that this one won't return any ordered extent that does not overlap the range.
1066 * And the difference against btrfs_lookup_ordered_extent() is, this function
1067 * ensures the first ordered extent gets returned.
1068 */
1069struct btrfs_ordered_extent *btrfs_lookup_first_ordered_range(
1070 struct btrfs_inode *inode, u64 file_offset, u64 len)
1071{
1072 struct rb_node *node;
1073 struct rb_node *cur;
1074 struct rb_node *prev;
1075 struct rb_node *next;
1076 struct btrfs_ordered_extent *entry = NULL;
1077
1078 spin_lock_irq(&inode->ordered_tree_lock);
1079 node = inode->ordered_tree.rb_node;
1080 /*
1081 * Here we don't want to use tree_search() which will use tree->last
1082 * and screw up the search order.
1083 * And __tree_search() can't return the adjacent ordered extents
1084 * either, thus here we do our own search.
1085 */
1086 while (node) {
1087 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1088
1089 if (file_offset < entry->file_offset) {
1090 node = node->rb_left;
1091 } else if (file_offset >= entry_end(entry)) {
1092 node = node->rb_right;
1093 } else {
1094 /*
1095 * Direct hit, got an ordered extent that starts at
1096 * @file_offset
1097 */
1098 goto out;
1099 }
1100 }
1101 if (!entry) {
1102 /* Empty tree */
1103 goto out;
1104 }
1105
1106 cur = &entry->rb_node;
1107 /* We got an entry around @file_offset, check adjacent entries */
1108 if (entry->file_offset < file_offset) {
1109 prev = cur;
1110 next = rb_next(cur);
1111 } else {
1112 prev = rb_prev(cur);
1113 next = cur;
1114 }
1115 if (prev) {
1116 entry = rb_entry(prev, struct btrfs_ordered_extent, rb_node);
1117 if (btrfs_range_overlaps(entry, file_offset, len))
1118 goto out;
1119 }
1120 if (next) {
1121 entry = rb_entry(next, struct btrfs_ordered_extent, rb_node);
1122 if (btrfs_range_overlaps(entry, file_offset, len))
1123 goto out;
1124 }
1125 /* No ordered extent in the range */
1126 entry = NULL;
1127out:
1128 if (entry) {
1129 refcount_inc(&entry->refs);
1130 trace_btrfs_ordered_extent_lookup_first_range(inode, entry);
1131 }
1132
1133 spin_unlock_irq(&inode->ordered_tree_lock);
1134 return entry;
1135}
1136
1137/*
1138 * Lock the passed range and ensures all pending ordered extents in it are run
1139 * to completion.
1140 *
1141 * @inode: Inode whose ordered tree is to be searched
1142 * @start: Beginning of range to flush
1143 * @end: Last byte of range to lock
1144 * @cached_state: If passed, will return the extent state responsible for the
1145 * locked range. It's the caller's responsibility to free the
1146 * cached state.
1147 *
1148 * Always return with the given range locked, ensuring after it's called no
1149 * order extent can be pending.
1150 */
1151void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
1152 u64 end,
1153 struct extent_state **cached_state)
1154{
1155 struct btrfs_ordered_extent *ordered;
1156 struct extent_state *cache = NULL;
1157 struct extent_state **cachedp = &cache;
1158
1159 if (cached_state)
1160 cachedp = cached_state;
1161
1162 while (1) {
1163 lock_extent(&inode->io_tree, start, end, cachedp);
1164 ordered = btrfs_lookup_ordered_range(inode, start,
1165 end - start + 1);
1166 if (!ordered) {
1167 /*
1168 * If no external cached_state has been passed then
1169 * decrement the extra ref taken for cachedp since we
1170 * aren't exposing it outside of this function
1171 */
1172 if (!cached_state)
1173 refcount_dec(&cache->refs);
1174 break;
1175 }
1176 unlock_extent(&inode->io_tree, start, end, cachedp);
1177 btrfs_start_ordered_extent(ordered);
1178 btrfs_put_ordered_extent(ordered);
1179 }
1180}
1181
1182/*
1183 * Lock the passed range and ensure all pending ordered extents in it are run
1184 * to completion in nowait mode.
1185 *
1186 * Return true if btrfs_lock_ordered_range does not return any extents,
1187 * otherwise false.
1188 */
1189bool btrfs_try_lock_ordered_range(struct btrfs_inode *inode, u64 start, u64 end,
1190 struct extent_state **cached_state)
1191{
1192 struct btrfs_ordered_extent *ordered;
1193
1194 if (!try_lock_extent(&inode->io_tree, start, end, cached_state))
1195 return false;
1196
1197 ordered = btrfs_lookup_ordered_range(inode, start, end - start + 1);
1198 if (!ordered)
1199 return true;
1200
1201 btrfs_put_ordered_extent(ordered);
1202 unlock_extent(&inode->io_tree, start, end, cached_state);
1203
1204 return false;
1205}
1206
1207/* Split out a new ordered extent for this first @len bytes of @ordered. */
1208struct btrfs_ordered_extent *btrfs_split_ordered_extent(
1209 struct btrfs_ordered_extent *ordered, u64 len)
1210{
1211 struct btrfs_inode *inode = ordered->inode;
1212 struct btrfs_root *root = inode->root;
1213 struct btrfs_fs_info *fs_info = root->fs_info;
1214 u64 file_offset = ordered->file_offset;
1215 u64 disk_bytenr = ordered->disk_bytenr;
1216 unsigned long flags = ordered->flags;
1217 struct btrfs_ordered_sum *sum, *tmpsum;
1218 struct btrfs_ordered_extent *new;
1219 struct rb_node *node;
1220 u64 offset = 0;
1221
1222 trace_btrfs_ordered_extent_split(inode, ordered);
1223
1224 ASSERT(!(flags & (1U << BTRFS_ORDERED_COMPRESSED)));
1225
1226 /*
1227 * The entire bio must be covered by the ordered extent, but we can't
1228 * reduce the original extent to a zero length either.
1229 */
1230 if (WARN_ON_ONCE(len >= ordered->num_bytes))
1231 return ERR_PTR(-EINVAL);
1232 /*
1233 * If our ordered extent had an error there's no point in continuing.
1234 * The error may have come from a transaction abort done either by this
1235 * task or some other concurrent task, and the transaction abort path
1236 * iterates over all existing ordered extents and sets the flag
1237 * BTRFS_ORDERED_IOERR on them.
1238 */
1239 if (unlikely(flags & (1U << BTRFS_ORDERED_IOERR))) {
1240 const int fs_error = BTRFS_FS_ERROR(fs_info);
1241
1242 return fs_error ? ERR_PTR(fs_error) : ERR_PTR(-EIO);
1243 }
1244 /* We cannot split partially completed ordered extents. */
1245 if (ordered->bytes_left) {
1246 ASSERT(!(flags & ~BTRFS_ORDERED_TYPE_FLAGS));
1247 if (WARN_ON_ONCE(ordered->bytes_left != ordered->disk_num_bytes))
1248 return ERR_PTR(-EINVAL);
1249 }
1250 /* We cannot split a compressed ordered extent. */
1251 if (WARN_ON_ONCE(ordered->disk_num_bytes != ordered->num_bytes))
1252 return ERR_PTR(-EINVAL);
1253
1254 new = alloc_ordered_extent(inode, file_offset, len, len, disk_bytenr,
1255 len, 0, flags, ordered->compress_type);
1256 if (IS_ERR(new))
1257 return new;
1258
1259 /* One ref for the tree. */
1260 refcount_inc(&new->refs);
1261
1262 /*
1263 * Take the root's ordered_extent_lock to avoid a race with
1264 * btrfs_wait_ordered_extents() when updating the disk_bytenr and
1265 * disk_num_bytes fields of the ordered extent below. And we disable
1266 * IRQs because the inode's ordered_tree_lock is used in IRQ context
1267 * elsewhere.
1268 *
1269 * There's no concern about a previous caller of
1270 * btrfs_wait_ordered_extents() getting the trimmed ordered extent
1271 * before we insert the new one, because even if it gets the ordered
1272 * extent before it's trimmed and the new one inserted, right before it
1273 * uses it or during its use, the ordered extent might have been
1274 * trimmed in the meanwhile, and it missed the new ordered extent.
1275 * There's no way around this and it's harmless for current use cases,
1276 * so we take the root's ordered_extent_lock to fix that race during
1277 * trimming and silence tools like KCSAN.
1278 */
1279 spin_lock_irq(&root->ordered_extent_lock);
1280 spin_lock(&inode->ordered_tree_lock);
1281
1282 /*
1283 * We don't have overlapping ordered extents (that would imply double
1284 * allocation of extents) and we checked above that the split length
1285 * does not cross the ordered extent's num_bytes field, so there's
1286 * no need to remove it and re-insert it in the tree.
1287 */
1288 ordered->file_offset += len;
1289 ordered->disk_bytenr += len;
1290 ordered->num_bytes -= len;
1291 ordered->disk_num_bytes -= len;
1292 ordered->ram_bytes -= len;
1293
1294 if (test_bit(BTRFS_ORDERED_IO_DONE, &ordered->flags)) {
1295 ASSERT(ordered->bytes_left == 0);
1296 new->bytes_left = 0;
1297 } else {
1298 ordered->bytes_left -= len;
1299 }
1300
1301 if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags)) {
1302 if (ordered->truncated_len > len) {
1303 ordered->truncated_len -= len;
1304 } else {
1305 new->truncated_len = ordered->truncated_len;
1306 ordered->truncated_len = 0;
1307 }
1308 }
1309
1310 list_for_each_entry_safe(sum, tmpsum, &ordered->list, list) {
1311 if (offset == len)
1312 break;
1313 list_move_tail(&sum->list, &new->list);
1314 offset += sum->len;
1315 }
1316
1317 node = tree_insert(&inode->ordered_tree, new->file_offset, &new->rb_node);
1318 if (unlikely(node))
1319 btrfs_panic(fs_info, -EEXIST,
1320 "inconsistency in ordered tree at offset %llu after split",
1321 new->file_offset);
1322 spin_unlock(&inode->ordered_tree_lock);
1323
1324 list_add_tail(&new->root_extent_list, &root->ordered_extents);
1325 root->nr_ordered_extents++;
1326 spin_unlock_irq(&root->ordered_extent_lock);
1327 return new;
1328}
1329
1330int __init ordered_data_init(void)
1331{
1332 btrfs_ordered_extent_cache = KMEM_CACHE(btrfs_ordered_extent, 0);
1333 if (!btrfs_ordered_extent_cache)
1334 return -ENOMEM;
1335
1336 return 0;
1337}
1338
1339void __cold ordered_data_exit(void)
1340{
1341 kmem_cache_destroy(btrfs_ordered_extent_cache);
1342}