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 "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}