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