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