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/*
2 * Copyright (C) 2007 Oracle. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/slab.h>
20#include <linux/blkdev.h>
21#include <linux/writeback.h>
22#include <linux/pagevec.h>
23#include "ctree.h"
24#include "transaction.h"
25#include "btrfs_inode.h"
26#include "extent_io.h"
27#include "disk-io.h"
28
29static struct kmem_cache *btrfs_ordered_extent_cache;
30
31static u64 entry_end(struct btrfs_ordered_extent *entry)
32{
33 if (entry->file_offset + entry->len < entry->file_offset)
34 return (u64)-1;
35 return entry->file_offset + entry->len;
36}
37
38/* returns NULL if the insertion worked, or it returns the node it did find
39 * in the tree
40 */
41static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
42 struct rb_node *node)
43{
44 struct rb_node **p = &root->rb_node;
45 struct rb_node *parent = NULL;
46 struct btrfs_ordered_extent *entry;
47
48 while (*p) {
49 parent = *p;
50 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
51
52 if (file_offset < entry->file_offset)
53 p = &(*p)->rb_left;
54 else if (file_offset >= entry_end(entry))
55 p = &(*p)->rb_right;
56 else
57 return parent;
58 }
59
60 rb_link_node(node, parent, p);
61 rb_insert_color(node, root);
62 return NULL;
63}
64
65static void ordered_data_tree_panic(struct inode *inode, int errno,
66 u64 offset)
67{
68 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
69 btrfs_panic(fs_info, errno, "Inconsistency in ordered tree at offset "
70 "%llu\n", offset);
71}
72
73/*
74 * look for a given offset in the tree, and if it can't be found return the
75 * first lesser offset
76 */
77static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
78 struct rb_node **prev_ret)
79{
80 struct rb_node *n = root->rb_node;
81 struct rb_node *prev = NULL;
82 struct rb_node *test;
83 struct btrfs_ordered_extent *entry;
84 struct btrfs_ordered_extent *prev_entry = NULL;
85
86 while (n) {
87 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
88 prev = n;
89 prev_entry = entry;
90
91 if (file_offset < entry->file_offset)
92 n = n->rb_left;
93 else if (file_offset >= entry_end(entry))
94 n = n->rb_right;
95 else
96 return n;
97 }
98 if (!prev_ret)
99 return NULL;
100
101 while (prev && file_offset >= entry_end(prev_entry)) {
102 test = rb_next(prev);
103 if (!test)
104 break;
105 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
106 rb_node);
107 if (file_offset < entry_end(prev_entry))
108 break;
109
110 prev = test;
111 }
112 if (prev)
113 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
114 rb_node);
115 while (prev && file_offset < entry_end(prev_entry)) {
116 test = rb_prev(prev);
117 if (!test)
118 break;
119 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
120 rb_node);
121 prev = test;
122 }
123 *prev_ret = prev;
124 return NULL;
125}
126
127/*
128 * helper to check if a given offset is inside a given entry
129 */
130static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
131{
132 if (file_offset < entry->file_offset ||
133 entry->file_offset + entry->len <= file_offset)
134 return 0;
135 return 1;
136}
137
138static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
139 u64 len)
140{
141 if (file_offset + len <= entry->file_offset ||
142 entry->file_offset + entry->len <= file_offset)
143 return 0;
144 return 1;
145}
146
147/*
148 * look find the first ordered struct that has this offset, otherwise
149 * the first one less than this offset
150 */
151static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
152 u64 file_offset)
153{
154 struct rb_root *root = &tree->tree;
155 struct rb_node *prev = NULL;
156 struct rb_node *ret;
157 struct btrfs_ordered_extent *entry;
158
159 if (tree->last) {
160 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
161 rb_node);
162 if (offset_in_entry(entry, file_offset))
163 return tree->last;
164 }
165 ret = __tree_search(root, file_offset, &prev);
166 if (!ret)
167 ret = prev;
168 if (ret)
169 tree->last = ret;
170 return ret;
171}
172
173/* allocate and add a new ordered_extent into the per-inode tree.
174 * file_offset is the logical offset in the file
175 *
176 * start is the disk block number of an extent already reserved in the
177 * extent allocation tree
178 *
179 * len is the length of the extent
180 *
181 * The tree is given a single reference on the ordered extent that was
182 * inserted.
183 */
184static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
185 u64 start, u64 len, u64 disk_len,
186 int type, int dio, int compress_type)
187{
188 struct btrfs_root *root = BTRFS_I(inode)->root;
189 struct btrfs_ordered_inode_tree *tree;
190 struct rb_node *node;
191 struct btrfs_ordered_extent *entry;
192
193 tree = &BTRFS_I(inode)->ordered_tree;
194 entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
195 if (!entry)
196 return -ENOMEM;
197
198 entry->file_offset = file_offset;
199 entry->start = start;
200 entry->len = len;
201 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM) &&
202 !(type == BTRFS_ORDERED_NOCOW))
203 entry->csum_bytes_left = disk_len;
204 entry->disk_len = disk_len;
205 entry->bytes_left = len;
206 entry->inode = igrab(inode);
207 entry->compress_type = compress_type;
208 entry->truncated_len = (u64)-1;
209 if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
210 set_bit(type, &entry->flags);
211
212 if (dio)
213 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
214
215 /* one ref for the tree */
216 atomic_set(&entry->refs, 1);
217 init_waitqueue_head(&entry->wait);
218 INIT_LIST_HEAD(&entry->list);
219 INIT_LIST_HEAD(&entry->root_extent_list);
220 INIT_LIST_HEAD(&entry->work_list);
221 init_completion(&entry->completion);
222 INIT_LIST_HEAD(&entry->log_list);
223
224 trace_btrfs_ordered_extent_add(inode, entry);
225
226 spin_lock_irq(&tree->lock);
227 node = tree_insert(&tree->tree, file_offset,
228 &entry->rb_node);
229 if (node)
230 ordered_data_tree_panic(inode, -EEXIST, file_offset);
231 spin_unlock_irq(&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(&root->fs_info->ordered_root_lock);
239 BUG_ON(!list_empty(&root->ordered_root));
240 list_add_tail(&root->ordered_root,
241 &root->fs_info->ordered_roots);
242 spin_unlock(&root->fs_info->ordered_root_lock);
243 }
244 spin_unlock(&root->ordered_extent_lock);
245
246 return 0;
247}
248
249int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
250 u64 start, u64 len, u64 disk_len, int type)
251{
252 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
253 disk_len, type, 0,
254 BTRFS_COMPRESS_NONE);
255}
256
257int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
258 u64 start, u64 len, u64 disk_len, int type)
259{
260 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
261 disk_len, type, 1,
262 BTRFS_COMPRESS_NONE);
263}
264
265int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
266 u64 start, u64 len, u64 disk_len,
267 int type, int compress_type)
268{
269 return __btrfs_add_ordered_extent(inode, file_offset, start, len,
270 disk_len, type, 0,
271 compress_type);
272}
273
274/*
275 * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
276 * when an ordered extent is finished. If the list covers more than one
277 * ordered extent, it is split across multiples.
278 */
279void btrfs_add_ordered_sum(struct inode *inode,
280 struct btrfs_ordered_extent *entry,
281 struct btrfs_ordered_sum *sum)
282{
283 struct btrfs_ordered_inode_tree *tree;
284
285 tree = &BTRFS_I(inode)->ordered_tree;
286 spin_lock_irq(&tree->lock);
287 list_add_tail(&sum->list, &entry->list);
288 WARN_ON(entry->csum_bytes_left < sum->len);
289 entry->csum_bytes_left -= sum->len;
290 if (entry->csum_bytes_left == 0)
291 wake_up(&entry->wait);
292 spin_unlock_irq(&tree->lock);
293}
294
295/*
296 * this is used to account for finished IO across a given range
297 * of the file. The IO may span ordered extents. If
298 * a given ordered_extent is completely done, 1 is returned, otherwise
299 * 0.
300 *
301 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
302 * to make sure this function only returns 1 once for a given ordered extent.
303 *
304 * file_offset is updated to one byte past the range that is recorded as
305 * complete. This allows you to walk forward in the file.
306 */
307int btrfs_dec_test_first_ordered_pending(struct inode *inode,
308 struct btrfs_ordered_extent **cached,
309 u64 *file_offset, u64 io_size, int uptodate)
310{
311 struct btrfs_ordered_inode_tree *tree;
312 struct rb_node *node;
313 struct btrfs_ordered_extent *entry = NULL;
314 int ret;
315 unsigned long flags;
316 u64 dec_end;
317 u64 dec_start;
318 u64 to_dec;
319
320 tree = &BTRFS_I(inode)->ordered_tree;
321 spin_lock_irqsave(&tree->lock, flags);
322 node = tree_search(tree, *file_offset);
323 if (!node) {
324 ret = 1;
325 goto out;
326 }
327
328 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
329 if (!offset_in_entry(entry, *file_offset)) {
330 ret = 1;
331 goto out;
332 }
333
334 dec_start = max(*file_offset, entry->file_offset);
335 dec_end = min(*file_offset + io_size, entry->file_offset +
336 entry->len);
337 *file_offset = dec_end;
338 if (dec_start > dec_end) {
339 btrfs_crit(BTRFS_I(inode)->root->fs_info,
340 "bad ordering dec_start %llu end %llu", dec_start, dec_end);
341 }
342 to_dec = dec_end - dec_start;
343 if (to_dec > entry->bytes_left) {
344 btrfs_crit(BTRFS_I(inode)->root->fs_info,
345 "bad ordered accounting left %llu size %llu",
346 entry->bytes_left, to_dec);
347 }
348 entry->bytes_left -= to_dec;
349 if (!uptodate)
350 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
351
352 if (entry->bytes_left == 0) {
353 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
354 if (waitqueue_active(&entry->wait))
355 wake_up(&entry->wait);
356 } else {
357 ret = 1;
358 }
359out:
360 if (!ret && cached && entry) {
361 *cached = entry;
362 atomic_inc(&entry->refs);
363 }
364 spin_unlock_irqrestore(&tree->lock, flags);
365 return ret == 0;
366}
367
368/*
369 * this is used to account for finished IO across a given range
370 * of the file. The IO should not span ordered extents. If
371 * a given ordered_extent is completely done, 1 is returned, otherwise
372 * 0.
373 *
374 * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
375 * to make sure this function only returns 1 once for a given ordered extent.
376 */
377int btrfs_dec_test_ordered_pending(struct inode *inode,
378 struct btrfs_ordered_extent **cached,
379 u64 file_offset, u64 io_size, int uptodate)
380{
381 struct btrfs_ordered_inode_tree *tree;
382 struct rb_node *node;
383 struct btrfs_ordered_extent *entry = NULL;
384 unsigned long flags;
385 int ret;
386
387 tree = &BTRFS_I(inode)->ordered_tree;
388 spin_lock_irqsave(&tree->lock, flags);
389 if (cached && *cached) {
390 entry = *cached;
391 goto have_entry;
392 }
393
394 node = tree_search(tree, file_offset);
395 if (!node) {
396 ret = 1;
397 goto out;
398 }
399
400 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
401have_entry:
402 if (!offset_in_entry(entry, file_offset)) {
403 ret = 1;
404 goto out;
405 }
406
407 if (io_size > entry->bytes_left) {
408 btrfs_crit(BTRFS_I(inode)->root->fs_info,
409 "bad ordered accounting left %llu size %llu",
410 entry->bytes_left, io_size);
411 }
412 entry->bytes_left -= io_size;
413 if (!uptodate)
414 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
415
416 if (entry->bytes_left == 0) {
417 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
418 if (waitqueue_active(&entry->wait))
419 wake_up(&entry->wait);
420 } else {
421 ret = 1;
422 }
423out:
424 if (!ret && cached && entry) {
425 *cached = entry;
426 atomic_inc(&entry->refs);
427 }
428 spin_unlock_irqrestore(&tree->lock, flags);
429 return ret == 0;
430}
431
432/* Needs to either be called under a log transaction or the log_mutex */
433void btrfs_get_logged_extents(struct inode *inode,
434 struct list_head *logged_list)
435{
436 struct btrfs_ordered_inode_tree *tree;
437 struct btrfs_ordered_extent *ordered;
438 struct rb_node *n;
439
440 tree = &BTRFS_I(inode)->ordered_tree;
441 spin_lock_irq(&tree->lock);
442 for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
443 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
444 if (!list_empty(&ordered->log_list))
445 continue;
446 list_add_tail(&ordered->log_list, logged_list);
447 atomic_inc(&ordered->refs);
448 }
449 spin_unlock_irq(&tree->lock);
450}
451
452void btrfs_put_logged_extents(struct list_head *logged_list)
453{
454 struct btrfs_ordered_extent *ordered;
455
456 while (!list_empty(logged_list)) {
457 ordered = list_first_entry(logged_list,
458 struct btrfs_ordered_extent,
459 log_list);
460 list_del_init(&ordered->log_list);
461 btrfs_put_ordered_extent(ordered);
462 }
463}
464
465void btrfs_submit_logged_extents(struct list_head *logged_list,
466 struct btrfs_root *log)
467{
468 int index = log->log_transid % 2;
469
470 spin_lock_irq(&log->log_extents_lock[index]);
471 list_splice_tail(logged_list, &log->logged_list[index]);
472 spin_unlock_irq(&log->log_extents_lock[index]);
473}
474
475void btrfs_wait_logged_extents(struct btrfs_root *log, u64 transid)
476{
477 struct btrfs_ordered_extent *ordered;
478 int index = transid % 2;
479
480 spin_lock_irq(&log->log_extents_lock[index]);
481 while (!list_empty(&log->logged_list[index])) {
482 ordered = list_first_entry(&log->logged_list[index],
483 struct btrfs_ordered_extent,
484 log_list);
485 list_del_init(&ordered->log_list);
486 spin_unlock_irq(&log->log_extents_lock[index]);
487 wait_event(ordered->wait, test_bit(BTRFS_ORDERED_IO_DONE,
488 &ordered->flags));
489 btrfs_put_ordered_extent(ordered);
490 spin_lock_irq(&log->log_extents_lock[index]);
491 }
492 spin_unlock_irq(&log->log_extents_lock[index]);
493}
494
495void btrfs_free_logged_extents(struct btrfs_root *log, u64 transid)
496{
497 struct btrfs_ordered_extent *ordered;
498 int index = transid % 2;
499
500 spin_lock_irq(&log->log_extents_lock[index]);
501 while (!list_empty(&log->logged_list[index])) {
502 ordered = list_first_entry(&log->logged_list[index],
503 struct btrfs_ordered_extent,
504 log_list);
505 list_del_init(&ordered->log_list);
506 spin_unlock_irq(&log->log_extents_lock[index]);
507 btrfs_put_ordered_extent(ordered);
508 spin_lock_irq(&log->log_extents_lock[index]);
509 }
510 spin_unlock_irq(&log->log_extents_lock[index]);
511}
512
513/*
514 * used to drop a reference on an ordered extent. This will free
515 * the extent if the last reference is dropped
516 */
517void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
518{
519 struct list_head *cur;
520 struct btrfs_ordered_sum *sum;
521
522 trace_btrfs_ordered_extent_put(entry->inode, entry);
523
524 if (atomic_dec_and_test(&entry->refs)) {
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 kfree(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 inode *inode,
542 struct btrfs_ordered_extent *entry)
543{
544 struct btrfs_ordered_inode_tree *tree;
545 struct btrfs_root *root = BTRFS_I(inode)->root;
546 struct rb_node *node;
547
548 tree = &BTRFS_I(inode)->ordered_tree;
549 spin_lock_irq(&tree->lock);
550 node = &entry->rb_node;
551 rb_erase(node, &tree->tree);
552 if (tree->last == node)
553 tree->last = NULL;
554 set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
555 spin_unlock_irq(&tree->lock);
556
557 spin_lock(&root->ordered_extent_lock);
558 list_del_init(&entry->root_extent_list);
559 root->nr_ordered_extents--;
560
561 trace_btrfs_ordered_extent_remove(inode, entry);
562
563 /*
564 * we have no more ordered extents for this inode and
565 * no dirty pages. We can safely remove it from the
566 * list of ordered extents
567 */
568 if (RB_EMPTY_ROOT(&tree->tree) &&
569 !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
570 spin_lock(&root->fs_info->ordered_root_lock);
571 list_del_init(&BTRFS_I(inode)->ordered_operations);
572 spin_unlock(&root->fs_info->ordered_root_lock);
573 }
574
575 if (!root->nr_ordered_extents) {
576 spin_lock(&root->fs_info->ordered_root_lock);
577 BUG_ON(list_empty(&root->ordered_root));
578 list_del_init(&root->ordered_root);
579 spin_unlock(&root->fs_info->ordered_root_lock);
580 }
581 spin_unlock(&root->ordered_extent_lock);
582 wake_up(&entry->wait);
583}
584
585static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
586{
587 struct btrfs_ordered_extent *ordered;
588
589 ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
590 btrfs_start_ordered_extent(ordered->inode, ordered, 1);
591 complete(&ordered->completion);
592}
593
594/*
595 * wait for all the ordered extents in a root. This is done when balancing
596 * space between drives.
597 */
598int btrfs_wait_ordered_extents(struct btrfs_root *root, int nr)
599{
600 struct list_head splice, works;
601 struct btrfs_ordered_extent *ordered, *next;
602 int count = 0;
603
604 INIT_LIST_HEAD(&splice);
605 INIT_LIST_HEAD(&works);
606
607 mutex_lock(&root->ordered_extent_mutex);
608 spin_lock(&root->ordered_extent_lock);
609 list_splice_init(&root->ordered_extents, &splice);
610 while (!list_empty(&splice) && nr) {
611 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
612 root_extent_list);
613 list_move_tail(&ordered->root_extent_list,
614 &root->ordered_extents);
615 atomic_inc(&ordered->refs);
616 spin_unlock(&root->ordered_extent_lock);
617
618 btrfs_init_work(&ordered->flush_work,
619 btrfs_run_ordered_extent_work, NULL, NULL);
620 list_add_tail(&ordered->work_list, &works);
621 btrfs_queue_work(root->fs_info->flush_workers,
622 &ordered->flush_work);
623
624 cond_resched();
625 spin_lock(&root->ordered_extent_lock);
626 if (nr != -1)
627 nr--;
628 count++;
629 }
630 list_splice_tail(&splice, &root->ordered_extents);
631 spin_unlock(&root->ordered_extent_lock);
632
633 list_for_each_entry_safe(ordered, next, &works, work_list) {
634 list_del_init(&ordered->work_list);
635 wait_for_completion(&ordered->completion);
636 btrfs_put_ordered_extent(ordered);
637 cond_resched();
638 }
639 mutex_unlock(&root->ordered_extent_mutex);
640
641 return count;
642}
643
644void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, int nr)
645{
646 struct btrfs_root *root;
647 struct list_head splice;
648 int done;
649
650 INIT_LIST_HEAD(&splice);
651
652 mutex_lock(&fs_info->ordered_operations_mutex);
653 spin_lock(&fs_info->ordered_root_lock);
654 list_splice_init(&fs_info->ordered_roots, &splice);
655 while (!list_empty(&splice) && nr) {
656 root = list_first_entry(&splice, struct btrfs_root,
657 ordered_root);
658 root = btrfs_grab_fs_root(root);
659 BUG_ON(!root);
660 list_move_tail(&root->ordered_root,
661 &fs_info->ordered_roots);
662 spin_unlock(&fs_info->ordered_root_lock);
663
664 done = btrfs_wait_ordered_extents(root, nr);
665 btrfs_put_fs_root(root);
666
667 spin_lock(&fs_info->ordered_root_lock);
668 if (nr != -1) {
669 nr -= done;
670 WARN_ON(nr < 0);
671 }
672 }
673 list_splice_tail(&splice, &fs_info->ordered_roots);
674 spin_unlock(&fs_info->ordered_root_lock);
675 mutex_unlock(&fs_info->ordered_operations_mutex);
676}
677
678/*
679 * this is used during transaction commit to write all the inodes
680 * added to the ordered operation list. These files must be fully on
681 * disk before the transaction commits.
682 *
683 * we have two modes here, one is to just start the IO via filemap_flush
684 * and the other is to wait for all the io. When we wait, we have an
685 * extra check to make sure the ordered operation list really is empty
686 * before we return
687 */
688int btrfs_run_ordered_operations(struct btrfs_trans_handle *trans,
689 struct btrfs_root *root, int wait)
690{
691 struct btrfs_inode *btrfs_inode;
692 struct inode *inode;
693 struct btrfs_transaction *cur_trans = trans->transaction;
694 struct list_head splice;
695 struct list_head works;
696 struct btrfs_delalloc_work *work, *next;
697 int ret = 0;
698
699 INIT_LIST_HEAD(&splice);
700 INIT_LIST_HEAD(&works);
701
702 mutex_lock(&root->fs_info->ordered_extent_flush_mutex);
703 spin_lock(&root->fs_info->ordered_root_lock);
704 list_splice_init(&cur_trans->ordered_operations, &splice);
705 while (!list_empty(&splice)) {
706 btrfs_inode = list_entry(splice.next, struct btrfs_inode,
707 ordered_operations);
708 inode = &btrfs_inode->vfs_inode;
709
710 list_del_init(&btrfs_inode->ordered_operations);
711
712 /*
713 * the inode may be getting freed (in sys_unlink path).
714 */
715 inode = igrab(inode);
716 if (!inode)
717 continue;
718
719 if (!wait)
720 list_add_tail(&BTRFS_I(inode)->ordered_operations,
721 &cur_trans->ordered_operations);
722 spin_unlock(&root->fs_info->ordered_root_lock);
723
724 work = btrfs_alloc_delalloc_work(inode, wait, 1);
725 if (!work) {
726 spin_lock(&root->fs_info->ordered_root_lock);
727 if (list_empty(&BTRFS_I(inode)->ordered_operations))
728 list_add_tail(&btrfs_inode->ordered_operations,
729 &splice);
730 list_splice_tail(&splice,
731 &cur_trans->ordered_operations);
732 spin_unlock(&root->fs_info->ordered_root_lock);
733 ret = -ENOMEM;
734 goto out;
735 }
736 list_add_tail(&work->list, &works);
737 btrfs_queue_work(root->fs_info->flush_workers,
738 &work->work);
739
740 cond_resched();
741 spin_lock(&root->fs_info->ordered_root_lock);
742 }
743 spin_unlock(&root->fs_info->ordered_root_lock);
744out:
745 list_for_each_entry_safe(work, next, &works, list) {
746 list_del_init(&work->list);
747 btrfs_wait_and_free_delalloc_work(work);
748 }
749 mutex_unlock(&root->fs_info->ordered_extent_flush_mutex);
750 return ret;
751}
752
753/*
754 * Used to start IO or wait for a given ordered extent to finish.
755 *
756 * If wait is one, this effectively waits on page writeback for all the pages
757 * in the extent, and it waits on the io completion code to insert
758 * metadata into the btree corresponding to the extent
759 */
760void btrfs_start_ordered_extent(struct inode *inode,
761 struct btrfs_ordered_extent *entry,
762 int wait)
763{
764 u64 start = entry->file_offset;
765 u64 end = start + entry->len - 1;
766
767 trace_btrfs_ordered_extent_start(inode, entry);
768
769 /*
770 * pages in the range can be dirty, clean or writeback. We
771 * start IO on any dirty ones so the wait doesn't stall waiting
772 * for the flusher thread to find them
773 */
774 if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
775 filemap_fdatawrite_range(inode->i_mapping, start, end);
776 if (wait) {
777 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
778 &entry->flags));
779 }
780}
781
782/*
783 * Used to wait on ordered extents across a large range of bytes.
784 */
785int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
786{
787 int ret = 0;
788 u64 end;
789 u64 orig_end;
790 struct btrfs_ordered_extent *ordered;
791
792 if (start + len < start) {
793 orig_end = INT_LIMIT(loff_t);
794 } else {
795 orig_end = start + len - 1;
796 if (orig_end > INT_LIMIT(loff_t))
797 orig_end = INT_LIMIT(loff_t);
798 }
799
800 /* start IO across the range first to instantiate any delalloc
801 * extents
802 */
803 ret = filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
804 if (ret)
805 return ret;
806 /*
807 * So with compression we will find and lock a dirty page and clear the
808 * first one as dirty, setup an async extent, and immediately return
809 * with the entire range locked but with nobody actually marked with
810 * writeback. So we can't just filemap_write_and_wait_range() and
811 * expect it to work since it will just kick off a thread to do the
812 * actual work. So we need to call filemap_fdatawrite_range _again_
813 * since it will wait on the page lock, which won't be unlocked until
814 * after the pages have been marked as writeback and so we're good to go
815 * from there. We have to do this otherwise we'll miss the ordered
816 * extents and that results in badness. Please Josef, do not think you
817 * know better and pull this out at some point in the future, it is
818 * right and you are wrong.
819 */
820 if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
821 &BTRFS_I(inode)->runtime_flags)) {
822 ret = filemap_fdatawrite_range(inode->i_mapping, start,
823 orig_end);
824 if (ret)
825 return ret;
826 }
827 ret = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
828 if (ret)
829 return ret;
830
831 end = orig_end;
832 while (1) {
833 ordered = btrfs_lookup_first_ordered_extent(inode, end);
834 if (!ordered)
835 break;
836 if (ordered->file_offset > orig_end) {
837 btrfs_put_ordered_extent(ordered);
838 break;
839 }
840 if (ordered->file_offset + ordered->len <= start) {
841 btrfs_put_ordered_extent(ordered);
842 break;
843 }
844 btrfs_start_ordered_extent(inode, ordered, 1);
845 end = ordered->file_offset;
846 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
847 ret = -EIO;
848 btrfs_put_ordered_extent(ordered);
849 if (ret || end == 0 || end == start)
850 break;
851 end--;
852 }
853 return ret;
854}
855
856/*
857 * find an ordered extent corresponding to file_offset. return NULL if
858 * nothing is found, otherwise take a reference on the extent and return it
859 */
860struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
861 u64 file_offset)
862{
863 struct btrfs_ordered_inode_tree *tree;
864 struct rb_node *node;
865 struct btrfs_ordered_extent *entry = NULL;
866
867 tree = &BTRFS_I(inode)->ordered_tree;
868 spin_lock_irq(&tree->lock);
869 node = tree_search(tree, file_offset);
870 if (!node)
871 goto out;
872
873 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
874 if (!offset_in_entry(entry, file_offset))
875 entry = NULL;
876 if (entry)
877 atomic_inc(&entry->refs);
878out:
879 spin_unlock_irq(&tree->lock);
880 return entry;
881}
882
883/* Since the DIO code tries to lock a wide area we need to look for any ordered
884 * extents that exist in the range, rather than just the start of the range.
885 */
886struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
887 u64 file_offset,
888 u64 len)
889{
890 struct btrfs_ordered_inode_tree *tree;
891 struct rb_node *node;
892 struct btrfs_ordered_extent *entry = NULL;
893
894 tree = &BTRFS_I(inode)->ordered_tree;
895 spin_lock_irq(&tree->lock);
896 node = tree_search(tree, file_offset);
897 if (!node) {
898 node = tree_search(tree, file_offset + len);
899 if (!node)
900 goto out;
901 }
902
903 while (1) {
904 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
905 if (range_overlaps(entry, file_offset, len))
906 break;
907
908 if (entry->file_offset >= file_offset + len) {
909 entry = NULL;
910 break;
911 }
912 entry = NULL;
913 node = rb_next(node);
914 if (!node)
915 break;
916 }
917out:
918 if (entry)
919 atomic_inc(&entry->refs);
920 spin_unlock_irq(&tree->lock);
921 return entry;
922}
923
924/*
925 * lookup and return any extent before 'file_offset'. NULL is returned
926 * if none is found
927 */
928struct btrfs_ordered_extent *
929btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
930{
931 struct btrfs_ordered_inode_tree *tree;
932 struct rb_node *node;
933 struct btrfs_ordered_extent *entry = NULL;
934
935 tree = &BTRFS_I(inode)->ordered_tree;
936 spin_lock_irq(&tree->lock);
937 node = tree_search(tree, file_offset);
938 if (!node)
939 goto out;
940
941 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
942 atomic_inc(&entry->refs);
943out:
944 spin_unlock_irq(&tree->lock);
945 return entry;
946}
947
948/*
949 * After an extent is done, call this to conditionally update the on disk
950 * i_size. i_size is updated to cover any fully written part of the file.
951 */
952int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
953 struct btrfs_ordered_extent *ordered)
954{
955 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
956 u64 disk_i_size;
957 u64 new_i_size;
958 u64 i_size = i_size_read(inode);
959 struct rb_node *node;
960 struct rb_node *prev = NULL;
961 struct btrfs_ordered_extent *test;
962 int ret = 1;
963
964 spin_lock_irq(&tree->lock);
965 if (ordered) {
966 offset = entry_end(ordered);
967 if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags))
968 offset = min(offset,
969 ordered->file_offset +
970 ordered->truncated_len);
971 } else {
972 offset = ALIGN(offset, BTRFS_I(inode)->root->sectorsize);
973 }
974 disk_i_size = BTRFS_I(inode)->disk_i_size;
975
976 /* truncate file */
977 if (disk_i_size > i_size) {
978 BTRFS_I(inode)->disk_i_size = i_size;
979 ret = 0;
980 goto out;
981 }
982
983 /*
984 * if the disk i_size is already at the inode->i_size, or
985 * this ordered extent is inside the disk i_size, we're done
986 */
987 if (disk_i_size == i_size)
988 goto out;
989
990 /*
991 * We still need to update disk_i_size if outstanding_isize is greater
992 * than disk_i_size.
993 */
994 if (offset <= disk_i_size &&
995 (!ordered || ordered->outstanding_isize <= disk_i_size))
996 goto out;
997
998 /*
999 * walk backward from this ordered extent to disk_i_size.
1000 * if we find an ordered extent then we can't update disk i_size
1001 * yet
1002 */
1003 if (ordered) {
1004 node = rb_prev(&ordered->rb_node);
1005 } else {
1006 prev = tree_search(tree, offset);
1007 /*
1008 * we insert file extents without involving ordered struct,
1009 * so there should be no ordered struct cover this offset
1010 */
1011 if (prev) {
1012 test = rb_entry(prev, struct btrfs_ordered_extent,
1013 rb_node);
1014 BUG_ON(offset_in_entry(test, offset));
1015 }
1016 node = prev;
1017 }
1018 for (; node; node = rb_prev(node)) {
1019 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
1020
1021 /* We treat this entry as if it doesnt exist */
1022 if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
1023 continue;
1024 if (test->file_offset + test->len <= disk_i_size)
1025 break;
1026 if (test->file_offset >= i_size)
1027 break;
1028 if (entry_end(test) > disk_i_size) {
1029 /*
1030 * we don't update disk_i_size now, so record this
1031 * undealt i_size. Or we will not know the real
1032 * i_size.
1033 */
1034 if (test->outstanding_isize < offset)
1035 test->outstanding_isize = offset;
1036 if (ordered &&
1037 ordered->outstanding_isize >
1038 test->outstanding_isize)
1039 test->outstanding_isize =
1040 ordered->outstanding_isize;
1041 goto out;
1042 }
1043 }
1044 new_i_size = min_t(u64, offset, i_size);
1045
1046 /*
1047 * Some ordered extents may completed before the current one, and
1048 * we hold the real i_size in ->outstanding_isize.
1049 */
1050 if (ordered && ordered->outstanding_isize > new_i_size)
1051 new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
1052 BTRFS_I(inode)->disk_i_size = new_i_size;
1053 ret = 0;
1054out:
1055 /*
1056 * We need to do this because we can't remove ordered extents until
1057 * after the i_disk_size has been updated and then the inode has been
1058 * updated to reflect the change, so we need to tell anybody who finds
1059 * this ordered extent that we've already done all the real work, we
1060 * just haven't completed all the other work.
1061 */
1062 if (ordered)
1063 set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
1064 spin_unlock_irq(&tree->lock);
1065 return ret;
1066}
1067
1068/*
1069 * search the ordered extents for one corresponding to 'offset' and
1070 * try to find a checksum. This is used because we allow pages to
1071 * be reclaimed before their checksum is actually put into the btree
1072 */
1073int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
1074 u32 *sum, int len)
1075{
1076 struct btrfs_ordered_sum *ordered_sum;
1077 struct btrfs_ordered_extent *ordered;
1078 struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1079 unsigned long num_sectors;
1080 unsigned long i;
1081 u32 sectorsize = BTRFS_I(inode)->root->sectorsize;
1082 int index = 0;
1083
1084 ordered = btrfs_lookup_ordered_extent(inode, offset);
1085 if (!ordered)
1086 return 0;
1087
1088 spin_lock_irq(&tree->lock);
1089 list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
1090 if (disk_bytenr >= ordered_sum->bytenr &&
1091 disk_bytenr < ordered_sum->bytenr + ordered_sum->len) {
1092 i = (disk_bytenr - ordered_sum->bytenr) >>
1093 inode->i_sb->s_blocksize_bits;
1094 num_sectors = ordered_sum->len >>
1095 inode->i_sb->s_blocksize_bits;
1096 num_sectors = min_t(int, len - index, num_sectors - i);
1097 memcpy(sum + index, ordered_sum->sums + i,
1098 num_sectors);
1099
1100 index += (int)num_sectors;
1101 if (index == len)
1102 goto out;
1103 disk_bytenr += num_sectors * sectorsize;
1104 }
1105 }
1106out:
1107 spin_unlock_irq(&tree->lock);
1108 btrfs_put_ordered_extent(ordered);
1109 return index;
1110}
1111
1112
1113/*
1114 * add a given inode to the list of inodes that must be fully on
1115 * disk before a transaction commit finishes.
1116 *
1117 * This basically gives us the ext3 style data=ordered mode, and it is mostly
1118 * used to make sure renamed files are fully on disk.
1119 *
1120 * It is a noop if the inode is already fully on disk.
1121 *
1122 * If trans is not null, we'll do a friendly check for a transaction that
1123 * is already flushing things and force the IO down ourselves.
1124 */
1125void btrfs_add_ordered_operation(struct btrfs_trans_handle *trans,
1126 struct btrfs_root *root, struct inode *inode)
1127{
1128 struct btrfs_transaction *cur_trans = trans->transaction;
1129 u64 last_mod;
1130
1131 last_mod = max(BTRFS_I(inode)->generation, BTRFS_I(inode)->last_trans);
1132
1133 /*
1134 * if this file hasn't been changed since the last transaction
1135 * commit, we can safely return without doing anything
1136 */
1137 if (last_mod <= root->fs_info->last_trans_committed)
1138 return;
1139
1140 spin_lock(&root->fs_info->ordered_root_lock);
1141 if (list_empty(&BTRFS_I(inode)->ordered_operations)) {
1142 list_add_tail(&BTRFS_I(inode)->ordered_operations,
1143 &cur_trans->ordered_operations);
1144 }
1145 spin_unlock(&root->fs_info->ordered_root_lock);
1146}
1147
1148int __init ordered_data_init(void)
1149{
1150 btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1151 sizeof(struct btrfs_ordered_extent), 0,
1152 SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
1153 NULL);
1154 if (!btrfs_ordered_extent_cache)
1155 return -ENOMEM;
1156
1157 return 0;
1158}
1159
1160void ordered_data_exit(void)
1161{
1162 if (btrfs_ordered_extent_cache)
1163 kmem_cache_destroy(btrfs_ordered_extent_cache);
1164}