Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2007 Oracle. All rights reserved.
4 */
5
6#include <linux/fs.h>
7#include <linux/pagemap.h>
8#include <linux/time.h>
9#include <linux/init.h>
10#include <linux/string.h>
11#include <linux/backing-dev.h>
12#include <linux/falloc.h>
13#include <linux/writeback.h>
14#include <linux/compat.h>
15#include <linux/slab.h>
16#include <linux/btrfs.h>
17#include <linux/uio.h>
18#include <linux/iversion.h>
19#include "ctree.h"
20#include "disk-io.h"
21#include "transaction.h"
22#include "btrfs_inode.h"
23#include "print-tree.h"
24#include "tree-log.h"
25#include "locking.h"
26#include "volumes.h"
27#include "qgroup.h"
28#include "compression.h"
29#include "delalloc-space.h"
30#include "reflink.h"
31#include "subpage.h"
32
33static struct kmem_cache *btrfs_inode_defrag_cachep;
34/*
35 * when auto defrag is enabled we
36 * queue up these defrag structs to remember which
37 * inodes need defragging passes
38 */
39struct inode_defrag {
40 struct rb_node rb_node;
41 /* objectid */
42 u64 ino;
43 /*
44 * transid where the defrag was added, we search for
45 * extents newer than this
46 */
47 u64 transid;
48
49 /* root objectid */
50 u64 root;
51
52 /* last offset we were able to defrag */
53 u64 last_offset;
54
55 /* if we've wrapped around back to zero once already */
56 int cycled;
57};
58
59static int __compare_inode_defrag(struct inode_defrag *defrag1,
60 struct inode_defrag *defrag2)
61{
62 if (defrag1->root > defrag2->root)
63 return 1;
64 else if (defrag1->root < defrag2->root)
65 return -1;
66 else if (defrag1->ino > defrag2->ino)
67 return 1;
68 else if (defrag1->ino < defrag2->ino)
69 return -1;
70 else
71 return 0;
72}
73
74/* pop a record for an inode into the defrag tree. The lock
75 * must be held already
76 *
77 * If you're inserting a record for an older transid than an
78 * existing record, the transid already in the tree is lowered
79 *
80 * If an existing record is found the defrag item you
81 * pass in is freed
82 */
83static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
84 struct inode_defrag *defrag)
85{
86 struct btrfs_fs_info *fs_info = inode->root->fs_info;
87 struct inode_defrag *entry;
88 struct rb_node **p;
89 struct rb_node *parent = NULL;
90 int ret;
91
92 p = &fs_info->defrag_inodes.rb_node;
93 while (*p) {
94 parent = *p;
95 entry = rb_entry(parent, struct inode_defrag, rb_node);
96
97 ret = __compare_inode_defrag(defrag, entry);
98 if (ret < 0)
99 p = &parent->rb_left;
100 else if (ret > 0)
101 p = &parent->rb_right;
102 else {
103 /* if we're reinserting an entry for
104 * an old defrag run, make sure to
105 * lower the transid of our existing record
106 */
107 if (defrag->transid < entry->transid)
108 entry->transid = defrag->transid;
109 if (defrag->last_offset > entry->last_offset)
110 entry->last_offset = defrag->last_offset;
111 return -EEXIST;
112 }
113 }
114 set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
115 rb_link_node(&defrag->rb_node, parent, p);
116 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
117 return 0;
118}
119
120static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
121{
122 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
123 return 0;
124
125 if (btrfs_fs_closing(fs_info))
126 return 0;
127
128 return 1;
129}
130
131/*
132 * insert a defrag record for this inode if auto defrag is
133 * enabled
134 */
135int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
136 struct btrfs_inode *inode)
137{
138 struct btrfs_root *root = inode->root;
139 struct btrfs_fs_info *fs_info = root->fs_info;
140 struct inode_defrag *defrag;
141 u64 transid;
142 int ret;
143
144 if (!__need_auto_defrag(fs_info))
145 return 0;
146
147 if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
148 return 0;
149
150 if (trans)
151 transid = trans->transid;
152 else
153 transid = inode->root->last_trans;
154
155 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
156 if (!defrag)
157 return -ENOMEM;
158
159 defrag->ino = btrfs_ino(inode);
160 defrag->transid = transid;
161 defrag->root = root->root_key.objectid;
162
163 spin_lock(&fs_info->defrag_inodes_lock);
164 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
165 /*
166 * If we set IN_DEFRAG flag and evict the inode from memory,
167 * and then re-read this inode, this new inode doesn't have
168 * IN_DEFRAG flag. At the case, we may find the existed defrag.
169 */
170 ret = __btrfs_add_inode_defrag(inode, defrag);
171 if (ret)
172 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
173 } else {
174 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
175 }
176 spin_unlock(&fs_info->defrag_inodes_lock);
177 return 0;
178}
179
180/*
181 * Requeue the defrag object. If there is a defrag object that points to
182 * the same inode in the tree, we will merge them together (by
183 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
184 */
185static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
186 struct inode_defrag *defrag)
187{
188 struct btrfs_fs_info *fs_info = inode->root->fs_info;
189 int ret;
190
191 if (!__need_auto_defrag(fs_info))
192 goto out;
193
194 /*
195 * Here we don't check the IN_DEFRAG flag, because we need merge
196 * them together.
197 */
198 spin_lock(&fs_info->defrag_inodes_lock);
199 ret = __btrfs_add_inode_defrag(inode, defrag);
200 spin_unlock(&fs_info->defrag_inodes_lock);
201 if (ret)
202 goto out;
203 return;
204out:
205 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
206}
207
208/*
209 * pick the defragable inode that we want, if it doesn't exist, we will get
210 * the next one.
211 */
212static struct inode_defrag *
213btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
214{
215 struct inode_defrag *entry = NULL;
216 struct inode_defrag tmp;
217 struct rb_node *p;
218 struct rb_node *parent = NULL;
219 int ret;
220
221 tmp.ino = ino;
222 tmp.root = root;
223
224 spin_lock(&fs_info->defrag_inodes_lock);
225 p = fs_info->defrag_inodes.rb_node;
226 while (p) {
227 parent = p;
228 entry = rb_entry(parent, struct inode_defrag, rb_node);
229
230 ret = __compare_inode_defrag(&tmp, entry);
231 if (ret < 0)
232 p = parent->rb_left;
233 else if (ret > 0)
234 p = parent->rb_right;
235 else
236 goto out;
237 }
238
239 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
240 parent = rb_next(parent);
241 if (parent)
242 entry = rb_entry(parent, struct inode_defrag, rb_node);
243 else
244 entry = NULL;
245 }
246out:
247 if (entry)
248 rb_erase(parent, &fs_info->defrag_inodes);
249 spin_unlock(&fs_info->defrag_inodes_lock);
250 return entry;
251}
252
253void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
254{
255 struct inode_defrag *defrag;
256 struct rb_node *node;
257
258 spin_lock(&fs_info->defrag_inodes_lock);
259 node = rb_first(&fs_info->defrag_inodes);
260 while (node) {
261 rb_erase(node, &fs_info->defrag_inodes);
262 defrag = rb_entry(node, struct inode_defrag, rb_node);
263 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
264
265 cond_resched_lock(&fs_info->defrag_inodes_lock);
266
267 node = rb_first(&fs_info->defrag_inodes);
268 }
269 spin_unlock(&fs_info->defrag_inodes_lock);
270}
271
272#define BTRFS_DEFRAG_BATCH 1024
273
274static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
275 struct inode_defrag *defrag)
276{
277 struct btrfs_root *inode_root;
278 struct inode *inode;
279 struct btrfs_ioctl_defrag_range_args range;
280 int num_defrag;
281 int ret;
282
283 /* get the inode */
284 inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
285 if (IS_ERR(inode_root)) {
286 ret = PTR_ERR(inode_root);
287 goto cleanup;
288 }
289
290 inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
291 btrfs_put_root(inode_root);
292 if (IS_ERR(inode)) {
293 ret = PTR_ERR(inode);
294 goto cleanup;
295 }
296
297 /* do a chunk of defrag */
298 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
299 memset(&range, 0, sizeof(range));
300 range.len = (u64)-1;
301 range.start = defrag->last_offset;
302
303 sb_start_write(fs_info->sb);
304 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
305 BTRFS_DEFRAG_BATCH);
306 sb_end_write(fs_info->sb);
307 /*
308 * if we filled the whole defrag batch, there
309 * must be more work to do. Queue this defrag
310 * again
311 */
312 if (num_defrag == BTRFS_DEFRAG_BATCH) {
313 defrag->last_offset = range.start;
314 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
315 } else if (defrag->last_offset && !defrag->cycled) {
316 /*
317 * we didn't fill our defrag batch, but
318 * we didn't start at zero. Make sure we loop
319 * around to the start of the file.
320 */
321 defrag->last_offset = 0;
322 defrag->cycled = 1;
323 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
324 } else {
325 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
326 }
327
328 iput(inode);
329 return 0;
330cleanup:
331 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
332 return ret;
333}
334
335/*
336 * run through the list of inodes in the FS that need
337 * defragging
338 */
339int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
340{
341 struct inode_defrag *defrag;
342 u64 first_ino = 0;
343 u64 root_objectid = 0;
344
345 atomic_inc(&fs_info->defrag_running);
346 while (1) {
347 /* Pause the auto defragger. */
348 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
349 &fs_info->fs_state))
350 break;
351
352 if (!__need_auto_defrag(fs_info))
353 break;
354
355 /* find an inode to defrag */
356 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
357 first_ino);
358 if (!defrag) {
359 if (root_objectid || first_ino) {
360 root_objectid = 0;
361 first_ino = 0;
362 continue;
363 } else {
364 break;
365 }
366 }
367
368 first_ino = defrag->ino + 1;
369 root_objectid = defrag->root;
370
371 __btrfs_run_defrag_inode(fs_info, defrag);
372 }
373 atomic_dec(&fs_info->defrag_running);
374
375 /*
376 * during unmount, we use the transaction_wait queue to
377 * wait for the defragger to stop
378 */
379 wake_up(&fs_info->transaction_wait);
380 return 0;
381}
382
383/* simple helper to fault in pages and copy. This should go away
384 * and be replaced with calls into generic code.
385 */
386static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
387 struct page **prepared_pages,
388 struct iov_iter *i)
389{
390 size_t copied = 0;
391 size_t total_copied = 0;
392 int pg = 0;
393 int offset = offset_in_page(pos);
394
395 while (write_bytes > 0) {
396 size_t count = min_t(size_t,
397 PAGE_SIZE - offset, write_bytes);
398 struct page *page = prepared_pages[pg];
399 /*
400 * Copy data from userspace to the current page
401 */
402 copied = copy_page_from_iter_atomic(page, offset, count, i);
403
404 /* Flush processor's dcache for this page */
405 flush_dcache_page(page);
406
407 /*
408 * if we get a partial write, we can end up with
409 * partially up to date pages. These add
410 * a lot of complexity, so make sure they don't
411 * happen by forcing this copy to be retried.
412 *
413 * The rest of the btrfs_file_write code will fall
414 * back to page at a time copies after we return 0.
415 */
416 if (unlikely(copied < count)) {
417 if (!PageUptodate(page)) {
418 iov_iter_revert(i, copied);
419 copied = 0;
420 }
421 if (!copied)
422 break;
423 }
424
425 write_bytes -= copied;
426 total_copied += copied;
427 offset += copied;
428 if (offset == PAGE_SIZE) {
429 pg++;
430 offset = 0;
431 }
432 }
433 return total_copied;
434}
435
436/*
437 * unlocks pages after btrfs_file_write is done with them
438 */
439static void btrfs_drop_pages(struct page **pages, size_t num_pages)
440{
441 size_t i;
442 for (i = 0; i < num_pages; i++) {
443 /* page checked is some magic around finding pages that
444 * have been modified without going through btrfs_set_page_dirty
445 * clear it here. There should be no need to mark the pages
446 * accessed as prepare_pages should have marked them accessed
447 * in prepare_pages via find_or_create_page()
448 */
449 ClearPageChecked(pages[i]);
450 unlock_page(pages[i]);
451 put_page(pages[i]);
452 }
453}
454
455/*
456 * After btrfs_copy_from_user(), update the following things for delalloc:
457 * - Mark newly dirtied pages as DELALLOC in the io tree.
458 * Used to advise which range is to be written back.
459 * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
460 * - Update inode size for past EOF write
461 */
462int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
463 size_t num_pages, loff_t pos, size_t write_bytes,
464 struct extent_state **cached, bool noreserve)
465{
466 struct btrfs_fs_info *fs_info = inode->root->fs_info;
467 int err = 0;
468 int i;
469 u64 num_bytes;
470 u64 start_pos;
471 u64 end_of_last_block;
472 u64 end_pos = pos + write_bytes;
473 loff_t isize = i_size_read(&inode->vfs_inode);
474 unsigned int extra_bits = 0;
475
476 if (write_bytes == 0)
477 return 0;
478
479 if (noreserve)
480 extra_bits |= EXTENT_NORESERVE;
481
482 start_pos = round_down(pos, fs_info->sectorsize);
483 num_bytes = round_up(write_bytes + pos - start_pos,
484 fs_info->sectorsize);
485 ASSERT(num_bytes <= U32_MAX);
486
487 end_of_last_block = start_pos + num_bytes - 1;
488
489 /*
490 * The pages may have already been dirty, clear out old accounting so
491 * we can set things up properly
492 */
493 clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
494 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
495 0, 0, cached);
496
497 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
498 extra_bits, cached);
499 if (err)
500 return err;
501
502 for (i = 0; i < num_pages; i++) {
503 struct page *p = pages[i];
504
505 btrfs_page_clamp_set_uptodate(fs_info, p, start_pos, num_bytes);
506 ClearPageChecked(p);
507 btrfs_page_clamp_set_dirty(fs_info, p, start_pos, num_bytes);
508 }
509
510 /*
511 * we've only changed i_size in ram, and we haven't updated
512 * the disk i_size. There is no need to log the inode
513 * at this time.
514 */
515 if (end_pos > isize)
516 i_size_write(&inode->vfs_inode, end_pos);
517 return 0;
518}
519
520/*
521 * this drops all the extents in the cache that intersect the range
522 * [start, end]. Existing extents are split as required.
523 */
524void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
525 int skip_pinned)
526{
527 struct extent_map *em;
528 struct extent_map *split = NULL;
529 struct extent_map *split2 = NULL;
530 struct extent_map_tree *em_tree = &inode->extent_tree;
531 u64 len = end - start + 1;
532 u64 gen;
533 int ret;
534 int testend = 1;
535 unsigned long flags;
536 int compressed = 0;
537 bool modified;
538
539 WARN_ON(end < start);
540 if (end == (u64)-1) {
541 len = (u64)-1;
542 testend = 0;
543 }
544 while (1) {
545 int no_splits = 0;
546
547 modified = false;
548 if (!split)
549 split = alloc_extent_map();
550 if (!split2)
551 split2 = alloc_extent_map();
552 if (!split || !split2)
553 no_splits = 1;
554
555 write_lock(&em_tree->lock);
556 em = lookup_extent_mapping(em_tree, start, len);
557 if (!em) {
558 write_unlock(&em_tree->lock);
559 break;
560 }
561 flags = em->flags;
562 gen = em->generation;
563 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
564 if (testend && em->start + em->len >= start + len) {
565 free_extent_map(em);
566 write_unlock(&em_tree->lock);
567 break;
568 }
569 start = em->start + em->len;
570 if (testend)
571 len = start + len - (em->start + em->len);
572 free_extent_map(em);
573 write_unlock(&em_tree->lock);
574 continue;
575 }
576 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
577 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
578 clear_bit(EXTENT_FLAG_LOGGING, &flags);
579 modified = !list_empty(&em->list);
580 if (no_splits)
581 goto next;
582
583 if (em->start < start) {
584 split->start = em->start;
585 split->len = start - em->start;
586
587 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
588 split->orig_start = em->orig_start;
589 split->block_start = em->block_start;
590
591 if (compressed)
592 split->block_len = em->block_len;
593 else
594 split->block_len = split->len;
595 split->orig_block_len = max(split->block_len,
596 em->orig_block_len);
597 split->ram_bytes = em->ram_bytes;
598 } else {
599 split->orig_start = split->start;
600 split->block_len = 0;
601 split->block_start = em->block_start;
602 split->orig_block_len = 0;
603 split->ram_bytes = split->len;
604 }
605
606 split->generation = gen;
607 split->flags = flags;
608 split->compress_type = em->compress_type;
609 replace_extent_mapping(em_tree, em, split, modified);
610 free_extent_map(split);
611 split = split2;
612 split2 = NULL;
613 }
614 if (testend && em->start + em->len > start + len) {
615 u64 diff = start + len - em->start;
616
617 split->start = start + len;
618 split->len = em->start + em->len - (start + len);
619 split->flags = flags;
620 split->compress_type = em->compress_type;
621 split->generation = gen;
622
623 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
624 split->orig_block_len = max(em->block_len,
625 em->orig_block_len);
626
627 split->ram_bytes = em->ram_bytes;
628 if (compressed) {
629 split->block_len = em->block_len;
630 split->block_start = em->block_start;
631 split->orig_start = em->orig_start;
632 } else {
633 split->block_len = split->len;
634 split->block_start = em->block_start
635 + diff;
636 split->orig_start = em->orig_start;
637 }
638 } else {
639 split->ram_bytes = split->len;
640 split->orig_start = split->start;
641 split->block_len = 0;
642 split->block_start = em->block_start;
643 split->orig_block_len = 0;
644 }
645
646 if (extent_map_in_tree(em)) {
647 replace_extent_mapping(em_tree, em, split,
648 modified);
649 } else {
650 ret = add_extent_mapping(em_tree, split,
651 modified);
652 ASSERT(ret == 0); /* Logic error */
653 }
654 free_extent_map(split);
655 split = NULL;
656 }
657next:
658 if (extent_map_in_tree(em))
659 remove_extent_mapping(em_tree, em);
660 write_unlock(&em_tree->lock);
661
662 /* once for us */
663 free_extent_map(em);
664 /* once for the tree*/
665 free_extent_map(em);
666 }
667 if (split)
668 free_extent_map(split);
669 if (split2)
670 free_extent_map(split2);
671}
672
673/*
674 * this is very complex, but the basic idea is to drop all extents
675 * in the range start - end. hint_block is filled in with a block number
676 * that would be a good hint to the block allocator for this file.
677 *
678 * If an extent intersects the range but is not entirely inside the range
679 * it is either truncated or split. Anything entirely inside the range
680 * is deleted from the tree.
681 *
682 * Note: the VFS' inode number of bytes is not updated, it's up to the caller
683 * to deal with that. We set the field 'bytes_found' of the arguments structure
684 * with the number of allocated bytes found in the target range, so that the
685 * caller can update the inode's number of bytes in an atomic way when
686 * replacing extents in a range to avoid races with stat(2).
687 */
688int btrfs_drop_extents(struct btrfs_trans_handle *trans,
689 struct btrfs_root *root, struct btrfs_inode *inode,
690 struct btrfs_drop_extents_args *args)
691{
692 struct btrfs_fs_info *fs_info = root->fs_info;
693 struct extent_buffer *leaf;
694 struct btrfs_file_extent_item *fi;
695 struct btrfs_ref ref = { 0 };
696 struct btrfs_key key;
697 struct btrfs_key new_key;
698 u64 ino = btrfs_ino(inode);
699 u64 search_start = args->start;
700 u64 disk_bytenr = 0;
701 u64 num_bytes = 0;
702 u64 extent_offset = 0;
703 u64 extent_end = 0;
704 u64 last_end = args->start;
705 int del_nr = 0;
706 int del_slot = 0;
707 int extent_type;
708 int recow;
709 int ret;
710 int modify_tree = -1;
711 int update_refs;
712 int found = 0;
713 int leafs_visited = 0;
714 struct btrfs_path *path = args->path;
715
716 args->bytes_found = 0;
717 args->extent_inserted = false;
718
719 /* Must always have a path if ->replace_extent is true */
720 ASSERT(!(args->replace_extent && !args->path));
721
722 if (!path) {
723 path = btrfs_alloc_path();
724 if (!path) {
725 ret = -ENOMEM;
726 goto out;
727 }
728 }
729
730 if (args->drop_cache)
731 btrfs_drop_extent_cache(inode, args->start, args->end - 1, 0);
732
733 if (args->start >= inode->disk_i_size && !args->replace_extent)
734 modify_tree = 0;
735
736 update_refs = (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID);
737 while (1) {
738 recow = 0;
739 ret = btrfs_lookup_file_extent(trans, root, path, ino,
740 search_start, modify_tree);
741 if (ret < 0)
742 break;
743 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
744 leaf = path->nodes[0];
745 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
746 if (key.objectid == ino &&
747 key.type == BTRFS_EXTENT_DATA_KEY)
748 path->slots[0]--;
749 }
750 ret = 0;
751 leafs_visited++;
752next_slot:
753 leaf = path->nodes[0];
754 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
755 BUG_ON(del_nr > 0);
756 ret = btrfs_next_leaf(root, path);
757 if (ret < 0)
758 break;
759 if (ret > 0) {
760 ret = 0;
761 break;
762 }
763 leafs_visited++;
764 leaf = path->nodes[0];
765 recow = 1;
766 }
767
768 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
769
770 if (key.objectid > ino)
771 break;
772 if (WARN_ON_ONCE(key.objectid < ino) ||
773 key.type < BTRFS_EXTENT_DATA_KEY) {
774 ASSERT(del_nr == 0);
775 path->slots[0]++;
776 goto next_slot;
777 }
778 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
779 break;
780
781 fi = btrfs_item_ptr(leaf, path->slots[0],
782 struct btrfs_file_extent_item);
783 extent_type = btrfs_file_extent_type(leaf, fi);
784
785 if (extent_type == BTRFS_FILE_EXTENT_REG ||
786 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
787 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
788 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
789 extent_offset = btrfs_file_extent_offset(leaf, fi);
790 extent_end = key.offset +
791 btrfs_file_extent_num_bytes(leaf, fi);
792 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
793 extent_end = key.offset +
794 btrfs_file_extent_ram_bytes(leaf, fi);
795 } else {
796 /* can't happen */
797 BUG();
798 }
799
800 /*
801 * Don't skip extent items representing 0 byte lengths. They
802 * used to be created (bug) if while punching holes we hit
803 * -ENOSPC condition. So if we find one here, just ensure we
804 * delete it, otherwise we would insert a new file extent item
805 * with the same key (offset) as that 0 bytes length file
806 * extent item in the call to setup_items_for_insert() later
807 * in this function.
808 */
809 if (extent_end == key.offset && extent_end >= search_start) {
810 last_end = extent_end;
811 goto delete_extent_item;
812 }
813
814 if (extent_end <= search_start) {
815 path->slots[0]++;
816 goto next_slot;
817 }
818
819 found = 1;
820 search_start = max(key.offset, args->start);
821 if (recow || !modify_tree) {
822 modify_tree = -1;
823 btrfs_release_path(path);
824 continue;
825 }
826
827 /*
828 * | - range to drop - |
829 * | -------- extent -------- |
830 */
831 if (args->start > key.offset && args->end < extent_end) {
832 BUG_ON(del_nr > 0);
833 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
834 ret = -EOPNOTSUPP;
835 break;
836 }
837
838 memcpy(&new_key, &key, sizeof(new_key));
839 new_key.offset = args->start;
840 ret = btrfs_duplicate_item(trans, root, path,
841 &new_key);
842 if (ret == -EAGAIN) {
843 btrfs_release_path(path);
844 continue;
845 }
846 if (ret < 0)
847 break;
848
849 leaf = path->nodes[0];
850 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
851 struct btrfs_file_extent_item);
852 btrfs_set_file_extent_num_bytes(leaf, fi,
853 args->start - key.offset);
854
855 fi = btrfs_item_ptr(leaf, path->slots[0],
856 struct btrfs_file_extent_item);
857
858 extent_offset += args->start - key.offset;
859 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
860 btrfs_set_file_extent_num_bytes(leaf, fi,
861 extent_end - args->start);
862 btrfs_mark_buffer_dirty(leaf);
863
864 if (update_refs && disk_bytenr > 0) {
865 btrfs_init_generic_ref(&ref,
866 BTRFS_ADD_DELAYED_REF,
867 disk_bytenr, num_bytes, 0);
868 btrfs_init_data_ref(&ref,
869 root->root_key.objectid,
870 new_key.objectid,
871 args->start - extent_offset);
872 ret = btrfs_inc_extent_ref(trans, &ref);
873 BUG_ON(ret); /* -ENOMEM */
874 }
875 key.offset = args->start;
876 }
877 /*
878 * From here on out we will have actually dropped something, so
879 * last_end can be updated.
880 */
881 last_end = extent_end;
882
883 /*
884 * | ---- range to drop ----- |
885 * | -------- extent -------- |
886 */
887 if (args->start <= key.offset && args->end < extent_end) {
888 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
889 ret = -EOPNOTSUPP;
890 break;
891 }
892
893 memcpy(&new_key, &key, sizeof(new_key));
894 new_key.offset = args->end;
895 btrfs_set_item_key_safe(fs_info, path, &new_key);
896
897 extent_offset += args->end - key.offset;
898 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
899 btrfs_set_file_extent_num_bytes(leaf, fi,
900 extent_end - args->end);
901 btrfs_mark_buffer_dirty(leaf);
902 if (update_refs && disk_bytenr > 0)
903 args->bytes_found += args->end - key.offset;
904 break;
905 }
906
907 search_start = extent_end;
908 /*
909 * | ---- range to drop ----- |
910 * | -------- extent -------- |
911 */
912 if (args->start > key.offset && args->end >= extent_end) {
913 BUG_ON(del_nr > 0);
914 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
915 ret = -EOPNOTSUPP;
916 break;
917 }
918
919 btrfs_set_file_extent_num_bytes(leaf, fi,
920 args->start - key.offset);
921 btrfs_mark_buffer_dirty(leaf);
922 if (update_refs && disk_bytenr > 0)
923 args->bytes_found += extent_end - args->start;
924 if (args->end == extent_end)
925 break;
926
927 path->slots[0]++;
928 goto next_slot;
929 }
930
931 /*
932 * | ---- range to drop ----- |
933 * | ------ extent ------ |
934 */
935 if (args->start <= key.offset && args->end >= extent_end) {
936delete_extent_item:
937 if (del_nr == 0) {
938 del_slot = path->slots[0];
939 del_nr = 1;
940 } else {
941 BUG_ON(del_slot + del_nr != path->slots[0]);
942 del_nr++;
943 }
944
945 if (update_refs &&
946 extent_type == BTRFS_FILE_EXTENT_INLINE) {
947 args->bytes_found += extent_end - key.offset;
948 extent_end = ALIGN(extent_end,
949 fs_info->sectorsize);
950 } else if (update_refs && disk_bytenr > 0) {
951 btrfs_init_generic_ref(&ref,
952 BTRFS_DROP_DELAYED_REF,
953 disk_bytenr, num_bytes, 0);
954 btrfs_init_data_ref(&ref,
955 root->root_key.objectid,
956 key.objectid,
957 key.offset - extent_offset);
958 ret = btrfs_free_extent(trans, &ref);
959 BUG_ON(ret); /* -ENOMEM */
960 args->bytes_found += extent_end - key.offset;
961 }
962
963 if (args->end == extent_end)
964 break;
965
966 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
967 path->slots[0]++;
968 goto next_slot;
969 }
970
971 ret = btrfs_del_items(trans, root, path, del_slot,
972 del_nr);
973 if (ret) {
974 btrfs_abort_transaction(trans, ret);
975 break;
976 }
977
978 del_nr = 0;
979 del_slot = 0;
980
981 btrfs_release_path(path);
982 continue;
983 }
984
985 BUG();
986 }
987
988 if (!ret && del_nr > 0) {
989 /*
990 * Set path->slots[0] to first slot, so that after the delete
991 * if items are move off from our leaf to its immediate left or
992 * right neighbor leafs, we end up with a correct and adjusted
993 * path->slots[0] for our insertion (if args->replace_extent).
994 */
995 path->slots[0] = del_slot;
996 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
997 if (ret)
998 btrfs_abort_transaction(trans, ret);
999 }
1000
1001 leaf = path->nodes[0];
1002 /*
1003 * If btrfs_del_items() was called, it might have deleted a leaf, in
1004 * which case it unlocked our path, so check path->locks[0] matches a
1005 * write lock.
1006 */
1007 if (!ret && args->replace_extent && leafs_visited == 1 &&
1008 path->locks[0] == BTRFS_WRITE_LOCK &&
1009 btrfs_leaf_free_space(leaf) >=
1010 sizeof(struct btrfs_item) + args->extent_item_size) {
1011
1012 key.objectid = ino;
1013 key.type = BTRFS_EXTENT_DATA_KEY;
1014 key.offset = args->start;
1015 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1016 struct btrfs_key slot_key;
1017
1018 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1019 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1020 path->slots[0]++;
1021 }
1022 setup_items_for_insert(root, path, &key,
1023 &args->extent_item_size, 1);
1024 args->extent_inserted = true;
1025 }
1026
1027 if (!args->path)
1028 btrfs_free_path(path);
1029 else if (!args->extent_inserted)
1030 btrfs_release_path(path);
1031out:
1032 args->drop_end = found ? min(args->end, last_end) : args->end;
1033
1034 return ret;
1035}
1036
1037static int extent_mergeable(struct extent_buffer *leaf, int slot,
1038 u64 objectid, u64 bytenr, u64 orig_offset,
1039 u64 *start, u64 *end)
1040{
1041 struct btrfs_file_extent_item *fi;
1042 struct btrfs_key key;
1043 u64 extent_end;
1044
1045 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1046 return 0;
1047
1048 btrfs_item_key_to_cpu(leaf, &key, slot);
1049 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1050 return 0;
1051
1052 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1053 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1054 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1055 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1056 btrfs_file_extent_compression(leaf, fi) ||
1057 btrfs_file_extent_encryption(leaf, fi) ||
1058 btrfs_file_extent_other_encoding(leaf, fi))
1059 return 0;
1060
1061 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1062 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1063 return 0;
1064
1065 *start = key.offset;
1066 *end = extent_end;
1067 return 1;
1068}
1069
1070/*
1071 * Mark extent in the range start - end as written.
1072 *
1073 * This changes extent type from 'pre-allocated' to 'regular'. If only
1074 * part of extent is marked as written, the extent will be split into
1075 * two or three.
1076 */
1077int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1078 struct btrfs_inode *inode, u64 start, u64 end)
1079{
1080 struct btrfs_fs_info *fs_info = trans->fs_info;
1081 struct btrfs_root *root = inode->root;
1082 struct extent_buffer *leaf;
1083 struct btrfs_path *path;
1084 struct btrfs_file_extent_item *fi;
1085 struct btrfs_ref ref = { 0 };
1086 struct btrfs_key key;
1087 struct btrfs_key new_key;
1088 u64 bytenr;
1089 u64 num_bytes;
1090 u64 extent_end;
1091 u64 orig_offset;
1092 u64 other_start;
1093 u64 other_end;
1094 u64 split;
1095 int del_nr = 0;
1096 int del_slot = 0;
1097 int recow;
1098 int ret = 0;
1099 u64 ino = btrfs_ino(inode);
1100
1101 path = btrfs_alloc_path();
1102 if (!path)
1103 return -ENOMEM;
1104again:
1105 recow = 0;
1106 split = start;
1107 key.objectid = ino;
1108 key.type = BTRFS_EXTENT_DATA_KEY;
1109 key.offset = split;
1110
1111 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1112 if (ret < 0)
1113 goto out;
1114 if (ret > 0 && path->slots[0] > 0)
1115 path->slots[0]--;
1116
1117 leaf = path->nodes[0];
1118 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1119 if (key.objectid != ino ||
1120 key.type != BTRFS_EXTENT_DATA_KEY) {
1121 ret = -EINVAL;
1122 btrfs_abort_transaction(trans, ret);
1123 goto out;
1124 }
1125 fi = btrfs_item_ptr(leaf, path->slots[0],
1126 struct btrfs_file_extent_item);
1127 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1128 ret = -EINVAL;
1129 btrfs_abort_transaction(trans, ret);
1130 goto out;
1131 }
1132 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1133 if (key.offset > start || extent_end < end) {
1134 ret = -EINVAL;
1135 btrfs_abort_transaction(trans, ret);
1136 goto out;
1137 }
1138
1139 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1140 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1141 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1142 memcpy(&new_key, &key, sizeof(new_key));
1143
1144 if (start == key.offset && end < extent_end) {
1145 other_start = 0;
1146 other_end = start;
1147 if (extent_mergeable(leaf, path->slots[0] - 1,
1148 ino, bytenr, orig_offset,
1149 &other_start, &other_end)) {
1150 new_key.offset = end;
1151 btrfs_set_item_key_safe(fs_info, path, &new_key);
1152 fi = btrfs_item_ptr(leaf, path->slots[0],
1153 struct btrfs_file_extent_item);
1154 btrfs_set_file_extent_generation(leaf, fi,
1155 trans->transid);
1156 btrfs_set_file_extent_num_bytes(leaf, fi,
1157 extent_end - end);
1158 btrfs_set_file_extent_offset(leaf, fi,
1159 end - orig_offset);
1160 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1161 struct btrfs_file_extent_item);
1162 btrfs_set_file_extent_generation(leaf, fi,
1163 trans->transid);
1164 btrfs_set_file_extent_num_bytes(leaf, fi,
1165 end - other_start);
1166 btrfs_mark_buffer_dirty(leaf);
1167 goto out;
1168 }
1169 }
1170
1171 if (start > key.offset && end == extent_end) {
1172 other_start = end;
1173 other_end = 0;
1174 if (extent_mergeable(leaf, path->slots[0] + 1,
1175 ino, bytenr, orig_offset,
1176 &other_start, &other_end)) {
1177 fi = btrfs_item_ptr(leaf, path->slots[0],
1178 struct btrfs_file_extent_item);
1179 btrfs_set_file_extent_num_bytes(leaf, fi,
1180 start - key.offset);
1181 btrfs_set_file_extent_generation(leaf, fi,
1182 trans->transid);
1183 path->slots[0]++;
1184 new_key.offset = start;
1185 btrfs_set_item_key_safe(fs_info, path, &new_key);
1186
1187 fi = btrfs_item_ptr(leaf, path->slots[0],
1188 struct btrfs_file_extent_item);
1189 btrfs_set_file_extent_generation(leaf, fi,
1190 trans->transid);
1191 btrfs_set_file_extent_num_bytes(leaf, fi,
1192 other_end - start);
1193 btrfs_set_file_extent_offset(leaf, fi,
1194 start - orig_offset);
1195 btrfs_mark_buffer_dirty(leaf);
1196 goto out;
1197 }
1198 }
1199
1200 while (start > key.offset || end < extent_end) {
1201 if (key.offset == start)
1202 split = end;
1203
1204 new_key.offset = split;
1205 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1206 if (ret == -EAGAIN) {
1207 btrfs_release_path(path);
1208 goto again;
1209 }
1210 if (ret < 0) {
1211 btrfs_abort_transaction(trans, ret);
1212 goto out;
1213 }
1214
1215 leaf = path->nodes[0];
1216 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1217 struct btrfs_file_extent_item);
1218 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1219 btrfs_set_file_extent_num_bytes(leaf, fi,
1220 split - key.offset);
1221
1222 fi = btrfs_item_ptr(leaf, path->slots[0],
1223 struct btrfs_file_extent_item);
1224
1225 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1226 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1227 btrfs_set_file_extent_num_bytes(leaf, fi,
1228 extent_end - split);
1229 btrfs_mark_buffer_dirty(leaf);
1230
1231 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1232 num_bytes, 0);
1233 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1234 orig_offset);
1235 ret = btrfs_inc_extent_ref(trans, &ref);
1236 if (ret) {
1237 btrfs_abort_transaction(trans, ret);
1238 goto out;
1239 }
1240
1241 if (split == start) {
1242 key.offset = start;
1243 } else {
1244 if (start != key.offset) {
1245 ret = -EINVAL;
1246 btrfs_abort_transaction(trans, ret);
1247 goto out;
1248 }
1249 path->slots[0]--;
1250 extent_end = end;
1251 }
1252 recow = 1;
1253 }
1254
1255 other_start = end;
1256 other_end = 0;
1257 btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1258 num_bytes, 0);
1259 btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1260 if (extent_mergeable(leaf, path->slots[0] + 1,
1261 ino, bytenr, orig_offset,
1262 &other_start, &other_end)) {
1263 if (recow) {
1264 btrfs_release_path(path);
1265 goto again;
1266 }
1267 extent_end = other_end;
1268 del_slot = path->slots[0] + 1;
1269 del_nr++;
1270 ret = btrfs_free_extent(trans, &ref);
1271 if (ret) {
1272 btrfs_abort_transaction(trans, ret);
1273 goto out;
1274 }
1275 }
1276 other_start = 0;
1277 other_end = start;
1278 if (extent_mergeable(leaf, path->slots[0] - 1,
1279 ino, bytenr, orig_offset,
1280 &other_start, &other_end)) {
1281 if (recow) {
1282 btrfs_release_path(path);
1283 goto again;
1284 }
1285 key.offset = other_start;
1286 del_slot = path->slots[0];
1287 del_nr++;
1288 ret = btrfs_free_extent(trans, &ref);
1289 if (ret) {
1290 btrfs_abort_transaction(trans, ret);
1291 goto out;
1292 }
1293 }
1294 if (del_nr == 0) {
1295 fi = btrfs_item_ptr(leaf, path->slots[0],
1296 struct btrfs_file_extent_item);
1297 btrfs_set_file_extent_type(leaf, fi,
1298 BTRFS_FILE_EXTENT_REG);
1299 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1300 btrfs_mark_buffer_dirty(leaf);
1301 } else {
1302 fi = btrfs_item_ptr(leaf, del_slot - 1,
1303 struct btrfs_file_extent_item);
1304 btrfs_set_file_extent_type(leaf, fi,
1305 BTRFS_FILE_EXTENT_REG);
1306 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1307 btrfs_set_file_extent_num_bytes(leaf, fi,
1308 extent_end - key.offset);
1309 btrfs_mark_buffer_dirty(leaf);
1310
1311 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1312 if (ret < 0) {
1313 btrfs_abort_transaction(trans, ret);
1314 goto out;
1315 }
1316 }
1317out:
1318 btrfs_free_path(path);
1319 return ret;
1320}
1321
1322/*
1323 * on error we return an unlocked page and the error value
1324 * on success we return a locked page and 0
1325 */
1326static int prepare_uptodate_page(struct inode *inode,
1327 struct page *page, u64 pos,
1328 bool force_uptodate)
1329{
1330 int ret = 0;
1331
1332 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1333 !PageUptodate(page)) {
1334 ret = btrfs_readpage(NULL, page);
1335 if (ret)
1336 return ret;
1337 lock_page(page);
1338 if (!PageUptodate(page)) {
1339 unlock_page(page);
1340 return -EIO;
1341 }
1342 if (page->mapping != inode->i_mapping) {
1343 unlock_page(page);
1344 return -EAGAIN;
1345 }
1346 }
1347 return 0;
1348}
1349
1350/*
1351 * this just gets pages into the page cache and locks them down.
1352 */
1353static noinline int prepare_pages(struct inode *inode, struct page **pages,
1354 size_t num_pages, loff_t pos,
1355 size_t write_bytes, bool force_uptodate)
1356{
1357 int i;
1358 unsigned long index = pos >> PAGE_SHIFT;
1359 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1360 int err = 0;
1361 int faili;
1362
1363 for (i = 0; i < num_pages; i++) {
1364again:
1365 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1366 mask | __GFP_WRITE);
1367 if (!pages[i]) {
1368 faili = i - 1;
1369 err = -ENOMEM;
1370 goto fail;
1371 }
1372
1373 err = set_page_extent_mapped(pages[i]);
1374 if (err < 0) {
1375 faili = i;
1376 goto fail;
1377 }
1378
1379 if (i == 0)
1380 err = prepare_uptodate_page(inode, pages[i], pos,
1381 force_uptodate);
1382 if (!err && i == num_pages - 1)
1383 err = prepare_uptodate_page(inode, pages[i],
1384 pos + write_bytes, false);
1385 if (err) {
1386 put_page(pages[i]);
1387 if (err == -EAGAIN) {
1388 err = 0;
1389 goto again;
1390 }
1391 faili = i - 1;
1392 goto fail;
1393 }
1394 wait_on_page_writeback(pages[i]);
1395 }
1396
1397 return 0;
1398fail:
1399 while (faili >= 0) {
1400 unlock_page(pages[faili]);
1401 put_page(pages[faili]);
1402 faili--;
1403 }
1404 return err;
1405
1406}
1407
1408/*
1409 * This function locks the extent and properly waits for data=ordered extents
1410 * to finish before allowing the pages to be modified if need.
1411 *
1412 * The return value:
1413 * 1 - the extent is locked
1414 * 0 - the extent is not locked, and everything is OK
1415 * -EAGAIN - need re-prepare the pages
1416 * the other < 0 number - Something wrong happens
1417 */
1418static noinline int
1419lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1420 size_t num_pages, loff_t pos,
1421 size_t write_bytes,
1422 u64 *lockstart, u64 *lockend,
1423 struct extent_state **cached_state)
1424{
1425 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1426 u64 start_pos;
1427 u64 last_pos;
1428 int i;
1429 int ret = 0;
1430
1431 start_pos = round_down(pos, fs_info->sectorsize);
1432 last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1433
1434 if (start_pos < inode->vfs_inode.i_size) {
1435 struct btrfs_ordered_extent *ordered;
1436
1437 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1438 cached_state);
1439 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1440 last_pos - start_pos + 1);
1441 if (ordered &&
1442 ordered->file_offset + ordered->num_bytes > start_pos &&
1443 ordered->file_offset <= last_pos) {
1444 unlock_extent_cached(&inode->io_tree, start_pos,
1445 last_pos, cached_state);
1446 for (i = 0; i < num_pages; i++) {
1447 unlock_page(pages[i]);
1448 put_page(pages[i]);
1449 }
1450 btrfs_start_ordered_extent(ordered, 1);
1451 btrfs_put_ordered_extent(ordered);
1452 return -EAGAIN;
1453 }
1454 if (ordered)
1455 btrfs_put_ordered_extent(ordered);
1456
1457 *lockstart = start_pos;
1458 *lockend = last_pos;
1459 ret = 1;
1460 }
1461
1462 /*
1463 * We should be called after prepare_pages() which should have locked
1464 * all pages in the range.
1465 */
1466 for (i = 0; i < num_pages; i++)
1467 WARN_ON(!PageLocked(pages[i]));
1468
1469 return ret;
1470}
1471
1472static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1473 size_t *write_bytes, bool nowait)
1474{
1475 struct btrfs_fs_info *fs_info = inode->root->fs_info;
1476 struct btrfs_root *root = inode->root;
1477 u64 lockstart, lockend;
1478 u64 num_bytes;
1479 int ret;
1480
1481 if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1482 return 0;
1483
1484 if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1485 return -EAGAIN;
1486
1487 lockstart = round_down(pos, fs_info->sectorsize);
1488 lockend = round_up(pos + *write_bytes,
1489 fs_info->sectorsize) - 1;
1490 num_bytes = lockend - lockstart + 1;
1491
1492 if (nowait) {
1493 struct btrfs_ordered_extent *ordered;
1494
1495 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1496 return -EAGAIN;
1497
1498 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1499 num_bytes);
1500 if (ordered) {
1501 btrfs_put_ordered_extent(ordered);
1502 ret = -EAGAIN;
1503 goto out_unlock;
1504 }
1505 } else {
1506 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1507 lockend, NULL);
1508 }
1509
1510 ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1511 NULL, NULL, NULL, false);
1512 if (ret <= 0) {
1513 ret = 0;
1514 if (!nowait)
1515 btrfs_drew_write_unlock(&root->snapshot_lock);
1516 } else {
1517 *write_bytes = min_t(size_t, *write_bytes ,
1518 num_bytes - pos + lockstart);
1519 }
1520out_unlock:
1521 unlock_extent(&inode->io_tree, lockstart, lockend);
1522
1523 return ret;
1524}
1525
1526static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1527 size_t *write_bytes)
1528{
1529 return check_can_nocow(inode, pos, write_bytes, true);
1530}
1531
1532/*
1533 * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1534 *
1535 * @pos: File offset
1536 * @write_bytes: The length to write, will be updated to the nocow writeable
1537 * range
1538 *
1539 * This function will flush ordered extents in the range to ensure proper
1540 * nocow checks.
1541 *
1542 * Return:
1543 * >0 and update @write_bytes if we can do nocow write
1544 * 0 if we can't do nocow write
1545 * -EAGAIN if we can't get the needed lock or there are ordered extents
1546 * for * (nowait == true) case
1547 * <0 if other error happened
1548 *
1549 * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1550 */
1551int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1552 size_t *write_bytes)
1553{
1554 return check_can_nocow(inode, pos, write_bytes, false);
1555}
1556
1557void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1558{
1559 btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1560}
1561
1562static void update_time_for_write(struct inode *inode)
1563{
1564 struct timespec64 now;
1565
1566 if (IS_NOCMTIME(inode))
1567 return;
1568
1569 now = current_time(inode);
1570 if (!timespec64_equal(&inode->i_mtime, &now))
1571 inode->i_mtime = now;
1572
1573 if (!timespec64_equal(&inode->i_ctime, &now))
1574 inode->i_ctime = now;
1575
1576 if (IS_I_VERSION(inode))
1577 inode_inc_iversion(inode);
1578}
1579
1580static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1581 size_t count)
1582{
1583 struct file *file = iocb->ki_filp;
1584 struct inode *inode = file_inode(file);
1585 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1586 loff_t pos = iocb->ki_pos;
1587 int ret;
1588 loff_t oldsize;
1589 loff_t start_pos;
1590
1591 if (iocb->ki_flags & IOCB_NOWAIT) {
1592 size_t nocow_bytes = count;
1593
1594 /* We will allocate space in case nodatacow is not set, so bail */
1595 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes) <= 0)
1596 return -EAGAIN;
1597 /*
1598 * There are holes in the range or parts of the range that must
1599 * be COWed (shared extents, RO block groups, etc), so just bail
1600 * out.
1601 */
1602 if (nocow_bytes < count)
1603 return -EAGAIN;
1604 }
1605
1606 current->backing_dev_info = inode_to_bdi(inode);
1607 ret = file_remove_privs(file);
1608 if (ret)
1609 return ret;
1610
1611 /*
1612 * We reserve space for updating the inode when we reserve space for the
1613 * extent we are going to write, so we will enospc out there. We don't
1614 * need to start yet another transaction to update the inode as we will
1615 * update the inode when we finish writing whatever data we write.
1616 */
1617 update_time_for_write(inode);
1618
1619 start_pos = round_down(pos, fs_info->sectorsize);
1620 oldsize = i_size_read(inode);
1621 if (start_pos > oldsize) {
1622 /* Expand hole size to cover write data, preventing empty gap */
1623 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1624
1625 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1626 if (ret) {
1627 current->backing_dev_info = NULL;
1628 return ret;
1629 }
1630 }
1631
1632 return 0;
1633}
1634
1635static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1636 struct iov_iter *i)
1637{
1638 struct file *file = iocb->ki_filp;
1639 loff_t pos;
1640 struct inode *inode = file_inode(file);
1641 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1642 struct page **pages = NULL;
1643 struct extent_changeset *data_reserved = NULL;
1644 u64 release_bytes = 0;
1645 u64 lockstart;
1646 u64 lockend;
1647 size_t num_written = 0;
1648 int nrptrs;
1649 ssize_t ret;
1650 bool only_release_metadata = false;
1651 bool force_page_uptodate = false;
1652 loff_t old_isize = i_size_read(inode);
1653 unsigned int ilock_flags = 0;
1654
1655 if (iocb->ki_flags & IOCB_NOWAIT)
1656 ilock_flags |= BTRFS_ILOCK_TRY;
1657
1658 ret = btrfs_inode_lock(inode, ilock_flags);
1659 if (ret < 0)
1660 return ret;
1661
1662 ret = generic_write_checks(iocb, i);
1663 if (ret <= 0)
1664 goto out;
1665
1666 ret = btrfs_write_check(iocb, i, ret);
1667 if (ret < 0)
1668 goto out;
1669
1670 pos = iocb->ki_pos;
1671 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1672 PAGE_SIZE / (sizeof(struct page *)));
1673 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1674 nrptrs = max(nrptrs, 8);
1675 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1676 if (!pages) {
1677 ret = -ENOMEM;
1678 goto out;
1679 }
1680
1681 while (iov_iter_count(i) > 0) {
1682 struct extent_state *cached_state = NULL;
1683 size_t offset = offset_in_page(pos);
1684 size_t sector_offset;
1685 size_t write_bytes = min(iov_iter_count(i),
1686 nrptrs * (size_t)PAGE_SIZE -
1687 offset);
1688 size_t num_pages;
1689 size_t reserve_bytes;
1690 size_t dirty_pages;
1691 size_t copied;
1692 size_t dirty_sectors;
1693 size_t num_sectors;
1694 int extents_locked;
1695
1696 /*
1697 * Fault pages before locking them in prepare_pages
1698 * to avoid recursive lock
1699 */
1700 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1701 ret = -EFAULT;
1702 break;
1703 }
1704
1705 only_release_metadata = false;
1706 sector_offset = pos & (fs_info->sectorsize - 1);
1707
1708 extent_changeset_release(data_reserved);
1709 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1710 &data_reserved, pos,
1711 write_bytes);
1712 if (ret < 0) {
1713 /*
1714 * If we don't have to COW at the offset, reserve
1715 * metadata only. write_bytes may get smaller than
1716 * requested here.
1717 */
1718 if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1719 &write_bytes) > 0)
1720 only_release_metadata = true;
1721 else
1722 break;
1723 }
1724
1725 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1726 WARN_ON(num_pages > nrptrs);
1727 reserve_bytes = round_up(write_bytes + sector_offset,
1728 fs_info->sectorsize);
1729 WARN_ON(reserve_bytes == 0);
1730 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1731 reserve_bytes);
1732 if (ret) {
1733 if (!only_release_metadata)
1734 btrfs_free_reserved_data_space(BTRFS_I(inode),
1735 data_reserved, pos,
1736 write_bytes);
1737 else
1738 btrfs_check_nocow_unlock(BTRFS_I(inode));
1739 break;
1740 }
1741
1742 release_bytes = reserve_bytes;
1743again:
1744 /*
1745 * This is going to setup the pages array with the number of
1746 * pages we want, so we don't really need to worry about the
1747 * contents of pages from loop to loop
1748 */
1749 ret = prepare_pages(inode, pages, num_pages,
1750 pos, write_bytes,
1751 force_page_uptodate);
1752 if (ret) {
1753 btrfs_delalloc_release_extents(BTRFS_I(inode),
1754 reserve_bytes);
1755 break;
1756 }
1757
1758 extents_locked = lock_and_cleanup_extent_if_need(
1759 BTRFS_I(inode), pages,
1760 num_pages, pos, write_bytes, &lockstart,
1761 &lockend, &cached_state);
1762 if (extents_locked < 0) {
1763 if (extents_locked == -EAGAIN)
1764 goto again;
1765 btrfs_delalloc_release_extents(BTRFS_I(inode),
1766 reserve_bytes);
1767 ret = extents_locked;
1768 break;
1769 }
1770
1771 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1772
1773 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1774 dirty_sectors = round_up(copied + sector_offset,
1775 fs_info->sectorsize);
1776 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1777
1778 /*
1779 * if we have trouble faulting in the pages, fall
1780 * back to one page at a time
1781 */
1782 if (copied < write_bytes)
1783 nrptrs = 1;
1784
1785 if (copied == 0) {
1786 force_page_uptodate = true;
1787 dirty_sectors = 0;
1788 dirty_pages = 0;
1789 } else {
1790 force_page_uptodate = false;
1791 dirty_pages = DIV_ROUND_UP(copied + offset,
1792 PAGE_SIZE);
1793 }
1794
1795 if (num_sectors > dirty_sectors) {
1796 /* release everything except the sectors we dirtied */
1797 release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1798 if (only_release_metadata) {
1799 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1800 release_bytes, true);
1801 } else {
1802 u64 __pos;
1803
1804 __pos = round_down(pos,
1805 fs_info->sectorsize) +
1806 (dirty_pages << PAGE_SHIFT);
1807 btrfs_delalloc_release_space(BTRFS_I(inode),
1808 data_reserved, __pos,
1809 release_bytes, true);
1810 }
1811 }
1812
1813 release_bytes = round_up(copied + sector_offset,
1814 fs_info->sectorsize);
1815
1816 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1817 dirty_pages, pos, copied,
1818 &cached_state, only_release_metadata);
1819
1820 /*
1821 * If we have not locked the extent range, because the range's
1822 * start offset is >= i_size, we might still have a non-NULL
1823 * cached extent state, acquired while marking the extent range
1824 * as delalloc through btrfs_dirty_pages(). Therefore free any
1825 * possible cached extent state to avoid a memory leak.
1826 */
1827 if (extents_locked)
1828 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1829 lockstart, lockend, &cached_state);
1830 else
1831 free_extent_state(cached_state);
1832
1833 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1834 if (ret) {
1835 btrfs_drop_pages(pages, num_pages);
1836 break;
1837 }
1838
1839 release_bytes = 0;
1840 if (only_release_metadata)
1841 btrfs_check_nocow_unlock(BTRFS_I(inode));
1842
1843 btrfs_drop_pages(pages, num_pages);
1844
1845 cond_resched();
1846
1847 balance_dirty_pages_ratelimited(inode->i_mapping);
1848
1849 pos += copied;
1850 num_written += copied;
1851 }
1852
1853 kfree(pages);
1854
1855 if (release_bytes) {
1856 if (only_release_metadata) {
1857 btrfs_check_nocow_unlock(BTRFS_I(inode));
1858 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1859 release_bytes, true);
1860 } else {
1861 btrfs_delalloc_release_space(BTRFS_I(inode),
1862 data_reserved,
1863 round_down(pos, fs_info->sectorsize),
1864 release_bytes, true);
1865 }
1866 }
1867
1868 extent_changeset_free(data_reserved);
1869 if (num_written > 0) {
1870 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1871 iocb->ki_pos += num_written;
1872 }
1873out:
1874 btrfs_inode_unlock(inode, ilock_flags);
1875 return num_written ? num_written : ret;
1876}
1877
1878static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1879 const struct iov_iter *iter, loff_t offset)
1880{
1881 const u32 blocksize_mask = fs_info->sectorsize - 1;
1882
1883 if (offset & blocksize_mask)
1884 return -EINVAL;
1885
1886 if (iov_iter_alignment(iter) & blocksize_mask)
1887 return -EINVAL;
1888
1889 return 0;
1890}
1891
1892static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1893{
1894 struct file *file = iocb->ki_filp;
1895 struct inode *inode = file_inode(file);
1896 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1897 loff_t pos;
1898 ssize_t written = 0;
1899 ssize_t written_buffered;
1900 loff_t endbyte;
1901 ssize_t err;
1902 unsigned int ilock_flags = 0;
1903 struct iomap_dio *dio = NULL;
1904
1905 if (iocb->ki_flags & IOCB_NOWAIT)
1906 ilock_flags |= BTRFS_ILOCK_TRY;
1907
1908 /* If the write DIO is within EOF, use a shared lock */
1909 if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1910 ilock_flags |= BTRFS_ILOCK_SHARED;
1911
1912relock:
1913 err = btrfs_inode_lock(inode, ilock_flags);
1914 if (err < 0)
1915 return err;
1916
1917 err = generic_write_checks(iocb, from);
1918 if (err <= 0) {
1919 btrfs_inode_unlock(inode, ilock_flags);
1920 return err;
1921 }
1922
1923 err = btrfs_write_check(iocb, from, err);
1924 if (err < 0) {
1925 btrfs_inode_unlock(inode, ilock_flags);
1926 goto out;
1927 }
1928
1929 pos = iocb->ki_pos;
1930 /*
1931 * Re-check since file size may have changed just before taking the
1932 * lock or pos may have changed because of O_APPEND in generic_write_check()
1933 */
1934 if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1935 pos + iov_iter_count(from) > i_size_read(inode)) {
1936 btrfs_inode_unlock(inode, ilock_flags);
1937 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1938 goto relock;
1939 }
1940
1941 if (check_direct_IO(fs_info, from, pos)) {
1942 btrfs_inode_unlock(inode, ilock_flags);
1943 goto buffered;
1944 }
1945
1946 dio = __iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
1947 0);
1948
1949 btrfs_inode_unlock(inode, ilock_flags);
1950
1951 if (IS_ERR_OR_NULL(dio)) {
1952 err = PTR_ERR_OR_ZERO(dio);
1953 if (err < 0 && err != -ENOTBLK)
1954 goto out;
1955 } else {
1956 written = iomap_dio_complete(dio);
1957 }
1958
1959 if (written < 0 || !iov_iter_count(from)) {
1960 err = written;
1961 goto out;
1962 }
1963
1964buffered:
1965 pos = iocb->ki_pos;
1966 written_buffered = btrfs_buffered_write(iocb, from);
1967 if (written_buffered < 0) {
1968 err = written_buffered;
1969 goto out;
1970 }
1971 /*
1972 * Ensure all data is persisted. We want the next direct IO read to be
1973 * able to read what was just written.
1974 */
1975 endbyte = pos + written_buffered - 1;
1976 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1977 if (err)
1978 goto out;
1979 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1980 if (err)
1981 goto out;
1982 written += written_buffered;
1983 iocb->ki_pos = pos + written_buffered;
1984 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1985 endbyte >> PAGE_SHIFT);
1986out:
1987 return written ? written : err;
1988}
1989
1990static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1991 struct iov_iter *from)
1992{
1993 struct file *file = iocb->ki_filp;
1994 struct btrfs_inode *inode = BTRFS_I(file_inode(file));
1995 ssize_t num_written = 0;
1996 const bool sync = iocb->ki_flags & IOCB_DSYNC;
1997
1998 /*
1999 * If the fs flips readonly due to some impossible error, although we
2000 * have opened a file as writable, we have to stop this write operation
2001 * to ensure consistency.
2002 */
2003 if (test_bit(BTRFS_FS_STATE_ERROR, &inode->root->fs_info->fs_state))
2004 return -EROFS;
2005
2006 if (!(iocb->ki_flags & IOCB_DIRECT) &&
2007 (iocb->ki_flags & IOCB_NOWAIT))
2008 return -EOPNOTSUPP;
2009
2010 if (sync)
2011 atomic_inc(&inode->sync_writers);
2012
2013 if (iocb->ki_flags & IOCB_DIRECT)
2014 num_written = btrfs_direct_write(iocb, from);
2015 else
2016 num_written = btrfs_buffered_write(iocb, from);
2017
2018 btrfs_set_inode_last_sub_trans(inode);
2019
2020 if (num_written > 0)
2021 num_written = generic_write_sync(iocb, num_written);
2022
2023 if (sync)
2024 atomic_dec(&inode->sync_writers);
2025
2026 current->backing_dev_info = NULL;
2027 return num_written;
2028}
2029
2030int btrfs_release_file(struct inode *inode, struct file *filp)
2031{
2032 struct btrfs_file_private *private = filp->private_data;
2033
2034 if (private && private->filldir_buf)
2035 kfree(private->filldir_buf);
2036 kfree(private);
2037 filp->private_data = NULL;
2038
2039 /*
2040 * Set by setattr when we are about to truncate a file from a non-zero
2041 * size to a zero size. This tries to flush down new bytes that may
2042 * have been written if the application were using truncate to replace
2043 * a file in place.
2044 */
2045 if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2046 &BTRFS_I(inode)->runtime_flags))
2047 filemap_flush(inode->i_mapping);
2048 return 0;
2049}
2050
2051static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2052{
2053 int ret;
2054 struct blk_plug plug;
2055
2056 /*
2057 * This is only called in fsync, which would do synchronous writes, so
2058 * a plug can merge adjacent IOs as much as possible. Esp. in case of
2059 * multiple disks using raid profile, a large IO can be split to
2060 * several segments of stripe length (currently 64K).
2061 */
2062 blk_start_plug(&plug);
2063 atomic_inc(&BTRFS_I(inode)->sync_writers);
2064 ret = btrfs_fdatawrite_range(inode, start, end);
2065 atomic_dec(&BTRFS_I(inode)->sync_writers);
2066 blk_finish_plug(&plug);
2067
2068 return ret;
2069}
2070
2071static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
2072{
2073 struct btrfs_inode *inode = BTRFS_I(ctx->inode);
2074 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2075
2076 if (btrfs_inode_in_log(inode, fs_info->generation) &&
2077 list_empty(&ctx->ordered_extents))
2078 return true;
2079
2080 /*
2081 * If we are doing a fast fsync we can not bail out if the inode's
2082 * last_trans is <= then the last committed transaction, because we only
2083 * update the last_trans of the inode during ordered extent completion,
2084 * and for a fast fsync we don't wait for that, we only wait for the
2085 * writeback to complete.
2086 */
2087 if (inode->last_trans <= fs_info->last_trans_committed &&
2088 (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
2089 list_empty(&ctx->ordered_extents)))
2090 return true;
2091
2092 return false;
2093}
2094
2095/*
2096 * fsync call for both files and directories. This logs the inode into
2097 * the tree log instead of forcing full commits whenever possible.
2098 *
2099 * It needs to call filemap_fdatawait so that all ordered extent updates are
2100 * in the metadata btree are up to date for copying to the log.
2101 *
2102 * It drops the inode mutex before doing the tree log commit. This is an
2103 * important optimization for directories because holding the mutex prevents
2104 * new operations on the dir while we write to disk.
2105 */
2106int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2107{
2108 struct dentry *dentry = file_dentry(file);
2109 struct inode *inode = d_inode(dentry);
2110 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2111 struct btrfs_root *root = BTRFS_I(inode)->root;
2112 struct btrfs_trans_handle *trans;
2113 struct btrfs_log_ctx ctx;
2114 int ret = 0, err;
2115 u64 len;
2116 bool full_sync;
2117
2118 trace_btrfs_sync_file(file, datasync);
2119
2120 btrfs_init_log_ctx(&ctx, inode);
2121
2122 /*
2123 * Always set the range to a full range, otherwise we can get into
2124 * several problems, from missing file extent items to represent holes
2125 * when not using the NO_HOLES feature, to log tree corruption due to
2126 * races between hole detection during logging and completion of ordered
2127 * extents outside the range, to missing checksums due to ordered extents
2128 * for which we flushed only a subset of their pages.
2129 */
2130 start = 0;
2131 end = LLONG_MAX;
2132 len = (u64)LLONG_MAX + 1;
2133
2134 /*
2135 * We write the dirty pages in the range and wait until they complete
2136 * out of the ->i_mutex. If so, we can flush the dirty pages by
2137 * multi-task, and make the performance up. See
2138 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2139 */
2140 ret = start_ordered_ops(inode, start, end);
2141 if (ret)
2142 goto out;
2143
2144 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2145
2146 atomic_inc(&root->log_batch);
2147
2148 /*
2149 * Always check for the full sync flag while holding the inode's lock,
2150 * to avoid races with other tasks. The flag must be either set all the
2151 * time during logging or always off all the time while logging.
2152 */
2153 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2154 &BTRFS_I(inode)->runtime_flags);
2155
2156 /*
2157 * Before we acquired the inode's lock and the mmap lock, someone may
2158 * have dirtied more pages in the target range. We need to make sure
2159 * that writeback for any such pages does not start while we are logging
2160 * the inode, because if it does, any of the following might happen when
2161 * we are not doing a full inode sync:
2162 *
2163 * 1) We log an extent after its writeback finishes but before its
2164 * checksums are added to the csum tree, leading to -EIO errors
2165 * when attempting to read the extent after a log replay.
2166 *
2167 * 2) We can end up logging an extent before its writeback finishes.
2168 * Therefore after the log replay we will have a file extent item
2169 * pointing to an unwritten extent (and no data checksums as well).
2170 *
2171 * So trigger writeback for any eventual new dirty pages and then we
2172 * wait for all ordered extents to complete below.
2173 */
2174 ret = start_ordered_ops(inode, start, end);
2175 if (ret) {
2176 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2177 goto out;
2178 }
2179
2180 /*
2181 * We have to do this here to avoid the priority inversion of waiting on
2182 * IO of a lower priority task while holding a transaction open.
2183 *
2184 * For a full fsync we wait for the ordered extents to complete while
2185 * for a fast fsync we wait just for writeback to complete, and then
2186 * attach the ordered extents to the transaction so that a transaction
2187 * commit waits for their completion, to avoid data loss if we fsync,
2188 * the current transaction commits before the ordered extents complete
2189 * and a power failure happens right after that.
2190 *
2191 * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
2192 * logical address recorded in the ordered extent may change. We need
2193 * to wait for the IO to stabilize the logical address.
2194 */
2195 if (full_sync || btrfs_is_zoned(fs_info)) {
2196 ret = btrfs_wait_ordered_range(inode, start, len);
2197 } else {
2198 /*
2199 * Get our ordered extents as soon as possible to avoid doing
2200 * checksum lookups in the csum tree, and use instead the
2201 * checksums attached to the ordered extents.
2202 */
2203 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2204 &ctx.ordered_extents);
2205 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2206 }
2207
2208 if (ret)
2209 goto out_release_extents;
2210
2211 atomic_inc(&root->log_batch);
2212
2213 smp_mb();
2214 if (skip_inode_logging(&ctx)) {
2215 /*
2216 * We've had everything committed since the last time we were
2217 * modified so clear this flag in case it was set for whatever
2218 * reason, it's no longer relevant.
2219 */
2220 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2221 &BTRFS_I(inode)->runtime_flags);
2222 /*
2223 * An ordered extent might have started before and completed
2224 * already with io errors, in which case the inode was not
2225 * updated and we end up here. So check the inode's mapping
2226 * for any errors that might have happened since we last
2227 * checked called fsync.
2228 */
2229 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2230 goto out_release_extents;
2231 }
2232
2233 /*
2234 * We use start here because we will need to wait on the IO to complete
2235 * in btrfs_sync_log, which could require joining a transaction (for
2236 * example checking cross references in the nocow path). If we use join
2237 * here we could get into a situation where we're waiting on IO to
2238 * happen that is blocked on a transaction trying to commit. With start
2239 * we inc the extwriter counter, so we wait for all extwriters to exit
2240 * before we start blocking joiners. This comment is to keep somebody
2241 * from thinking they are super smart and changing this to
2242 * btrfs_join_transaction *cough*Josef*cough*.
2243 */
2244 trans = btrfs_start_transaction(root, 0);
2245 if (IS_ERR(trans)) {
2246 ret = PTR_ERR(trans);
2247 goto out_release_extents;
2248 }
2249 trans->in_fsync = true;
2250
2251 ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2252 btrfs_release_log_ctx_extents(&ctx);
2253 if (ret < 0) {
2254 /* Fallthrough and commit/free transaction. */
2255 ret = 1;
2256 }
2257
2258 /* we've logged all the items and now have a consistent
2259 * version of the file in the log. It is possible that
2260 * someone will come in and modify the file, but that's
2261 * fine because the log is consistent on disk, and we
2262 * have references to all of the file's extents
2263 *
2264 * It is possible that someone will come in and log the
2265 * file again, but that will end up using the synchronization
2266 * inside btrfs_sync_log to keep things safe.
2267 */
2268 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2269
2270 if (ret != BTRFS_NO_LOG_SYNC) {
2271 if (!ret) {
2272 ret = btrfs_sync_log(trans, root, &ctx);
2273 if (!ret) {
2274 ret = btrfs_end_transaction(trans);
2275 goto out;
2276 }
2277 }
2278 if (!full_sync) {
2279 ret = btrfs_wait_ordered_range(inode, start, len);
2280 if (ret) {
2281 btrfs_end_transaction(trans);
2282 goto out;
2283 }
2284 }
2285 ret = btrfs_commit_transaction(trans);
2286 } else {
2287 ret = btrfs_end_transaction(trans);
2288 }
2289out:
2290 ASSERT(list_empty(&ctx.list));
2291 err = file_check_and_advance_wb_err(file);
2292 if (!ret)
2293 ret = err;
2294 return ret > 0 ? -EIO : ret;
2295
2296out_release_extents:
2297 btrfs_release_log_ctx_extents(&ctx);
2298 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2299 goto out;
2300}
2301
2302static const struct vm_operations_struct btrfs_file_vm_ops = {
2303 .fault = filemap_fault,
2304 .map_pages = filemap_map_pages,
2305 .page_mkwrite = btrfs_page_mkwrite,
2306};
2307
2308static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2309{
2310 struct address_space *mapping = filp->f_mapping;
2311
2312 if (!mapping->a_ops->readpage)
2313 return -ENOEXEC;
2314
2315 file_accessed(filp);
2316 vma->vm_ops = &btrfs_file_vm_ops;
2317
2318 return 0;
2319}
2320
2321static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2322 int slot, u64 start, u64 end)
2323{
2324 struct btrfs_file_extent_item *fi;
2325 struct btrfs_key key;
2326
2327 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2328 return 0;
2329
2330 btrfs_item_key_to_cpu(leaf, &key, slot);
2331 if (key.objectid != btrfs_ino(inode) ||
2332 key.type != BTRFS_EXTENT_DATA_KEY)
2333 return 0;
2334
2335 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2336
2337 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2338 return 0;
2339
2340 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2341 return 0;
2342
2343 if (key.offset == end)
2344 return 1;
2345 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2346 return 1;
2347 return 0;
2348}
2349
2350static int fill_holes(struct btrfs_trans_handle *trans,
2351 struct btrfs_inode *inode,
2352 struct btrfs_path *path, u64 offset, u64 end)
2353{
2354 struct btrfs_fs_info *fs_info = trans->fs_info;
2355 struct btrfs_root *root = inode->root;
2356 struct extent_buffer *leaf;
2357 struct btrfs_file_extent_item *fi;
2358 struct extent_map *hole_em;
2359 struct extent_map_tree *em_tree = &inode->extent_tree;
2360 struct btrfs_key key;
2361 int ret;
2362
2363 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2364 goto out;
2365
2366 key.objectid = btrfs_ino(inode);
2367 key.type = BTRFS_EXTENT_DATA_KEY;
2368 key.offset = offset;
2369
2370 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2371 if (ret <= 0) {
2372 /*
2373 * We should have dropped this offset, so if we find it then
2374 * something has gone horribly wrong.
2375 */
2376 if (ret == 0)
2377 ret = -EINVAL;
2378 return ret;
2379 }
2380
2381 leaf = path->nodes[0];
2382 if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2383 u64 num_bytes;
2384
2385 path->slots[0]--;
2386 fi = btrfs_item_ptr(leaf, path->slots[0],
2387 struct btrfs_file_extent_item);
2388 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2389 end - offset;
2390 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2391 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2392 btrfs_set_file_extent_offset(leaf, fi, 0);
2393 btrfs_mark_buffer_dirty(leaf);
2394 goto out;
2395 }
2396
2397 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2398 u64 num_bytes;
2399
2400 key.offset = offset;
2401 btrfs_set_item_key_safe(fs_info, path, &key);
2402 fi = btrfs_item_ptr(leaf, path->slots[0],
2403 struct btrfs_file_extent_item);
2404 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2405 offset;
2406 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2407 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2408 btrfs_set_file_extent_offset(leaf, fi, 0);
2409 btrfs_mark_buffer_dirty(leaf);
2410 goto out;
2411 }
2412 btrfs_release_path(path);
2413
2414 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2415 offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2416 if (ret)
2417 return ret;
2418
2419out:
2420 btrfs_release_path(path);
2421
2422 hole_em = alloc_extent_map();
2423 if (!hole_em) {
2424 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2425 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2426 } else {
2427 hole_em->start = offset;
2428 hole_em->len = end - offset;
2429 hole_em->ram_bytes = hole_em->len;
2430 hole_em->orig_start = offset;
2431
2432 hole_em->block_start = EXTENT_MAP_HOLE;
2433 hole_em->block_len = 0;
2434 hole_em->orig_block_len = 0;
2435 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2436 hole_em->generation = trans->transid;
2437
2438 do {
2439 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2440 write_lock(&em_tree->lock);
2441 ret = add_extent_mapping(em_tree, hole_em, 1);
2442 write_unlock(&em_tree->lock);
2443 } while (ret == -EEXIST);
2444 free_extent_map(hole_em);
2445 if (ret)
2446 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2447 &inode->runtime_flags);
2448 }
2449
2450 return 0;
2451}
2452
2453/*
2454 * Find a hole extent on given inode and change start/len to the end of hole
2455 * extent.(hole/vacuum extent whose em->start <= start &&
2456 * em->start + em->len > start)
2457 * When a hole extent is found, return 1 and modify start/len.
2458 */
2459static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2460{
2461 struct btrfs_fs_info *fs_info = inode->root->fs_info;
2462 struct extent_map *em;
2463 int ret = 0;
2464
2465 em = btrfs_get_extent(inode, NULL, 0,
2466 round_down(*start, fs_info->sectorsize),
2467 round_up(*len, fs_info->sectorsize));
2468 if (IS_ERR(em))
2469 return PTR_ERR(em);
2470
2471 /* Hole or vacuum extent(only exists in no-hole mode) */
2472 if (em->block_start == EXTENT_MAP_HOLE) {
2473 ret = 1;
2474 *len = em->start + em->len > *start + *len ?
2475 0 : *start + *len - em->start - em->len;
2476 *start = em->start + em->len;
2477 }
2478 free_extent_map(em);
2479 return ret;
2480}
2481
2482static int btrfs_punch_hole_lock_range(struct inode *inode,
2483 const u64 lockstart,
2484 const u64 lockend,
2485 struct extent_state **cached_state)
2486{
2487 /*
2488 * For subpage case, if the range is not at page boundary, we could
2489 * have pages at the leading/tailing part of the range.
2490 * This could lead to dead loop since filemap_range_has_page()
2491 * will always return true.
2492 * So here we need to do extra page alignment for
2493 * filemap_range_has_page().
2494 */
2495 const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2496 const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2497
2498 while (1) {
2499 struct btrfs_ordered_extent *ordered;
2500 int ret;
2501
2502 truncate_pagecache_range(inode, lockstart, lockend);
2503
2504 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2505 cached_state);
2506 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2507 lockend);
2508
2509 /*
2510 * We need to make sure we have no ordered extents in this range
2511 * and nobody raced in and read a page in this range, if we did
2512 * we need to try again.
2513 */
2514 if ((!ordered ||
2515 (ordered->file_offset + ordered->num_bytes <= lockstart ||
2516 ordered->file_offset > lockend)) &&
2517 !filemap_range_has_page(inode->i_mapping,
2518 page_lockstart, page_lockend)) {
2519 if (ordered)
2520 btrfs_put_ordered_extent(ordered);
2521 break;
2522 }
2523 if (ordered)
2524 btrfs_put_ordered_extent(ordered);
2525 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2526 lockend, cached_state);
2527 ret = btrfs_wait_ordered_range(inode, lockstart,
2528 lockend - lockstart + 1);
2529 if (ret)
2530 return ret;
2531 }
2532 return 0;
2533}
2534
2535static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2536 struct btrfs_inode *inode,
2537 struct btrfs_path *path,
2538 struct btrfs_replace_extent_info *extent_info,
2539 const u64 replace_len,
2540 const u64 bytes_to_drop)
2541{
2542 struct btrfs_fs_info *fs_info = trans->fs_info;
2543 struct btrfs_root *root = inode->root;
2544 struct btrfs_file_extent_item *extent;
2545 struct extent_buffer *leaf;
2546 struct btrfs_key key;
2547 int slot;
2548 struct btrfs_ref ref = { 0 };
2549 int ret;
2550
2551 if (replace_len == 0)
2552 return 0;
2553
2554 if (extent_info->disk_offset == 0 &&
2555 btrfs_fs_incompat(fs_info, NO_HOLES)) {
2556 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2557 return 0;
2558 }
2559
2560 key.objectid = btrfs_ino(inode);
2561 key.type = BTRFS_EXTENT_DATA_KEY;
2562 key.offset = extent_info->file_offset;
2563 ret = btrfs_insert_empty_item(trans, root, path, &key,
2564 sizeof(struct btrfs_file_extent_item));
2565 if (ret)
2566 return ret;
2567 leaf = path->nodes[0];
2568 slot = path->slots[0];
2569 write_extent_buffer(leaf, extent_info->extent_buf,
2570 btrfs_item_ptr_offset(leaf, slot),
2571 sizeof(struct btrfs_file_extent_item));
2572 extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2573 ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2574 btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2575 btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2576 if (extent_info->is_new_extent)
2577 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2578 btrfs_mark_buffer_dirty(leaf);
2579 btrfs_release_path(path);
2580
2581 ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2582 replace_len);
2583 if (ret)
2584 return ret;
2585
2586 /* If it's a hole, nothing more needs to be done. */
2587 if (extent_info->disk_offset == 0) {
2588 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2589 return 0;
2590 }
2591
2592 btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2593
2594 if (extent_info->is_new_extent && extent_info->insertions == 0) {
2595 key.objectid = extent_info->disk_offset;
2596 key.type = BTRFS_EXTENT_ITEM_KEY;
2597 key.offset = extent_info->disk_len;
2598 ret = btrfs_alloc_reserved_file_extent(trans, root,
2599 btrfs_ino(inode),
2600 extent_info->file_offset,
2601 extent_info->qgroup_reserved,
2602 &key);
2603 } else {
2604 u64 ref_offset;
2605
2606 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2607 extent_info->disk_offset,
2608 extent_info->disk_len, 0);
2609 ref_offset = extent_info->file_offset - extent_info->data_offset;
2610 btrfs_init_data_ref(&ref, root->root_key.objectid,
2611 btrfs_ino(inode), ref_offset);
2612 ret = btrfs_inc_extent_ref(trans, &ref);
2613 }
2614
2615 extent_info->insertions++;
2616
2617 return ret;
2618}
2619
2620/*
2621 * The respective range must have been previously locked, as well as the inode.
2622 * The end offset is inclusive (last byte of the range).
2623 * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2624 * the file range with an extent.
2625 * When not punching a hole, we don't want to end up in a state where we dropped
2626 * extents without inserting a new one, so we must abort the transaction to avoid
2627 * a corruption.
2628 */
2629int btrfs_replace_file_extents(struct btrfs_inode *inode,
2630 struct btrfs_path *path, const u64 start,
2631 const u64 end,
2632 struct btrfs_replace_extent_info *extent_info,
2633 struct btrfs_trans_handle **trans_out)
2634{
2635 struct btrfs_drop_extents_args drop_args = { 0 };
2636 struct btrfs_root *root = inode->root;
2637 struct btrfs_fs_info *fs_info = root->fs_info;
2638 u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2639 u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2640 struct btrfs_trans_handle *trans = NULL;
2641 struct btrfs_block_rsv *rsv;
2642 unsigned int rsv_count;
2643 u64 cur_offset;
2644 u64 len = end - start;
2645 int ret = 0;
2646
2647 if (end <= start)
2648 return -EINVAL;
2649
2650 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2651 if (!rsv) {
2652 ret = -ENOMEM;
2653 goto out;
2654 }
2655 rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2656 rsv->failfast = 1;
2657
2658 /*
2659 * 1 - update the inode
2660 * 1 - removing the extents in the range
2661 * 1 - adding the hole extent if no_holes isn't set or if we are
2662 * replacing the range with a new extent
2663 */
2664 if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2665 rsv_count = 3;
2666 else
2667 rsv_count = 2;
2668
2669 trans = btrfs_start_transaction(root, rsv_count);
2670 if (IS_ERR(trans)) {
2671 ret = PTR_ERR(trans);
2672 trans = NULL;
2673 goto out_free;
2674 }
2675
2676 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2677 min_size, false);
2678 BUG_ON(ret);
2679 trans->block_rsv = rsv;
2680
2681 cur_offset = start;
2682 drop_args.path = path;
2683 drop_args.end = end + 1;
2684 drop_args.drop_cache = true;
2685 while (cur_offset < end) {
2686 drop_args.start = cur_offset;
2687 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2688 /* If we are punching a hole decrement the inode's byte count */
2689 if (!extent_info)
2690 btrfs_update_inode_bytes(inode, 0,
2691 drop_args.bytes_found);
2692 if (ret != -ENOSPC) {
2693 /*
2694 * The only time we don't want to abort is if we are
2695 * attempting to clone a partial inline extent, in which
2696 * case we'll get EOPNOTSUPP. However if we aren't
2697 * clone we need to abort no matter what, because if we
2698 * got EOPNOTSUPP via prealloc then we messed up and
2699 * need to abort.
2700 */
2701 if (ret &&
2702 (ret != -EOPNOTSUPP ||
2703 (extent_info && extent_info->is_new_extent)))
2704 btrfs_abort_transaction(trans, ret);
2705 break;
2706 }
2707
2708 trans->block_rsv = &fs_info->trans_block_rsv;
2709
2710 if (!extent_info && cur_offset < drop_args.drop_end &&
2711 cur_offset < ino_size) {
2712 ret = fill_holes(trans, inode, path, cur_offset,
2713 drop_args.drop_end);
2714 if (ret) {
2715 /*
2716 * If we failed then we didn't insert our hole
2717 * entries for the area we dropped, so now the
2718 * fs is corrupted, so we must abort the
2719 * transaction.
2720 */
2721 btrfs_abort_transaction(trans, ret);
2722 break;
2723 }
2724 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2725 /*
2726 * We are past the i_size here, but since we didn't
2727 * insert holes we need to clear the mapped area so we
2728 * know to not set disk_i_size in this area until a new
2729 * file extent is inserted here.
2730 */
2731 ret = btrfs_inode_clear_file_extent_range(inode,
2732 cur_offset,
2733 drop_args.drop_end - cur_offset);
2734 if (ret) {
2735 /*
2736 * We couldn't clear our area, so we could
2737 * presumably adjust up and corrupt the fs, so
2738 * we need to abort.
2739 */
2740 btrfs_abort_transaction(trans, ret);
2741 break;
2742 }
2743 }
2744
2745 if (extent_info &&
2746 drop_args.drop_end > extent_info->file_offset) {
2747 u64 replace_len = drop_args.drop_end -
2748 extent_info->file_offset;
2749
2750 ret = btrfs_insert_replace_extent(trans, inode, path,
2751 extent_info, replace_len,
2752 drop_args.bytes_found);
2753 if (ret) {
2754 btrfs_abort_transaction(trans, ret);
2755 break;
2756 }
2757 extent_info->data_len -= replace_len;
2758 extent_info->data_offset += replace_len;
2759 extent_info->file_offset += replace_len;
2760 }
2761
2762 ret = btrfs_update_inode(trans, root, inode);
2763 if (ret)
2764 break;
2765
2766 btrfs_end_transaction(trans);
2767 btrfs_btree_balance_dirty(fs_info);
2768
2769 trans = btrfs_start_transaction(root, rsv_count);
2770 if (IS_ERR(trans)) {
2771 ret = PTR_ERR(trans);
2772 trans = NULL;
2773 break;
2774 }
2775
2776 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2777 rsv, min_size, false);
2778 BUG_ON(ret); /* shouldn't happen */
2779 trans->block_rsv = rsv;
2780
2781 cur_offset = drop_args.drop_end;
2782 len = end - cur_offset;
2783 if (!extent_info && len) {
2784 ret = find_first_non_hole(inode, &cur_offset, &len);
2785 if (unlikely(ret < 0))
2786 break;
2787 if (ret && !len) {
2788 ret = 0;
2789 break;
2790 }
2791 }
2792 }
2793
2794 /*
2795 * If we were cloning, force the next fsync to be a full one since we
2796 * we replaced (or just dropped in the case of cloning holes when
2797 * NO_HOLES is enabled) file extent items and did not setup new extent
2798 * maps for the replacement extents (or holes).
2799 */
2800 if (extent_info && !extent_info->is_new_extent)
2801 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2802
2803 if (ret)
2804 goto out_trans;
2805
2806 trans->block_rsv = &fs_info->trans_block_rsv;
2807 /*
2808 * If we are using the NO_HOLES feature we might have had already an
2809 * hole that overlaps a part of the region [lockstart, lockend] and
2810 * ends at (or beyond) lockend. Since we have no file extent items to
2811 * represent holes, drop_end can be less than lockend and so we must
2812 * make sure we have an extent map representing the existing hole (the
2813 * call to __btrfs_drop_extents() might have dropped the existing extent
2814 * map representing the existing hole), otherwise the fast fsync path
2815 * will not record the existence of the hole region
2816 * [existing_hole_start, lockend].
2817 */
2818 if (drop_args.drop_end <= end)
2819 drop_args.drop_end = end + 1;
2820 /*
2821 * Don't insert file hole extent item if it's for a range beyond eof
2822 * (because it's useless) or if it represents a 0 bytes range (when
2823 * cur_offset == drop_end).
2824 */
2825 if (!extent_info && cur_offset < ino_size &&
2826 cur_offset < drop_args.drop_end) {
2827 ret = fill_holes(trans, inode, path, cur_offset,
2828 drop_args.drop_end);
2829 if (ret) {
2830 /* Same comment as above. */
2831 btrfs_abort_transaction(trans, ret);
2832 goto out_trans;
2833 }
2834 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2835 /* See the comment in the loop above for the reasoning here. */
2836 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2837 drop_args.drop_end - cur_offset);
2838 if (ret) {
2839 btrfs_abort_transaction(trans, ret);
2840 goto out_trans;
2841 }
2842
2843 }
2844 if (extent_info) {
2845 ret = btrfs_insert_replace_extent(trans, inode, path,
2846 extent_info, extent_info->data_len,
2847 drop_args.bytes_found);
2848 if (ret) {
2849 btrfs_abort_transaction(trans, ret);
2850 goto out_trans;
2851 }
2852 }
2853
2854out_trans:
2855 if (!trans)
2856 goto out_free;
2857
2858 trans->block_rsv = &fs_info->trans_block_rsv;
2859 if (ret)
2860 btrfs_end_transaction(trans);
2861 else
2862 *trans_out = trans;
2863out_free:
2864 btrfs_free_block_rsv(fs_info, rsv);
2865out:
2866 return ret;
2867}
2868
2869static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2870{
2871 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2872 struct btrfs_root *root = BTRFS_I(inode)->root;
2873 struct extent_state *cached_state = NULL;
2874 struct btrfs_path *path;
2875 struct btrfs_trans_handle *trans = NULL;
2876 u64 lockstart;
2877 u64 lockend;
2878 u64 tail_start;
2879 u64 tail_len;
2880 u64 orig_start = offset;
2881 int ret = 0;
2882 bool same_block;
2883 u64 ino_size;
2884 bool truncated_block = false;
2885 bool updated_inode = false;
2886
2887 ret = btrfs_wait_ordered_range(inode, offset, len);
2888 if (ret)
2889 return ret;
2890
2891 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2892 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2893 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2894 if (ret < 0)
2895 goto out_only_mutex;
2896 if (ret && !len) {
2897 /* Already in a large hole */
2898 ret = 0;
2899 goto out_only_mutex;
2900 }
2901
2902 lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2903 lockend = round_down(offset + len,
2904 btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2905 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2906 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2907 /*
2908 * We needn't truncate any block which is beyond the end of the file
2909 * because we are sure there is no data there.
2910 */
2911 /*
2912 * Only do this if we are in the same block and we aren't doing the
2913 * entire block.
2914 */
2915 if (same_block && len < fs_info->sectorsize) {
2916 if (offset < ino_size) {
2917 truncated_block = true;
2918 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2919 0);
2920 } else {
2921 ret = 0;
2922 }
2923 goto out_only_mutex;
2924 }
2925
2926 /* zero back part of the first block */
2927 if (offset < ino_size) {
2928 truncated_block = true;
2929 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2930 if (ret) {
2931 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2932 return ret;
2933 }
2934 }
2935
2936 /* Check the aligned pages after the first unaligned page,
2937 * if offset != orig_start, which means the first unaligned page
2938 * including several following pages are already in holes,
2939 * the extra check can be skipped */
2940 if (offset == orig_start) {
2941 /* after truncate page, check hole again */
2942 len = offset + len - lockstart;
2943 offset = lockstart;
2944 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2945 if (ret < 0)
2946 goto out_only_mutex;
2947 if (ret && !len) {
2948 ret = 0;
2949 goto out_only_mutex;
2950 }
2951 lockstart = offset;
2952 }
2953
2954 /* Check the tail unaligned part is in a hole */
2955 tail_start = lockend + 1;
2956 tail_len = offset + len - tail_start;
2957 if (tail_len) {
2958 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2959 if (unlikely(ret < 0))
2960 goto out_only_mutex;
2961 if (!ret) {
2962 /* zero the front end of the last page */
2963 if (tail_start + tail_len < ino_size) {
2964 truncated_block = true;
2965 ret = btrfs_truncate_block(BTRFS_I(inode),
2966 tail_start + tail_len,
2967 0, 1);
2968 if (ret)
2969 goto out_only_mutex;
2970 }
2971 }
2972 }
2973
2974 if (lockend < lockstart) {
2975 ret = 0;
2976 goto out_only_mutex;
2977 }
2978
2979 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2980 &cached_state);
2981 if (ret)
2982 goto out_only_mutex;
2983
2984 path = btrfs_alloc_path();
2985 if (!path) {
2986 ret = -ENOMEM;
2987 goto out;
2988 }
2989
2990 ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
2991 lockend, NULL, &trans);
2992 btrfs_free_path(path);
2993 if (ret)
2994 goto out;
2995
2996 ASSERT(trans != NULL);
2997 inode_inc_iversion(inode);
2998 inode->i_mtime = inode->i_ctime = current_time(inode);
2999 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3000 updated_inode = true;
3001 btrfs_end_transaction(trans);
3002 btrfs_btree_balance_dirty(fs_info);
3003out:
3004 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3005 &cached_state);
3006out_only_mutex:
3007 if (!updated_inode && truncated_block && !ret) {
3008 /*
3009 * If we only end up zeroing part of a page, we still need to
3010 * update the inode item, so that all the time fields are
3011 * updated as well as the necessary btrfs inode in memory fields
3012 * for detecting, at fsync time, if the inode isn't yet in the
3013 * log tree or it's there but not up to date.
3014 */
3015 struct timespec64 now = current_time(inode);
3016
3017 inode_inc_iversion(inode);
3018 inode->i_mtime = now;
3019 inode->i_ctime = now;
3020 trans = btrfs_start_transaction(root, 1);
3021 if (IS_ERR(trans)) {
3022 ret = PTR_ERR(trans);
3023 } else {
3024 int ret2;
3025
3026 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3027 ret2 = btrfs_end_transaction(trans);
3028 if (!ret)
3029 ret = ret2;
3030 }
3031 }
3032 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3033 return ret;
3034}
3035
3036/* Helper structure to record which range is already reserved */
3037struct falloc_range {
3038 struct list_head list;
3039 u64 start;
3040 u64 len;
3041};
3042
3043/*
3044 * Helper function to add falloc range
3045 *
3046 * Caller should have locked the larger range of extent containing
3047 * [start, len)
3048 */
3049static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3050{
3051 struct falloc_range *range = NULL;
3052
3053 if (!list_empty(head)) {
3054 /*
3055 * As fallocate iterates by bytenr order, we only need to check
3056 * the last range.
3057 */
3058 range = list_last_entry(head, struct falloc_range, list);
3059 if (range->start + range->len == start) {
3060 range->len += len;
3061 return 0;
3062 }
3063 }
3064
3065 range = kmalloc(sizeof(*range), GFP_KERNEL);
3066 if (!range)
3067 return -ENOMEM;
3068 range->start = start;
3069 range->len = len;
3070 list_add_tail(&range->list, head);
3071 return 0;
3072}
3073
3074static int btrfs_fallocate_update_isize(struct inode *inode,
3075 const u64 end,
3076 const int mode)
3077{
3078 struct btrfs_trans_handle *trans;
3079 struct btrfs_root *root = BTRFS_I(inode)->root;
3080 int ret;
3081 int ret2;
3082
3083 if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3084 return 0;
3085
3086 trans = btrfs_start_transaction(root, 1);
3087 if (IS_ERR(trans))
3088 return PTR_ERR(trans);
3089
3090 inode->i_ctime = current_time(inode);
3091 i_size_write(inode, end);
3092 btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
3093 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3094 ret2 = btrfs_end_transaction(trans);
3095
3096 return ret ? ret : ret2;
3097}
3098
3099enum {
3100 RANGE_BOUNDARY_WRITTEN_EXTENT,
3101 RANGE_BOUNDARY_PREALLOC_EXTENT,
3102 RANGE_BOUNDARY_HOLE,
3103};
3104
3105static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3106 u64 offset)
3107{
3108 const u64 sectorsize = btrfs_inode_sectorsize(inode);
3109 struct extent_map *em;
3110 int ret;
3111
3112 offset = round_down(offset, sectorsize);
3113 em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3114 if (IS_ERR(em))
3115 return PTR_ERR(em);
3116
3117 if (em->block_start == EXTENT_MAP_HOLE)
3118 ret = RANGE_BOUNDARY_HOLE;
3119 else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3120 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3121 else
3122 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3123
3124 free_extent_map(em);
3125 return ret;
3126}
3127
3128static int btrfs_zero_range(struct inode *inode,
3129 loff_t offset,
3130 loff_t len,
3131 const int mode)
3132{
3133 struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3134 struct extent_map *em;
3135 struct extent_changeset *data_reserved = NULL;
3136 int ret;
3137 u64 alloc_hint = 0;
3138 const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3139 u64 alloc_start = round_down(offset, sectorsize);
3140 u64 alloc_end = round_up(offset + len, sectorsize);
3141 u64 bytes_to_reserve = 0;
3142 bool space_reserved = false;
3143
3144 inode_dio_wait(inode);
3145
3146 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3147 alloc_end - alloc_start);
3148 if (IS_ERR(em)) {
3149 ret = PTR_ERR(em);
3150 goto out;
3151 }
3152
3153 /*
3154 * Avoid hole punching and extent allocation for some cases. More cases
3155 * could be considered, but these are unlikely common and we keep things
3156 * as simple as possible for now. Also, intentionally, if the target
3157 * range contains one or more prealloc extents together with regular
3158 * extents and holes, we drop all the existing extents and allocate a
3159 * new prealloc extent, so that we get a larger contiguous disk extent.
3160 */
3161 if (em->start <= alloc_start &&
3162 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3163 const u64 em_end = em->start + em->len;
3164
3165 if (em_end >= offset + len) {
3166 /*
3167 * The whole range is already a prealloc extent,
3168 * do nothing except updating the inode's i_size if
3169 * needed.
3170 */
3171 free_extent_map(em);
3172 ret = btrfs_fallocate_update_isize(inode, offset + len,
3173 mode);
3174 goto out;
3175 }
3176 /*
3177 * Part of the range is already a prealloc extent, so operate
3178 * only on the remaining part of the range.
3179 */
3180 alloc_start = em_end;
3181 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3182 len = offset + len - alloc_start;
3183 offset = alloc_start;
3184 alloc_hint = em->block_start + em->len;
3185 }
3186 free_extent_map(em);
3187
3188 if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3189 BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3190 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3191 sectorsize);
3192 if (IS_ERR(em)) {
3193 ret = PTR_ERR(em);
3194 goto out;
3195 }
3196
3197 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3198 free_extent_map(em);
3199 ret = btrfs_fallocate_update_isize(inode, offset + len,
3200 mode);
3201 goto out;
3202 }
3203 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3204 free_extent_map(em);
3205 ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3206 0);
3207 if (!ret)
3208 ret = btrfs_fallocate_update_isize(inode,
3209 offset + len,
3210 mode);
3211 return ret;
3212 }
3213 free_extent_map(em);
3214 alloc_start = round_down(offset, sectorsize);
3215 alloc_end = alloc_start + sectorsize;
3216 goto reserve_space;
3217 }
3218
3219 alloc_start = round_up(offset, sectorsize);
3220 alloc_end = round_down(offset + len, sectorsize);
3221
3222 /*
3223 * For unaligned ranges, check the pages at the boundaries, they might
3224 * map to an extent, in which case we need to partially zero them, or
3225 * they might map to a hole, in which case we need our allocation range
3226 * to cover them.
3227 */
3228 if (!IS_ALIGNED(offset, sectorsize)) {
3229 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3230 offset);
3231 if (ret < 0)
3232 goto out;
3233 if (ret == RANGE_BOUNDARY_HOLE) {
3234 alloc_start = round_down(offset, sectorsize);
3235 ret = 0;
3236 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3237 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3238 if (ret)
3239 goto out;
3240 } else {
3241 ret = 0;
3242 }
3243 }
3244
3245 if (!IS_ALIGNED(offset + len, sectorsize)) {
3246 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3247 offset + len);
3248 if (ret < 0)
3249 goto out;
3250 if (ret == RANGE_BOUNDARY_HOLE) {
3251 alloc_end = round_up(offset + len, sectorsize);
3252 ret = 0;
3253 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3254 ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
3255 0, 1);
3256 if (ret)
3257 goto out;
3258 } else {
3259 ret = 0;
3260 }
3261 }
3262
3263reserve_space:
3264 if (alloc_start < alloc_end) {
3265 struct extent_state *cached_state = NULL;
3266 const u64 lockstart = alloc_start;
3267 const u64 lockend = alloc_end - 1;
3268
3269 bytes_to_reserve = alloc_end - alloc_start;
3270 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3271 bytes_to_reserve);
3272 if (ret < 0)
3273 goto out;
3274 space_reserved = true;
3275 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3276 &cached_state);
3277 if (ret)
3278 goto out;
3279 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3280 alloc_start, bytes_to_reserve);
3281 if (ret) {
3282 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3283 lockend, &cached_state);
3284 goto out;
3285 }
3286 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3287 alloc_end - alloc_start,
3288 i_blocksize(inode),
3289 offset + len, &alloc_hint);
3290 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3291 lockend, &cached_state);
3292 /* btrfs_prealloc_file_range releases reserved space on error */
3293 if (ret) {
3294 space_reserved = false;
3295 goto out;
3296 }
3297 }
3298 ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3299 out:
3300 if (ret && space_reserved)
3301 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3302 alloc_start, bytes_to_reserve);
3303 extent_changeset_free(data_reserved);
3304
3305 return ret;
3306}
3307
3308static long btrfs_fallocate(struct file *file, int mode,
3309 loff_t offset, loff_t len)
3310{
3311 struct inode *inode = file_inode(file);
3312 struct extent_state *cached_state = NULL;
3313 struct extent_changeset *data_reserved = NULL;
3314 struct falloc_range *range;
3315 struct falloc_range *tmp;
3316 struct list_head reserve_list;
3317 u64 cur_offset;
3318 u64 last_byte;
3319 u64 alloc_start;
3320 u64 alloc_end;
3321 u64 alloc_hint = 0;
3322 u64 locked_end;
3323 u64 actual_end = 0;
3324 struct extent_map *em;
3325 int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3326 int ret;
3327
3328 /* Do not allow fallocate in ZONED mode */
3329 if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3330 return -EOPNOTSUPP;
3331
3332 alloc_start = round_down(offset, blocksize);
3333 alloc_end = round_up(offset + len, blocksize);
3334 cur_offset = alloc_start;
3335
3336 /* Make sure we aren't being give some crap mode */
3337 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3338 FALLOC_FL_ZERO_RANGE))
3339 return -EOPNOTSUPP;
3340
3341 if (mode & FALLOC_FL_PUNCH_HOLE)
3342 return btrfs_punch_hole(inode, offset, len);
3343
3344 /*
3345 * Only trigger disk allocation, don't trigger qgroup reserve
3346 *
3347 * For qgroup space, it will be checked later.
3348 */
3349 if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3350 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3351 alloc_end - alloc_start);
3352 if (ret < 0)
3353 return ret;
3354 }
3355
3356 btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
3357
3358 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3359 ret = inode_newsize_ok(inode, offset + len);
3360 if (ret)
3361 goto out;
3362 }
3363
3364 /*
3365 * TODO: Move these two operations after we have checked
3366 * accurate reserved space, or fallocate can still fail but
3367 * with page truncated or size expanded.
3368 *
3369 * But that's a minor problem and won't do much harm BTW.
3370 */
3371 if (alloc_start > inode->i_size) {
3372 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3373 alloc_start);
3374 if (ret)
3375 goto out;
3376 } else if (offset + len > inode->i_size) {
3377 /*
3378 * If we are fallocating from the end of the file onward we
3379 * need to zero out the end of the block if i_size lands in the
3380 * middle of a block.
3381 */
3382 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3383 if (ret)
3384 goto out;
3385 }
3386
3387 /*
3388 * wait for ordered IO before we have any locks. We'll loop again
3389 * below with the locks held.
3390 */
3391 ret = btrfs_wait_ordered_range(inode, alloc_start,
3392 alloc_end - alloc_start);
3393 if (ret)
3394 goto out;
3395
3396 if (mode & FALLOC_FL_ZERO_RANGE) {
3397 ret = btrfs_zero_range(inode, offset, len, mode);
3398 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3399 return ret;
3400 }
3401
3402 locked_end = alloc_end - 1;
3403 while (1) {
3404 struct btrfs_ordered_extent *ordered;
3405
3406 /* the extent lock is ordered inside the running
3407 * transaction
3408 */
3409 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3410 locked_end, &cached_state);
3411 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3412 locked_end);
3413
3414 if (ordered &&
3415 ordered->file_offset + ordered->num_bytes > alloc_start &&
3416 ordered->file_offset < alloc_end) {
3417 btrfs_put_ordered_extent(ordered);
3418 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3419 alloc_start, locked_end,
3420 &cached_state);
3421 /*
3422 * we can't wait on the range with the transaction
3423 * running or with the extent lock held
3424 */
3425 ret = btrfs_wait_ordered_range(inode, alloc_start,
3426 alloc_end - alloc_start);
3427 if (ret)
3428 goto out;
3429 } else {
3430 if (ordered)
3431 btrfs_put_ordered_extent(ordered);
3432 break;
3433 }
3434 }
3435
3436 /* First, check if we exceed the qgroup limit */
3437 INIT_LIST_HEAD(&reserve_list);
3438 while (cur_offset < alloc_end) {
3439 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3440 alloc_end - cur_offset);
3441 if (IS_ERR(em)) {
3442 ret = PTR_ERR(em);
3443 break;
3444 }
3445 last_byte = min(extent_map_end(em), alloc_end);
3446 actual_end = min_t(u64, extent_map_end(em), offset + len);
3447 last_byte = ALIGN(last_byte, blocksize);
3448 if (em->block_start == EXTENT_MAP_HOLE ||
3449 (cur_offset >= inode->i_size &&
3450 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3451 ret = add_falloc_range(&reserve_list, cur_offset,
3452 last_byte - cur_offset);
3453 if (ret < 0) {
3454 free_extent_map(em);
3455 break;
3456 }
3457 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3458 &data_reserved, cur_offset,
3459 last_byte - cur_offset);
3460 if (ret < 0) {
3461 cur_offset = last_byte;
3462 free_extent_map(em);
3463 break;
3464 }
3465 } else {
3466 /*
3467 * Do not need to reserve unwritten extent for this
3468 * range, free reserved data space first, otherwise
3469 * it'll result in false ENOSPC error.
3470 */
3471 btrfs_free_reserved_data_space(BTRFS_I(inode),
3472 data_reserved, cur_offset,
3473 last_byte - cur_offset);
3474 }
3475 free_extent_map(em);
3476 cur_offset = last_byte;
3477 }
3478
3479 /*
3480 * If ret is still 0, means we're OK to fallocate.
3481 * Or just cleanup the list and exit.
3482 */
3483 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3484 if (!ret)
3485 ret = btrfs_prealloc_file_range(inode, mode,
3486 range->start,
3487 range->len, i_blocksize(inode),
3488 offset + len, &alloc_hint);
3489 else
3490 btrfs_free_reserved_data_space(BTRFS_I(inode),
3491 data_reserved, range->start,
3492 range->len);
3493 list_del(&range->list);
3494 kfree(range);
3495 }
3496 if (ret < 0)
3497 goto out_unlock;
3498
3499 /*
3500 * We didn't need to allocate any more space, but we still extended the
3501 * size of the file so we need to update i_size and the inode item.
3502 */
3503 ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3504out_unlock:
3505 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3506 &cached_state);
3507out:
3508 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3509 /* Let go of our reservation. */
3510 if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3511 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3512 cur_offset, alloc_end - cur_offset);
3513 extent_changeset_free(data_reserved);
3514 return ret;
3515}
3516
3517static loff_t find_desired_extent(struct btrfs_inode *inode, loff_t offset,
3518 int whence)
3519{
3520 struct btrfs_fs_info *fs_info = inode->root->fs_info;
3521 struct extent_map *em = NULL;
3522 struct extent_state *cached_state = NULL;
3523 loff_t i_size = inode->vfs_inode.i_size;
3524 u64 lockstart;
3525 u64 lockend;
3526 u64 start;
3527 u64 len;
3528 int ret = 0;
3529
3530 if (i_size == 0 || offset >= i_size)
3531 return -ENXIO;
3532
3533 /*
3534 * offset can be negative, in this case we start finding DATA/HOLE from
3535 * the very start of the file.
3536 */
3537 start = max_t(loff_t, 0, offset);
3538
3539 lockstart = round_down(start, fs_info->sectorsize);
3540 lockend = round_up(i_size, fs_info->sectorsize);
3541 if (lockend <= lockstart)
3542 lockend = lockstart + fs_info->sectorsize;
3543 lockend--;
3544 len = lockend - lockstart + 1;
3545
3546 lock_extent_bits(&inode->io_tree, lockstart, lockend, &cached_state);
3547
3548 while (start < i_size) {
3549 em = btrfs_get_extent_fiemap(inode, start, len);
3550 if (IS_ERR(em)) {
3551 ret = PTR_ERR(em);
3552 em = NULL;
3553 break;
3554 }
3555
3556 if (whence == SEEK_HOLE &&
3557 (em->block_start == EXTENT_MAP_HOLE ||
3558 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3559 break;
3560 else if (whence == SEEK_DATA &&
3561 (em->block_start != EXTENT_MAP_HOLE &&
3562 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3563 break;
3564
3565 start = em->start + em->len;
3566 free_extent_map(em);
3567 em = NULL;
3568 cond_resched();
3569 }
3570 free_extent_map(em);
3571 unlock_extent_cached(&inode->io_tree, lockstart, lockend,
3572 &cached_state);
3573 if (ret) {
3574 offset = ret;
3575 } else {
3576 if (whence == SEEK_DATA && start >= i_size)
3577 offset = -ENXIO;
3578 else
3579 offset = min_t(loff_t, start, i_size);
3580 }
3581
3582 return offset;
3583}
3584
3585static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3586{
3587 struct inode *inode = file->f_mapping->host;
3588
3589 switch (whence) {
3590 default:
3591 return generic_file_llseek(file, offset, whence);
3592 case SEEK_DATA:
3593 case SEEK_HOLE:
3594 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3595 offset = find_desired_extent(BTRFS_I(inode), offset, whence);
3596 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3597 break;
3598 }
3599
3600 if (offset < 0)
3601 return offset;
3602
3603 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3604}
3605
3606static int btrfs_file_open(struct inode *inode, struct file *filp)
3607{
3608 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3609 return generic_file_open(inode, filp);
3610}
3611
3612static int check_direct_read(struct btrfs_fs_info *fs_info,
3613 const struct iov_iter *iter, loff_t offset)
3614{
3615 int ret;
3616 int i, seg;
3617
3618 ret = check_direct_IO(fs_info, iter, offset);
3619 if (ret < 0)
3620 return ret;
3621
3622 if (!iter_is_iovec(iter))
3623 return 0;
3624
3625 for (seg = 0; seg < iter->nr_segs; seg++)
3626 for (i = seg + 1; i < iter->nr_segs; i++)
3627 if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3628 return -EINVAL;
3629 return 0;
3630}
3631
3632static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3633{
3634 struct inode *inode = file_inode(iocb->ki_filp);
3635 ssize_t ret;
3636
3637 if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3638 return 0;
3639
3640 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3641 ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops, 0);
3642 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3643 return ret;
3644}
3645
3646static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3647{
3648 ssize_t ret = 0;
3649
3650 if (iocb->ki_flags & IOCB_DIRECT) {
3651 ret = btrfs_direct_read(iocb, to);
3652 if (ret < 0 || !iov_iter_count(to) ||
3653 iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3654 return ret;
3655 }
3656
3657 return filemap_read(iocb, to, ret);
3658}
3659
3660const struct file_operations btrfs_file_operations = {
3661 .llseek = btrfs_file_llseek,
3662 .read_iter = btrfs_file_read_iter,
3663 .splice_read = generic_file_splice_read,
3664 .write_iter = btrfs_file_write_iter,
3665 .splice_write = iter_file_splice_write,
3666 .mmap = btrfs_file_mmap,
3667 .open = btrfs_file_open,
3668 .release = btrfs_release_file,
3669 .fsync = btrfs_sync_file,
3670 .fallocate = btrfs_fallocate,
3671 .unlocked_ioctl = btrfs_ioctl,
3672#ifdef CONFIG_COMPAT
3673 .compat_ioctl = btrfs_compat_ioctl,
3674#endif
3675 .remap_file_range = btrfs_remap_file_range,
3676};
3677
3678void __cold btrfs_auto_defrag_exit(void)
3679{
3680 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3681}
3682
3683int __init btrfs_auto_defrag_init(void)
3684{
3685 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3686 sizeof(struct inode_defrag), 0,
3687 SLAB_MEM_SPREAD,
3688 NULL);
3689 if (!btrfs_inode_defrag_cachep)
3690 return -ENOMEM;
3691
3692 return 0;
3693}
3694
3695int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3696{
3697 int ret;
3698
3699 /*
3700 * So with compression we will find and lock a dirty page and clear the
3701 * first one as dirty, setup an async extent, and immediately return
3702 * with the entire range locked but with nobody actually marked with
3703 * writeback. So we can't just filemap_write_and_wait_range() and
3704 * expect it to work since it will just kick off a thread to do the
3705 * actual work. So we need to call filemap_fdatawrite_range _again_
3706 * since it will wait on the page lock, which won't be unlocked until
3707 * after the pages have been marked as writeback and so we're good to go
3708 * from there. We have to do this otherwise we'll miss the ordered
3709 * extents and that results in badness. Please Josef, do not think you
3710 * know better and pull this out at some point in the future, it is
3711 * right and you are wrong.
3712 */
3713 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3714 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3715 &BTRFS_I(inode)->runtime_flags))
3716 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3717
3718 return ret;
3719}
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/fs.h>
20#include <linux/pagemap.h>
21#include <linux/highmem.h>
22#include <linux/time.h>
23#include <linux/init.h>
24#include <linux/string.h>
25#include <linux/backing-dev.h>
26#include <linux/mpage.h>
27#include <linux/falloc.h>
28#include <linux/swap.h>
29#include <linux/writeback.h>
30#include <linux/compat.h>
31#include <linux/slab.h>
32#include <linux/btrfs.h>
33#include <linux/uio.h>
34#include "ctree.h"
35#include "disk-io.h"
36#include "transaction.h"
37#include "btrfs_inode.h"
38#include "print-tree.h"
39#include "tree-log.h"
40#include "locking.h"
41#include "volumes.h"
42#include "qgroup.h"
43#include "compression.h"
44
45static struct kmem_cache *btrfs_inode_defrag_cachep;
46/*
47 * when auto defrag is enabled we
48 * queue up these defrag structs to remember which
49 * inodes need defragging passes
50 */
51struct inode_defrag {
52 struct rb_node rb_node;
53 /* objectid */
54 u64 ino;
55 /*
56 * transid where the defrag was added, we search for
57 * extents newer than this
58 */
59 u64 transid;
60
61 /* root objectid */
62 u64 root;
63
64 /* last offset we were able to defrag */
65 u64 last_offset;
66
67 /* if we've wrapped around back to zero once already */
68 int cycled;
69};
70
71static int __compare_inode_defrag(struct inode_defrag *defrag1,
72 struct inode_defrag *defrag2)
73{
74 if (defrag1->root > defrag2->root)
75 return 1;
76 else if (defrag1->root < defrag2->root)
77 return -1;
78 else if (defrag1->ino > defrag2->ino)
79 return 1;
80 else if (defrag1->ino < defrag2->ino)
81 return -1;
82 else
83 return 0;
84}
85
86/* pop a record for an inode into the defrag tree. The lock
87 * must be held already
88 *
89 * If you're inserting a record for an older transid than an
90 * existing record, the transid already in the tree is lowered
91 *
92 * If an existing record is found the defrag item you
93 * pass in is freed
94 */
95static int __btrfs_add_inode_defrag(struct inode *inode,
96 struct inode_defrag *defrag)
97{
98 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
99 struct inode_defrag *entry;
100 struct rb_node **p;
101 struct rb_node *parent = NULL;
102 int ret;
103
104 p = &fs_info->defrag_inodes.rb_node;
105 while (*p) {
106 parent = *p;
107 entry = rb_entry(parent, struct inode_defrag, rb_node);
108
109 ret = __compare_inode_defrag(defrag, entry);
110 if (ret < 0)
111 p = &parent->rb_left;
112 else if (ret > 0)
113 p = &parent->rb_right;
114 else {
115 /* if we're reinserting an entry for
116 * an old defrag run, make sure to
117 * lower the transid of our existing record
118 */
119 if (defrag->transid < entry->transid)
120 entry->transid = defrag->transid;
121 if (defrag->last_offset > entry->last_offset)
122 entry->last_offset = defrag->last_offset;
123 return -EEXIST;
124 }
125 }
126 set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
127 rb_link_node(&defrag->rb_node, parent, p);
128 rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
129 return 0;
130}
131
132static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
133{
134 if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
135 return 0;
136
137 if (btrfs_fs_closing(fs_info))
138 return 0;
139
140 return 1;
141}
142
143/*
144 * insert a defrag record for this inode if auto defrag is
145 * enabled
146 */
147int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
148 struct inode *inode)
149{
150 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
151 struct btrfs_root *root = BTRFS_I(inode)->root;
152 struct inode_defrag *defrag;
153 u64 transid;
154 int ret;
155
156 if (!__need_auto_defrag(fs_info))
157 return 0;
158
159 if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
160 return 0;
161
162 if (trans)
163 transid = trans->transid;
164 else
165 transid = BTRFS_I(inode)->root->last_trans;
166
167 defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
168 if (!defrag)
169 return -ENOMEM;
170
171 defrag->ino = btrfs_ino(inode);
172 defrag->transid = transid;
173 defrag->root = root->root_key.objectid;
174
175 spin_lock(&fs_info->defrag_inodes_lock);
176 if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
177 /*
178 * If we set IN_DEFRAG flag and evict the inode from memory,
179 * and then re-read this inode, this new inode doesn't have
180 * IN_DEFRAG flag. At the case, we may find the existed defrag.
181 */
182 ret = __btrfs_add_inode_defrag(inode, defrag);
183 if (ret)
184 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
185 } else {
186 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
187 }
188 spin_unlock(&fs_info->defrag_inodes_lock);
189 return 0;
190}
191
192/*
193 * Requeue the defrag object. If there is a defrag object that points to
194 * the same inode in the tree, we will merge them together (by
195 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
196 */
197static void btrfs_requeue_inode_defrag(struct inode *inode,
198 struct inode_defrag *defrag)
199{
200 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
201 int ret;
202
203 if (!__need_auto_defrag(fs_info))
204 goto out;
205
206 /*
207 * Here we don't check the IN_DEFRAG flag, because we need merge
208 * them together.
209 */
210 spin_lock(&fs_info->defrag_inodes_lock);
211 ret = __btrfs_add_inode_defrag(inode, defrag);
212 spin_unlock(&fs_info->defrag_inodes_lock);
213 if (ret)
214 goto out;
215 return;
216out:
217 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
218}
219
220/*
221 * pick the defragable inode that we want, if it doesn't exist, we will get
222 * the next one.
223 */
224static struct inode_defrag *
225btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
226{
227 struct inode_defrag *entry = NULL;
228 struct inode_defrag tmp;
229 struct rb_node *p;
230 struct rb_node *parent = NULL;
231 int ret;
232
233 tmp.ino = ino;
234 tmp.root = root;
235
236 spin_lock(&fs_info->defrag_inodes_lock);
237 p = fs_info->defrag_inodes.rb_node;
238 while (p) {
239 parent = p;
240 entry = rb_entry(parent, struct inode_defrag, rb_node);
241
242 ret = __compare_inode_defrag(&tmp, entry);
243 if (ret < 0)
244 p = parent->rb_left;
245 else if (ret > 0)
246 p = parent->rb_right;
247 else
248 goto out;
249 }
250
251 if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
252 parent = rb_next(parent);
253 if (parent)
254 entry = rb_entry(parent, struct inode_defrag, rb_node);
255 else
256 entry = NULL;
257 }
258out:
259 if (entry)
260 rb_erase(parent, &fs_info->defrag_inodes);
261 spin_unlock(&fs_info->defrag_inodes_lock);
262 return entry;
263}
264
265void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
266{
267 struct inode_defrag *defrag;
268 struct rb_node *node;
269
270 spin_lock(&fs_info->defrag_inodes_lock);
271 node = rb_first(&fs_info->defrag_inodes);
272 while (node) {
273 rb_erase(node, &fs_info->defrag_inodes);
274 defrag = rb_entry(node, struct inode_defrag, rb_node);
275 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
276
277 cond_resched_lock(&fs_info->defrag_inodes_lock);
278
279 node = rb_first(&fs_info->defrag_inodes);
280 }
281 spin_unlock(&fs_info->defrag_inodes_lock);
282}
283
284#define BTRFS_DEFRAG_BATCH 1024
285
286static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
287 struct inode_defrag *defrag)
288{
289 struct btrfs_root *inode_root;
290 struct inode *inode;
291 struct btrfs_key key;
292 struct btrfs_ioctl_defrag_range_args range;
293 int num_defrag;
294 int index;
295 int ret;
296
297 /* get the inode */
298 key.objectid = defrag->root;
299 key.type = BTRFS_ROOT_ITEM_KEY;
300 key.offset = (u64)-1;
301
302 index = srcu_read_lock(&fs_info->subvol_srcu);
303
304 inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
305 if (IS_ERR(inode_root)) {
306 ret = PTR_ERR(inode_root);
307 goto cleanup;
308 }
309
310 key.objectid = defrag->ino;
311 key.type = BTRFS_INODE_ITEM_KEY;
312 key.offset = 0;
313 inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
314 if (IS_ERR(inode)) {
315 ret = PTR_ERR(inode);
316 goto cleanup;
317 }
318 srcu_read_unlock(&fs_info->subvol_srcu, index);
319
320 /* do a chunk of defrag */
321 clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
322 memset(&range, 0, sizeof(range));
323 range.len = (u64)-1;
324 range.start = defrag->last_offset;
325
326 sb_start_write(fs_info->sb);
327 num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
328 BTRFS_DEFRAG_BATCH);
329 sb_end_write(fs_info->sb);
330 /*
331 * if we filled the whole defrag batch, there
332 * must be more work to do. Queue this defrag
333 * again
334 */
335 if (num_defrag == BTRFS_DEFRAG_BATCH) {
336 defrag->last_offset = range.start;
337 btrfs_requeue_inode_defrag(inode, defrag);
338 } else if (defrag->last_offset && !defrag->cycled) {
339 /*
340 * we didn't fill our defrag batch, but
341 * we didn't start at zero. Make sure we loop
342 * around to the start of the file.
343 */
344 defrag->last_offset = 0;
345 defrag->cycled = 1;
346 btrfs_requeue_inode_defrag(inode, defrag);
347 } else {
348 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
349 }
350
351 iput(inode);
352 return 0;
353cleanup:
354 srcu_read_unlock(&fs_info->subvol_srcu, index);
355 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
356 return ret;
357}
358
359/*
360 * run through the list of inodes in the FS that need
361 * defragging
362 */
363int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
364{
365 struct inode_defrag *defrag;
366 u64 first_ino = 0;
367 u64 root_objectid = 0;
368
369 atomic_inc(&fs_info->defrag_running);
370 while (1) {
371 /* Pause the auto defragger. */
372 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
373 &fs_info->fs_state))
374 break;
375
376 if (!__need_auto_defrag(fs_info))
377 break;
378
379 /* find an inode to defrag */
380 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
381 first_ino);
382 if (!defrag) {
383 if (root_objectid || first_ino) {
384 root_objectid = 0;
385 first_ino = 0;
386 continue;
387 } else {
388 break;
389 }
390 }
391
392 first_ino = defrag->ino + 1;
393 root_objectid = defrag->root;
394
395 __btrfs_run_defrag_inode(fs_info, defrag);
396 }
397 atomic_dec(&fs_info->defrag_running);
398
399 /*
400 * during unmount, we use the transaction_wait queue to
401 * wait for the defragger to stop
402 */
403 wake_up(&fs_info->transaction_wait);
404 return 0;
405}
406
407/* simple helper to fault in pages and copy. This should go away
408 * and be replaced with calls into generic code.
409 */
410static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
411 struct page **prepared_pages,
412 struct iov_iter *i)
413{
414 size_t copied = 0;
415 size_t total_copied = 0;
416 int pg = 0;
417 int offset = pos & (PAGE_SIZE - 1);
418
419 while (write_bytes > 0) {
420 size_t count = min_t(size_t,
421 PAGE_SIZE - offset, write_bytes);
422 struct page *page = prepared_pages[pg];
423 /*
424 * Copy data from userspace to the current page
425 */
426 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
427
428 /* Flush processor's dcache for this page */
429 flush_dcache_page(page);
430
431 /*
432 * if we get a partial write, we can end up with
433 * partially up to date pages. These add
434 * a lot of complexity, so make sure they don't
435 * happen by forcing this copy to be retried.
436 *
437 * The rest of the btrfs_file_write code will fall
438 * back to page at a time copies after we return 0.
439 */
440 if (!PageUptodate(page) && copied < count)
441 copied = 0;
442
443 iov_iter_advance(i, copied);
444 write_bytes -= copied;
445 total_copied += copied;
446
447 /* Return to btrfs_file_write_iter to fault page */
448 if (unlikely(copied == 0))
449 break;
450
451 if (copied < PAGE_SIZE - offset) {
452 offset += copied;
453 } else {
454 pg++;
455 offset = 0;
456 }
457 }
458 return total_copied;
459}
460
461/*
462 * unlocks pages after btrfs_file_write is done with them
463 */
464static void btrfs_drop_pages(struct page **pages, size_t num_pages)
465{
466 size_t i;
467 for (i = 0; i < num_pages; i++) {
468 /* page checked is some magic around finding pages that
469 * have been modified without going through btrfs_set_page_dirty
470 * clear it here. There should be no need to mark the pages
471 * accessed as prepare_pages should have marked them accessed
472 * in prepare_pages via find_or_create_page()
473 */
474 ClearPageChecked(pages[i]);
475 unlock_page(pages[i]);
476 put_page(pages[i]);
477 }
478}
479
480/*
481 * after copy_from_user, pages need to be dirtied and we need to make
482 * sure holes are created between the current EOF and the start of
483 * any next extents (if required).
484 *
485 * this also makes the decision about creating an inline extent vs
486 * doing real data extents, marking pages dirty and delalloc as required.
487 */
488int btrfs_dirty_pages(struct inode *inode, struct page **pages,
489 size_t num_pages, loff_t pos, size_t write_bytes,
490 struct extent_state **cached)
491{
492 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
493 int err = 0;
494 int i;
495 u64 num_bytes;
496 u64 start_pos;
497 u64 end_of_last_block;
498 u64 end_pos = pos + write_bytes;
499 loff_t isize = i_size_read(inode);
500
501 start_pos = pos & ~((u64) fs_info->sectorsize - 1);
502 num_bytes = round_up(write_bytes + pos - start_pos,
503 fs_info->sectorsize);
504
505 end_of_last_block = start_pos + num_bytes - 1;
506 err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
507 cached, 0);
508 if (err)
509 return err;
510
511 for (i = 0; i < num_pages; i++) {
512 struct page *p = pages[i];
513 SetPageUptodate(p);
514 ClearPageChecked(p);
515 set_page_dirty(p);
516 }
517
518 /*
519 * we've only changed i_size in ram, and we haven't updated
520 * the disk i_size. There is no need to log the inode
521 * at this time.
522 */
523 if (end_pos > isize)
524 i_size_write(inode, end_pos);
525 return 0;
526}
527
528/*
529 * this drops all the extents in the cache that intersect the range
530 * [start, end]. Existing extents are split as required.
531 */
532void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
533 int skip_pinned)
534{
535 struct extent_map *em;
536 struct extent_map *split = NULL;
537 struct extent_map *split2 = NULL;
538 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
539 u64 len = end - start + 1;
540 u64 gen;
541 int ret;
542 int testend = 1;
543 unsigned long flags;
544 int compressed = 0;
545 bool modified;
546
547 WARN_ON(end < start);
548 if (end == (u64)-1) {
549 len = (u64)-1;
550 testend = 0;
551 }
552 while (1) {
553 int no_splits = 0;
554
555 modified = false;
556 if (!split)
557 split = alloc_extent_map();
558 if (!split2)
559 split2 = alloc_extent_map();
560 if (!split || !split2)
561 no_splits = 1;
562
563 write_lock(&em_tree->lock);
564 em = lookup_extent_mapping(em_tree, start, len);
565 if (!em) {
566 write_unlock(&em_tree->lock);
567 break;
568 }
569 flags = em->flags;
570 gen = em->generation;
571 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
572 if (testend && em->start + em->len >= start + len) {
573 free_extent_map(em);
574 write_unlock(&em_tree->lock);
575 break;
576 }
577 start = em->start + em->len;
578 if (testend)
579 len = start + len - (em->start + em->len);
580 free_extent_map(em);
581 write_unlock(&em_tree->lock);
582 continue;
583 }
584 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
585 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
586 clear_bit(EXTENT_FLAG_LOGGING, &flags);
587 modified = !list_empty(&em->list);
588 if (no_splits)
589 goto next;
590
591 if (em->start < start) {
592 split->start = em->start;
593 split->len = start - em->start;
594
595 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
596 split->orig_start = em->orig_start;
597 split->block_start = em->block_start;
598
599 if (compressed)
600 split->block_len = em->block_len;
601 else
602 split->block_len = split->len;
603 split->orig_block_len = max(split->block_len,
604 em->orig_block_len);
605 split->ram_bytes = em->ram_bytes;
606 } else {
607 split->orig_start = split->start;
608 split->block_len = 0;
609 split->block_start = em->block_start;
610 split->orig_block_len = 0;
611 split->ram_bytes = split->len;
612 }
613
614 split->generation = gen;
615 split->bdev = em->bdev;
616 split->flags = flags;
617 split->compress_type = em->compress_type;
618 replace_extent_mapping(em_tree, em, split, modified);
619 free_extent_map(split);
620 split = split2;
621 split2 = NULL;
622 }
623 if (testend && em->start + em->len > start + len) {
624 u64 diff = start + len - em->start;
625
626 split->start = start + len;
627 split->len = em->start + em->len - (start + len);
628 split->bdev = em->bdev;
629 split->flags = flags;
630 split->compress_type = em->compress_type;
631 split->generation = gen;
632
633 if (em->block_start < EXTENT_MAP_LAST_BYTE) {
634 split->orig_block_len = max(em->block_len,
635 em->orig_block_len);
636
637 split->ram_bytes = em->ram_bytes;
638 if (compressed) {
639 split->block_len = em->block_len;
640 split->block_start = em->block_start;
641 split->orig_start = em->orig_start;
642 } else {
643 split->block_len = split->len;
644 split->block_start = em->block_start
645 + diff;
646 split->orig_start = em->orig_start;
647 }
648 } else {
649 split->ram_bytes = split->len;
650 split->orig_start = split->start;
651 split->block_len = 0;
652 split->block_start = em->block_start;
653 split->orig_block_len = 0;
654 }
655
656 if (extent_map_in_tree(em)) {
657 replace_extent_mapping(em_tree, em, split,
658 modified);
659 } else {
660 ret = add_extent_mapping(em_tree, split,
661 modified);
662 ASSERT(ret == 0); /* Logic error */
663 }
664 free_extent_map(split);
665 split = NULL;
666 }
667next:
668 if (extent_map_in_tree(em))
669 remove_extent_mapping(em_tree, em);
670 write_unlock(&em_tree->lock);
671
672 /* once for us */
673 free_extent_map(em);
674 /* once for the tree*/
675 free_extent_map(em);
676 }
677 if (split)
678 free_extent_map(split);
679 if (split2)
680 free_extent_map(split2);
681}
682
683/*
684 * this is very complex, but the basic idea is to drop all extents
685 * in the range start - end. hint_block is filled in with a block number
686 * that would be a good hint to the block allocator for this file.
687 *
688 * If an extent intersects the range but is not entirely inside the range
689 * it is either truncated or split. Anything entirely inside the range
690 * is deleted from the tree.
691 */
692int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
693 struct btrfs_root *root, struct inode *inode,
694 struct btrfs_path *path, u64 start, u64 end,
695 u64 *drop_end, int drop_cache,
696 int replace_extent,
697 u32 extent_item_size,
698 int *key_inserted)
699{
700 struct btrfs_fs_info *fs_info = root->fs_info;
701 struct extent_buffer *leaf;
702 struct btrfs_file_extent_item *fi;
703 struct btrfs_key key;
704 struct btrfs_key new_key;
705 u64 ino = btrfs_ino(inode);
706 u64 search_start = start;
707 u64 disk_bytenr = 0;
708 u64 num_bytes = 0;
709 u64 extent_offset = 0;
710 u64 extent_end = 0;
711 u64 last_end = start;
712 int del_nr = 0;
713 int del_slot = 0;
714 int extent_type;
715 int recow;
716 int ret;
717 int modify_tree = -1;
718 int update_refs;
719 int found = 0;
720 int leafs_visited = 0;
721
722 if (drop_cache)
723 btrfs_drop_extent_cache(inode, start, end - 1, 0);
724
725 if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
726 modify_tree = 0;
727
728 update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
729 root == fs_info->tree_root);
730 while (1) {
731 recow = 0;
732 ret = btrfs_lookup_file_extent(trans, root, path, ino,
733 search_start, modify_tree);
734 if (ret < 0)
735 break;
736 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
737 leaf = path->nodes[0];
738 btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
739 if (key.objectid == ino &&
740 key.type == BTRFS_EXTENT_DATA_KEY)
741 path->slots[0]--;
742 }
743 ret = 0;
744 leafs_visited++;
745next_slot:
746 leaf = path->nodes[0];
747 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
748 BUG_ON(del_nr > 0);
749 ret = btrfs_next_leaf(root, path);
750 if (ret < 0)
751 break;
752 if (ret > 0) {
753 ret = 0;
754 break;
755 }
756 leafs_visited++;
757 leaf = path->nodes[0];
758 recow = 1;
759 }
760
761 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
762
763 if (key.objectid > ino)
764 break;
765 if (WARN_ON_ONCE(key.objectid < ino) ||
766 key.type < BTRFS_EXTENT_DATA_KEY) {
767 ASSERT(del_nr == 0);
768 path->slots[0]++;
769 goto next_slot;
770 }
771 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
772 break;
773
774 fi = btrfs_item_ptr(leaf, path->slots[0],
775 struct btrfs_file_extent_item);
776 extent_type = btrfs_file_extent_type(leaf, fi);
777
778 if (extent_type == BTRFS_FILE_EXTENT_REG ||
779 extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
780 disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
781 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
782 extent_offset = btrfs_file_extent_offset(leaf, fi);
783 extent_end = key.offset +
784 btrfs_file_extent_num_bytes(leaf, fi);
785 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
786 extent_end = key.offset +
787 btrfs_file_extent_inline_len(leaf,
788 path->slots[0], fi);
789 } else {
790 /* can't happen */
791 BUG();
792 }
793
794 /*
795 * Don't skip extent items representing 0 byte lengths. They
796 * used to be created (bug) if while punching holes we hit
797 * -ENOSPC condition. So if we find one here, just ensure we
798 * delete it, otherwise we would insert a new file extent item
799 * with the same key (offset) as that 0 bytes length file
800 * extent item in the call to setup_items_for_insert() later
801 * in this function.
802 */
803 if (extent_end == key.offset && extent_end >= search_start) {
804 last_end = extent_end;
805 goto delete_extent_item;
806 }
807
808 if (extent_end <= search_start) {
809 path->slots[0]++;
810 goto next_slot;
811 }
812
813 found = 1;
814 search_start = max(key.offset, start);
815 if (recow || !modify_tree) {
816 modify_tree = -1;
817 btrfs_release_path(path);
818 continue;
819 }
820
821 /*
822 * | - range to drop - |
823 * | -------- extent -------- |
824 */
825 if (start > key.offset && end < extent_end) {
826 BUG_ON(del_nr > 0);
827 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
828 ret = -EOPNOTSUPP;
829 break;
830 }
831
832 memcpy(&new_key, &key, sizeof(new_key));
833 new_key.offset = start;
834 ret = btrfs_duplicate_item(trans, root, path,
835 &new_key);
836 if (ret == -EAGAIN) {
837 btrfs_release_path(path);
838 continue;
839 }
840 if (ret < 0)
841 break;
842
843 leaf = path->nodes[0];
844 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
845 struct btrfs_file_extent_item);
846 btrfs_set_file_extent_num_bytes(leaf, fi,
847 start - key.offset);
848
849 fi = btrfs_item_ptr(leaf, path->slots[0],
850 struct btrfs_file_extent_item);
851
852 extent_offset += start - key.offset;
853 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
854 btrfs_set_file_extent_num_bytes(leaf, fi,
855 extent_end - start);
856 btrfs_mark_buffer_dirty(leaf);
857
858 if (update_refs && disk_bytenr > 0) {
859 ret = btrfs_inc_extent_ref(trans, fs_info,
860 disk_bytenr, num_bytes, 0,
861 root->root_key.objectid,
862 new_key.objectid,
863 start - extent_offset);
864 BUG_ON(ret); /* -ENOMEM */
865 }
866 key.offset = start;
867 }
868 /*
869 * From here on out we will have actually dropped something, so
870 * last_end can be updated.
871 */
872 last_end = extent_end;
873
874 /*
875 * | ---- range to drop ----- |
876 * | -------- extent -------- |
877 */
878 if (start <= key.offset && end < extent_end) {
879 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
880 ret = -EOPNOTSUPP;
881 break;
882 }
883
884 memcpy(&new_key, &key, sizeof(new_key));
885 new_key.offset = end;
886 btrfs_set_item_key_safe(fs_info, path, &new_key);
887
888 extent_offset += end - key.offset;
889 btrfs_set_file_extent_offset(leaf, fi, extent_offset);
890 btrfs_set_file_extent_num_bytes(leaf, fi,
891 extent_end - end);
892 btrfs_mark_buffer_dirty(leaf);
893 if (update_refs && disk_bytenr > 0)
894 inode_sub_bytes(inode, end - key.offset);
895 break;
896 }
897
898 search_start = extent_end;
899 /*
900 * | ---- range to drop ----- |
901 * | -------- extent -------- |
902 */
903 if (start > key.offset && end >= extent_end) {
904 BUG_ON(del_nr > 0);
905 if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
906 ret = -EOPNOTSUPP;
907 break;
908 }
909
910 btrfs_set_file_extent_num_bytes(leaf, fi,
911 start - key.offset);
912 btrfs_mark_buffer_dirty(leaf);
913 if (update_refs && disk_bytenr > 0)
914 inode_sub_bytes(inode, extent_end - start);
915 if (end == extent_end)
916 break;
917
918 path->slots[0]++;
919 goto next_slot;
920 }
921
922 /*
923 * | ---- range to drop ----- |
924 * | ------ extent ------ |
925 */
926 if (start <= key.offset && end >= extent_end) {
927delete_extent_item:
928 if (del_nr == 0) {
929 del_slot = path->slots[0];
930 del_nr = 1;
931 } else {
932 BUG_ON(del_slot + del_nr != path->slots[0]);
933 del_nr++;
934 }
935
936 if (update_refs &&
937 extent_type == BTRFS_FILE_EXTENT_INLINE) {
938 inode_sub_bytes(inode,
939 extent_end - key.offset);
940 extent_end = ALIGN(extent_end,
941 fs_info->sectorsize);
942 } else if (update_refs && disk_bytenr > 0) {
943 ret = btrfs_free_extent(trans, fs_info,
944 disk_bytenr, num_bytes, 0,
945 root->root_key.objectid,
946 key.objectid, key.offset -
947 extent_offset);
948 BUG_ON(ret); /* -ENOMEM */
949 inode_sub_bytes(inode,
950 extent_end - key.offset);
951 }
952
953 if (end == extent_end)
954 break;
955
956 if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
957 path->slots[0]++;
958 goto next_slot;
959 }
960
961 ret = btrfs_del_items(trans, root, path, del_slot,
962 del_nr);
963 if (ret) {
964 btrfs_abort_transaction(trans, ret);
965 break;
966 }
967
968 del_nr = 0;
969 del_slot = 0;
970
971 btrfs_release_path(path);
972 continue;
973 }
974
975 BUG_ON(1);
976 }
977
978 if (!ret && del_nr > 0) {
979 /*
980 * Set path->slots[0] to first slot, so that after the delete
981 * if items are move off from our leaf to its immediate left or
982 * right neighbor leafs, we end up with a correct and adjusted
983 * path->slots[0] for our insertion (if replace_extent != 0).
984 */
985 path->slots[0] = del_slot;
986 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
987 if (ret)
988 btrfs_abort_transaction(trans, ret);
989 }
990
991 leaf = path->nodes[0];
992 /*
993 * If btrfs_del_items() was called, it might have deleted a leaf, in
994 * which case it unlocked our path, so check path->locks[0] matches a
995 * write lock.
996 */
997 if (!ret && replace_extent && leafs_visited == 1 &&
998 (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
999 path->locks[0] == BTRFS_WRITE_LOCK) &&
1000 btrfs_leaf_free_space(fs_info, leaf) >=
1001 sizeof(struct btrfs_item) + extent_item_size) {
1002
1003 key.objectid = ino;
1004 key.type = BTRFS_EXTENT_DATA_KEY;
1005 key.offset = start;
1006 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1007 struct btrfs_key slot_key;
1008
1009 btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1010 if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1011 path->slots[0]++;
1012 }
1013 setup_items_for_insert(root, path, &key,
1014 &extent_item_size,
1015 extent_item_size,
1016 sizeof(struct btrfs_item) +
1017 extent_item_size, 1);
1018 *key_inserted = 1;
1019 }
1020
1021 if (!replace_extent || !(*key_inserted))
1022 btrfs_release_path(path);
1023 if (drop_end)
1024 *drop_end = found ? min(end, last_end) : end;
1025 return ret;
1026}
1027
1028int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1029 struct btrfs_root *root, struct inode *inode, u64 start,
1030 u64 end, int drop_cache)
1031{
1032 struct btrfs_path *path;
1033 int ret;
1034
1035 path = btrfs_alloc_path();
1036 if (!path)
1037 return -ENOMEM;
1038 ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1039 drop_cache, 0, 0, NULL);
1040 btrfs_free_path(path);
1041 return ret;
1042}
1043
1044static int extent_mergeable(struct extent_buffer *leaf, int slot,
1045 u64 objectid, u64 bytenr, u64 orig_offset,
1046 u64 *start, u64 *end)
1047{
1048 struct btrfs_file_extent_item *fi;
1049 struct btrfs_key key;
1050 u64 extent_end;
1051
1052 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1053 return 0;
1054
1055 btrfs_item_key_to_cpu(leaf, &key, slot);
1056 if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1057 return 0;
1058
1059 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1060 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1061 btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1062 btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1063 btrfs_file_extent_compression(leaf, fi) ||
1064 btrfs_file_extent_encryption(leaf, fi) ||
1065 btrfs_file_extent_other_encoding(leaf, fi))
1066 return 0;
1067
1068 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1069 if ((*start && *start != key.offset) || (*end && *end != extent_end))
1070 return 0;
1071
1072 *start = key.offset;
1073 *end = extent_end;
1074 return 1;
1075}
1076
1077/*
1078 * Mark extent in the range start - end as written.
1079 *
1080 * This changes extent type from 'pre-allocated' to 'regular'. If only
1081 * part of extent is marked as written, the extent will be split into
1082 * two or three.
1083 */
1084int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1085 struct inode *inode, u64 start, u64 end)
1086{
1087 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1088 struct btrfs_root *root = BTRFS_I(inode)->root;
1089 struct extent_buffer *leaf;
1090 struct btrfs_path *path;
1091 struct btrfs_file_extent_item *fi;
1092 struct btrfs_key key;
1093 struct btrfs_key new_key;
1094 u64 bytenr;
1095 u64 num_bytes;
1096 u64 extent_end;
1097 u64 orig_offset;
1098 u64 other_start;
1099 u64 other_end;
1100 u64 split;
1101 int del_nr = 0;
1102 int del_slot = 0;
1103 int recow;
1104 int ret;
1105 u64 ino = btrfs_ino(inode);
1106
1107 path = btrfs_alloc_path();
1108 if (!path)
1109 return -ENOMEM;
1110again:
1111 recow = 0;
1112 split = start;
1113 key.objectid = ino;
1114 key.type = BTRFS_EXTENT_DATA_KEY;
1115 key.offset = split;
1116
1117 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1118 if (ret < 0)
1119 goto out;
1120 if (ret > 0 && path->slots[0] > 0)
1121 path->slots[0]--;
1122
1123 leaf = path->nodes[0];
1124 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1125 if (key.objectid != ino ||
1126 key.type != BTRFS_EXTENT_DATA_KEY) {
1127 ret = -EINVAL;
1128 btrfs_abort_transaction(trans, ret);
1129 goto out;
1130 }
1131 fi = btrfs_item_ptr(leaf, path->slots[0],
1132 struct btrfs_file_extent_item);
1133 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1134 ret = -EINVAL;
1135 btrfs_abort_transaction(trans, ret);
1136 goto out;
1137 }
1138 extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1139 if (key.offset > start || extent_end < end) {
1140 ret = -EINVAL;
1141 btrfs_abort_transaction(trans, ret);
1142 goto out;
1143 }
1144
1145 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1146 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1147 orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1148 memcpy(&new_key, &key, sizeof(new_key));
1149
1150 if (start == key.offset && end < extent_end) {
1151 other_start = 0;
1152 other_end = start;
1153 if (extent_mergeable(leaf, path->slots[0] - 1,
1154 ino, bytenr, orig_offset,
1155 &other_start, &other_end)) {
1156 new_key.offset = end;
1157 btrfs_set_item_key_safe(fs_info, path, &new_key);
1158 fi = btrfs_item_ptr(leaf, path->slots[0],
1159 struct btrfs_file_extent_item);
1160 btrfs_set_file_extent_generation(leaf, fi,
1161 trans->transid);
1162 btrfs_set_file_extent_num_bytes(leaf, fi,
1163 extent_end - end);
1164 btrfs_set_file_extent_offset(leaf, fi,
1165 end - orig_offset);
1166 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1167 struct btrfs_file_extent_item);
1168 btrfs_set_file_extent_generation(leaf, fi,
1169 trans->transid);
1170 btrfs_set_file_extent_num_bytes(leaf, fi,
1171 end - other_start);
1172 btrfs_mark_buffer_dirty(leaf);
1173 goto out;
1174 }
1175 }
1176
1177 if (start > key.offset && end == extent_end) {
1178 other_start = end;
1179 other_end = 0;
1180 if (extent_mergeable(leaf, path->slots[0] + 1,
1181 ino, bytenr, orig_offset,
1182 &other_start, &other_end)) {
1183 fi = btrfs_item_ptr(leaf, path->slots[0],
1184 struct btrfs_file_extent_item);
1185 btrfs_set_file_extent_num_bytes(leaf, fi,
1186 start - key.offset);
1187 btrfs_set_file_extent_generation(leaf, fi,
1188 trans->transid);
1189 path->slots[0]++;
1190 new_key.offset = start;
1191 btrfs_set_item_key_safe(fs_info, path, &new_key);
1192
1193 fi = btrfs_item_ptr(leaf, path->slots[0],
1194 struct btrfs_file_extent_item);
1195 btrfs_set_file_extent_generation(leaf, fi,
1196 trans->transid);
1197 btrfs_set_file_extent_num_bytes(leaf, fi,
1198 other_end - start);
1199 btrfs_set_file_extent_offset(leaf, fi,
1200 start - orig_offset);
1201 btrfs_mark_buffer_dirty(leaf);
1202 goto out;
1203 }
1204 }
1205
1206 while (start > key.offset || end < extent_end) {
1207 if (key.offset == start)
1208 split = end;
1209
1210 new_key.offset = split;
1211 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1212 if (ret == -EAGAIN) {
1213 btrfs_release_path(path);
1214 goto again;
1215 }
1216 if (ret < 0) {
1217 btrfs_abort_transaction(trans, ret);
1218 goto out;
1219 }
1220
1221 leaf = path->nodes[0];
1222 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1223 struct btrfs_file_extent_item);
1224 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1225 btrfs_set_file_extent_num_bytes(leaf, fi,
1226 split - key.offset);
1227
1228 fi = btrfs_item_ptr(leaf, path->slots[0],
1229 struct btrfs_file_extent_item);
1230
1231 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1232 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1233 btrfs_set_file_extent_num_bytes(leaf, fi,
1234 extent_end - split);
1235 btrfs_mark_buffer_dirty(leaf);
1236
1237 ret = btrfs_inc_extent_ref(trans, fs_info, bytenr, num_bytes,
1238 0, root->root_key.objectid,
1239 ino, orig_offset);
1240 if (ret) {
1241 btrfs_abort_transaction(trans, ret);
1242 goto out;
1243 }
1244
1245 if (split == start) {
1246 key.offset = start;
1247 } else {
1248 if (start != key.offset) {
1249 ret = -EINVAL;
1250 btrfs_abort_transaction(trans, ret);
1251 goto out;
1252 }
1253 path->slots[0]--;
1254 extent_end = end;
1255 }
1256 recow = 1;
1257 }
1258
1259 other_start = end;
1260 other_end = 0;
1261 if (extent_mergeable(leaf, path->slots[0] + 1,
1262 ino, bytenr, orig_offset,
1263 &other_start, &other_end)) {
1264 if (recow) {
1265 btrfs_release_path(path);
1266 goto again;
1267 }
1268 extent_end = other_end;
1269 del_slot = path->slots[0] + 1;
1270 del_nr++;
1271 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1272 0, root->root_key.objectid,
1273 ino, orig_offset);
1274 if (ret) {
1275 btrfs_abort_transaction(trans, ret);
1276 goto out;
1277 }
1278 }
1279 other_start = 0;
1280 other_end = start;
1281 if (extent_mergeable(leaf, path->slots[0] - 1,
1282 ino, bytenr, orig_offset,
1283 &other_start, &other_end)) {
1284 if (recow) {
1285 btrfs_release_path(path);
1286 goto again;
1287 }
1288 key.offset = other_start;
1289 del_slot = path->slots[0];
1290 del_nr++;
1291 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1292 0, root->root_key.objectid,
1293 ino, orig_offset);
1294 if (ret) {
1295 btrfs_abort_transaction(trans, ret);
1296 goto out;
1297 }
1298 }
1299 if (del_nr == 0) {
1300 fi = btrfs_item_ptr(leaf, path->slots[0],
1301 struct btrfs_file_extent_item);
1302 btrfs_set_file_extent_type(leaf, fi,
1303 BTRFS_FILE_EXTENT_REG);
1304 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1305 btrfs_mark_buffer_dirty(leaf);
1306 } else {
1307 fi = btrfs_item_ptr(leaf, del_slot - 1,
1308 struct btrfs_file_extent_item);
1309 btrfs_set_file_extent_type(leaf, fi,
1310 BTRFS_FILE_EXTENT_REG);
1311 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1312 btrfs_set_file_extent_num_bytes(leaf, fi,
1313 extent_end - key.offset);
1314 btrfs_mark_buffer_dirty(leaf);
1315
1316 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1317 if (ret < 0) {
1318 btrfs_abort_transaction(trans, ret);
1319 goto out;
1320 }
1321 }
1322out:
1323 btrfs_free_path(path);
1324 return 0;
1325}
1326
1327/*
1328 * on error we return an unlocked page and the error value
1329 * on success we return a locked page and 0
1330 */
1331static int prepare_uptodate_page(struct inode *inode,
1332 struct page *page, u64 pos,
1333 bool force_uptodate)
1334{
1335 int ret = 0;
1336
1337 if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1338 !PageUptodate(page)) {
1339 ret = btrfs_readpage(NULL, page);
1340 if (ret)
1341 return ret;
1342 lock_page(page);
1343 if (!PageUptodate(page)) {
1344 unlock_page(page);
1345 return -EIO;
1346 }
1347 if (page->mapping != inode->i_mapping) {
1348 unlock_page(page);
1349 return -EAGAIN;
1350 }
1351 }
1352 return 0;
1353}
1354
1355/*
1356 * this just gets pages into the page cache and locks them down.
1357 */
1358static noinline int prepare_pages(struct inode *inode, struct page **pages,
1359 size_t num_pages, loff_t pos,
1360 size_t write_bytes, bool force_uptodate)
1361{
1362 int i;
1363 unsigned long index = pos >> PAGE_SHIFT;
1364 gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1365 int err = 0;
1366 int faili;
1367
1368 for (i = 0; i < num_pages; i++) {
1369again:
1370 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1371 mask | __GFP_WRITE);
1372 if (!pages[i]) {
1373 faili = i - 1;
1374 err = -ENOMEM;
1375 goto fail;
1376 }
1377
1378 if (i == 0)
1379 err = prepare_uptodate_page(inode, pages[i], pos,
1380 force_uptodate);
1381 if (!err && i == num_pages - 1)
1382 err = prepare_uptodate_page(inode, pages[i],
1383 pos + write_bytes, false);
1384 if (err) {
1385 put_page(pages[i]);
1386 if (err == -EAGAIN) {
1387 err = 0;
1388 goto again;
1389 }
1390 faili = i - 1;
1391 goto fail;
1392 }
1393 wait_on_page_writeback(pages[i]);
1394 }
1395
1396 return 0;
1397fail:
1398 while (faili >= 0) {
1399 unlock_page(pages[faili]);
1400 put_page(pages[faili]);
1401 faili--;
1402 }
1403 return err;
1404
1405}
1406
1407/*
1408 * This function locks the extent and properly waits for data=ordered extents
1409 * to finish before allowing the pages to be modified if need.
1410 *
1411 * The return value:
1412 * 1 - the extent is locked
1413 * 0 - the extent is not locked, and everything is OK
1414 * -EAGAIN - need re-prepare the pages
1415 * the other < 0 number - Something wrong happens
1416 */
1417static noinline int
1418lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1419 size_t num_pages, loff_t pos,
1420 size_t write_bytes,
1421 u64 *lockstart, u64 *lockend,
1422 struct extent_state **cached_state)
1423{
1424 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1425 u64 start_pos;
1426 u64 last_pos;
1427 int i;
1428 int ret = 0;
1429
1430 start_pos = round_down(pos, fs_info->sectorsize);
1431 last_pos = start_pos
1432 + round_up(pos + write_bytes - start_pos,
1433 fs_info->sectorsize) - 1;
1434
1435 if (start_pos < inode->i_size) {
1436 struct btrfs_ordered_extent *ordered;
1437 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1438 start_pos, last_pos, cached_state);
1439 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1440 last_pos - start_pos + 1);
1441 if (ordered &&
1442 ordered->file_offset + ordered->len > start_pos &&
1443 ordered->file_offset <= last_pos) {
1444 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1445 start_pos, last_pos,
1446 cached_state, GFP_NOFS);
1447 for (i = 0; i < num_pages; i++) {
1448 unlock_page(pages[i]);
1449 put_page(pages[i]);
1450 }
1451 btrfs_start_ordered_extent(inode, ordered, 1);
1452 btrfs_put_ordered_extent(ordered);
1453 return -EAGAIN;
1454 }
1455 if (ordered)
1456 btrfs_put_ordered_extent(ordered);
1457
1458 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1459 last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1460 EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1461 0, 0, cached_state, GFP_NOFS);
1462 *lockstart = start_pos;
1463 *lockend = last_pos;
1464 ret = 1;
1465 }
1466
1467 for (i = 0; i < num_pages; i++) {
1468 if (clear_page_dirty_for_io(pages[i]))
1469 account_page_redirty(pages[i]);
1470 set_page_extent_mapped(pages[i]);
1471 WARN_ON(!PageLocked(pages[i]));
1472 }
1473
1474 return ret;
1475}
1476
1477static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1478 size_t *write_bytes)
1479{
1480 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1481 struct btrfs_root *root = BTRFS_I(inode)->root;
1482 struct btrfs_ordered_extent *ordered;
1483 u64 lockstart, lockend;
1484 u64 num_bytes;
1485 int ret;
1486
1487 ret = btrfs_start_write_no_snapshoting(root);
1488 if (!ret)
1489 return -ENOSPC;
1490
1491 lockstart = round_down(pos, fs_info->sectorsize);
1492 lockend = round_up(pos + *write_bytes,
1493 fs_info->sectorsize) - 1;
1494
1495 while (1) {
1496 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1497 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1498 lockend - lockstart + 1);
1499 if (!ordered) {
1500 break;
1501 }
1502 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1503 btrfs_start_ordered_extent(inode, ordered, 1);
1504 btrfs_put_ordered_extent(ordered);
1505 }
1506
1507 num_bytes = lockend - lockstart + 1;
1508 ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1509 if (ret <= 0) {
1510 ret = 0;
1511 btrfs_end_write_no_snapshoting(root);
1512 } else {
1513 *write_bytes = min_t(size_t, *write_bytes ,
1514 num_bytes - pos + lockstart);
1515 }
1516
1517 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1518
1519 return ret;
1520}
1521
1522static noinline ssize_t __btrfs_buffered_write(struct file *file,
1523 struct iov_iter *i,
1524 loff_t pos)
1525{
1526 struct inode *inode = file_inode(file);
1527 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1528 struct btrfs_root *root = BTRFS_I(inode)->root;
1529 struct page **pages = NULL;
1530 struct extent_state *cached_state = NULL;
1531 u64 release_bytes = 0;
1532 u64 lockstart;
1533 u64 lockend;
1534 size_t num_written = 0;
1535 int nrptrs;
1536 int ret = 0;
1537 bool only_release_metadata = false;
1538 bool force_page_uptodate = false;
1539 bool need_unlock;
1540
1541 nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1542 PAGE_SIZE / (sizeof(struct page *)));
1543 nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1544 nrptrs = max(nrptrs, 8);
1545 pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1546 if (!pages)
1547 return -ENOMEM;
1548
1549 while (iov_iter_count(i) > 0) {
1550 size_t offset = pos & (PAGE_SIZE - 1);
1551 size_t sector_offset;
1552 size_t write_bytes = min(iov_iter_count(i),
1553 nrptrs * (size_t)PAGE_SIZE -
1554 offset);
1555 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1556 PAGE_SIZE);
1557 size_t reserve_bytes;
1558 size_t dirty_pages;
1559 size_t copied;
1560 size_t dirty_sectors;
1561 size_t num_sectors;
1562
1563 WARN_ON(num_pages > nrptrs);
1564
1565 /*
1566 * Fault pages before locking them in prepare_pages
1567 * to avoid recursive lock
1568 */
1569 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1570 ret = -EFAULT;
1571 break;
1572 }
1573
1574 sector_offset = pos & (fs_info->sectorsize - 1);
1575 reserve_bytes = round_up(write_bytes + sector_offset,
1576 fs_info->sectorsize);
1577
1578 ret = btrfs_check_data_free_space(inode, pos, write_bytes);
1579 if (ret < 0) {
1580 if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1581 BTRFS_INODE_PREALLOC)) &&
1582 check_can_nocow(inode, pos, &write_bytes) > 0) {
1583 /*
1584 * For nodata cow case, no need to reserve
1585 * data space.
1586 */
1587 only_release_metadata = true;
1588 /*
1589 * our prealloc extent may be smaller than
1590 * write_bytes, so scale down.
1591 */
1592 num_pages = DIV_ROUND_UP(write_bytes + offset,
1593 PAGE_SIZE);
1594 reserve_bytes = round_up(write_bytes +
1595 sector_offset,
1596 fs_info->sectorsize);
1597 } else {
1598 break;
1599 }
1600 }
1601
1602 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1603 if (ret) {
1604 if (!only_release_metadata)
1605 btrfs_free_reserved_data_space(inode, pos,
1606 write_bytes);
1607 else
1608 btrfs_end_write_no_snapshoting(root);
1609 break;
1610 }
1611
1612 release_bytes = reserve_bytes;
1613 need_unlock = false;
1614again:
1615 /*
1616 * This is going to setup the pages array with the number of
1617 * pages we want, so we don't really need to worry about the
1618 * contents of pages from loop to loop
1619 */
1620 ret = prepare_pages(inode, pages, num_pages,
1621 pos, write_bytes,
1622 force_page_uptodate);
1623 if (ret)
1624 break;
1625
1626 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1627 pos, write_bytes, &lockstart,
1628 &lockend, &cached_state);
1629 if (ret < 0) {
1630 if (ret == -EAGAIN)
1631 goto again;
1632 break;
1633 } else if (ret > 0) {
1634 need_unlock = true;
1635 ret = 0;
1636 }
1637
1638 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1639
1640 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1641 dirty_sectors = round_up(copied + sector_offset,
1642 fs_info->sectorsize);
1643 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1644
1645 /*
1646 * if we have trouble faulting in the pages, fall
1647 * back to one page at a time
1648 */
1649 if (copied < write_bytes)
1650 nrptrs = 1;
1651
1652 if (copied == 0) {
1653 force_page_uptodate = true;
1654 dirty_sectors = 0;
1655 dirty_pages = 0;
1656 } else {
1657 force_page_uptodate = false;
1658 dirty_pages = DIV_ROUND_UP(copied + offset,
1659 PAGE_SIZE);
1660 }
1661
1662 /*
1663 * If we had a short copy we need to release the excess delaloc
1664 * bytes we reserved. We need to increment outstanding_extents
1665 * because btrfs_delalloc_release_space and
1666 * btrfs_delalloc_release_metadata will decrement it, but
1667 * we still have an outstanding extent for the chunk we actually
1668 * managed to copy.
1669 */
1670 if (num_sectors > dirty_sectors) {
1671 /* release everything except the sectors we dirtied */
1672 release_bytes -= dirty_sectors <<
1673 fs_info->sb->s_blocksize_bits;
1674 if (copied > 0) {
1675 spin_lock(&BTRFS_I(inode)->lock);
1676 BTRFS_I(inode)->outstanding_extents++;
1677 spin_unlock(&BTRFS_I(inode)->lock);
1678 }
1679 if (only_release_metadata) {
1680 btrfs_delalloc_release_metadata(inode,
1681 release_bytes);
1682 } else {
1683 u64 __pos;
1684
1685 __pos = round_down(pos,
1686 fs_info->sectorsize) +
1687 (dirty_pages << PAGE_SHIFT);
1688 btrfs_delalloc_release_space(inode, __pos,
1689 release_bytes);
1690 }
1691 }
1692
1693 release_bytes = round_up(copied + sector_offset,
1694 fs_info->sectorsize);
1695
1696 if (copied > 0)
1697 ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1698 pos, copied, NULL);
1699 if (need_unlock)
1700 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1701 lockstart, lockend, &cached_state,
1702 GFP_NOFS);
1703 if (ret) {
1704 btrfs_drop_pages(pages, num_pages);
1705 break;
1706 }
1707
1708 release_bytes = 0;
1709 if (only_release_metadata)
1710 btrfs_end_write_no_snapshoting(root);
1711
1712 if (only_release_metadata && copied > 0) {
1713 lockstart = round_down(pos,
1714 fs_info->sectorsize);
1715 lockend = round_up(pos + copied,
1716 fs_info->sectorsize) - 1;
1717
1718 set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1719 lockend, EXTENT_NORESERVE, NULL,
1720 NULL, GFP_NOFS);
1721 only_release_metadata = false;
1722 }
1723
1724 btrfs_drop_pages(pages, num_pages);
1725
1726 cond_resched();
1727
1728 balance_dirty_pages_ratelimited(inode->i_mapping);
1729 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1730 btrfs_btree_balance_dirty(fs_info);
1731
1732 pos += copied;
1733 num_written += copied;
1734 }
1735
1736 kfree(pages);
1737
1738 if (release_bytes) {
1739 if (only_release_metadata) {
1740 btrfs_end_write_no_snapshoting(root);
1741 btrfs_delalloc_release_metadata(inode, release_bytes);
1742 } else {
1743 btrfs_delalloc_release_space(inode,
1744 round_down(pos, fs_info->sectorsize),
1745 release_bytes);
1746 }
1747 }
1748
1749 return num_written ? num_written : ret;
1750}
1751
1752static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1753{
1754 struct file *file = iocb->ki_filp;
1755 struct inode *inode = file_inode(file);
1756 loff_t pos = iocb->ki_pos;
1757 ssize_t written;
1758 ssize_t written_buffered;
1759 loff_t endbyte;
1760 int err;
1761
1762 written = generic_file_direct_write(iocb, from);
1763
1764 if (written < 0 || !iov_iter_count(from))
1765 return written;
1766
1767 pos += written;
1768 written_buffered = __btrfs_buffered_write(file, from, pos);
1769 if (written_buffered < 0) {
1770 err = written_buffered;
1771 goto out;
1772 }
1773 /*
1774 * Ensure all data is persisted. We want the next direct IO read to be
1775 * able to read what was just written.
1776 */
1777 endbyte = pos + written_buffered - 1;
1778 err = btrfs_fdatawrite_range(inode, pos, endbyte);
1779 if (err)
1780 goto out;
1781 err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1782 if (err)
1783 goto out;
1784 written += written_buffered;
1785 iocb->ki_pos = pos + written_buffered;
1786 invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1787 endbyte >> PAGE_SHIFT);
1788out:
1789 return written ? written : err;
1790}
1791
1792static void update_time_for_write(struct inode *inode)
1793{
1794 struct timespec now;
1795
1796 if (IS_NOCMTIME(inode))
1797 return;
1798
1799 now = current_time(inode);
1800 if (!timespec_equal(&inode->i_mtime, &now))
1801 inode->i_mtime = now;
1802
1803 if (!timespec_equal(&inode->i_ctime, &now))
1804 inode->i_ctime = now;
1805
1806 if (IS_I_VERSION(inode))
1807 inode_inc_iversion(inode);
1808}
1809
1810static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1811 struct iov_iter *from)
1812{
1813 struct file *file = iocb->ki_filp;
1814 struct inode *inode = file_inode(file);
1815 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1816 struct btrfs_root *root = BTRFS_I(inode)->root;
1817 u64 start_pos;
1818 u64 end_pos;
1819 ssize_t num_written = 0;
1820 bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1821 ssize_t err;
1822 loff_t pos;
1823 size_t count;
1824 loff_t oldsize;
1825 int clean_page = 0;
1826
1827 inode_lock(inode);
1828 err = generic_write_checks(iocb, from);
1829 if (err <= 0) {
1830 inode_unlock(inode);
1831 return err;
1832 }
1833
1834 current->backing_dev_info = inode_to_bdi(inode);
1835 err = file_remove_privs(file);
1836 if (err) {
1837 inode_unlock(inode);
1838 goto out;
1839 }
1840
1841 /*
1842 * If BTRFS flips readonly due to some impossible error
1843 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1844 * although we have opened a file as writable, we have
1845 * to stop this write operation to ensure FS consistency.
1846 */
1847 if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1848 inode_unlock(inode);
1849 err = -EROFS;
1850 goto out;
1851 }
1852
1853 /*
1854 * We reserve space for updating the inode when we reserve space for the
1855 * extent we are going to write, so we will enospc out there. We don't
1856 * need to start yet another transaction to update the inode as we will
1857 * update the inode when we finish writing whatever data we write.
1858 */
1859 update_time_for_write(inode);
1860
1861 pos = iocb->ki_pos;
1862 count = iov_iter_count(from);
1863 start_pos = round_down(pos, fs_info->sectorsize);
1864 oldsize = i_size_read(inode);
1865 if (start_pos > oldsize) {
1866 /* Expand hole size to cover write data, preventing empty gap */
1867 end_pos = round_up(pos + count,
1868 fs_info->sectorsize);
1869 err = btrfs_cont_expand(inode, oldsize, end_pos);
1870 if (err) {
1871 inode_unlock(inode);
1872 goto out;
1873 }
1874 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1875 clean_page = 1;
1876 }
1877
1878 if (sync)
1879 atomic_inc(&BTRFS_I(inode)->sync_writers);
1880
1881 if (iocb->ki_flags & IOCB_DIRECT) {
1882 num_written = __btrfs_direct_write(iocb, from);
1883 } else {
1884 num_written = __btrfs_buffered_write(file, from, pos);
1885 if (num_written > 0)
1886 iocb->ki_pos = pos + num_written;
1887 if (clean_page)
1888 pagecache_isize_extended(inode, oldsize,
1889 i_size_read(inode));
1890 }
1891
1892 inode_unlock(inode);
1893
1894 /*
1895 * We also have to set last_sub_trans to the current log transid,
1896 * otherwise subsequent syncs to a file that's been synced in this
1897 * transaction will appear to have already occurred.
1898 */
1899 spin_lock(&BTRFS_I(inode)->lock);
1900 BTRFS_I(inode)->last_sub_trans = root->log_transid;
1901 spin_unlock(&BTRFS_I(inode)->lock);
1902 if (num_written > 0)
1903 num_written = generic_write_sync(iocb, num_written);
1904
1905 if (sync)
1906 atomic_dec(&BTRFS_I(inode)->sync_writers);
1907out:
1908 current->backing_dev_info = NULL;
1909 return num_written ? num_written : err;
1910}
1911
1912int btrfs_release_file(struct inode *inode, struct file *filp)
1913{
1914 if (filp->private_data)
1915 btrfs_ioctl_trans_end(filp);
1916 /*
1917 * ordered_data_close is set by settattr when we are about to truncate
1918 * a file from a non-zero size to a zero size. This tries to
1919 * flush down new bytes that may have been written if the
1920 * application were using truncate to replace a file in place.
1921 */
1922 if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1923 &BTRFS_I(inode)->runtime_flags))
1924 filemap_flush(inode->i_mapping);
1925 return 0;
1926}
1927
1928static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1929{
1930 int ret;
1931
1932 atomic_inc(&BTRFS_I(inode)->sync_writers);
1933 ret = btrfs_fdatawrite_range(inode, start, end);
1934 atomic_dec(&BTRFS_I(inode)->sync_writers);
1935
1936 return ret;
1937}
1938
1939/*
1940 * fsync call for both files and directories. This logs the inode into
1941 * the tree log instead of forcing full commits whenever possible.
1942 *
1943 * It needs to call filemap_fdatawait so that all ordered extent updates are
1944 * in the metadata btree are up to date for copying to the log.
1945 *
1946 * It drops the inode mutex before doing the tree log commit. This is an
1947 * important optimization for directories because holding the mutex prevents
1948 * new operations on the dir while we write to disk.
1949 */
1950int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1951{
1952 struct dentry *dentry = file_dentry(file);
1953 struct inode *inode = d_inode(dentry);
1954 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1955 struct btrfs_root *root = BTRFS_I(inode)->root;
1956 struct btrfs_trans_handle *trans;
1957 struct btrfs_log_ctx ctx;
1958 int ret = 0;
1959 bool full_sync = 0;
1960 u64 len;
1961
1962 /*
1963 * The range length can be represented by u64, we have to do the typecasts
1964 * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
1965 */
1966 len = (u64)end - (u64)start + 1;
1967 trace_btrfs_sync_file(file, datasync);
1968
1969 /*
1970 * We write the dirty pages in the range and wait until they complete
1971 * out of the ->i_mutex. If so, we can flush the dirty pages by
1972 * multi-task, and make the performance up. See
1973 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1974 */
1975 ret = start_ordered_ops(inode, start, end);
1976 if (ret)
1977 return ret;
1978
1979 inode_lock(inode);
1980 atomic_inc(&root->log_batch);
1981 full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1982 &BTRFS_I(inode)->runtime_flags);
1983 /*
1984 * We might have have had more pages made dirty after calling
1985 * start_ordered_ops and before acquiring the inode's i_mutex.
1986 */
1987 if (full_sync) {
1988 /*
1989 * For a full sync, we need to make sure any ordered operations
1990 * start and finish before we start logging the inode, so that
1991 * all extents are persisted and the respective file extent
1992 * items are in the fs/subvol btree.
1993 */
1994 ret = btrfs_wait_ordered_range(inode, start, len);
1995 } else {
1996 /*
1997 * Start any new ordered operations before starting to log the
1998 * inode. We will wait for them to finish in btrfs_sync_log().
1999 *
2000 * Right before acquiring the inode's mutex, we might have new
2001 * writes dirtying pages, which won't immediately start the
2002 * respective ordered operations - that is done through the
2003 * fill_delalloc callbacks invoked from the writepage and
2004 * writepages address space operations. So make sure we start
2005 * all ordered operations before starting to log our inode. Not
2006 * doing this means that while logging the inode, writeback
2007 * could start and invoke writepage/writepages, which would call
2008 * the fill_delalloc callbacks (cow_file_range,
2009 * submit_compressed_extents). These callbacks add first an
2010 * extent map to the modified list of extents and then create
2011 * the respective ordered operation, which means in
2012 * tree-log.c:btrfs_log_inode() we might capture all existing
2013 * ordered operations (with btrfs_get_logged_extents()) before
2014 * the fill_delalloc callback adds its ordered operation, and by
2015 * the time we visit the modified list of extent maps (with
2016 * btrfs_log_changed_extents()), we see and process the extent
2017 * map they created. We then use the extent map to construct a
2018 * file extent item for logging without waiting for the
2019 * respective ordered operation to finish - this file extent
2020 * item points to a disk location that might not have yet been
2021 * written to, containing random data - so after a crash a log
2022 * replay will make our inode have file extent items that point
2023 * to disk locations containing invalid data, as we returned
2024 * success to userspace without waiting for the respective
2025 * ordered operation to finish, because it wasn't captured by
2026 * btrfs_get_logged_extents().
2027 */
2028 ret = start_ordered_ops(inode, start, end);
2029 }
2030 if (ret) {
2031 inode_unlock(inode);
2032 goto out;
2033 }
2034 atomic_inc(&root->log_batch);
2035
2036 /*
2037 * If the last transaction that changed this file was before the current
2038 * transaction and we have the full sync flag set in our inode, we can
2039 * bail out now without any syncing.
2040 *
2041 * Note that we can't bail out if the full sync flag isn't set. This is
2042 * because when the full sync flag is set we start all ordered extents
2043 * and wait for them to fully complete - when they complete they update
2044 * the inode's last_trans field through:
2045 *
2046 * btrfs_finish_ordered_io() ->
2047 * btrfs_update_inode_fallback() ->
2048 * btrfs_update_inode() ->
2049 * btrfs_set_inode_last_trans()
2050 *
2051 * So we are sure that last_trans is up to date and can do this check to
2052 * bail out safely. For the fast path, when the full sync flag is not
2053 * set in our inode, we can not do it because we start only our ordered
2054 * extents and don't wait for them to complete (that is when
2055 * btrfs_finish_ordered_io runs), so here at this point their last_trans
2056 * value might be less than or equals to fs_info->last_trans_committed,
2057 * and setting a speculative last_trans for an inode when a buffered
2058 * write is made (such as fs_info->generation + 1 for example) would not
2059 * be reliable since after setting the value and before fsync is called
2060 * any number of transactions can start and commit (transaction kthread
2061 * commits the current transaction periodically), and a transaction
2062 * commit does not start nor waits for ordered extents to complete.
2063 */
2064 smp_mb();
2065 if (btrfs_inode_in_log(inode, fs_info->generation) ||
2066 (full_sync && BTRFS_I(inode)->last_trans <=
2067 fs_info->last_trans_committed) ||
2068 (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2069 BTRFS_I(inode)->last_trans
2070 <= fs_info->last_trans_committed)) {
2071 /*
2072 * We've had everything committed since the last time we were
2073 * modified so clear this flag in case it was set for whatever
2074 * reason, it's no longer relevant.
2075 */
2076 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2077 &BTRFS_I(inode)->runtime_flags);
2078 /*
2079 * An ordered extent might have started before and completed
2080 * already with io errors, in which case the inode was not
2081 * updated and we end up here. So check the inode's mapping
2082 * flags for any errors that might have happened while doing
2083 * writeback of file data.
2084 */
2085 ret = filemap_check_errors(inode->i_mapping);
2086 inode_unlock(inode);
2087 goto out;
2088 }
2089
2090 /*
2091 * ok we haven't committed the transaction yet, lets do a commit
2092 */
2093 if (file->private_data)
2094 btrfs_ioctl_trans_end(file);
2095
2096 /*
2097 * We use start here because we will need to wait on the IO to complete
2098 * in btrfs_sync_log, which could require joining a transaction (for
2099 * example checking cross references in the nocow path). If we use join
2100 * here we could get into a situation where we're waiting on IO to
2101 * happen that is blocked on a transaction trying to commit. With start
2102 * we inc the extwriter counter, so we wait for all extwriters to exit
2103 * before we start blocking join'ers. This comment is to keep somebody
2104 * from thinking they are super smart and changing this to
2105 * btrfs_join_transaction *cough*Josef*cough*.
2106 */
2107 trans = btrfs_start_transaction(root, 0);
2108 if (IS_ERR(trans)) {
2109 ret = PTR_ERR(trans);
2110 inode_unlock(inode);
2111 goto out;
2112 }
2113 trans->sync = true;
2114
2115 btrfs_init_log_ctx(&ctx, inode);
2116
2117 ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2118 if (ret < 0) {
2119 /* Fallthrough and commit/free transaction. */
2120 ret = 1;
2121 }
2122
2123 /* we've logged all the items and now have a consistent
2124 * version of the file in the log. It is possible that
2125 * someone will come in and modify the file, but that's
2126 * fine because the log is consistent on disk, and we
2127 * have references to all of the file's extents
2128 *
2129 * It is possible that someone will come in and log the
2130 * file again, but that will end up using the synchronization
2131 * inside btrfs_sync_log to keep things safe.
2132 */
2133 inode_unlock(inode);
2134
2135 /*
2136 * If any of the ordered extents had an error, just return it to user
2137 * space, so that the application knows some writes didn't succeed and
2138 * can take proper action (retry for e.g.). Blindly committing the
2139 * transaction in this case, would fool userspace that everything was
2140 * successful. And we also want to make sure our log doesn't contain
2141 * file extent items pointing to extents that weren't fully written to -
2142 * just like in the non fast fsync path, where we check for the ordered
2143 * operation's error flag before writing to the log tree and return -EIO
2144 * if any of them had this flag set (btrfs_wait_ordered_range) -
2145 * therefore we need to check for errors in the ordered operations,
2146 * which are indicated by ctx.io_err.
2147 */
2148 if (ctx.io_err) {
2149 btrfs_end_transaction(trans);
2150 ret = ctx.io_err;
2151 goto out;
2152 }
2153
2154 if (ret != BTRFS_NO_LOG_SYNC) {
2155 if (!ret) {
2156 ret = btrfs_sync_log(trans, root, &ctx);
2157 if (!ret) {
2158 ret = btrfs_end_transaction(trans);
2159 goto out;
2160 }
2161 }
2162 if (!full_sync) {
2163 ret = btrfs_wait_ordered_range(inode, start, len);
2164 if (ret) {
2165 btrfs_end_transaction(trans);
2166 goto out;
2167 }
2168 }
2169 ret = btrfs_commit_transaction(trans);
2170 } else {
2171 ret = btrfs_end_transaction(trans);
2172 }
2173out:
2174 return ret > 0 ? -EIO : ret;
2175}
2176
2177static const struct vm_operations_struct btrfs_file_vm_ops = {
2178 .fault = filemap_fault,
2179 .map_pages = filemap_map_pages,
2180 .page_mkwrite = btrfs_page_mkwrite,
2181};
2182
2183static int btrfs_file_mmap(struct file *filp, struct vm_area_struct *vma)
2184{
2185 struct address_space *mapping = filp->f_mapping;
2186
2187 if (!mapping->a_ops->readpage)
2188 return -ENOEXEC;
2189
2190 file_accessed(filp);
2191 vma->vm_ops = &btrfs_file_vm_ops;
2192
2193 return 0;
2194}
2195
2196static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2197 int slot, u64 start, u64 end)
2198{
2199 struct btrfs_file_extent_item *fi;
2200 struct btrfs_key key;
2201
2202 if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2203 return 0;
2204
2205 btrfs_item_key_to_cpu(leaf, &key, slot);
2206 if (key.objectid != btrfs_ino(inode) ||
2207 key.type != BTRFS_EXTENT_DATA_KEY)
2208 return 0;
2209
2210 fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2211
2212 if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2213 return 0;
2214
2215 if (btrfs_file_extent_disk_bytenr(leaf, fi))
2216 return 0;
2217
2218 if (key.offset == end)
2219 return 1;
2220 if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2221 return 1;
2222 return 0;
2223}
2224
2225static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2226 struct btrfs_path *path, u64 offset, u64 end)
2227{
2228 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2229 struct btrfs_root *root = BTRFS_I(inode)->root;
2230 struct extent_buffer *leaf;
2231 struct btrfs_file_extent_item *fi;
2232 struct extent_map *hole_em;
2233 struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2234 struct btrfs_key key;
2235 int ret;
2236
2237 if (btrfs_fs_incompat(fs_info, NO_HOLES))
2238 goto out;
2239
2240 key.objectid = btrfs_ino(inode);
2241 key.type = BTRFS_EXTENT_DATA_KEY;
2242 key.offset = offset;
2243
2244 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2245 if (ret <= 0) {
2246 /*
2247 * We should have dropped this offset, so if we find it then
2248 * something has gone horribly wrong.
2249 */
2250 if (ret == 0)
2251 ret = -EINVAL;
2252 return ret;
2253 }
2254
2255 leaf = path->nodes[0];
2256 if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2257 u64 num_bytes;
2258
2259 path->slots[0]--;
2260 fi = btrfs_item_ptr(leaf, path->slots[0],
2261 struct btrfs_file_extent_item);
2262 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2263 end - offset;
2264 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2265 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2266 btrfs_set_file_extent_offset(leaf, fi, 0);
2267 btrfs_mark_buffer_dirty(leaf);
2268 goto out;
2269 }
2270
2271 if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2272 u64 num_bytes;
2273
2274 key.offset = offset;
2275 btrfs_set_item_key_safe(fs_info, path, &key);
2276 fi = btrfs_item_ptr(leaf, path->slots[0],
2277 struct btrfs_file_extent_item);
2278 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2279 offset;
2280 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2281 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2282 btrfs_set_file_extent_offset(leaf, fi, 0);
2283 btrfs_mark_buffer_dirty(leaf);
2284 goto out;
2285 }
2286 btrfs_release_path(path);
2287
2288 ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2289 0, 0, end - offset, 0, end - offset,
2290 0, 0, 0);
2291 if (ret)
2292 return ret;
2293
2294out:
2295 btrfs_release_path(path);
2296
2297 hole_em = alloc_extent_map();
2298 if (!hole_em) {
2299 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2300 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2301 &BTRFS_I(inode)->runtime_flags);
2302 } else {
2303 hole_em->start = offset;
2304 hole_em->len = end - offset;
2305 hole_em->ram_bytes = hole_em->len;
2306 hole_em->orig_start = offset;
2307
2308 hole_em->block_start = EXTENT_MAP_HOLE;
2309 hole_em->block_len = 0;
2310 hole_em->orig_block_len = 0;
2311 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2312 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2313 hole_em->generation = trans->transid;
2314
2315 do {
2316 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2317 write_lock(&em_tree->lock);
2318 ret = add_extent_mapping(em_tree, hole_em, 1);
2319 write_unlock(&em_tree->lock);
2320 } while (ret == -EEXIST);
2321 free_extent_map(hole_em);
2322 if (ret)
2323 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2324 &BTRFS_I(inode)->runtime_flags);
2325 }
2326
2327 return 0;
2328}
2329
2330/*
2331 * Find a hole extent on given inode and change start/len to the end of hole
2332 * extent.(hole/vacuum extent whose em->start <= start &&
2333 * em->start + em->len > start)
2334 * When a hole extent is found, return 1 and modify start/len.
2335 */
2336static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2337{
2338 struct extent_map *em;
2339 int ret = 0;
2340
2341 em = btrfs_get_extent(inode, NULL, 0, *start, *len, 0);
2342 if (IS_ERR_OR_NULL(em)) {
2343 if (!em)
2344 ret = -ENOMEM;
2345 else
2346 ret = PTR_ERR(em);
2347 return ret;
2348 }
2349
2350 /* Hole or vacuum extent(only exists in no-hole mode) */
2351 if (em->block_start == EXTENT_MAP_HOLE) {
2352 ret = 1;
2353 *len = em->start + em->len > *start + *len ?
2354 0 : *start + *len - em->start - em->len;
2355 *start = em->start + em->len;
2356 }
2357 free_extent_map(em);
2358 return ret;
2359}
2360
2361static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2362{
2363 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2364 struct btrfs_root *root = BTRFS_I(inode)->root;
2365 struct extent_state *cached_state = NULL;
2366 struct btrfs_path *path;
2367 struct btrfs_block_rsv *rsv;
2368 struct btrfs_trans_handle *trans;
2369 u64 lockstart;
2370 u64 lockend;
2371 u64 tail_start;
2372 u64 tail_len;
2373 u64 orig_start = offset;
2374 u64 cur_offset;
2375 u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2376 u64 drop_end;
2377 int ret = 0;
2378 int err = 0;
2379 unsigned int rsv_count;
2380 bool same_block;
2381 bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2382 u64 ino_size;
2383 bool truncated_block = false;
2384 bool updated_inode = false;
2385
2386 ret = btrfs_wait_ordered_range(inode, offset, len);
2387 if (ret)
2388 return ret;
2389
2390 inode_lock(inode);
2391 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2392 ret = find_first_non_hole(inode, &offset, &len);
2393 if (ret < 0)
2394 goto out_only_mutex;
2395 if (ret && !len) {
2396 /* Already in a large hole */
2397 ret = 0;
2398 goto out_only_mutex;
2399 }
2400
2401 lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2402 lockend = round_down(offset + len,
2403 btrfs_inode_sectorsize(inode)) - 1;
2404 same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2405 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2406 /*
2407 * We needn't truncate any block which is beyond the end of the file
2408 * because we are sure there is no data there.
2409 */
2410 /*
2411 * Only do this if we are in the same block and we aren't doing the
2412 * entire block.
2413 */
2414 if (same_block && len < fs_info->sectorsize) {
2415 if (offset < ino_size) {
2416 truncated_block = true;
2417 ret = btrfs_truncate_block(inode, offset, len, 0);
2418 } else {
2419 ret = 0;
2420 }
2421 goto out_only_mutex;
2422 }
2423
2424 /* zero back part of the first block */
2425 if (offset < ino_size) {
2426 truncated_block = true;
2427 ret = btrfs_truncate_block(inode, offset, 0, 0);
2428 if (ret) {
2429 inode_unlock(inode);
2430 return ret;
2431 }
2432 }
2433
2434 /* Check the aligned pages after the first unaligned page,
2435 * if offset != orig_start, which means the first unaligned page
2436 * including several following pages are already in holes,
2437 * the extra check can be skipped */
2438 if (offset == orig_start) {
2439 /* after truncate page, check hole again */
2440 len = offset + len - lockstart;
2441 offset = lockstart;
2442 ret = find_first_non_hole(inode, &offset, &len);
2443 if (ret < 0)
2444 goto out_only_mutex;
2445 if (ret && !len) {
2446 ret = 0;
2447 goto out_only_mutex;
2448 }
2449 lockstart = offset;
2450 }
2451
2452 /* Check the tail unaligned part is in a hole */
2453 tail_start = lockend + 1;
2454 tail_len = offset + len - tail_start;
2455 if (tail_len) {
2456 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2457 if (unlikely(ret < 0))
2458 goto out_only_mutex;
2459 if (!ret) {
2460 /* zero the front end of the last page */
2461 if (tail_start + tail_len < ino_size) {
2462 truncated_block = true;
2463 ret = btrfs_truncate_block(inode,
2464 tail_start + tail_len,
2465 0, 1);
2466 if (ret)
2467 goto out_only_mutex;
2468 }
2469 }
2470 }
2471
2472 if (lockend < lockstart) {
2473 ret = 0;
2474 goto out_only_mutex;
2475 }
2476
2477 while (1) {
2478 struct btrfs_ordered_extent *ordered;
2479
2480 truncate_pagecache_range(inode, lockstart, lockend);
2481
2482 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2483 &cached_state);
2484 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2485
2486 /*
2487 * We need to make sure we have no ordered extents in this range
2488 * and nobody raced in and read a page in this range, if we did
2489 * we need to try again.
2490 */
2491 if ((!ordered ||
2492 (ordered->file_offset + ordered->len <= lockstart ||
2493 ordered->file_offset > lockend)) &&
2494 !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2495 if (ordered)
2496 btrfs_put_ordered_extent(ordered);
2497 break;
2498 }
2499 if (ordered)
2500 btrfs_put_ordered_extent(ordered);
2501 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2502 lockend, &cached_state, GFP_NOFS);
2503 ret = btrfs_wait_ordered_range(inode, lockstart,
2504 lockend - lockstart + 1);
2505 if (ret) {
2506 inode_unlock(inode);
2507 return ret;
2508 }
2509 }
2510
2511 path = btrfs_alloc_path();
2512 if (!path) {
2513 ret = -ENOMEM;
2514 goto out;
2515 }
2516
2517 rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2518 if (!rsv) {
2519 ret = -ENOMEM;
2520 goto out_free;
2521 }
2522 rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2523 rsv->failfast = 1;
2524
2525 /*
2526 * 1 - update the inode
2527 * 1 - removing the extents in the range
2528 * 1 - adding the hole extent if no_holes isn't set
2529 */
2530 rsv_count = no_holes ? 2 : 3;
2531 trans = btrfs_start_transaction(root, rsv_count);
2532 if (IS_ERR(trans)) {
2533 err = PTR_ERR(trans);
2534 goto out_free;
2535 }
2536
2537 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2538 min_size, 0);
2539 BUG_ON(ret);
2540 trans->block_rsv = rsv;
2541
2542 cur_offset = lockstart;
2543 len = lockend - cur_offset;
2544 while (cur_offset < lockend) {
2545 ret = __btrfs_drop_extents(trans, root, inode, path,
2546 cur_offset, lockend + 1,
2547 &drop_end, 1, 0, 0, NULL);
2548 if (ret != -ENOSPC)
2549 break;
2550
2551 trans->block_rsv = &fs_info->trans_block_rsv;
2552
2553 if (cur_offset < drop_end && cur_offset < ino_size) {
2554 ret = fill_holes(trans, inode, path, cur_offset,
2555 drop_end);
2556 if (ret) {
2557 /*
2558 * If we failed then we didn't insert our hole
2559 * entries for the area we dropped, so now the
2560 * fs is corrupted, so we must abort the
2561 * transaction.
2562 */
2563 btrfs_abort_transaction(trans, ret);
2564 err = ret;
2565 break;
2566 }
2567 }
2568
2569 cur_offset = drop_end;
2570
2571 ret = btrfs_update_inode(trans, root, inode);
2572 if (ret) {
2573 err = ret;
2574 break;
2575 }
2576
2577 btrfs_end_transaction(trans);
2578 btrfs_btree_balance_dirty(fs_info);
2579
2580 trans = btrfs_start_transaction(root, rsv_count);
2581 if (IS_ERR(trans)) {
2582 ret = PTR_ERR(trans);
2583 trans = NULL;
2584 break;
2585 }
2586
2587 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2588 rsv, min_size, 0);
2589 BUG_ON(ret); /* shouldn't happen */
2590 trans->block_rsv = rsv;
2591
2592 ret = find_first_non_hole(inode, &cur_offset, &len);
2593 if (unlikely(ret < 0))
2594 break;
2595 if (ret && !len) {
2596 ret = 0;
2597 break;
2598 }
2599 }
2600
2601 if (ret) {
2602 err = ret;
2603 goto out_trans;
2604 }
2605
2606 trans->block_rsv = &fs_info->trans_block_rsv;
2607 /*
2608 * If we are using the NO_HOLES feature we might have had already an
2609 * hole that overlaps a part of the region [lockstart, lockend] and
2610 * ends at (or beyond) lockend. Since we have no file extent items to
2611 * represent holes, drop_end can be less than lockend and so we must
2612 * make sure we have an extent map representing the existing hole (the
2613 * call to __btrfs_drop_extents() might have dropped the existing extent
2614 * map representing the existing hole), otherwise the fast fsync path
2615 * will not record the existence of the hole region
2616 * [existing_hole_start, lockend].
2617 */
2618 if (drop_end <= lockend)
2619 drop_end = lockend + 1;
2620 /*
2621 * Don't insert file hole extent item if it's for a range beyond eof
2622 * (because it's useless) or if it represents a 0 bytes range (when
2623 * cur_offset == drop_end).
2624 */
2625 if (cur_offset < ino_size && cur_offset < drop_end) {
2626 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2627 if (ret) {
2628 /* Same comment as above. */
2629 btrfs_abort_transaction(trans, ret);
2630 err = ret;
2631 goto out_trans;
2632 }
2633 }
2634
2635out_trans:
2636 if (!trans)
2637 goto out_free;
2638
2639 inode_inc_iversion(inode);
2640 inode->i_mtime = inode->i_ctime = current_time(inode);
2641
2642 trans->block_rsv = &fs_info->trans_block_rsv;
2643 ret = btrfs_update_inode(trans, root, inode);
2644 updated_inode = true;
2645 btrfs_end_transaction(trans);
2646 btrfs_btree_balance_dirty(fs_info);
2647out_free:
2648 btrfs_free_path(path);
2649 btrfs_free_block_rsv(fs_info, rsv);
2650out:
2651 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2652 &cached_state, GFP_NOFS);
2653out_only_mutex:
2654 if (!updated_inode && truncated_block && !ret && !err) {
2655 /*
2656 * If we only end up zeroing part of a page, we still need to
2657 * update the inode item, so that all the time fields are
2658 * updated as well as the necessary btrfs inode in memory fields
2659 * for detecting, at fsync time, if the inode isn't yet in the
2660 * log tree or it's there but not up to date.
2661 */
2662 trans = btrfs_start_transaction(root, 1);
2663 if (IS_ERR(trans)) {
2664 err = PTR_ERR(trans);
2665 } else {
2666 err = btrfs_update_inode(trans, root, inode);
2667 ret = btrfs_end_transaction(trans);
2668 }
2669 }
2670 inode_unlock(inode);
2671 if (ret && !err)
2672 err = ret;
2673 return err;
2674}
2675
2676/* Helper structure to record which range is already reserved */
2677struct falloc_range {
2678 struct list_head list;
2679 u64 start;
2680 u64 len;
2681};
2682
2683/*
2684 * Helper function to add falloc range
2685 *
2686 * Caller should have locked the larger range of extent containing
2687 * [start, len)
2688 */
2689static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2690{
2691 struct falloc_range *prev = NULL;
2692 struct falloc_range *range = NULL;
2693
2694 if (list_empty(head))
2695 goto insert;
2696
2697 /*
2698 * As fallocate iterate by bytenr order, we only need to check
2699 * the last range.
2700 */
2701 prev = list_entry(head->prev, struct falloc_range, list);
2702 if (prev->start + prev->len == start) {
2703 prev->len += len;
2704 return 0;
2705 }
2706insert:
2707 range = kmalloc(sizeof(*range), GFP_KERNEL);
2708 if (!range)
2709 return -ENOMEM;
2710 range->start = start;
2711 range->len = len;
2712 list_add_tail(&range->list, head);
2713 return 0;
2714}
2715
2716static long btrfs_fallocate(struct file *file, int mode,
2717 loff_t offset, loff_t len)
2718{
2719 struct inode *inode = file_inode(file);
2720 struct extent_state *cached_state = NULL;
2721 struct falloc_range *range;
2722 struct falloc_range *tmp;
2723 struct list_head reserve_list;
2724 u64 cur_offset;
2725 u64 last_byte;
2726 u64 alloc_start;
2727 u64 alloc_end;
2728 u64 alloc_hint = 0;
2729 u64 locked_end;
2730 u64 actual_end = 0;
2731 struct extent_map *em;
2732 int blocksize = btrfs_inode_sectorsize(inode);
2733 int ret;
2734
2735 alloc_start = round_down(offset, blocksize);
2736 alloc_end = round_up(offset + len, blocksize);
2737 cur_offset = alloc_start;
2738
2739 /* Make sure we aren't being give some crap mode */
2740 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2741 return -EOPNOTSUPP;
2742
2743 if (mode & FALLOC_FL_PUNCH_HOLE)
2744 return btrfs_punch_hole(inode, offset, len);
2745
2746 /*
2747 * Only trigger disk allocation, don't trigger qgroup reserve
2748 *
2749 * For qgroup space, it will be checked later.
2750 */
2751 ret = btrfs_alloc_data_chunk_ondemand(inode, alloc_end - alloc_start);
2752 if (ret < 0)
2753 return ret;
2754
2755 inode_lock(inode);
2756
2757 if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2758 ret = inode_newsize_ok(inode, offset + len);
2759 if (ret)
2760 goto out;
2761 }
2762
2763 /*
2764 * TODO: Move these two operations after we have checked
2765 * accurate reserved space, or fallocate can still fail but
2766 * with page truncated or size expanded.
2767 *
2768 * But that's a minor problem and won't do much harm BTW.
2769 */
2770 if (alloc_start > inode->i_size) {
2771 ret = btrfs_cont_expand(inode, i_size_read(inode),
2772 alloc_start);
2773 if (ret)
2774 goto out;
2775 } else if (offset + len > inode->i_size) {
2776 /*
2777 * If we are fallocating from the end of the file onward we
2778 * need to zero out the end of the block if i_size lands in the
2779 * middle of a block.
2780 */
2781 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2782 if (ret)
2783 goto out;
2784 }
2785
2786 /*
2787 * wait for ordered IO before we have any locks. We'll loop again
2788 * below with the locks held.
2789 */
2790 ret = btrfs_wait_ordered_range(inode, alloc_start,
2791 alloc_end - alloc_start);
2792 if (ret)
2793 goto out;
2794
2795 locked_end = alloc_end - 1;
2796 while (1) {
2797 struct btrfs_ordered_extent *ordered;
2798
2799 /* the extent lock is ordered inside the running
2800 * transaction
2801 */
2802 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2803 locked_end, &cached_state);
2804 ordered = btrfs_lookup_first_ordered_extent(inode,
2805 alloc_end - 1);
2806 if (ordered &&
2807 ordered->file_offset + ordered->len > alloc_start &&
2808 ordered->file_offset < alloc_end) {
2809 btrfs_put_ordered_extent(ordered);
2810 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2811 alloc_start, locked_end,
2812 &cached_state, GFP_KERNEL);
2813 /*
2814 * we can't wait on the range with the transaction
2815 * running or with the extent lock held
2816 */
2817 ret = btrfs_wait_ordered_range(inode, alloc_start,
2818 alloc_end - alloc_start);
2819 if (ret)
2820 goto out;
2821 } else {
2822 if (ordered)
2823 btrfs_put_ordered_extent(ordered);
2824 break;
2825 }
2826 }
2827
2828 /* First, check if we exceed the qgroup limit */
2829 INIT_LIST_HEAD(&reserve_list);
2830 while (1) {
2831 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2832 alloc_end - cur_offset, 0);
2833 if (IS_ERR_OR_NULL(em)) {
2834 if (!em)
2835 ret = -ENOMEM;
2836 else
2837 ret = PTR_ERR(em);
2838 break;
2839 }
2840 last_byte = min(extent_map_end(em), alloc_end);
2841 actual_end = min_t(u64, extent_map_end(em), offset + len);
2842 last_byte = ALIGN(last_byte, blocksize);
2843 if (em->block_start == EXTENT_MAP_HOLE ||
2844 (cur_offset >= inode->i_size &&
2845 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2846 ret = add_falloc_range(&reserve_list, cur_offset,
2847 last_byte - cur_offset);
2848 if (ret < 0) {
2849 free_extent_map(em);
2850 break;
2851 }
2852 ret = btrfs_qgroup_reserve_data(inode, cur_offset,
2853 last_byte - cur_offset);
2854 if (ret < 0)
2855 break;
2856 } else {
2857 /*
2858 * Do not need to reserve unwritten extent for this
2859 * range, free reserved data space first, otherwise
2860 * it'll result in false ENOSPC error.
2861 */
2862 btrfs_free_reserved_data_space(inode, cur_offset,
2863 last_byte - cur_offset);
2864 }
2865 free_extent_map(em);
2866 cur_offset = last_byte;
2867 if (cur_offset >= alloc_end)
2868 break;
2869 }
2870
2871 /*
2872 * If ret is still 0, means we're OK to fallocate.
2873 * Or just cleanup the list and exit.
2874 */
2875 list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2876 if (!ret)
2877 ret = btrfs_prealloc_file_range(inode, mode,
2878 range->start,
2879 range->len, 1 << inode->i_blkbits,
2880 offset + len, &alloc_hint);
2881 else
2882 btrfs_free_reserved_data_space(inode, range->start,
2883 range->len);
2884 list_del(&range->list);
2885 kfree(range);
2886 }
2887 if (ret < 0)
2888 goto out_unlock;
2889
2890 if (actual_end > inode->i_size &&
2891 !(mode & FALLOC_FL_KEEP_SIZE)) {
2892 struct btrfs_trans_handle *trans;
2893 struct btrfs_root *root = BTRFS_I(inode)->root;
2894
2895 /*
2896 * We didn't need to allocate any more space, but we
2897 * still extended the size of the file so we need to
2898 * update i_size and the inode item.
2899 */
2900 trans = btrfs_start_transaction(root, 1);
2901 if (IS_ERR(trans)) {
2902 ret = PTR_ERR(trans);
2903 } else {
2904 inode->i_ctime = current_time(inode);
2905 i_size_write(inode, actual_end);
2906 btrfs_ordered_update_i_size(inode, actual_end, NULL);
2907 ret = btrfs_update_inode(trans, root, inode);
2908 if (ret)
2909 btrfs_end_transaction(trans);
2910 else
2911 ret = btrfs_end_transaction(trans);
2912 }
2913 }
2914out_unlock:
2915 unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2916 &cached_state, GFP_KERNEL);
2917out:
2918 inode_unlock(inode);
2919 /* Let go of our reservation. */
2920 if (ret != 0)
2921 btrfs_free_reserved_data_space(inode, alloc_start,
2922 alloc_end - cur_offset);
2923 return ret;
2924}
2925
2926static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2927{
2928 struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2929 struct extent_map *em = NULL;
2930 struct extent_state *cached_state = NULL;
2931 u64 lockstart;
2932 u64 lockend;
2933 u64 start;
2934 u64 len;
2935 int ret = 0;
2936
2937 if (inode->i_size == 0)
2938 return -ENXIO;
2939
2940 /*
2941 * *offset can be negative, in this case we start finding DATA/HOLE from
2942 * the very start of the file.
2943 */
2944 start = max_t(loff_t, 0, *offset);
2945
2946 lockstart = round_down(start, fs_info->sectorsize);
2947 lockend = round_up(i_size_read(inode),
2948 fs_info->sectorsize);
2949 if (lockend <= lockstart)
2950 lockend = lockstart + fs_info->sectorsize;
2951 lockend--;
2952 len = lockend - lockstart + 1;
2953
2954 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2955 &cached_state);
2956
2957 while (start < inode->i_size) {
2958 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2959 if (IS_ERR(em)) {
2960 ret = PTR_ERR(em);
2961 em = NULL;
2962 break;
2963 }
2964
2965 if (whence == SEEK_HOLE &&
2966 (em->block_start == EXTENT_MAP_HOLE ||
2967 test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2968 break;
2969 else if (whence == SEEK_DATA &&
2970 (em->block_start != EXTENT_MAP_HOLE &&
2971 !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2972 break;
2973
2974 start = em->start + em->len;
2975 free_extent_map(em);
2976 em = NULL;
2977 cond_resched();
2978 }
2979 free_extent_map(em);
2980 if (!ret) {
2981 if (whence == SEEK_DATA && start >= inode->i_size)
2982 ret = -ENXIO;
2983 else
2984 *offset = min_t(loff_t, start, inode->i_size);
2985 }
2986 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2987 &cached_state, GFP_NOFS);
2988 return ret;
2989}
2990
2991static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2992{
2993 struct inode *inode = file->f_mapping->host;
2994 int ret;
2995
2996 inode_lock(inode);
2997 switch (whence) {
2998 case SEEK_END:
2999 case SEEK_CUR:
3000 offset = generic_file_llseek(file, offset, whence);
3001 goto out;
3002 case SEEK_DATA:
3003 case SEEK_HOLE:
3004 if (offset >= i_size_read(inode)) {
3005 inode_unlock(inode);
3006 return -ENXIO;
3007 }
3008
3009 ret = find_desired_extent(inode, &offset, whence);
3010 if (ret) {
3011 inode_unlock(inode);
3012 return ret;
3013 }
3014 }
3015
3016 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3017out:
3018 inode_unlock(inode);
3019 return offset;
3020}
3021
3022const struct file_operations btrfs_file_operations = {
3023 .llseek = btrfs_file_llseek,
3024 .read_iter = generic_file_read_iter,
3025 .splice_read = generic_file_splice_read,
3026 .write_iter = btrfs_file_write_iter,
3027 .mmap = btrfs_file_mmap,
3028 .open = generic_file_open,
3029 .release = btrfs_release_file,
3030 .fsync = btrfs_sync_file,
3031 .fallocate = btrfs_fallocate,
3032 .unlocked_ioctl = btrfs_ioctl,
3033#ifdef CONFIG_COMPAT
3034 .compat_ioctl = btrfs_compat_ioctl,
3035#endif
3036 .clone_file_range = btrfs_clone_file_range,
3037 .dedupe_file_range = btrfs_dedupe_file_range,
3038};
3039
3040void btrfs_auto_defrag_exit(void)
3041{
3042 kmem_cache_destroy(btrfs_inode_defrag_cachep);
3043}
3044
3045int btrfs_auto_defrag_init(void)
3046{
3047 btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3048 sizeof(struct inode_defrag), 0,
3049 SLAB_MEM_SPREAD,
3050 NULL);
3051 if (!btrfs_inode_defrag_cachep)
3052 return -ENOMEM;
3053
3054 return 0;
3055}
3056
3057int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3058{
3059 int ret;
3060
3061 /*
3062 * So with compression we will find and lock a dirty page and clear the
3063 * first one as dirty, setup an async extent, and immediately return
3064 * with the entire range locked but with nobody actually marked with
3065 * writeback. So we can't just filemap_write_and_wait_range() and
3066 * expect it to work since it will just kick off a thread to do the
3067 * actual work. So we need to call filemap_fdatawrite_range _again_
3068 * since it will wait on the page lock, which won't be unlocked until
3069 * after the pages have been marked as writeback and so we're good to go
3070 * from there. We have to do this otherwise we'll miss the ordered
3071 * extents and that results in badness. Please Josef, do not think you
3072 * know better and pull this out at some point in the future, it is
3073 * right and you are wrong.
3074 */
3075 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3076 if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3077 &BTRFS_I(inode)->runtime_flags))
3078 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3079
3080 return ret;
3081}