xref: /linux/fs/btrfs/defrag.c (revision 9bce95edb1b4d2802de9273b5170bfcff3090d24)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5 
6 #include <linux/sched.h>
7 #include "ctree.h"
8 #include "disk-io.h"
9 #include "transaction.h"
10 #include "locking.h"
11 #include "accessors.h"
12 #include "messages.h"
13 #include "delalloc-space.h"
14 #include "subpage.h"
15 #include "defrag.h"
16 #include "file-item.h"
17 #include "super.h"
18 #include "compression.h"
19 
20 static struct kmem_cache *btrfs_inode_defrag_cachep;
21 
22 /*
23  * When auto defrag is enabled we queue up these defrag structs to remember
24  * which inodes need defragging passes.
25  */
26 struct inode_defrag {
27 	struct rb_node rb_node;
28 	/* Inode number */
29 	u64 ino;
30 	/*
31 	 * Transid where the defrag was added, we search for extents newer than
32 	 * this.
33 	 */
34 	u64 transid;
35 
36 	/* Root objectid */
37 	u64 root;
38 
39 	/*
40 	 * The extent size threshold for autodefrag.
41 	 *
42 	 * This value is different for compressed/non-compressed extents, thus
43 	 * needs to be passed from higher layer.
44 	 * (aka, inode_should_defrag())
45 	 */
46 	u32 extent_thresh;
47 };
48 
49 static int compare_inode_defrag(const struct inode_defrag *defrag1,
50 				const struct inode_defrag *defrag2)
51 {
52 	if (defrag1->root > defrag2->root)
53 		return 1;
54 	else if (defrag1->root < defrag2->root)
55 		return -1;
56 	else if (defrag1->ino > defrag2->ino)
57 		return 1;
58 	else if (defrag1->ino < defrag2->ino)
59 		return -1;
60 	else
61 		return 0;
62 }
63 
64 static int inode_defrag_cmp(struct rb_node *new, const struct rb_node *existing)
65 {
66 	const struct inode_defrag *new_defrag = rb_entry(new, struct inode_defrag, rb_node);
67 	const struct inode_defrag *existing_defrag = rb_entry(existing, struct inode_defrag, rb_node);
68 
69 	return compare_inode_defrag(new_defrag, existing_defrag);
70 }
71 
72 /*
73  * Insert a record for an inode into the defrag tree.  The lock must be held
74  * already.
75  *
76  * If you're inserting a record for an older transid than an existing record,
77  * the transid already in the tree is lowered.
78  */
79 static int btrfs_insert_inode_defrag(struct btrfs_inode *inode,
80 				     struct inode_defrag *defrag)
81 {
82 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
83 	struct rb_node *node;
84 
85 	node = rb_find_add(&defrag->rb_node, &fs_info->defrag_inodes, inode_defrag_cmp);
86 	if (node) {
87 		struct inode_defrag *entry;
88 
89 		entry = rb_entry(node, struct inode_defrag, rb_node);
90 		/*
91 		 * If we're reinserting an entry for an old defrag run, make
92 		 * sure to lower the transid of our existing record.
93 		 */
94 		if (defrag->transid < entry->transid)
95 			entry->transid = defrag->transid;
96 		entry->extent_thresh = min(defrag->extent_thresh, entry->extent_thresh);
97 		return -EEXIST;
98 	}
99 	set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
100 	return 0;
101 }
102 
103 static inline bool need_auto_defrag(struct btrfs_fs_info *fs_info)
104 {
105 	if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
106 		return false;
107 
108 	if (btrfs_fs_closing(fs_info))
109 		return false;
110 
111 	return true;
112 }
113 
114 /*
115  * Insert a defrag record for this inode if auto defrag is enabled. No errors
116  * returned as they're not considered fatal.
117  */
118 void btrfs_add_inode_defrag(struct btrfs_inode *inode, u32 extent_thresh)
119 {
120 	struct btrfs_root *root = inode->root;
121 	struct btrfs_fs_info *fs_info = root->fs_info;
122 	struct inode_defrag *defrag;
123 	int ret;
124 
125 	if (!need_auto_defrag(fs_info))
126 		return;
127 
128 	if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
129 		return;
130 
131 	defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
132 	if (!defrag)
133 		return;
134 
135 	defrag->ino = btrfs_ino(inode);
136 	defrag->transid = btrfs_get_root_last_trans(root);
137 	defrag->root = btrfs_root_id(root);
138 	defrag->extent_thresh = extent_thresh;
139 
140 	spin_lock(&fs_info->defrag_inodes_lock);
141 	if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
142 		/*
143 		 * If we set IN_DEFRAG flag and evict the inode from memory,
144 		 * and then re-read this inode, this new inode doesn't have
145 		 * IN_DEFRAG flag. At the case, we may find the existed defrag.
146 		 */
147 		ret = btrfs_insert_inode_defrag(inode, defrag);
148 		if (ret)
149 			kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
150 	} else {
151 		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
152 	}
153 	spin_unlock(&fs_info->defrag_inodes_lock);
154 }
155 
156 /*
157  * Pick the defraggable inode that we want, if it doesn't exist, we will get the
158  * next one.
159  */
160 static struct inode_defrag *btrfs_pick_defrag_inode(
161 			struct btrfs_fs_info *fs_info, u64 root, u64 ino)
162 {
163 	struct inode_defrag *entry = NULL;
164 	struct inode_defrag tmp;
165 	struct rb_node *p;
166 	struct rb_node *parent = NULL;
167 	int ret;
168 
169 	tmp.ino = ino;
170 	tmp.root = root;
171 
172 	spin_lock(&fs_info->defrag_inodes_lock);
173 	p = fs_info->defrag_inodes.rb_node;
174 	while (p) {
175 		parent = p;
176 		entry = rb_entry(parent, struct inode_defrag, rb_node);
177 
178 		ret = compare_inode_defrag(&tmp, entry);
179 		if (ret < 0)
180 			p = parent->rb_left;
181 		else if (ret > 0)
182 			p = parent->rb_right;
183 		else
184 			goto out;
185 	}
186 
187 	if (parent && compare_inode_defrag(&tmp, entry) > 0) {
188 		parent = rb_next(parent);
189 		entry = rb_entry_safe(parent, struct inode_defrag, rb_node);
190 	}
191 out:
192 	if (entry)
193 		rb_erase(parent, &fs_info->defrag_inodes);
194 	spin_unlock(&fs_info->defrag_inodes_lock);
195 	return entry;
196 }
197 
198 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
199 {
200 	struct inode_defrag *defrag, *next;
201 
202 	spin_lock(&fs_info->defrag_inodes_lock);
203 
204 	rbtree_postorder_for_each_entry_safe(defrag, next,
205 					     &fs_info->defrag_inodes, rb_node)
206 		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
207 
208 	fs_info->defrag_inodes = RB_ROOT;
209 
210 	spin_unlock(&fs_info->defrag_inodes_lock);
211 }
212 
213 #define BTRFS_DEFRAG_BATCH	1024
214 
215 static int btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
216 				  struct inode_defrag *defrag,
217 				  struct file_ra_state *ra)
218 {
219 	struct btrfs_root *inode_root;
220 	struct btrfs_inode *inode;
221 	struct btrfs_ioctl_defrag_range_args range;
222 	int ret = 0;
223 	u64 cur = 0;
224 
225 again:
226 	if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
227 		goto cleanup;
228 	if (!need_auto_defrag(fs_info))
229 		goto cleanup;
230 
231 	/* Get the inode */
232 	inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
233 	if (IS_ERR(inode_root)) {
234 		ret = PTR_ERR(inode_root);
235 		goto cleanup;
236 	}
237 
238 	inode = btrfs_iget(defrag->ino, inode_root);
239 	btrfs_put_root(inode_root);
240 	if (IS_ERR(inode)) {
241 		ret = PTR_ERR(inode);
242 		goto cleanup;
243 	}
244 
245 	if (cur >= i_size_read(&inode->vfs_inode)) {
246 		iput(&inode->vfs_inode);
247 		goto cleanup;
248 	}
249 
250 	/* Do a chunk of defrag */
251 	clear_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
252 	memset(&range, 0, sizeof(range));
253 	range.len = (u64)-1;
254 	range.start = cur;
255 	range.extent_thresh = defrag->extent_thresh;
256 	file_ra_state_init(ra, inode->vfs_inode.i_mapping);
257 
258 	scoped_guard(super_write, fs_info->sb)
259 		ret = btrfs_defrag_file(inode, ra, &range,
260 					defrag->transid, BTRFS_DEFRAG_BATCH);
261 	iput(&inode->vfs_inode);
262 
263 	if (ret < 0)
264 		goto cleanup;
265 
266 	cur = max(cur + fs_info->sectorsize, range.start);
267 	goto again;
268 
269 cleanup:
270 	kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
271 	return ret;
272 }
273 
274 /*
275  * Run through the list of inodes in the FS that need defragging.
276  */
277 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
278 {
279 	struct inode_defrag *defrag;
280 	u64 first_ino = 0;
281 	u64 root_objectid = 0;
282 
283 	atomic_inc(&fs_info->defrag_running);
284 	while (1) {
285 		struct file_ra_state ra = { 0 };
286 
287 		/* Pause the auto defragger. */
288 		if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
289 			break;
290 
291 		if (!need_auto_defrag(fs_info))
292 			break;
293 
294 		/* find an inode to defrag */
295 		defrag = btrfs_pick_defrag_inode(fs_info, root_objectid, first_ino);
296 		if (!defrag) {
297 			if (root_objectid || first_ino) {
298 				root_objectid = 0;
299 				first_ino = 0;
300 				continue;
301 			} else {
302 				break;
303 			}
304 		}
305 
306 		first_ino = defrag->ino + 1;
307 		root_objectid = defrag->root;
308 
309 		btrfs_run_defrag_inode(fs_info, defrag, &ra);
310 	}
311 	atomic_dec(&fs_info->defrag_running);
312 
313 	/*
314 	 * During unmount, we use the transaction_wait queue to wait for the
315 	 * defragger to stop.
316 	 */
317 	wake_up(&fs_info->transaction_wait);
318 	return 0;
319 }
320 
321 /*
322  * Check if two blocks addresses are close, used by defrag.
323  */
324 static bool close_blocks(u64 blocknr, u64 other, u32 blocksize)
325 {
326 	if (blocknr < other && other - (blocknr + blocksize) < SZ_32K)
327 		return true;
328 	if (blocknr > other && blocknr - (other + blocksize) < SZ_32K)
329 		return true;
330 	return false;
331 }
332 
333 /*
334  * Go through all the leaves pointed to by a node and reallocate them so that
335  * disk order is close to key order.
336  */
337 static int btrfs_realloc_node(struct btrfs_trans_handle *trans,
338 			      struct btrfs_root *root,
339 			      struct extent_buffer *parent,
340 			      int start_slot, u64 *last_ret,
341 			      struct btrfs_key *progress)
342 {
343 	struct btrfs_fs_info *fs_info = root->fs_info;
344 	const u32 blocksize = fs_info->nodesize;
345 	const int end_slot = btrfs_header_nritems(parent) - 1;
346 	u64 search_start = *last_ret;
347 	u64 last_block = 0;
348 	int ret = 0;
349 	bool progress_passed = false;
350 
351 	/*
352 	 * COWing must happen through a running transaction, which always
353 	 * matches the current fs generation (it's a transaction with a state
354 	 * less than TRANS_STATE_UNBLOCKED). If it doesn't, then turn the fs
355 	 * into error state to prevent the commit of any transaction.
356 	 */
357 	if (unlikely(trans->transaction != fs_info->running_transaction ||
358 		     trans->transid != fs_info->generation)) {
359 		btrfs_abort_transaction(trans, -EUCLEAN);
360 		btrfs_crit(fs_info,
361 "unexpected transaction when attempting to reallocate parent %llu for root %llu, transaction %llu running transaction %llu fs generation %llu",
362 			   parent->start, btrfs_root_id(root), trans->transid,
363 			   fs_info->running_transaction->transid,
364 			   fs_info->generation);
365 		return -EUCLEAN;
366 	}
367 
368 	if (btrfs_header_nritems(parent) <= 1)
369 		return 0;
370 
371 	for (int i = start_slot; i <= end_slot; i++) {
372 		struct extent_buffer *cur;
373 		struct btrfs_disk_key disk_key;
374 		u64 blocknr;
375 		u64 other;
376 		bool close = true;
377 
378 		btrfs_node_key(parent, &disk_key, i);
379 		if (!progress_passed && btrfs_comp_keys(&disk_key, progress) < 0)
380 			continue;
381 
382 		progress_passed = true;
383 		blocknr = btrfs_node_blockptr(parent, i);
384 		if (last_block == 0)
385 			last_block = blocknr;
386 
387 		if (i > 0) {
388 			other = btrfs_node_blockptr(parent, i - 1);
389 			close = close_blocks(blocknr, other, blocksize);
390 		}
391 		if (!close && i < end_slot) {
392 			other = btrfs_node_blockptr(parent, i + 1);
393 			close = close_blocks(blocknr, other, blocksize);
394 		}
395 		if (close) {
396 			last_block = blocknr;
397 			continue;
398 		}
399 
400 		cur = btrfs_read_node_slot(parent, i);
401 		if (IS_ERR(cur))
402 			return PTR_ERR(cur);
403 		if (search_start == 0)
404 			search_start = last_block;
405 
406 		btrfs_tree_lock(cur);
407 		ret = btrfs_force_cow_block(trans, root, cur, parent, i,
408 					    &cur, search_start,
409 					    min(16 * blocksize,
410 						(end_slot - i) * blocksize),
411 					    BTRFS_NESTING_COW);
412 		if (ret) {
413 			btrfs_tree_unlock(cur);
414 			free_extent_buffer(cur);
415 			break;
416 		}
417 		search_start = cur->start;
418 		last_block = cur->start;
419 		*last_ret = search_start;
420 		btrfs_tree_unlock(cur);
421 		free_extent_buffer(cur);
422 	}
423 	return ret;
424 }
425 
426 /*
427  * Defrag all the leaves in a given btree.
428  * Read all the leaves and try to get key order to
429  * better reflect disk order
430  */
431 
432 static int btrfs_defrag_leaves(struct btrfs_trans_handle *trans,
433 			       struct btrfs_root *root)
434 {
435 	struct btrfs_path *path = NULL;
436 	struct btrfs_key key;
437 	int ret = 0;
438 	int wret;
439 	int level;
440 	int next_key_ret = 0;
441 	u64 last_ret = 0;
442 
443 	if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
444 		goto out;
445 
446 	path = btrfs_alloc_path();
447 	if (!path) {
448 		ret = -ENOMEM;
449 		goto out;
450 	}
451 
452 	level = btrfs_header_level(root->node);
453 
454 	if (level == 0)
455 		goto out;
456 
457 	if (root->defrag_progress.objectid == 0) {
458 		struct extent_buffer *root_node;
459 		u32 nritems;
460 
461 		root_node = btrfs_lock_root_node(root);
462 		nritems = btrfs_header_nritems(root_node);
463 		root->defrag_max.objectid = 0;
464 		/* from above we know this is not a leaf */
465 		btrfs_node_key_to_cpu(root_node, &root->defrag_max,
466 				      nritems - 1);
467 		btrfs_tree_unlock(root_node);
468 		free_extent_buffer(root_node);
469 		memset(&key, 0, sizeof(key));
470 	} else {
471 		memcpy(&key, &root->defrag_progress, sizeof(key));
472 	}
473 
474 	path->keep_locks = true;
475 
476 	ret = btrfs_search_forward(root, &key, path, BTRFS_OLDEST_GENERATION);
477 	if (ret < 0)
478 		goto out;
479 	if (ret > 0) {
480 		ret = 0;
481 		goto out;
482 	}
483 	btrfs_release_path(path);
484 	/*
485 	 * We don't need a lock on a leaf. btrfs_realloc_node() will lock all
486 	 * leafs from path->nodes[1], so set lowest_level to 1 to avoid later
487 	 * a deadlock (attempting to write lock an already write locked leaf).
488 	 */
489 	path->lowest_level = 1;
490 	wret = btrfs_search_slot(trans, root, &key, path, 0, 1);
491 
492 	if (wret < 0) {
493 		ret = wret;
494 		goto out;
495 	}
496 	if (!path->nodes[1]) {
497 		ret = 0;
498 		goto out;
499 	}
500 	/*
501 	 * The node at level 1 must always be locked when our path has
502 	 * keep_locks set and lowest_level is 1, regardless of the value of
503 	 * path->slots[1].
504 	 */
505 	ASSERT(path->locks[1] != 0);
506 	ret = btrfs_realloc_node(trans, root,
507 				 path->nodes[1], 0,
508 				 &last_ret,
509 				 &root->defrag_progress);
510 	if (ret) {
511 		WARN_ON(ret == -EAGAIN);
512 		goto out;
513 	}
514 	/*
515 	 * Now that we reallocated the node we can find the next key. Note that
516 	 * btrfs_find_next_key() can release our path and do another search
517 	 * without COWing, this is because even with path->keep_locks == true,
518 	 * btrfs_search_slot() / ctree.c:unlock_up() does not keeps a lock on a
519 	 * node when path->slots[node_level - 1] does not point to the last
520 	 * item or a slot beyond the last item (ctree.c:unlock_up()). Therefore
521 	 * we search for the next key after reallocating our node.
522 	 */
523 	path->slots[1] = btrfs_header_nritems(path->nodes[1]);
524 	next_key_ret = btrfs_find_next_key(root, path, &key, 1,
525 					   BTRFS_OLDEST_GENERATION);
526 	if (next_key_ret == 0) {
527 		memcpy(&root->defrag_progress, &key, sizeof(key));
528 		ret = -EAGAIN;
529 	}
530 out:
531 	btrfs_free_path(path);
532 	if (ret == -EAGAIN) {
533 		if (root->defrag_max.objectid > root->defrag_progress.objectid)
534 			goto done;
535 		if (root->defrag_max.type > root->defrag_progress.type)
536 			goto done;
537 		if (root->defrag_max.offset > root->defrag_progress.offset)
538 			goto done;
539 		ret = 0;
540 	}
541 done:
542 	if (ret != -EAGAIN)
543 		memset(&root->defrag_progress, 0,
544 		       sizeof(root->defrag_progress));
545 
546 	return ret;
547 }
548 
549 /*
550  * Defrag a given btree.  Every leaf in the btree is read and defragmented.
551  */
552 int btrfs_defrag_root(struct btrfs_root *root)
553 {
554 	struct btrfs_fs_info *fs_info = root->fs_info;
555 	int ret;
556 
557 	if (test_and_set_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state))
558 		return 0;
559 
560 	while (1) {
561 		struct btrfs_trans_handle *trans;
562 
563 		trans = btrfs_start_transaction(root, 0);
564 		if (IS_ERR(trans)) {
565 			ret = PTR_ERR(trans);
566 			break;
567 		}
568 
569 		ret = btrfs_defrag_leaves(trans, root);
570 
571 		btrfs_end_transaction(trans);
572 		btrfs_btree_balance_dirty(fs_info);
573 		cond_resched();
574 
575 		if (btrfs_fs_closing(fs_info) || ret != -EAGAIN)
576 			break;
577 
578 		if (btrfs_defrag_cancelled(fs_info)) {
579 			btrfs_debug(fs_info, "defrag_root cancelled");
580 			ret = -EAGAIN;
581 			break;
582 		}
583 	}
584 	clear_bit(BTRFS_ROOT_DEFRAG_RUNNING, &root->state);
585 	return ret;
586 }
587 
588 /*
589  * Defrag specific helper to get an extent map.
590  *
591  * Differences between this and btrfs_get_extent() are:
592  *
593  * - No extent_map will be added to inode->extent_tree
594  *   To reduce memory usage in the long run.
595  *
596  * - Extra optimization to skip file extents older than @newer_than
597  *   By using btrfs_search_forward() we can skip entire file ranges that
598  *   have extents created in past transactions, because btrfs_search_forward()
599  *   will not visit leaves and nodes with a generation smaller than given
600  *   minimal generation threshold (@newer_than).
601  *
602  * Return valid em if we find a file extent matching the requirement.
603  * Return NULL if we can not find a file extent matching the requirement.
604  *
605  * Return ERR_PTR() for error.
606  */
607 static struct extent_map *defrag_get_extent(struct btrfs_inode *inode,
608 					    u64 start, u64 newer_than)
609 {
610 	struct btrfs_root *root = inode->root;
611 	struct btrfs_file_extent_item *fi;
612 	BTRFS_PATH_AUTO_RELEASE(path);
613 	struct extent_map *em;
614 	struct btrfs_key key;
615 	u64 ino = btrfs_ino(inode);
616 	int ret;
617 
618 	em = btrfs_alloc_extent_map();
619 	if (!em) {
620 		ret = -ENOMEM;
621 		goto err;
622 	}
623 
624 	key.objectid = ino;
625 	key.type = BTRFS_EXTENT_DATA_KEY;
626 	key.offset = start;
627 
628 	if (newer_than) {
629 		ret = btrfs_search_forward(root, &key, &path, newer_than);
630 		if (ret < 0)
631 			goto err;
632 		/* Can't find anything newer */
633 		if (ret > 0)
634 			goto not_found;
635 	} else {
636 		ret = btrfs_search_slot(NULL, root, &key, &path, 0, 0);
637 		if (ret < 0)
638 			goto err;
639 	}
640 	if (path.slots[0] >= btrfs_header_nritems(path.nodes[0])) {
641 		/*
642 		 * If btrfs_search_slot() makes path to point beyond nritems,
643 		 * we should not have an empty leaf, as this inode must at
644 		 * least have its INODE_ITEM.
645 		 */
646 		ASSERT(btrfs_header_nritems(path.nodes[0]));
647 		path.slots[0] = btrfs_header_nritems(path.nodes[0]) - 1;
648 	}
649 	btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
650 	/* Perfect match, no need to go one slot back */
651 	if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY &&
652 	    key.offset == start)
653 		goto iterate;
654 
655 	/* We didn't find a perfect match, needs to go one slot back */
656 	if (path.slots[0] > 0) {
657 		btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
658 		if (key.objectid == ino && key.type == BTRFS_EXTENT_DATA_KEY)
659 			path.slots[0]--;
660 	}
661 
662 iterate:
663 	/* Iterate through the path to find a file extent covering @start */
664 	while (true) {
665 		u64 extent_end;
666 
667 		if (path.slots[0] >= btrfs_header_nritems(path.nodes[0]))
668 			goto next;
669 
670 		btrfs_item_key_to_cpu(path.nodes[0], &key, path.slots[0]);
671 
672 		/*
673 		 * We may go one slot back to INODE_REF/XATTR item, then
674 		 * need to go forward until we reach an EXTENT_DATA.
675 		 * But we should still has the correct ino as key.objectid.
676 		 */
677 		if (WARN_ON(key.objectid < ino) || key.type < BTRFS_EXTENT_DATA_KEY)
678 			goto next;
679 
680 		/* It's beyond our target range, definitely not extent found */
681 		if (key.objectid > ino || key.type > BTRFS_EXTENT_DATA_KEY)
682 			goto not_found;
683 
684 		/*
685 		 *	|	|<- File extent ->|
686 		 *	\- start
687 		 *
688 		 * This means there is a hole between start and key.offset.
689 		 */
690 		if (key.offset > start) {
691 			em->start = start;
692 			em->disk_bytenr = EXTENT_MAP_HOLE;
693 			em->disk_num_bytes = 0;
694 			em->ram_bytes = 0;
695 			em->offset = 0;
696 			em->len = key.offset - start;
697 			break;
698 		}
699 
700 		fi = btrfs_item_ptr(path.nodes[0], path.slots[0],
701 				    struct btrfs_file_extent_item);
702 		extent_end = btrfs_file_extent_end(&path);
703 
704 		/*
705 		 *	|<- file extent ->|	|
706 		 *				\- start
707 		 *
708 		 * We haven't reached start, search next slot.
709 		 */
710 		if (extent_end <= start)
711 			goto next;
712 
713 		/* Now this extent covers @start, convert it to em */
714 		btrfs_extent_item_to_extent_map(inode, &path, fi, em);
715 		break;
716 next:
717 		ret = btrfs_next_item(root, &path);
718 		if (ret < 0)
719 			goto err;
720 		if (ret > 0)
721 			goto not_found;
722 	}
723 	return em;
724 
725 not_found:
726 	btrfs_free_extent_map(em);
727 	return NULL;
728 
729 err:
730 	btrfs_free_extent_map(em);
731 	return ERR_PTR(ret);
732 }
733 
734 static struct extent_map *defrag_lookup_extent(struct inode *inode, u64 start,
735 					       u64 newer_than, bool locked)
736 {
737 	struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
738 	struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
739 	struct extent_map *em;
740 	const u32 sectorsize = BTRFS_I(inode)->root->fs_info->sectorsize;
741 
742 	/*
743 	 * Hopefully we have this extent in the tree already, try without the
744 	 * full extent lock.
745 	 */
746 	read_lock(&em_tree->lock);
747 	em = btrfs_lookup_extent_mapping(em_tree, start, sectorsize);
748 	read_unlock(&em_tree->lock);
749 
750 	/*
751 	 * We can get a merged extent, in that case, we need to re-search
752 	 * tree to get the original em for defrag.
753 	 *
754 	 * This is because even if we have adjacent extents that are contiguous
755 	 * and compatible (same type and flags), we still want to defrag them
756 	 * so that we use less metadata (extent items in the extent tree and
757 	 * file extent items in the inode's subvolume tree).
758 	 */
759 	if (em && (em->flags & EXTENT_FLAG_MERGED)) {
760 		btrfs_free_extent_map(em);
761 		em = NULL;
762 	}
763 
764 	if (!em) {
765 		struct extent_state *cached = NULL;
766 		u64 end = start + sectorsize - 1;
767 
768 		/* Get the big lock and read metadata off disk. */
769 		if (!locked)
770 			btrfs_lock_extent(io_tree, start, end, &cached);
771 		em = defrag_get_extent(BTRFS_I(inode), start, newer_than);
772 		if (!locked)
773 			btrfs_unlock_extent(io_tree, start, end, &cached);
774 
775 		if (IS_ERR(em))
776 			return NULL;
777 	}
778 
779 	return em;
780 }
781 
782 static u32 get_extent_max_capacity(const struct btrfs_fs_info *fs_info,
783 				   const struct extent_map *em)
784 {
785 	if (btrfs_extent_map_is_compressed(em))
786 		return BTRFS_MAX_COMPRESSED;
787 	return fs_info->max_extent_size;
788 }
789 
790 static bool defrag_check_next_extent(struct inode *inode, struct extent_map *em,
791 				     u32 extent_thresh, u64 newer_than, bool locked)
792 {
793 	struct btrfs_fs_info *fs_info = inode_to_fs_info(inode);
794 	struct extent_map *next;
795 	const u64 em_end = btrfs_extent_map_end(em);
796 	bool ret = false;
797 
798 	/* This is the last extent */
799 	if (em_end >= i_size_read(inode))
800 		return false;
801 
802 	/*
803 	 * Here we need to pass @newer_then when checking the next extent, or
804 	 * we will hit a case we mark current extent for defrag, but the next
805 	 * one will not be a target.
806 	 * This will just cause extra IO without really reducing the fragments.
807 	 */
808 	next = defrag_lookup_extent(inode, em_end, newer_than, locked);
809 	/* No more em or hole */
810 	if (!next || next->disk_bytenr >= EXTENT_MAP_LAST_BYTE)
811 		goto out;
812 	if (next->flags & EXTENT_FLAG_PREALLOC)
813 		goto out;
814 	/*
815 	 * If the next extent is at its max capacity, defragging current extent
816 	 * makes no sense, as the total number of extents won't change.
817 	 */
818 	if (next->len >= get_extent_max_capacity(fs_info, em))
819 		goto out;
820 	/* Skip older extent */
821 	if (next->generation < newer_than)
822 		goto out;
823 	/* Also check extent size */
824 	if (next->len >= extent_thresh)
825 		goto out;
826 
827 	ret = true;
828 out:
829 	btrfs_free_extent_map(next);
830 	return ret;
831 }
832 
833 /*
834  * Prepare one page to be defragged.
835  *
836  * This will ensure:
837  *
838  * - Returned page is locked and has been set up properly.
839  * - No ordered extent exists in the page.
840  * - The page is uptodate.
841  *
842  * NOTE: Caller should also wait for page writeback after the cluster is
843  * prepared, here we don't do writeback wait for each page.
844  */
845 static struct folio *defrag_prepare_one_folio(struct btrfs_inode *inode, pgoff_t index)
846 {
847 	struct address_space *mapping = inode->vfs_inode.i_mapping;
848 	gfp_t mask = btrfs_alloc_write_mask(mapping);
849 	u64 lock_start;
850 	u64 lock_end;
851 	struct extent_state *cached_state = NULL;
852 	struct folio *folio;
853 	int ret;
854 
855 again:
856 	/* TODO: Add order fgp order flags when large folios are fully enabled. */
857 	folio = __filemap_get_folio(mapping, index,
858 				    FGP_LOCK | FGP_ACCESSED | FGP_CREAT, mask);
859 	if (IS_ERR(folio))
860 		return folio;
861 
862 	ret = set_folio_extent_mapped(folio);
863 	if (ret < 0) {
864 		folio_unlock(folio);
865 		folio_put(folio);
866 		return ERR_PTR(ret);
867 	}
868 
869 	lock_start = folio_pos(folio);
870 	lock_end = folio_next_pos(folio) - 1;
871 	/* Wait for any existing ordered extent in the range */
872 	while (1) {
873 		struct btrfs_ordered_extent *ordered;
874 
875 		btrfs_lock_extent(&inode->io_tree, lock_start, lock_end, &cached_state);
876 		ordered = btrfs_lookup_ordered_range(inode, lock_start, folio_size(folio));
877 		btrfs_unlock_extent(&inode->io_tree, lock_start, lock_end, &cached_state);
878 		if (!ordered)
879 			break;
880 
881 		folio_unlock(folio);
882 		btrfs_start_ordered_extent(ordered);
883 		btrfs_put_ordered_extent(ordered);
884 		folio_lock(folio);
885 		/*
886 		 * We unlocked the folio above, so we need check if it was
887 		 * released or not.
888 		 */
889 		if (folio->mapping != mapping || !folio->private) {
890 			folio_unlock(folio);
891 			folio_put(folio);
892 			goto again;
893 		}
894 	}
895 
896 	/*
897 	 * Now the page range has no ordered extent any more.  Read the page to
898 	 * make it uptodate.
899 	 */
900 	if (!folio_test_uptodate(folio)) {
901 		btrfs_read_folio(NULL, folio);
902 		folio_lock(folio);
903 		if (folio->mapping != mapping || !folio->private) {
904 			folio_unlock(folio);
905 			folio_put(folio);
906 			goto again;
907 		}
908 		if (unlikely(!folio_test_uptodate(folio))) {
909 			folio_unlock(folio);
910 			folio_put(folio);
911 			return ERR_PTR(-EIO);
912 		}
913 	}
914 	return folio;
915 }
916 
917 struct defrag_target_range {
918 	struct list_head list;
919 	u64 start;
920 	u64 len;
921 };
922 
923 /*
924  * Collect all valid target extents.
925  *
926  * @start:	   file offset to lookup
927  * @len:	   length to lookup
928  * @extent_thresh: file extent size threshold, any extent size >= this value
929  *		   will be ignored
930  * @newer_than:    only defrag extents newer than this value
931  * @do_compress:   whether the defrag is doing compression or no-compression
932  *		   if true, @extent_thresh will be ignored and all regular
933  *		   file extents meeting @newer_than will be targets.
934  * @locked:	   if the range has already held extent lock
935  * @target_list:   list of targets file extents
936  */
937 static int defrag_collect_targets(struct btrfs_inode *inode,
938 				  u64 start, u64 len, u32 extent_thresh,
939 				  u64 newer_than, bool do_compress,
940 				  bool locked, struct list_head *target_list,
941 				  u64 *last_scanned_ret)
942 {
943 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
944 	bool last_is_target = false;
945 	u64 cur = start;
946 	int ret = 0;
947 
948 	while (cur < start + len) {
949 		struct extent_map *em;
950 		struct defrag_target_range *new;
951 		bool next_mergeable = true;
952 		u64 range_len;
953 
954 		last_is_target = false;
955 		em = defrag_lookup_extent(&inode->vfs_inode, cur, newer_than, locked);
956 		if (!em)
957 			break;
958 
959 		/*
960 		 * If the file extent is an inlined one, we may still want to
961 		 * defrag it (fallthrough) if it will cause a regular extent.
962 		 * This is for users who want to convert inline extents to
963 		 * regular ones through max_inline= mount option.
964 		 */
965 		if (em->disk_bytenr == EXTENT_MAP_INLINE &&
966 		    em->len <= inode->root->fs_info->max_inline)
967 			goto next;
968 
969 		/* Skip holes and preallocated extents. */
970 		if (em->disk_bytenr == EXTENT_MAP_HOLE ||
971 		    (em->flags & EXTENT_FLAG_PREALLOC))
972 			goto next;
973 
974 		/* Skip older extent */
975 		if (em->generation < newer_than)
976 			goto next;
977 
978 		/* This em is under writeback, no need to defrag */
979 		if (em->generation == (u64)-1)
980 			goto next;
981 
982 		/*
983 		 * Our start offset might be in the middle of an existing extent
984 		 * map, so take that into account.
985 		 */
986 		range_len = em->len - (cur - em->start);
987 		/*
988 		 * If this range of the extent map is already flagged for delalloc,
989 		 * skip it, because:
990 		 *
991 		 * 1) We could deadlock later, when trying to reserve space for
992 		 *    delalloc, because in case we can't immediately reserve space
993 		 *    the flusher can start delalloc and wait for the respective
994 		 *    ordered extents to complete. The deadlock would happen
995 		 *    because we do the space reservation while holding the range
996 		 *    locked, and starting writeback, or finishing an ordered
997 		 *    extent, requires locking the range;
998 		 *
999 		 * 2) If there's delalloc there, it means there's dirty pages for
1000 		 *    which writeback has not started yet (we clean the delalloc
1001 		 *    flag when starting writeback and after creating an ordered
1002 		 *    extent). If we mark pages in an adjacent range for defrag,
1003 		 *    then we will have a larger contiguous range for delalloc,
1004 		 *    very likely resulting in a larger extent after writeback is
1005 		 *    triggered (except in a case of free space fragmentation).
1006 		 */
1007 		if (btrfs_test_range_bit_exists(&inode->io_tree, cur, cur + range_len - 1,
1008 						EXTENT_DELALLOC))
1009 			goto next;
1010 
1011 		/*
1012 		 * For do_compress case, we want to compress all valid file
1013 		 * extents, thus no @extent_thresh or mergeable check.
1014 		 */
1015 		if (do_compress)
1016 			goto add;
1017 
1018 		/* Skip too large extent */
1019 		if (em->len >= extent_thresh)
1020 			goto next;
1021 
1022 		/*
1023 		 * Skip extents already at its max capacity, this is mostly for
1024 		 * compressed extents, which max cap is only 128K.
1025 		 */
1026 		if (em->len >= get_extent_max_capacity(fs_info, em))
1027 			goto next;
1028 
1029 		/*
1030 		 * Normally there are no more extents after an inline one, thus
1031 		 * @next_mergeable will normally be false and not defragged.
1032 		 * So if an inline extent passed all above checks, just add it
1033 		 * for defrag, and be converted to regular extents.
1034 		 */
1035 		if (em->disk_bytenr == EXTENT_MAP_INLINE)
1036 			goto add;
1037 
1038 		next_mergeable = defrag_check_next_extent(&inode->vfs_inode, em,
1039 						extent_thresh, newer_than, locked);
1040 		if (!next_mergeable) {
1041 			struct defrag_target_range *last;
1042 
1043 			/* Empty target list, no way to merge with last entry */
1044 			if (list_empty(target_list))
1045 				goto next;
1046 			last = list_last_entry(target_list,
1047 					       struct defrag_target_range, list);
1048 			/* Not mergeable with last entry */
1049 			if (last->start + last->len != cur)
1050 				goto next;
1051 
1052 			/* Mergeable, fall through to add it to @target_list. */
1053 		}
1054 
1055 add:
1056 		last_is_target = true;
1057 		range_len = min(btrfs_extent_map_end(em), start + len) - cur;
1058 		/*
1059 		 * This one is a good target, check if it can be merged into
1060 		 * last range of the target list.
1061 		 */
1062 		if (!list_empty(target_list)) {
1063 			struct defrag_target_range *last;
1064 
1065 			last = list_last_entry(target_list,
1066 					       struct defrag_target_range, list);
1067 			ASSERT(last->start + last->len <= cur);
1068 			if (last->start + last->len == cur) {
1069 				/* Mergeable, enlarge the last entry */
1070 				last->len += range_len;
1071 				goto next;
1072 			}
1073 			/* Fall through to allocate a new entry */
1074 		}
1075 
1076 		/* Allocate new defrag_target_range */
1077 		new = kmalloc_obj(*new, GFP_NOFS);
1078 		if (!new) {
1079 			btrfs_free_extent_map(em);
1080 			ret = -ENOMEM;
1081 			break;
1082 		}
1083 		new->start = cur;
1084 		new->len = range_len;
1085 		list_add_tail(&new->list, target_list);
1086 
1087 next:
1088 		cur = btrfs_extent_map_end(em);
1089 		btrfs_free_extent_map(em);
1090 	}
1091 	if (ret < 0) {
1092 		struct defrag_target_range *entry;
1093 		struct defrag_target_range *tmp;
1094 
1095 		list_for_each_entry_safe(entry, tmp, target_list, list) {
1096 			list_del_init(&entry->list);
1097 			kfree(entry);
1098 		}
1099 	}
1100 	if (!ret && last_scanned_ret) {
1101 		/*
1102 		 * If the last extent is not a target, the caller can skip to
1103 		 * the end of that extent.
1104 		 * Otherwise, we can only go the end of the specified range.
1105 		 */
1106 		if (!last_is_target)
1107 			*last_scanned_ret = max(cur, *last_scanned_ret);
1108 		else
1109 			*last_scanned_ret = max(start + len, *last_scanned_ret);
1110 	}
1111 	return ret;
1112 }
1113 
1114 #define CLUSTER_SIZE	(SZ_256K)
1115 static_assert(PAGE_ALIGNED(CLUSTER_SIZE));
1116 
1117 /*
1118  * Defrag one contiguous target range.
1119  *
1120  * @inode:	target inode
1121  * @target:	target range to defrag
1122  * @pages:	locked pages covering the defrag range
1123  * @nr_pages:	number of locked pages
1124  *
1125  * Caller should ensure:
1126  *
1127  * - Pages are prepared
1128  *   Pages should be locked, no ordered extent in the pages range,
1129  *   no writeback.
1130  *
1131  * - Extent bits are locked
1132  */
1133 static int defrag_one_locked_target(struct btrfs_inode *inode,
1134 				    struct defrag_target_range *target,
1135 				    struct folio **folios, int nr_pages,
1136 				    struct extent_state **cached_state)
1137 {
1138 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1139 	struct extent_changeset *data_reserved = NULL;
1140 	const u64 start = target->start;
1141 	const u64 len = target->len;
1142 	int ret = 0;
1143 
1144 	ret = btrfs_delalloc_reserve_space(inode, &data_reserved, start, len);
1145 	if (ret < 0)
1146 		return ret;
1147 	btrfs_clear_extent_bit(&inode->io_tree, start, start + len - 1,
1148 			       EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
1149 			       EXTENT_DEFRAG, cached_state);
1150 	btrfs_set_extent_bit(&inode->io_tree, start, start + len - 1,
1151 			     EXTENT_DELALLOC | EXTENT_DEFRAG, cached_state);
1152 
1153 	/*
1154 	 * Update the page status.
1155 	 * Due to possible large folios, we have to check all folios one by one.
1156 	 */
1157 	for (int i = 0; i < nr_pages && folios[i]; i++) {
1158 		struct folio *folio = folios[i];
1159 
1160 		if (!folio)
1161 			break;
1162 		if (start >= folio_next_pos(folio) ||
1163 		    start + len <= folio_pos(folio))
1164 			continue;
1165 		btrfs_folio_clamp_set_dirty(fs_info, folio, start, len);
1166 	}
1167 	btrfs_delalloc_release_extents(inode, len);
1168 	extent_changeset_free(data_reserved);
1169 
1170 	return ret;
1171 }
1172 
1173 static int defrag_one_range(struct btrfs_inode *inode, u64 start, u32 len,
1174 			    u32 extent_thresh, u64 newer_than, bool do_compress,
1175 			    u64 *last_scanned_ret)
1176 {
1177 	struct extent_state *cached_state = NULL;
1178 	struct defrag_target_range *entry;
1179 	struct defrag_target_range *tmp;
1180 	LIST_HEAD(target_list);
1181 	struct folio **folios;
1182 	const u32 sectorsize = inode->root->fs_info->sectorsize;
1183 	u64 cur = start;
1184 	const unsigned int nr_pages = ((start + len - 1) >> PAGE_SHIFT) -
1185 				      (start >> PAGE_SHIFT) + 1;
1186 	int ret = 0;
1187 
1188 	ASSERT(nr_pages <= CLUSTER_SIZE / PAGE_SIZE);
1189 	ASSERT(IS_ALIGNED(start, sectorsize) && IS_ALIGNED(len, sectorsize));
1190 
1191 	folios = kzalloc_objs(struct folio *, nr_pages, GFP_NOFS);
1192 	if (!folios)
1193 		return -ENOMEM;
1194 
1195 	/* Prepare all pages */
1196 	for (int i = 0; cur < start + len && i < nr_pages; i++) {
1197 		folios[i] = defrag_prepare_one_folio(inode, cur >> PAGE_SHIFT);
1198 		if (IS_ERR(folios[i])) {
1199 			ret = PTR_ERR(folios[i]);
1200 			folios[i] = NULL;
1201 			goto free_folios;
1202 		}
1203 		cur = folio_next_pos(folios[i]);
1204 	}
1205 	for (int i = 0; i < nr_pages; i++) {
1206 		if (!folios[i])
1207 			break;
1208 		folio_wait_writeback(folios[i]);
1209 	}
1210 
1211 	/* We should get at least one folio. */
1212 	ASSERT(folios[0]);
1213 	/* Lock the pages range */
1214 	btrfs_lock_extent(&inode->io_tree, folio_pos(folios[0]), cur - 1, &cached_state);
1215 	/*
1216 	 * Now we have a consistent view about the extent map, re-check
1217 	 * which range really needs to be defragged.
1218 	 *
1219 	 * And this time we have extent locked already, pass @locked = true
1220 	 * so that we won't relock the extent range and cause deadlock.
1221 	 */
1222 	ret = defrag_collect_targets(inode, start, len, extent_thresh,
1223 				     newer_than, do_compress, true,
1224 				     &target_list, last_scanned_ret);
1225 	if (ret < 0)
1226 		goto unlock_extent;
1227 
1228 	list_for_each_entry(entry, &target_list, list) {
1229 		ret = defrag_one_locked_target(inode, entry, folios, nr_pages,
1230 					       &cached_state);
1231 		if (ret < 0)
1232 			break;
1233 	}
1234 
1235 	list_for_each_entry_safe(entry, tmp, &target_list, list) {
1236 		list_del_init(&entry->list);
1237 		kfree(entry);
1238 	}
1239 unlock_extent:
1240 	btrfs_unlock_extent(&inode->io_tree, folio_pos(folios[0]), cur - 1, &cached_state);
1241 free_folios:
1242 	for (int i = 0; i < nr_pages; i++) {
1243 		if (!folios[i])
1244 			break;
1245 		folio_unlock(folios[i]);
1246 		folio_put(folios[i]);
1247 	}
1248 	kfree(folios);
1249 	return ret;
1250 }
1251 
1252 static int defrag_one_cluster(struct btrfs_inode *inode,
1253 			      struct file_ra_state *ra,
1254 			      u64 start, u32 len, u32 extent_thresh,
1255 			      u64 newer_than, bool do_compress,
1256 			      unsigned long *sectors_defragged,
1257 			      unsigned long max_sectors,
1258 			      u64 *last_scanned_ret)
1259 {
1260 	const u32 sectorsize = inode->root->fs_info->sectorsize;
1261 	struct defrag_target_range *entry;
1262 	struct defrag_target_range *tmp;
1263 	LIST_HEAD(target_list);
1264 	int ret;
1265 
1266 	ret = defrag_collect_targets(inode, start, len, extent_thresh,
1267 				     newer_than, do_compress, false,
1268 				     &target_list, NULL);
1269 	if (ret < 0)
1270 		goto out;
1271 
1272 	list_for_each_entry(entry, &target_list, list) {
1273 		u32 range_len = entry->len;
1274 
1275 		/* Reached or beyond the limit */
1276 		if (max_sectors && *sectors_defragged >= max_sectors) {
1277 			ret = 1;
1278 			break;
1279 		}
1280 
1281 		if (max_sectors)
1282 			range_len = min_t(u32, range_len,
1283 				(max_sectors - *sectors_defragged) * sectorsize);
1284 
1285 		/*
1286 		 * If defrag_one_range() has updated last_scanned_ret,
1287 		 * our range may already be invalid (e.g. hole punched).
1288 		 * Skip if our range is before last_scanned_ret, as there is
1289 		 * no need to defrag the range anymore.
1290 		 */
1291 		if (entry->start + range_len <= *last_scanned_ret)
1292 			continue;
1293 
1294 		page_cache_sync_readahead(inode->vfs_inode.i_mapping,
1295 				ra, NULL, entry->start >> PAGE_SHIFT,
1296 				((entry->start + range_len - 1) >> PAGE_SHIFT) -
1297 				(entry->start >> PAGE_SHIFT) + 1);
1298 		/*
1299 		 * Here we may not defrag any range if holes are punched before
1300 		 * we locked the pages.
1301 		 * But that's fine, it only affects the @sectors_defragged
1302 		 * accounting.
1303 		 */
1304 		ret = defrag_one_range(inode, entry->start, range_len,
1305 				       extent_thresh, newer_than, do_compress,
1306 				       last_scanned_ret);
1307 		if (ret < 0)
1308 			break;
1309 		*sectors_defragged += range_len >>
1310 				      inode->root->fs_info->sectorsize_bits;
1311 	}
1312 out:
1313 	list_for_each_entry_safe(entry, tmp, &target_list, list) {
1314 		list_del_init(&entry->list);
1315 		kfree(entry);
1316 	}
1317 	if (ret >= 0)
1318 		*last_scanned_ret = max(*last_scanned_ret, start + len);
1319 	return ret;
1320 }
1321 
1322 /*
1323  * Entry point to file defragmentation.
1324  *
1325  * @inode:	   inode to be defragged
1326  * @ra:		   readahead state
1327  * @range:	   defrag options including range and flags
1328  * @newer_than:	   minimum transid to defrag
1329  * @max_to_defrag: max number of sectors to be defragged, if 0, the whole inode
1330  *		   will be defragged.
1331  *
1332  * Return <0 for error.
1333  * Return >=0 for the number of sectors defragged, and range->start will be updated
1334  * to indicate the file offset where next defrag should be started at.
1335  * (Mostly for autodefrag, which sets @max_to_defrag thus we may exit early without
1336  *  defragging all the range).
1337  */
1338 int btrfs_defrag_file(struct btrfs_inode *inode, struct file_ra_state *ra,
1339 		      struct btrfs_ioctl_defrag_range_args *range,
1340 		      u64 newer_than, unsigned long max_to_defrag)
1341 {
1342 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1343 	unsigned long sectors_defragged = 0;
1344 	u64 isize = i_size_read(&inode->vfs_inode);
1345 	u64 cur;
1346 	u64 last_byte;
1347 	bool do_compress = (range->flags & BTRFS_DEFRAG_RANGE_COMPRESS);
1348 	bool no_compress = (range->flags & BTRFS_DEFRAG_RANGE_NOCOMPRESS);
1349 	int compress_type = BTRFS_COMPRESS_ZLIB;
1350 	int compress_level = 0;
1351 	int ret = 0;
1352 	u32 extent_thresh = range->extent_thresh;
1353 	pgoff_t start_index;
1354 
1355 	ASSERT(ra);
1356 
1357 	if (isize == 0)
1358 		return 0;
1359 
1360 	if (range->start >= isize)
1361 		return -EINVAL;
1362 
1363 	if (do_compress) {
1364 		if (range->flags & BTRFS_DEFRAG_RANGE_COMPRESS_LEVEL) {
1365 			if (range->compress.type >= BTRFS_NR_COMPRESS_TYPES)
1366 				return -EINVAL;
1367 			if (range->compress.type) {
1368 				compress_type  = range->compress.type;
1369 				compress_level = range->compress.level;
1370 				if (!btrfs_compress_level_valid(compress_type, compress_level))
1371 					return -EINVAL;
1372 			}
1373 		} else {
1374 			if (range->compress_type >= BTRFS_NR_COMPRESS_TYPES)
1375 				return -EINVAL;
1376 			if (range->compress_type)
1377 				compress_type = range->compress_type;
1378 		}
1379 	} else if (range->flags & BTRFS_DEFRAG_RANGE_NOCOMPRESS) {
1380 		compress_type = BTRFS_DEFRAG_DONT_COMPRESS;
1381 		compress_level = 1;
1382 	}
1383 
1384 	if (extent_thresh == 0)
1385 		extent_thresh = SZ_256K;
1386 
1387 	if (range->start + range->len > range->start) {
1388 		/* Got a specific range */
1389 		last_byte = min(isize, range->start + range->len);
1390 	} else {
1391 		/* Defrag until file end */
1392 		last_byte = isize;
1393 	}
1394 
1395 	/* Align the range */
1396 	cur = round_down(range->start, fs_info->sectorsize);
1397 	last_byte = round_up(last_byte, fs_info->sectorsize) - 1;
1398 
1399 	/*
1400 	 * Make writeback start from the beginning of the range, so that the
1401 	 * defrag range can be written sequentially.
1402 	 */
1403 	start_index = cur >> PAGE_SHIFT;
1404 	if (start_index < inode->vfs_inode.i_mapping->writeback_index)
1405 		inode->vfs_inode.i_mapping->writeback_index = start_index;
1406 
1407 	while (cur < last_byte) {
1408 		const unsigned long prev_sectors_defragged = sectors_defragged;
1409 		u64 last_scanned = cur;
1410 		u64 cluster_end;
1411 
1412 		if (btrfs_defrag_cancelled(fs_info)) {
1413 			ret = -EAGAIN;
1414 			break;
1415 		}
1416 
1417 		/* We want the cluster end at page boundary when possible */
1418 		cluster_end = (((cur >> PAGE_SHIFT) +
1419 			       (SZ_256K >> PAGE_SHIFT)) << PAGE_SHIFT) - 1;
1420 		cluster_end = min(cluster_end, last_byte);
1421 
1422 		btrfs_inode_lock(inode, 0);
1423 		if (IS_SWAPFILE(&inode->vfs_inode)) {
1424 			ret = -ETXTBSY;
1425 			btrfs_inode_unlock(inode, 0);
1426 			break;
1427 		}
1428 		if (!(inode->vfs_inode.i_sb->s_flags & SB_ACTIVE)) {
1429 			btrfs_inode_unlock(inode, 0);
1430 			break;
1431 		}
1432 		if (do_compress || no_compress) {
1433 			inode->defrag_compress = compress_type;
1434 			inode->defrag_compress_level = compress_level;
1435 		}
1436 		ret = defrag_one_cluster(inode, ra, cur,
1437 				cluster_end + 1 - cur, extent_thresh,
1438 				newer_than, do_compress || no_compress,
1439 				&sectors_defragged,
1440 				max_to_defrag, &last_scanned);
1441 
1442 		if (sectors_defragged > prev_sectors_defragged)
1443 			balance_dirty_pages_ratelimited(inode->vfs_inode.i_mapping);
1444 
1445 		btrfs_inode_unlock(inode, 0);
1446 		if (ret < 0)
1447 			break;
1448 		cur = max(cluster_end + 1, last_scanned);
1449 		if (ret > 0) {
1450 			ret = 0;
1451 			break;
1452 		}
1453 		cond_resched();
1454 	}
1455 
1456 	/*
1457 	 * Update range.start for autodefrag, this will indicate where to start
1458 	 * in next run.
1459 	 */
1460 	range->start = cur;
1461 	if (sectors_defragged) {
1462 		/*
1463 		 * We have defragged some sectors, for compression case they
1464 		 * need to be written back immediately.
1465 		 */
1466 		if (range->flags & BTRFS_DEFRAG_RANGE_START_IO) {
1467 			filemap_flush(inode->vfs_inode.i_mapping);
1468 			if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1469 				     &inode->runtime_flags))
1470 				filemap_flush(inode->vfs_inode.i_mapping);
1471 		}
1472 		if (range->compress_type == BTRFS_COMPRESS_LZO)
1473 			btrfs_set_fs_incompat(fs_info, COMPRESS_LZO);
1474 		else if (range->compress_type == BTRFS_COMPRESS_ZSTD)
1475 			btrfs_set_fs_incompat(fs_info, COMPRESS_ZSTD);
1476 		ret = sectors_defragged;
1477 	}
1478 	if (do_compress || no_compress) {
1479 		btrfs_inode_lock(inode, 0);
1480 		inode->defrag_compress = BTRFS_COMPRESS_NONE;
1481 		btrfs_inode_unlock(inode, 0);
1482 	}
1483 	return ret;
1484 }
1485 
1486 void __cold btrfs_auto_defrag_exit(void)
1487 {
1488 	kmem_cache_destroy(btrfs_inode_defrag_cachep);
1489 }
1490 
1491 int __init btrfs_auto_defrag_init(void)
1492 {
1493 	btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
1494 					sizeof(struct inode_defrag), 0, 0, NULL);
1495 	if (!btrfs_inode_defrag_cachep)
1496 		return -ENOMEM;
1497 
1498 	return 0;
1499 }
1500