xref: /linux/fs/btrfs/free-space-cache.c (revision 1e1159bb97cf4191849b4b03f77ab32977c2ece6)
1 /*
2  * Copyright (C) 2008 Red Hat.  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/pagemap.h>
20 #include <linux/sched.h>
21 #include <linux/slab.h>
22 #include <linux/math64.h>
23 #include <linux/ratelimit.h>
24 #include "ctree.h"
25 #include "free-space-cache.h"
26 #include "transaction.h"
27 #include "disk-io.h"
28 #include "extent_io.h"
29 #include "inode-map.h"
30 
31 #define BITS_PER_BITMAP		(PAGE_CACHE_SIZE * 8)
32 #define MAX_CACHE_BYTES_PER_GIG	(32 * 1024)
33 
34 static int link_free_space(struct btrfs_free_space_ctl *ctl,
35 			   struct btrfs_free_space *info);
36 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
37 			      struct btrfs_free_space *info);
38 
39 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
40 					       struct btrfs_path *path,
41 					       u64 offset)
42 {
43 	struct btrfs_key key;
44 	struct btrfs_key location;
45 	struct btrfs_disk_key disk_key;
46 	struct btrfs_free_space_header *header;
47 	struct extent_buffer *leaf;
48 	struct inode *inode = NULL;
49 	int ret;
50 
51 	key.objectid = BTRFS_FREE_SPACE_OBJECTID;
52 	key.offset = offset;
53 	key.type = 0;
54 
55 	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
56 	if (ret < 0)
57 		return ERR_PTR(ret);
58 	if (ret > 0) {
59 		btrfs_release_path(path);
60 		return ERR_PTR(-ENOENT);
61 	}
62 
63 	leaf = path->nodes[0];
64 	header = btrfs_item_ptr(leaf, path->slots[0],
65 				struct btrfs_free_space_header);
66 	btrfs_free_space_key(leaf, header, &disk_key);
67 	btrfs_disk_key_to_cpu(&location, &disk_key);
68 	btrfs_release_path(path);
69 
70 	inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
71 	if (!inode)
72 		return ERR_PTR(-ENOENT);
73 	if (IS_ERR(inode))
74 		return inode;
75 	if (is_bad_inode(inode)) {
76 		iput(inode);
77 		return ERR_PTR(-ENOENT);
78 	}
79 
80 	mapping_set_gfp_mask(inode->i_mapping,
81 			mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS);
82 
83 	return inode;
84 }
85 
86 struct inode *lookup_free_space_inode(struct btrfs_root *root,
87 				      struct btrfs_block_group_cache
88 				      *block_group, struct btrfs_path *path)
89 {
90 	struct inode *inode = NULL;
91 	u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
92 
93 	spin_lock(&block_group->lock);
94 	if (block_group->inode)
95 		inode = igrab(block_group->inode);
96 	spin_unlock(&block_group->lock);
97 	if (inode)
98 		return inode;
99 
100 	inode = __lookup_free_space_inode(root, path,
101 					  block_group->key.objectid);
102 	if (IS_ERR(inode))
103 		return inode;
104 
105 	spin_lock(&block_group->lock);
106 	if (!((BTRFS_I(inode)->flags & flags) == flags)) {
107 		btrfs_info(root->fs_info,
108 			"Old style space inode found, converting.");
109 		BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
110 			BTRFS_INODE_NODATACOW;
111 		block_group->disk_cache_state = BTRFS_DC_CLEAR;
112 	}
113 
114 	if (!block_group->iref) {
115 		block_group->inode = igrab(inode);
116 		block_group->iref = 1;
117 	}
118 	spin_unlock(&block_group->lock);
119 
120 	return inode;
121 }
122 
123 static int __create_free_space_inode(struct btrfs_root *root,
124 				     struct btrfs_trans_handle *trans,
125 				     struct btrfs_path *path,
126 				     u64 ino, u64 offset)
127 {
128 	struct btrfs_key key;
129 	struct btrfs_disk_key disk_key;
130 	struct btrfs_free_space_header *header;
131 	struct btrfs_inode_item *inode_item;
132 	struct extent_buffer *leaf;
133 	u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
134 	int ret;
135 
136 	ret = btrfs_insert_empty_inode(trans, root, path, ino);
137 	if (ret)
138 		return ret;
139 
140 	/* We inline crc's for the free disk space cache */
141 	if (ino != BTRFS_FREE_INO_OBJECTID)
142 		flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
143 
144 	leaf = path->nodes[0];
145 	inode_item = btrfs_item_ptr(leaf, path->slots[0],
146 				    struct btrfs_inode_item);
147 	btrfs_item_key(leaf, &disk_key, path->slots[0]);
148 	memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
149 			     sizeof(*inode_item));
150 	btrfs_set_inode_generation(leaf, inode_item, trans->transid);
151 	btrfs_set_inode_size(leaf, inode_item, 0);
152 	btrfs_set_inode_nbytes(leaf, inode_item, 0);
153 	btrfs_set_inode_uid(leaf, inode_item, 0);
154 	btrfs_set_inode_gid(leaf, inode_item, 0);
155 	btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
156 	btrfs_set_inode_flags(leaf, inode_item, flags);
157 	btrfs_set_inode_nlink(leaf, inode_item, 1);
158 	btrfs_set_inode_transid(leaf, inode_item, trans->transid);
159 	btrfs_set_inode_block_group(leaf, inode_item, offset);
160 	btrfs_mark_buffer_dirty(leaf);
161 	btrfs_release_path(path);
162 
163 	key.objectid = BTRFS_FREE_SPACE_OBJECTID;
164 	key.offset = offset;
165 	key.type = 0;
166 
167 	ret = btrfs_insert_empty_item(trans, root, path, &key,
168 				      sizeof(struct btrfs_free_space_header));
169 	if (ret < 0) {
170 		btrfs_release_path(path);
171 		return ret;
172 	}
173 	leaf = path->nodes[0];
174 	header = btrfs_item_ptr(leaf, path->slots[0],
175 				struct btrfs_free_space_header);
176 	memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
177 	btrfs_set_free_space_key(leaf, header, &disk_key);
178 	btrfs_mark_buffer_dirty(leaf);
179 	btrfs_release_path(path);
180 
181 	return 0;
182 }
183 
184 int create_free_space_inode(struct btrfs_root *root,
185 			    struct btrfs_trans_handle *trans,
186 			    struct btrfs_block_group_cache *block_group,
187 			    struct btrfs_path *path)
188 {
189 	int ret;
190 	u64 ino;
191 
192 	ret = btrfs_find_free_objectid(root, &ino);
193 	if (ret < 0)
194 		return ret;
195 
196 	return __create_free_space_inode(root, trans, path, ino,
197 					 block_group->key.objectid);
198 }
199 
200 int btrfs_check_trunc_cache_free_space(struct btrfs_root *root,
201 				       struct btrfs_block_rsv *rsv)
202 {
203 	u64 needed_bytes;
204 	int ret;
205 
206 	/* 1 for slack space, 1 for updating the inode */
207 	needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
208 		btrfs_calc_trans_metadata_size(root, 1);
209 
210 	spin_lock(&rsv->lock);
211 	if (rsv->reserved < needed_bytes)
212 		ret = -ENOSPC;
213 	else
214 		ret = 0;
215 	spin_unlock(&rsv->lock);
216 	return ret;
217 }
218 
219 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
220 				    struct btrfs_trans_handle *trans,
221 				    struct btrfs_path *path,
222 				    struct inode *inode)
223 {
224 	int ret = 0;
225 
226 	btrfs_i_size_write(inode, 0);
227 	truncate_pagecache(inode, 0);
228 
229 	/*
230 	 * We don't need an orphan item because truncating the free space cache
231 	 * will never be split across transactions.
232 	 */
233 	ret = btrfs_truncate_inode_items(trans, root, inode,
234 					 0, BTRFS_EXTENT_DATA_KEY);
235 	if (ret) {
236 		btrfs_abort_transaction(trans, root, ret);
237 		return ret;
238 	}
239 
240 	ret = btrfs_update_inode(trans, root, inode);
241 	if (ret)
242 		btrfs_abort_transaction(trans, root, ret);
243 
244 	return ret;
245 }
246 
247 static int readahead_cache(struct inode *inode)
248 {
249 	struct file_ra_state *ra;
250 	unsigned long last_index;
251 
252 	ra = kzalloc(sizeof(*ra), GFP_NOFS);
253 	if (!ra)
254 		return -ENOMEM;
255 
256 	file_ra_state_init(ra, inode->i_mapping);
257 	last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
258 
259 	page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
260 
261 	kfree(ra);
262 
263 	return 0;
264 }
265 
266 struct io_ctl {
267 	void *cur, *orig;
268 	struct page *page;
269 	struct page **pages;
270 	struct btrfs_root *root;
271 	unsigned long size;
272 	int index;
273 	int num_pages;
274 	unsigned check_crcs:1;
275 };
276 
277 static int io_ctl_init(struct io_ctl *io_ctl, struct inode *inode,
278 		       struct btrfs_root *root)
279 {
280 	memset(io_ctl, 0, sizeof(struct io_ctl));
281 	io_ctl->num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
282 		PAGE_CACHE_SHIFT;
283 	io_ctl->pages = kzalloc(sizeof(struct page *) * io_ctl->num_pages,
284 				GFP_NOFS);
285 	if (!io_ctl->pages)
286 		return -ENOMEM;
287 	io_ctl->root = root;
288 	if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
289 		io_ctl->check_crcs = 1;
290 	return 0;
291 }
292 
293 static void io_ctl_free(struct io_ctl *io_ctl)
294 {
295 	kfree(io_ctl->pages);
296 }
297 
298 static void io_ctl_unmap_page(struct io_ctl *io_ctl)
299 {
300 	if (io_ctl->cur) {
301 		kunmap(io_ctl->page);
302 		io_ctl->cur = NULL;
303 		io_ctl->orig = NULL;
304 	}
305 }
306 
307 static void io_ctl_map_page(struct io_ctl *io_ctl, int clear)
308 {
309 	ASSERT(io_ctl->index < io_ctl->num_pages);
310 	io_ctl->page = io_ctl->pages[io_ctl->index++];
311 	io_ctl->cur = kmap(io_ctl->page);
312 	io_ctl->orig = io_ctl->cur;
313 	io_ctl->size = PAGE_CACHE_SIZE;
314 	if (clear)
315 		memset(io_ctl->cur, 0, PAGE_CACHE_SIZE);
316 }
317 
318 static void io_ctl_drop_pages(struct io_ctl *io_ctl)
319 {
320 	int i;
321 
322 	io_ctl_unmap_page(io_ctl);
323 
324 	for (i = 0; i < io_ctl->num_pages; i++) {
325 		if (io_ctl->pages[i]) {
326 			ClearPageChecked(io_ctl->pages[i]);
327 			unlock_page(io_ctl->pages[i]);
328 			page_cache_release(io_ctl->pages[i]);
329 		}
330 	}
331 }
332 
333 static int io_ctl_prepare_pages(struct io_ctl *io_ctl, struct inode *inode,
334 				int uptodate)
335 {
336 	struct page *page;
337 	gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
338 	int i;
339 
340 	for (i = 0; i < io_ctl->num_pages; i++) {
341 		page = find_or_create_page(inode->i_mapping, i, mask);
342 		if (!page) {
343 			io_ctl_drop_pages(io_ctl);
344 			return -ENOMEM;
345 		}
346 		io_ctl->pages[i] = page;
347 		if (uptodate && !PageUptodate(page)) {
348 			btrfs_readpage(NULL, page);
349 			lock_page(page);
350 			if (!PageUptodate(page)) {
351 				printk(KERN_ERR "btrfs: error reading free "
352 				       "space cache\n");
353 				io_ctl_drop_pages(io_ctl);
354 				return -EIO;
355 			}
356 		}
357 	}
358 
359 	for (i = 0; i < io_ctl->num_pages; i++) {
360 		clear_page_dirty_for_io(io_ctl->pages[i]);
361 		set_page_extent_mapped(io_ctl->pages[i]);
362 	}
363 
364 	return 0;
365 }
366 
367 static void io_ctl_set_generation(struct io_ctl *io_ctl, u64 generation)
368 {
369 	__le64 *val;
370 
371 	io_ctl_map_page(io_ctl, 1);
372 
373 	/*
374 	 * Skip the csum areas.  If we don't check crcs then we just have a
375 	 * 64bit chunk at the front of the first page.
376 	 */
377 	if (io_ctl->check_crcs) {
378 		io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
379 		io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
380 	} else {
381 		io_ctl->cur += sizeof(u64);
382 		io_ctl->size -= sizeof(u64) * 2;
383 	}
384 
385 	val = io_ctl->cur;
386 	*val = cpu_to_le64(generation);
387 	io_ctl->cur += sizeof(u64);
388 }
389 
390 static int io_ctl_check_generation(struct io_ctl *io_ctl, u64 generation)
391 {
392 	__le64 *gen;
393 
394 	/*
395 	 * Skip the crc area.  If we don't check crcs then we just have a 64bit
396 	 * chunk at the front of the first page.
397 	 */
398 	if (io_ctl->check_crcs) {
399 		io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
400 		io_ctl->size -= sizeof(u64) +
401 			(sizeof(u32) * io_ctl->num_pages);
402 	} else {
403 		io_ctl->cur += sizeof(u64);
404 		io_ctl->size -= sizeof(u64) * 2;
405 	}
406 
407 	gen = io_ctl->cur;
408 	if (le64_to_cpu(*gen) != generation) {
409 		printk_ratelimited(KERN_ERR "btrfs: space cache generation "
410 				   "(%Lu) does not match inode (%Lu)\n", *gen,
411 				   generation);
412 		io_ctl_unmap_page(io_ctl);
413 		return -EIO;
414 	}
415 	io_ctl->cur += sizeof(u64);
416 	return 0;
417 }
418 
419 static void io_ctl_set_crc(struct io_ctl *io_ctl, int index)
420 {
421 	u32 *tmp;
422 	u32 crc = ~(u32)0;
423 	unsigned offset = 0;
424 
425 	if (!io_ctl->check_crcs) {
426 		io_ctl_unmap_page(io_ctl);
427 		return;
428 	}
429 
430 	if (index == 0)
431 		offset = sizeof(u32) * io_ctl->num_pages;
432 
433 	crc = btrfs_csum_data(io_ctl->orig + offset, crc,
434 			      PAGE_CACHE_SIZE - offset);
435 	btrfs_csum_final(crc, (char *)&crc);
436 	io_ctl_unmap_page(io_ctl);
437 	tmp = kmap(io_ctl->pages[0]);
438 	tmp += index;
439 	*tmp = crc;
440 	kunmap(io_ctl->pages[0]);
441 }
442 
443 static int io_ctl_check_crc(struct io_ctl *io_ctl, int index)
444 {
445 	u32 *tmp, val;
446 	u32 crc = ~(u32)0;
447 	unsigned offset = 0;
448 
449 	if (!io_ctl->check_crcs) {
450 		io_ctl_map_page(io_ctl, 0);
451 		return 0;
452 	}
453 
454 	if (index == 0)
455 		offset = sizeof(u32) * io_ctl->num_pages;
456 
457 	tmp = kmap(io_ctl->pages[0]);
458 	tmp += index;
459 	val = *tmp;
460 	kunmap(io_ctl->pages[0]);
461 
462 	io_ctl_map_page(io_ctl, 0);
463 	crc = btrfs_csum_data(io_ctl->orig + offset, crc,
464 			      PAGE_CACHE_SIZE - offset);
465 	btrfs_csum_final(crc, (char *)&crc);
466 	if (val != crc) {
467 		printk_ratelimited(KERN_ERR "btrfs: csum mismatch on free "
468 				   "space cache\n");
469 		io_ctl_unmap_page(io_ctl);
470 		return -EIO;
471 	}
472 
473 	return 0;
474 }
475 
476 static int io_ctl_add_entry(struct io_ctl *io_ctl, u64 offset, u64 bytes,
477 			    void *bitmap)
478 {
479 	struct btrfs_free_space_entry *entry;
480 
481 	if (!io_ctl->cur)
482 		return -ENOSPC;
483 
484 	entry = io_ctl->cur;
485 	entry->offset = cpu_to_le64(offset);
486 	entry->bytes = cpu_to_le64(bytes);
487 	entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
488 		BTRFS_FREE_SPACE_EXTENT;
489 	io_ctl->cur += sizeof(struct btrfs_free_space_entry);
490 	io_ctl->size -= sizeof(struct btrfs_free_space_entry);
491 
492 	if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
493 		return 0;
494 
495 	io_ctl_set_crc(io_ctl, io_ctl->index - 1);
496 
497 	/* No more pages to map */
498 	if (io_ctl->index >= io_ctl->num_pages)
499 		return 0;
500 
501 	/* map the next page */
502 	io_ctl_map_page(io_ctl, 1);
503 	return 0;
504 }
505 
506 static int io_ctl_add_bitmap(struct io_ctl *io_ctl, void *bitmap)
507 {
508 	if (!io_ctl->cur)
509 		return -ENOSPC;
510 
511 	/*
512 	 * If we aren't at the start of the current page, unmap this one and
513 	 * map the next one if there is any left.
514 	 */
515 	if (io_ctl->cur != io_ctl->orig) {
516 		io_ctl_set_crc(io_ctl, io_ctl->index - 1);
517 		if (io_ctl->index >= io_ctl->num_pages)
518 			return -ENOSPC;
519 		io_ctl_map_page(io_ctl, 0);
520 	}
521 
522 	memcpy(io_ctl->cur, bitmap, PAGE_CACHE_SIZE);
523 	io_ctl_set_crc(io_ctl, io_ctl->index - 1);
524 	if (io_ctl->index < io_ctl->num_pages)
525 		io_ctl_map_page(io_ctl, 0);
526 	return 0;
527 }
528 
529 static void io_ctl_zero_remaining_pages(struct io_ctl *io_ctl)
530 {
531 	/*
532 	 * If we're not on the boundary we know we've modified the page and we
533 	 * need to crc the page.
534 	 */
535 	if (io_ctl->cur != io_ctl->orig)
536 		io_ctl_set_crc(io_ctl, io_ctl->index - 1);
537 	else
538 		io_ctl_unmap_page(io_ctl);
539 
540 	while (io_ctl->index < io_ctl->num_pages) {
541 		io_ctl_map_page(io_ctl, 1);
542 		io_ctl_set_crc(io_ctl, io_ctl->index - 1);
543 	}
544 }
545 
546 static int io_ctl_read_entry(struct io_ctl *io_ctl,
547 			    struct btrfs_free_space *entry, u8 *type)
548 {
549 	struct btrfs_free_space_entry *e;
550 	int ret;
551 
552 	if (!io_ctl->cur) {
553 		ret = io_ctl_check_crc(io_ctl, io_ctl->index);
554 		if (ret)
555 			return ret;
556 	}
557 
558 	e = io_ctl->cur;
559 	entry->offset = le64_to_cpu(e->offset);
560 	entry->bytes = le64_to_cpu(e->bytes);
561 	*type = e->type;
562 	io_ctl->cur += sizeof(struct btrfs_free_space_entry);
563 	io_ctl->size -= sizeof(struct btrfs_free_space_entry);
564 
565 	if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
566 		return 0;
567 
568 	io_ctl_unmap_page(io_ctl);
569 
570 	return 0;
571 }
572 
573 static int io_ctl_read_bitmap(struct io_ctl *io_ctl,
574 			      struct btrfs_free_space *entry)
575 {
576 	int ret;
577 
578 	ret = io_ctl_check_crc(io_ctl, io_ctl->index);
579 	if (ret)
580 		return ret;
581 
582 	memcpy(entry->bitmap, io_ctl->cur, PAGE_CACHE_SIZE);
583 	io_ctl_unmap_page(io_ctl);
584 
585 	return 0;
586 }
587 
588 /*
589  * Since we attach pinned extents after the fact we can have contiguous sections
590  * of free space that are split up in entries.  This poses a problem with the
591  * tree logging stuff since it could have allocated across what appears to be 2
592  * entries since we would have merged the entries when adding the pinned extents
593  * back to the free space cache.  So run through the space cache that we just
594  * loaded and merge contiguous entries.  This will make the log replay stuff not
595  * blow up and it will make for nicer allocator behavior.
596  */
597 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
598 {
599 	struct btrfs_free_space *e, *prev = NULL;
600 	struct rb_node *n;
601 
602 again:
603 	spin_lock(&ctl->tree_lock);
604 	for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
605 		e = rb_entry(n, struct btrfs_free_space, offset_index);
606 		if (!prev)
607 			goto next;
608 		if (e->bitmap || prev->bitmap)
609 			goto next;
610 		if (prev->offset + prev->bytes == e->offset) {
611 			unlink_free_space(ctl, prev);
612 			unlink_free_space(ctl, e);
613 			prev->bytes += e->bytes;
614 			kmem_cache_free(btrfs_free_space_cachep, e);
615 			link_free_space(ctl, prev);
616 			prev = NULL;
617 			spin_unlock(&ctl->tree_lock);
618 			goto again;
619 		}
620 next:
621 		prev = e;
622 	}
623 	spin_unlock(&ctl->tree_lock);
624 }
625 
626 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
627 				   struct btrfs_free_space_ctl *ctl,
628 				   struct btrfs_path *path, u64 offset)
629 {
630 	struct btrfs_free_space_header *header;
631 	struct extent_buffer *leaf;
632 	struct io_ctl io_ctl;
633 	struct btrfs_key key;
634 	struct btrfs_free_space *e, *n;
635 	struct list_head bitmaps;
636 	u64 num_entries;
637 	u64 num_bitmaps;
638 	u64 generation;
639 	u8 type;
640 	int ret = 0;
641 
642 	INIT_LIST_HEAD(&bitmaps);
643 
644 	/* Nothing in the space cache, goodbye */
645 	if (!i_size_read(inode))
646 		return 0;
647 
648 	key.objectid = BTRFS_FREE_SPACE_OBJECTID;
649 	key.offset = offset;
650 	key.type = 0;
651 
652 	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
653 	if (ret < 0)
654 		return 0;
655 	else if (ret > 0) {
656 		btrfs_release_path(path);
657 		return 0;
658 	}
659 
660 	ret = -1;
661 
662 	leaf = path->nodes[0];
663 	header = btrfs_item_ptr(leaf, path->slots[0],
664 				struct btrfs_free_space_header);
665 	num_entries = btrfs_free_space_entries(leaf, header);
666 	num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
667 	generation = btrfs_free_space_generation(leaf, header);
668 	btrfs_release_path(path);
669 
670 	if (BTRFS_I(inode)->generation != generation) {
671 		btrfs_err(root->fs_info,
672 			"free space inode generation (%llu) "
673 			"did not match free space cache generation (%llu)",
674 			BTRFS_I(inode)->generation, generation);
675 		return 0;
676 	}
677 
678 	if (!num_entries)
679 		return 0;
680 
681 	ret = io_ctl_init(&io_ctl, inode, root);
682 	if (ret)
683 		return ret;
684 
685 	ret = readahead_cache(inode);
686 	if (ret)
687 		goto out;
688 
689 	ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
690 	if (ret)
691 		goto out;
692 
693 	ret = io_ctl_check_crc(&io_ctl, 0);
694 	if (ret)
695 		goto free_cache;
696 
697 	ret = io_ctl_check_generation(&io_ctl, generation);
698 	if (ret)
699 		goto free_cache;
700 
701 	while (num_entries) {
702 		e = kmem_cache_zalloc(btrfs_free_space_cachep,
703 				      GFP_NOFS);
704 		if (!e)
705 			goto free_cache;
706 
707 		ret = io_ctl_read_entry(&io_ctl, e, &type);
708 		if (ret) {
709 			kmem_cache_free(btrfs_free_space_cachep, e);
710 			goto free_cache;
711 		}
712 
713 		if (!e->bytes) {
714 			kmem_cache_free(btrfs_free_space_cachep, e);
715 			goto free_cache;
716 		}
717 
718 		if (type == BTRFS_FREE_SPACE_EXTENT) {
719 			spin_lock(&ctl->tree_lock);
720 			ret = link_free_space(ctl, e);
721 			spin_unlock(&ctl->tree_lock);
722 			if (ret) {
723 				btrfs_err(root->fs_info,
724 					"Duplicate entries in free space cache, dumping");
725 				kmem_cache_free(btrfs_free_space_cachep, e);
726 				goto free_cache;
727 			}
728 		} else {
729 			ASSERT(num_bitmaps);
730 			num_bitmaps--;
731 			e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
732 			if (!e->bitmap) {
733 				kmem_cache_free(
734 					btrfs_free_space_cachep, e);
735 				goto free_cache;
736 			}
737 			spin_lock(&ctl->tree_lock);
738 			ret = link_free_space(ctl, e);
739 			ctl->total_bitmaps++;
740 			ctl->op->recalc_thresholds(ctl);
741 			spin_unlock(&ctl->tree_lock);
742 			if (ret) {
743 				btrfs_err(root->fs_info,
744 					"Duplicate entries in free space cache, dumping");
745 				kmem_cache_free(btrfs_free_space_cachep, e);
746 				goto free_cache;
747 			}
748 			list_add_tail(&e->list, &bitmaps);
749 		}
750 
751 		num_entries--;
752 	}
753 
754 	io_ctl_unmap_page(&io_ctl);
755 
756 	/*
757 	 * We add the bitmaps at the end of the entries in order that
758 	 * the bitmap entries are added to the cache.
759 	 */
760 	list_for_each_entry_safe(e, n, &bitmaps, list) {
761 		list_del_init(&e->list);
762 		ret = io_ctl_read_bitmap(&io_ctl, e);
763 		if (ret)
764 			goto free_cache;
765 	}
766 
767 	io_ctl_drop_pages(&io_ctl);
768 	merge_space_tree(ctl);
769 	ret = 1;
770 out:
771 	io_ctl_free(&io_ctl);
772 	return ret;
773 free_cache:
774 	io_ctl_drop_pages(&io_ctl);
775 	__btrfs_remove_free_space_cache(ctl);
776 	goto out;
777 }
778 
779 int load_free_space_cache(struct btrfs_fs_info *fs_info,
780 			  struct btrfs_block_group_cache *block_group)
781 {
782 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
783 	struct btrfs_root *root = fs_info->tree_root;
784 	struct inode *inode;
785 	struct btrfs_path *path;
786 	int ret = 0;
787 	bool matched;
788 	u64 used = btrfs_block_group_used(&block_group->item);
789 
790 	/*
791 	 * If this block group has been marked to be cleared for one reason or
792 	 * another then we can't trust the on disk cache, so just return.
793 	 */
794 	spin_lock(&block_group->lock);
795 	if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
796 		spin_unlock(&block_group->lock);
797 		return 0;
798 	}
799 	spin_unlock(&block_group->lock);
800 
801 	path = btrfs_alloc_path();
802 	if (!path)
803 		return 0;
804 	path->search_commit_root = 1;
805 	path->skip_locking = 1;
806 
807 	inode = lookup_free_space_inode(root, block_group, path);
808 	if (IS_ERR(inode)) {
809 		btrfs_free_path(path);
810 		return 0;
811 	}
812 
813 	/* We may have converted the inode and made the cache invalid. */
814 	spin_lock(&block_group->lock);
815 	if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
816 		spin_unlock(&block_group->lock);
817 		btrfs_free_path(path);
818 		goto out;
819 	}
820 	spin_unlock(&block_group->lock);
821 
822 	ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
823 				      path, block_group->key.objectid);
824 	btrfs_free_path(path);
825 	if (ret <= 0)
826 		goto out;
827 
828 	spin_lock(&ctl->tree_lock);
829 	matched = (ctl->free_space == (block_group->key.offset - used -
830 				       block_group->bytes_super));
831 	spin_unlock(&ctl->tree_lock);
832 
833 	if (!matched) {
834 		__btrfs_remove_free_space_cache(ctl);
835 		btrfs_err(fs_info, "block group %llu has wrong amount of free space",
836 			block_group->key.objectid);
837 		ret = -1;
838 	}
839 out:
840 	if (ret < 0) {
841 		/* This cache is bogus, make sure it gets cleared */
842 		spin_lock(&block_group->lock);
843 		block_group->disk_cache_state = BTRFS_DC_CLEAR;
844 		spin_unlock(&block_group->lock);
845 		ret = 0;
846 
847 		btrfs_err(fs_info, "failed to load free space cache for block group %llu",
848 			block_group->key.objectid);
849 	}
850 
851 	iput(inode);
852 	return ret;
853 }
854 
855 /**
856  * __btrfs_write_out_cache - write out cached info to an inode
857  * @root - the root the inode belongs to
858  * @ctl - the free space cache we are going to write out
859  * @block_group - the block_group for this cache if it belongs to a block_group
860  * @trans - the trans handle
861  * @path - the path to use
862  * @offset - the offset for the key we'll insert
863  *
864  * This function writes out a free space cache struct to disk for quick recovery
865  * on mount.  This will return 0 if it was successfull in writing the cache out,
866  * and -1 if it was not.
867  */
868 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
869 				   struct btrfs_free_space_ctl *ctl,
870 				   struct btrfs_block_group_cache *block_group,
871 				   struct btrfs_trans_handle *trans,
872 				   struct btrfs_path *path, u64 offset)
873 {
874 	struct btrfs_free_space_header *header;
875 	struct extent_buffer *leaf;
876 	struct rb_node *node;
877 	struct list_head *pos, *n;
878 	struct extent_state *cached_state = NULL;
879 	struct btrfs_free_cluster *cluster = NULL;
880 	struct extent_io_tree *unpin = NULL;
881 	struct io_ctl io_ctl;
882 	struct list_head bitmap_list;
883 	struct btrfs_key key;
884 	u64 start, extent_start, extent_end, len;
885 	int entries = 0;
886 	int bitmaps = 0;
887 	int ret;
888 	int err = -1;
889 
890 	INIT_LIST_HEAD(&bitmap_list);
891 
892 	if (!i_size_read(inode))
893 		return -1;
894 
895 	ret = io_ctl_init(&io_ctl, inode, root);
896 	if (ret)
897 		return -1;
898 
899 	/* Get the cluster for this block_group if it exists */
900 	if (block_group && !list_empty(&block_group->cluster_list))
901 		cluster = list_entry(block_group->cluster_list.next,
902 				     struct btrfs_free_cluster,
903 				     block_group_list);
904 
905 	/* Lock all pages first so we can lock the extent safely. */
906 	io_ctl_prepare_pages(&io_ctl, inode, 0);
907 
908 	lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
909 			 0, &cached_state);
910 
911 	node = rb_first(&ctl->free_space_offset);
912 	if (!node && cluster) {
913 		node = rb_first(&cluster->root);
914 		cluster = NULL;
915 	}
916 
917 	/* Make sure we can fit our crcs into the first page */
918 	if (io_ctl.check_crcs &&
919 	    (io_ctl.num_pages * sizeof(u32)) >= PAGE_CACHE_SIZE)
920 		goto out_nospc;
921 
922 	io_ctl_set_generation(&io_ctl, trans->transid);
923 
924 	/* Write out the extent entries */
925 	while (node) {
926 		struct btrfs_free_space *e;
927 
928 		e = rb_entry(node, struct btrfs_free_space, offset_index);
929 		entries++;
930 
931 		ret = io_ctl_add_entry(&io_ctl, e->offset, e->bytes,
932 				       e->bitmap);
933 		if (ret)
934 			goto out_nospc;
935 
936 		if (e->bitmap) {
937 			list_add_tail(&e->list, &bitmap_list);
938 			bitmaps++;
939 		}
940 		node = rb_next(node);
941 		if (!node && cluster) {
942 			node = rb_first(&cluster->root);
943 			cluster = NULL;
944 		}
945 	}
946 
947 	/*
948 	 * We want to add any pinned extents to our free space cache
949 	 * so we don't leak the space
950 	 */
951 
952 	/*
953 	 * We shouldn't have switched the pinned extents yet so this is the
954 	 * right one
955 	 */
956 	unpin = root->fs_info->pinned_extents;
957 
958 	if (block_group)
959 		start = block_group->key.objectid;
960 
961 	while (block_group && (start < block_group->key.objectid +
962 			       block_group->key.offset)) {
963 		ret = find_first_extent_bit(unpin, start,
964 					    &extent_start, &extent_end,
965 					    EXTENT_DIRTY, NULL);
966 		if (ret) {
967 			ret = 0;
968 			break;
969 		}
970 
971 		/* This pinned extent is out of our range */
972 		if (extent_start >= block_group->key.objectid +
973 		    block_group->key.offset)
974 			break;
975 
976 		extent_start = max(extent_start, start);
977 		extent_end = min(block_group->key.objectid +
978 				 block_group->key.offset, extent_end + 1);
979 		len = extent_end - extent_start;
980 
981 		entries++;
982 		ret = io_ctl_add_entry(&io_ctl, extent_start, len, NULL);
983 		if (ret)
984 			goto out_nospc;
985 
986 		start = extent_end;
987 	}
988 
989 	/* Write out the bitmaps */
990 	list_for_each_safe(pos, n, &bitmap_list) {
991 		struct btrfs_free_space *entry =
992 			list_entry(pos, struct btrfs_free_space, list);
993 
994 		ret = io_ctl_add_bitmap(&io_ctl, entry->bitmap);
995 		if (ret)
996 			goto out_nospc;
997 		list_del_init(&entry->list);
998 	}
999 
1000 	/* Zero out the rest of the pages just to make sure */
1001 	io_ctl_zero_remaining_pages(&io_ctl);
1002 
1003 	ret = btrfs_dirty_pages(root, inode, io_ctl.pages, io_ctl.num_pages,
1004 				0, i_size_read(inode), &cached_state);
1005 	io_ctl_drop_pages(&io_ctl);
1006 	unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1007 			     i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1008 
1009 	if (ret)
1010 		goto out;
1011 
1012 
1013 	btrfs_wait_ordered_range(inode, 0, (u64)-1);
1014 
1015 	key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1016 	key.offset = offset;
1017 	key.type = 0;
1018 
1019 	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1020 	if (ret < 0) {
1021 		clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1022 				 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1023 				 GFP_NOFS);
1024 		goto out;
1025 	}
1026 	leaf = path->nodes[0];
1027 	if (ret > 0) {
1028 		struct btrfs_key found_key;
1029 		ASSERT(path->slots[0]);
1030 		path->slots[0]--;
1031 		btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1032 		if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1033 		    found_key.offset != offset) {
1034 			clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1035 					 inode->i_size - 1,
1036 					 EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1037 					 NULL, GFP_NOFS);
1038 			btrfs_release_path(path);
1039 			goto out;
1040 		}
1041 	}
1042 
1043 	BTRFS_I(inode)->generation = trans->transid;
1044 	header = btrfs_item_ptr(leaf, path->slots[0],
1045 				struct btrfs_free_space_header);
1046 	btrfs_set_free_space_entries(leaf, header, entries);
1047 	btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1048 	btrfs_set_free_space_generation(leaf, header, trans->transid);
1049 	btrfs_mark_buffer_dirty(leaf);
1050 	btrfs_release_path(path);
1051 
1052 	err = 0;
1053 out:
1054 	io_ctl_free(&io_ctl);
1055 	if (err) {
1056 		invalidate_inode_pages2(inode->i_mapping);
1057 		BTRFS_I(inode)->generation = 0;
1058 	}
1059 	btrfs_update_inode(trans, root, inode);
1060 	return err;
1061 
1062 out_nospc:
1063 	list_for_each_safe(pos, n, &bitmap_list) {
1064 		struct btrfs_free_space *entry =
1065 			list_entry(pos, struct btrfs_free_space, list);
1066 		list_del_init(&entry->list);
1067 	}
1068 	io_ctl_drop_pages(&io_ctl);
1069 	unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1070 			     i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1071 	goto out;
1072 }
1073 
1074 int btrfs_write_out_cache(struct btrfs_root *root,
1075 			  struct btrfs_trans_handle *trans,
1076 			  struct btrfs_block_group_cache *block_group,
1077 			  struct btrfs_path *path)
1078 {
1079 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1080 	struct inode *inode;
1081 	int ret = 0;
1082 
1083 	root = root->fs_info->tree_root;
1084 
1085 	spin_lock(&block_group->lock);
1086 	if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1087 		spin_unlock(&block_group->lock);
1088 		return 0;
1089 	}
1090 	spin_unlock(&block_group->lock);
1091 
1092 	inode = lookup_free_space_inode(root, block_group, path);
1093 	if (IS_ERR(inode))
1094 		return 0;
1095 
1096 	ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
1097 				      path, block_group->key.objectid);
1098 	if (ret) {
1099 		spin_lock(&block_group->lock);
1100 		block_group->disk_cache_state = BTRFS_DC_ERROR;
1101 		spin_unlock(&block_group->lock);
1102 		ret = 0;
1103 #ifdef DEBUG
1104 		btrfs_err(root->fs_info,
1105 			"failed to write free space cache for block group %llu",
1106 			block_group->key.objectid);
1107 #endif
1108 	}
1109 
1110 	iput(inode);
1111 	return ret;
1112 }
1113 
1114 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1115 					  u64 offset)
1116 {
1117 	ASSERT(offset >= bitmap_start);
1118 	offset -= bitmap_start;
1119 	return (unsigned long)(div_u64(offset, unit));
1120 }
1121 
1122 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1123 {
1124 	return (unsigned long)(div_u64(bytes, unit));
1125 }
1126 
1127 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1128 				   u64 offset)
1129 {
1130 	u64 bitmap_start;
1131 	u64 bytes_per_bitmap;
1132 
1133 	bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1134 	bitmap_start = offset - ctl->start;
1135 	bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1136 	bitmap_start *= bytes_per_bitmap;
1137 	bitmap_start += ctl->start;
1138 
1139 	return bitmap_start;
1140 }
1141 
1142 static int tree_insert_offset(struct rb_root *root, u64 offset,
1143 			      struct rb_node *node, int bitmap)
1144 {
1145 	struct rb_node **p = &root->rb_node;
1146 	struct rb_node *parent = NULL;
1147 	struct btrfs_free_space *info;
1148 
1149 	while (*p) {
1150 		parent = *p;
1151 		info = rb_entry(parent, struct btrfs_free_space, offset_index);
1152 
1153 		if (offset < info->offset) {
1154 			p = &(*p)->rb_left;
1155 		} else if (offset > info->offset) {
1156 			p = &(*p)->rb_right;
1157 		} else {
1158 			/*
1159 			 * we could have a bitmap entry and an extent entry
1160 			 * share the same offset.  If this is the case, we want
1161 			 * the extent entry to always be found first if we do a
1162 			 * linear search through the tree, since we want to have
1163 			 * the quickest allocation time, and allocating from an
1164 			 * extent is faster than allocating from a bitmap.  So
1165 			 * if we're inserting a bitmap and we find an entry at
1166 			 * this offset, we want to go right, or after this entry
1167 			 * logically.  If we are inserting an extent and we've
1168 			 * found a bitmap, we want to go left, or before
1169 			 * logically.
1170 			 */
1171 			if (bitmap) {
1172 				if (info->bitmap) {
1173 					WARN_ON_ONCE(1);
1174 					return -EEXIST;
1175 				}
1176 				p = &(*p)->rb_right;
1177 			} else {
1178 				if (!info->bitmap) {
1179 					WARN_ON_ONCE(1);
1180 					return -EEXIST;
1181 				}
1182 				p = &(*p)->rb_left;
1183 			}
1184 		}
1185 	}
1186 
1187 	rb_link_node(node, parent, p);
1188 	rb_insert_color(node, root);
1189 
1190 	return 0;
1191 }
1192 
1193 /*
1194  * searches the tree for the given offset.
1195  *
1196  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1197  * want a section that has at least bytes size and comes at or after the given
1198  * offset.
1199  */
1200 static struct btrfs_free_space *
1201 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1202 		   u64 offset, int bitmap_only, int fuzzy)
1203 {
1204 	struct rb_node *n = ctl->free_space_offset.rb_node;
1205 	struct btrfs_free_space *entry, *prev = NULL;
1206 
1207 	/* find entry that is closest to the 'offset' */
1208 	while (1) {
1209 		if (!n) {
1210 			entry = NULL;
1211 			break;
1212 		}
1213 
1214 		entry = rb_entry(n, struct btrfs_free_space, offset_index);
1215 		prev = entry;
1216 
1217 		if (offset < entry->offset)
1218 			n = n->rb_left;
1219 		else if (offset > entry->offset)
1220 			n = n->rb_right;
1221 		else
1222 			break;
1223 	}
1224 
1225 	if (bitmap_only) {
1226 		if (!entry)
1227 			return NULL;
1228 		if (entry->bitmap)
1229 			return entry;
1230 
1231 		/*
1232 		 * bitmap entry and extent entry may share same offset,
1233 		 * in that case, bitmap entry comes after extent entry.
1234 		 */
1235 		n = rb_next(n);
1236 		if (!n)
1237 			return NULL;
1238 		entry = rb_entry(n, struct btrfs_free_space, offset_index);
1239 		if (entry->offset != offset)
1240 			return NULL;
1241 
1242 		WARN_ON(!entry->bitmap);
1243 		return entry;
1244 	} else if (entry) {
1245 		if (entry->bitmap) {
1246 			/*
1247 			 * if previous extent entry covers the offset,
1248 			 * we should return it instead of the bitmap entry
1249 			 */
1250 			n = rb_prev(&entry->offset_index);
1251 			if (n) {
1252 				prev = rb_entry(n, struct btrfs_free_space,
1253 						offset_index);
1254 				if (!prev->bitmap &&
1255 				    prev->offset + prev->bytes > offset)
1256 					entry = prev;
1257 			}
1258 		}
1259 		return entry;
1260 	}
1261 
1262 	if (!prev)
1263 		return NULL;
1264 
1265 	/* find last entry before the 'offset' */
1266 	entry = prev;
1267 	if (entry->offset > offset) {
1268 		n = rb_prev(&entry->offset_index);
1269 		if (n) {
1270 			entry = rb_entry(n, struct btrfs_free_space,
1271 					offset_index);
1272 			ASSERT(entry->offset <= offset);
1273 		} else {
1274 			if (fuzzy)
1275 				return entry;
1276 			else
1277 				return NULL;
1278 		}
1279 	}
1280 
1281 	if (entry->bitmap) {
1282 		n = rb_prev(&entry->offset_index);
1283 		if (n) {
1284 			prev = rb_entry(n, struct btrfs_free_space,
1285 					offset_index);
1286 			if (!prev->bitmap &&
1287 			    prev->offset + prev->bytes > offset)
1288 				return prev;
1289 		}
1290 		if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1291 			return entry;
1292 	} else if (entry->offset + entry->bytes > offset)
1293 		return entry;
1294 
1295 	if (!fuzzy)
1296 		return NULL;
1297 
1298 	while (1) {
1299 		if (entry->bitmap) {
1300 			if (entry->offset + BITS_PER_BITMAP *
1301 			    ctl->unit > offset)
1302 				break;
1303 		} else {
1304 			if (entry->offset + entry->bytes > offset)
1305 				break;
1306 		}
1307 
1308 		n = rb_next(&entry->offset_index);
1309 		if (!n)
1310 			return NULL;
1311 		entry = rb_entry(n, struct btrfs_free_space, offset_index);
1312 	}
1313 	return entry;
1314 }
1315 
1316 static inline void
1317 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1318 		    struct btrfs_free_space *info)
1319 {
1320 	rb_erase(&info->offset_index, &ctl->free_space_offset);
1321 	ctl->free_extents--;
1322 }
1323 
1324 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1325 			      struct btrfs_free_space *info)
1326 {
1327 	__unlink_free_space(ctl, info);
1328 	ctl->free_space -= info->bytes;
1329 }
1330 
1331 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1332 			   struct btrfs_free_space *info)
1333 {
1334 	int ret = 0;
1335 
1336 	ASSERT(info->bytes || info->bitmap);
1337 	ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1338 				 &info->offset_index, (info->bitmap != NULL));
1339 	if (ret)
1340 		return ret;
1341 
1342 	ctl->free_space += info->bytes;
1343 	ctl->free_extents++;
1344 	return ret;
1345 }
1346 
1347 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1348 {
1349 	struct btrfs_block_group_cache *block_group = ctl->private;
1350 	u64 max_bytes;
1351 	u64 bitmap_bytes;
1352 	u64 extent_bytes;
1353 	u64 size = block_group->key.offset;
1354 	u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1355 	int max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1356 
1357 	max_bitmaps = max(max_bitmaps, 1);
1358 
1359 	ASSERT(ctl->total_bitmaps <= max_bitmaps);
1360 
1361 	/*
1362 	 * The goal is to keep the total amount of memory used per 1gb of space
1363 	 * at or below 32k, so we need to adjust how much memory we allow to be
1364 	 * used by extent based free space tracking
1365 	 */
1366 	if (size < 1024 * 1024 * 1024)
1367 		max_bytes = MAX_CACHE_BYTES_PER_GIG;
1368 	else
1369 		max_bytes = MAX_CACHE_BYTES_PER_GIG *
1370 			div64_u64(size, 1024 * 1024 * 1024);
1371 
1372 	/*
1373 	 * we want to account for 1 more bitmap than what we have so we can make
1374 	 * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1375 	 * we add more bitmaps.
1376 	 */
1377 	bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1378 
1379 	if (bitmap_bytes >= max_bytes) {
1380 		ctl->extents_thresh = 0;
1381 		return;
1382 	}
1383 
1384 	/*
1385 	 * we want the extent entry threshold to always be at most 1/2 the maxw
1386 	 * bytes we can have, or whatever is less than that.
1387 	 */
1388 	extent_bytes = max_bytes - bitmap_bytes;
1389 	extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1390 
1391 	ctl->extents_thresh =
1392 		div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1393 }
1394 
1395 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1396 				       struct btrfs_free_space *info,
1397 				       u64 offset, u64 bytes)
1398 {
1399 	unsigned long start, count;
1400 
1401 	start = offset_to_bit(info->offset, ctl->unit, offset);
1402 	count = bytes_to_bits(bytes, ctl->unit);
1403 	ASSERT(start + count <= BITS_PER_BITMAP);
1404 
1405 	bitmap_clear(info->bitmap, start, count);
1406 
1407 	info->bytes -= bytes;
1408 }
1409 
1410 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1411 			      struct btrfs_free_space *info, u64 offset,
1412 			      u64 bytes)
1413 {
1414 	__bitmap_clear_bits(ctl, info, offset, bytes);
1415 	ctl->free_space -= bytes;
1416 }
1417 
1418 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1419 			    struct btrfs_free_space *info, u64 offset,
1420 			    u64 bytes)
1421 {
1422 	unsigned long start, count;
1423 
1424 	start = offset_to_bit(info->offset, ctl->unit, offset);
1425 	count = bytes_to_bits(bytes, ctl->unit);
1426 	ASSERT(start + count <= BITS_PER_BITMAP);
1427 
1428 	bitmap_set(info->bitmap, start, count);
1429 
1430 	info->bytes += bytes;
1431 	ctl->free_space += bytes;
1432 }
1433 
1434 /*
1435  * If we can not find suitable extent, we will use bytes to record
1436  * the size of the max extent.
1437  */
1438 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1439 			 struct btrfs_free_space *bitmap_info, u64 *offset,
1440 			 u64 *bytes)
1441 {
1442 	unsigned long found_bits = 0;
1443 	unsigned long max_bits = 0;
1444 	unsigned long bits, i;
1445 	unsigned long next_zero;
1446 	unsigned long extent_bits;
1447 
1448 	i = offset_to_bit(bitmap_info->offset, ctl->unit,
1449 			  max_t(u64, *offset, bitmap_info->offset));
1450 	bits = bytes_to_bits(*bytes, ctl->unit);
1451 
1452 	for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1453 		next_zero = find_next_zero_bit(bitmap_info->bitmap,
1454 					       BITS_PER_BITMAP, i);
1455 		extent_bits = next_zero - i;
1456 		if (extent_bits >= bits) {
1457 			found_bits = extent_bits;
1458 			break;
1459 		} else if (extent_bits > max_bits) {
1460 			max_bits = extent_bits;
1461 		}
1462 		i = next_zero;
1463 	}
1464 
1465 	if (found_bits) {
1466 		*offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1467 		*bytes = (u64)(found_bits) * ctl->unit;
1468 		return 0;
1469 	}
1470 
1471 	*bytes = (u64)(max_bits) * ctl->unit;
1472 	return -1;
1473 }
1474 
1475 /* Cache the size of the max extent in bytes */
1476 static struct btrfs_free_space *
1477 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1478 		unsigned long align, u64 *max_extent_size)
1479 {
1480 	struct btrfs_free_space *entry;
1481 	struct rb_node *node;
1482 	u64 tmp;
1483 	u64 align_off;
1484 	int ret;
1485 
1486 	if (!ctl->free_space_offset.rb_node)
1487 		goto out;
1488 
1489 	entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1490 	if (!entry)
1491 		goto out;
1492 
1493 	for (node = &entry->offset_index; node; node = rb_next(node)) {
1494 		entry = rb_entry(node, struct btrfs_free_space, offset_index);
1495 		if (entry->bytes < *bytes) {
1496 			if (entry->bytes > *max_extent_size)
1497 				*max_extent_size = entry->bytes;
1498 			continue;
1499 		}
1500 
1501 		/* make sure the space returned is big enough
1502 		 * to match our requested alignment
1503 		 */
1504 		if (*bytes >= align) {
1505 			tmp = entry->offset - ctl->start + align - 1;
1506 			do_div(tmp, align);
1507 			tmp = tmp * align + ctl->start;
1508 			align_off = tmp - entry->offset;
1509 		} else {
1510 			align_off = 0;
1511 			tmp = entry->offset;
1512 		}
1513 
1514 		if (entry->bytes < *bytes + align_off) {
1515 			if (entry->bytes > *max_extent_size)
1516 				*max_extent_size = entry->bytes;
1517 			continue;
1518 		}
1519 
1520 		if (entry->bitmap) {
1521 			u64 size = *bytes;
1522 
1523 			ret = search_bitmap(ctl, entry, &tmp, &size);
1524 			if (!ret) {
1525 				*offset = tmp;
1526 				*bytes = size;
1527 				return entry;
1528 			} else if (size > *max_extent_size) {
1529 				*max_extent_size = size;
1530 			}
1531 			continue;
1532 		}
1533 
1534 		*offset = tmp;
1535 		*bytes = entry->bytes - align_off;
1536 		return entry;
1537 	}
1538 out:
1539 	return NULL;
1540 }
1541 
1542 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1543 			   struct btrfs_free_space *info, u64 offset)
1544 {
1545 	info->offset = offset_to_bitmap(ctl, offset);
1546 	info->bytes = 0;
1547 	INIT_LIST_HEAD(&info->list);
1548 	link_free_space(ctl, info);
1549 	ctl->total_bitmaps++;
1550 
1551 	ctl->op->recalc_thresholds(ctl);
1552 }
1553 
1554 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1555 			struct btrfs_free_space *bitmap_info)
1556 {
1557 	unlink_free_space(ctl, bitmap_info);
1558 	kfree(bitmap_info->bitmap);
1559 	kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1560 	ctl->total_bitmaps--;
1561 	ctl->op->recalc_thresholds(ctl);
1562 }
1563 
1564 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1565 			      struct btrfs_free_space *bitmap_info,
1566 			      u64 *offset, u64 *bytes)
1567 {
1568 	u64 end;
1569 	u64 search_start, search_bytes;
1570 	int ret;
1571 
1572 again:
1573 	end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1574 
1575 	/*
1576 	 * We need to search for bits in this bitmap.  We could only cover some
1577 	 * of the extent in this bitmap thanks to how we add space, so we need
1578 	 * to search for as much as it as we can and clear that amount, and then
1579 	 * go searching for the next bit.
1580 	 */
1581 	search_start = *offset;
1582 	search_bytes = ctl->unit;
1583 	search_bytes = min(search_bytes, end - search_start + 1);
1584 	ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1585 	if (ret < 0 || search_start != *offset)
1586 		return -EINVAL;
1587 
1588 	/* We may have found more bits than what we need */
1589 	search_bytes = min(search_bytes, *bytes);
1590 
1591 	/* Cannot clear past the end of the bitmap */
1592 	search_bytes = min(search_bytes, end - search_start + 1);
1593 
1594 	bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1595 	*offset += search_bytes;
1596 	*bytes -= search_bytes;
1597 
1598 	if (*bytes) {
1599 		struct rb_node *next = rb_next(&bitmap_info->offset_index);
1600 		if (!bitmap_info->bytes)
1601 			free_bitmap(ctl, bitmap_info);
1602 
1603 		/*
1604 		 * no entry after this bitmap, but we still have bytes to
1605 		 * remove, so something has gone wrong.
1606 		 */
1607 		if (!next)
1608 			return -EINVAL;
1609 
1610 		bitmap_info = rb_entry(next, struct btrfs_free_space,
1611 				       offset_index);
1612 
1613 		/*
1614 		 * if the next entry isn't a bitmap we need to return to let the
1615 		 * extent stuff do its work.
1616 		 */
1617 		if (!bitmap_info->bitmap)
1618 			return -EAGAIN;
1619 
1620 		/*
1621 		 * Ok the next item is a bitmap, but it may not actually hold
1622 		 * the information for the rest of this free space stuff, so
1623 		 * look for it, and if we don't find it return so we can try
1624 		 * everything over again.
1625 		 */
1626 		search_start = *offset;
1627 		search_bytes = ctl->unit;
1628 		ret = search_bitmap(ctl, bitmap_info, &search_start,
1629 				    &search_bytes);
1630 		if (ret < 0 || search_start != *offset)
1631 			return -EAGAIN;
1632 
1633 		goto again;
1634 	} else if (!bitmap_info->bytes)
1635 		free_bitmap(ctl, bitmap_info);
1636 
1637 	return 0;
1638 }
1639 
1640 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1641 			       struct btrfs_free_space *info, u64 offset,
1642 			       u64 bytes)
1643 {
1644 	u64 bytes_to_set = 0;
1645 	u64 end;
1646 
1647 	end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1648 
1649 	bytes_to_set = min(end - offset, bytes);
1650 
1651 	bitmap_set_bits(ctl, info, offset, bytes_to_set);
1652 
1653 	return bytes_to_set;
1654 
1655 }
1656 
1657 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1658 		      struct btrfs_free_space *info)
1659 {
1660 	struct btrfs_block_group_cache *block_group = ctl->private;
1661 
1662 	/*
1663 	 * If we are below the extents threshold then we can add this as an
1664 	 * extent, and don't have to deal with the bitmap
1665 	 */
1666 	if (ctl->free_extents < ctl->extents_thresh) {
1667 		/*
1668 		 * If this block group has some small extents we don't want to
1669 		 * use up all of our free slots in the cache with them, we want
1670 		 * to reserve them to larger extents, however if we have plent
1671 		 * of cache left then go ahead an dadd them, no sense in adding
1672 		 * the overhead of a bitmap if we don't have to.
1673 		 */
1674 		if (info->bytes <= block_group->sectorsize * 4) {
1675 			if (ctl->free_extents * 2 <= ctl->extents_thresh)
1676 				return false;
1677 		} else {
1678 			return false;
1679 		}
1680 	}
1681 
1682 	/*
1683 	 * The original block groups from mkfs can be really small, like 8
1684 	 * megabytes, so don't bother with a bitmap for those entries.  However
1685 	 * some block groups can be smaller than what a bitmap would cover but
1686 	 * are still large enough that they could overflow the 32k memory limit,
1687 	 * so allow those block groups to still be allowed to have a bitmap
1688 	 * entry.
1689 	 */
1690 	if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
1691 		return false;
1692 
1693 	return true;
1694 }
1695 
1696 static struct btrfs_free_space_op free_space_op = {
1697 	.recalc_thresholds	= recalculate_thresholds,
1698 	.use_bitmap		= use_bitmap,
1699 };
1700 
1701 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1702 			      struct btrfs_free_space *info)
1703 {
1704 	struct btrfs_free_space *bitmap_info;
1705 	struct btrfs_block_group_cache *block_group = NULL;
1706 	int added = 0;
1707 	u64 bytes, offset, bytes_added;
1708 	int ret;
1709 
1710 	bytes = info->bytes;
1711 	offset = info->offset;
1712 
1713 	if (!ctl->op->use_bitmap(ctl, info))
1714 		return 0;
1715 
1716 	if (ctl->op == &free_space_op)
1717 		block_group = ctl->private;
1718 again:
1719 	/*
1720 	 * Since we link bitmaps right into the cluster we need to see if we
1721 	 * have a cluster here, and if so and it has our bitmap we need to add
1722 	 * the free space to that bitmap.
1723 	 */
1724 	if (block_group && !list_empty(&block_group->cluster_list)) {
1725 		struct btrfs_free_cluster *cluster;
1726 		struct rb_node *node;
1727 		struct btrfs_free_space *entry;
1728 
1729 		cluster = list_entry(block_group->cluster_list.next,
1730 				     struct btrfs_free_cluster,
1731 				     block_group_list);
1732 		spin_lock(&cluster->lock);
1733 		node = rb_first(&cluster->root);
1734 		if (!node) {
1735 			spin_unlock(&cluster->lock);
1736 			goto no_cluster_bitmap;
1737 		}
1738 
1739 		entry = rb_entry(node, struct btrfs_free_space, offset_index);
1740 		if (!entry->bitmap) {
1741 			spin_unlock(&cluster->lock);
1742 			goto no_cluster_bitmap;
1743 		}
1744 
1745 		if (entry->offset == offset_to_bitmap(ctl, offset)) {
1746 			bytes_added = add_bytes_to_bitmap(ctl, entry,
1747 							  offset, bytes);
1748 			bytes -= bytes_added;
1749 			offset += bytes_added;
1750 		}
1751 		spin_unlock(&cluster->lock);
1752 		if (!bytes) {
1753 			ret = 1;
1754 			goto out;
1755 		}
1756 	}
1757 
1758 no_cluster_bitmap:
1759 	bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1760 					 1, 0);
1761 	if (!bitmap_info) {
1762 		ASSERT(added == 0);
1763 		goto new_bitmap;
1764 	}
1765 
1766 	bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
1767 	bytes -= bytes_added;
1768 	offset += bytes_added;
1769 	added = 0;
1770 
1771 	if (!bytes) {
1772 		ret = 1;
1773 		goto out;
1774 	} else
1775 		goto again;
1776 
1777 new_bitmap:
1778 	if (info && info->bitmap) {
1779 		add_new_bitmap(ctl, info, offset);
1780 		added = 1;
1781 		info = NULL;
1782 		goto again;
1783 	} else {
1784 		spin_unlock(&ctl->tree_lock);
1785 
1786 		/* no pre-allocated info, allocate a new one */
1787 		if (!info) {
1788 			info = kmem_cache_zalloc(btrfs_free_space_cachep,
1789 						 GFP_NOFS);
1790 			if (!info) {
1791 				spin_lock(&ctl->tree_lock);
1792 				ret = -ENOMEM;
1793 				goto out;
1794 			}
1795 		}
1796 
1797 		/* allocate the bitmap */
1798 		info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1799 		spin_lock(&ctl->tree_lock);
1800 		if (!info->bitmap) {
1801 			ret = -ENOMEM;
1802 			goto out;
1803 		}
1804 		goto again;
1805 	}
1806 
1807 out:
1808 	if (info) {
1809 		if (info->bitmap)
1810 			kfree(info->bitmap);
1811 		kmem_cache_free(btrfs_free_space_cachep, info);
1812 	}
1813 
1814 	return ret;
1815 }
1816 
1817 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
1818 			  struct btrfs_free_space *info, bool update_stat)
1819 {
1820 	struct btrfs_free_space *left_info;
1821 	struct btrfs_free_space *right_info;
1822 	bool merged = false;
1823 	u64 offset = info->offset;
1824 	u64 bytes = info->bytes;
1825 
1826 	/*
1827 	 * first we want to see if there is free space adjacent to the range we
1828 	 * are adding, if there is remove that struct and add a new one to
1829 	 * cover the entire range
1830 	 */
1831 	right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
1832 	if (right_info && rb_prev(&right_info->offset_index))
1833 		left_info = rb_entry(rb_prev(&right_info->offset_index),
1834 				     struct btrfs_free_space, offset_index);
1835 	else
1836 		left_info = tree_search_offset(ctl, offset - 1, 0, 0);
1837 
1838 	if (right_info && !right_info->bitmap) {
1839 		if (update_stat)
1840 			unlink_free_space(ctl, right_info);
1841 		else
1842 			__unlink_free_space(ctl, right_info);
1843 		info->bytes += right_info->bytes;
1844 		kmem_cache_free(btrfs_free_space_cachep, right_info);
1845 		merged = true;
1846 	}
1847 
1848 	if (left_info && !left_info->bitmap &&
1849 	    left_info->offset + left_info->bytes == offset) {
1850 		if (update_stat)
1851 			unlink_free_space(ctl, left_info);
1852 		else
1853 			__unlink_free_space(ctl, left_info);
1854 		info->offset = left_info->offset;
1855 		info->bytes += left_info->bytes;
1856 		kmem_cache_free(btrfs_free_space_cachep, left_info);
1857 		merged = true;
1858 	}
1859 
1860 	return merged;
1861 }
1862 
1863 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
1864 			   u64 offset, u64 bytes)
1865 {
1866 	struct btrfs_free_space *info;
1867 	int ret = 0;
1868 
1869 	info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1870 	if (!info)
1871 		return -ENOMEM;
1872 
1873 	info->offset = offset;
1874 	info->bytes = bytes;
1875 
1876 	spin_lock(&ctl->tree_lock);
1877 
1878 	if (try_merge_free_space(ctl, info, true))
1879 		goto link;
1880 
1881 	/*
1882 	 * There was no extent directly to the left or right of this new
1883 	 * extent then we know we're going to have to allocate a new extent, so
1884 	 * before we do that see if we need to drop this into a bitmap
1885 	 */
1886 	ret = insert_into_bitmap(ctl, info);
1887 	if (ret < 0) {
1888 		goto out;
1889 	} else if (ret) {
1890 		ret = 0;
1891 		goto out;
1892 	}
1893 link:
1894 	ret = link_free_space(ctl, info);
1895 	if (ret)
1896 		kmem_cache_free(btrfs_free_space_cachep, info);
1897 out:
1898 	spin_unlock(&ctl->tree_lock);
1899 
1900 	if (ret) {
1901 		printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1902 		ASSERT(ret != -EEXIST);
1903 	}
1904 
1905 	return ret;
1906 }
1907 
1908 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1909 			    u64 offset, u64 bytes)
1910 {
1911 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1912 	struct btrfs_free_space *info;
1913 	int ret;
1914 	bool re_search = false;
1915 
1916 	spin_lock(&ctl->tree_lock);
1917 
1918 again:
1919 	ret = 0;
1920 	if (!bytes)
1921 		goto out_lock;
1922 
1923 	info = tree_search_offset(ctl, offset, 0, 0);
1924 	if (!info) {
1925 		/*
1926 		 * oops didn't find an extent that matched the space we wanted
1927 		 * to remove, look for a bitmap instead
1928 		 */
1929 		info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1930 					  1, 0);
1931 		if (!info) {
1932 			/*
1933 			 * If we found a partial bit of our free space in a
1934 			 * bitmap but then couldn't find the other part this may
1935 			 * be a problem, so WARN about it.
1936 			 */
1937 			WARN_ON(re_search);
1938 			goto out_lock;
1939 		}
1940 	}
1941 
1942 	re_search = false;
1943 	if (!info->bitmap) {
1944 		unlink_free_space(ctl, info);
1945 		if (offset == info->offset) {
1946 			u64 to_free = min(bytes, info->bytes);
1947 
1948 			info->bytes -= to_free;
1949 			info->offset += to_free;
1950 			if (info->bytes) {
1951 				ret = link_free_space(ctl, info);
1952 				WARN_ON(ret);
1953 			} else {
1954 				kmem_cache_free(btrfs_free_space_cachep, info);
1955 			}
1956 
1957 			offset += to_free;
1958 			bytes -= to_free;
1959 			goto again;
1960 		} else {
1961 			u64 old_end = info->bytes + info->offset;
1962 
1963 			info->bytes = offset - info->offset;
1964 			ret = link_free_space(ctl, info);
1965 			WARN_ON(ret);
1966 			if (ret)
1967 				goto out_lock;
1968 
1969 			/* Not enough bytes in this entry to satisfy us */
1970 			if (old_end < offset + bytes) {
1971 				bytes -= old_end - offset;
1972 				offset = old_end;
1973 				goto again;
1974 			} else if (old_end == offset + bytes) {
1975 				/* all done */
1976 				goto out_lock;
1977 			}
1978 			spin_unlock(&ctl->tree_lock);
1979 
1980 			ret = btrfs_add_free_space(block_group, offset + bytes,
1981 						   old_end - (offset + bytes));
1982 			WARN_ON(ret);
1983 			goto out;
1984 		}
1985 	}
1986 
1987 	ret = remove_from_bitmap(ctl, info, &offset, &bytes);
1988 	if (ret == -EAGAIN) {
1989 		re_search = true;
1990 		goto again;
1991 	}
1992 out_lock:
1993 	spin_unlock(&ctl->tree_lock);
1994 out:
1995 	return ret;
1996 }
1997 
1998 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1999 			   u64 bytes)
2000 {
2001 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2002 	struct btrfs_free_space *info;
2003 	struct rb_node *n;
2004 	int count = 0;
2005 
2006 	for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2007 		info = rb_entry(n, struct btrfs_free_space, offset_index);
2008 		if (info->bytes >= bytes && !block_group->ro)
2009 			count++;
2010 		printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
2011 		       info->offset, info->bytes,
2012 		       (info->bitmap) ? "yes" : "no");
2013 	}
2014 	printk(KERN_INFO "block group has cluster?: %s\n",
2015 	       list_empty(&block_group->cluster_list) ? "no" : "yes");
2016 	printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
2017 	       "\n", count);
2018 }
2019 
2020 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2021 {
2022 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2023 
2024 	spin_lock_init(&ctl->tree_lock);
2025 	ctl->unit = block_group->sectorsize;
2026 	ctl->start = block_group->key.objectid;
2027 	ctl->private = block_group;
2028 	ctl->op = &free_space_op;
2029 
2030 	/*
2031 	 * we only want to have 32k of ram per block group for keeping
2032 	 * track of free space, and if we pass 1/2 of that we want to
2033 	 * start converting things over to using bitmaps
2034 	 */
2035 	ctl->extents_thresh = ((1024 * 32) / 2) /
2036 				sizeof(struct btrfs_free_space);
2037 }
2038 
2039 /*
2040  * for a given cluster, put all of its extents back into the free
2041  * space cache.  If the block group passed doesn't match the block group
2042  * pointed to by the cluster, someone else raced in and freed the
2043  * cluster already.  In that case, we just return without changing anything
2044  */
2045 static int
2046 __btrfs_return_cluster_to_free_space(
2047 			     struct btrfs_block_group_cache *block_group,
2048 			     struct btrfs_free_cluster *cluster)
2049 {
2050 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2051 	struct btrfs_free_space *entry;
2052 	struct rb_node *node;
2053 
2054 	spin_lock(&cluster->lock);
2055 	if (cluster->block_group != block_group)
2056 		goto out;
2057 
2058 	cluster->block_group = NULL;
2059 	cluster->window_start = 0;
2060 	list_del_init(&cluster->block_group_list);
2061 
2062 	node = rb_first(&cluster->root);
2063 	while (node) {
2064 		bool bitmap;
2065 
2066 		entry = rb_entry(node, struct btrfs_free_space, offset_index);
2067 		node = rb_next(&entry->offset_index);
2068 		rb_erase(&entry->offset_index, &cluster->root);
2069 
2070 		bitmap = (entry->bitmap != NULL);
2071 		if (!bitmap)
2072 			try_merge_free_space(ctl, entry, false);
2073 		tree_insert_offset(&ctl->free_space_offset,
2074 				   entry->offset, &entry->offset_index, bitmap);
2075 	}
2076 	cluster->root = RB_ROOT;
2077 
2078 out:
2079 	spin_unlock(&cluster->lock);
2080 	btrfs_put_block_group(block_group);
2081 	return 0;
2082 }
2083 
2084 static void __btrfs_remove_free_space_cache_locked(
2085 				struct btrfs_free_space_ctl *ctl)
2086 {
2087 	struct btrfs_free_space *info;
2088 	struct rb_node *node;
2089 
2090 	while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2091 		info = rb_entry(node, struct btrfs_free_space, offset_index);
2092 		if (!info->bitmap) {
2093 			unlink_free_space(ctl, info);
2094 			kmem_cache_free(btrfs_free_space_cachep, info);
2095 		} else {
2096 			free_bitmap(ctl, info);
2097 		}
2098 		if (need_resched()) {
2099 			spin_unlock(&ctl->tree_lock);
2100 			cond_resched();
2101 			spin_lock(&ctl->tree_lock);
2102 		}
2103 	}
2104 }
2105 
2106 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2107 {
2108 	spin_lock(&ctl->tree_lock);
2109 	__btrfs_remove_free_space_cache_locked(ctl);
2110 	spin_unlock(&ctl->tree_lock);
2111 }
2112 
2113 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2114 {
2115 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2116 	struct btrfs_free_cluster *cluster;
2117 	struct list_head *head;
2118 
2119 	spin_lock(&ctl->tree_lock);
2120 	while ((head = block_group->cluster_list.next) !=
2121 	       &block_group->cluster_list) {
2122 		cluster = list_entry(head, struct btrfs_free_cluster,
2123 				     block_group_list);
2124 
2125 		WARN_ON(cluster->block_group != block_group);
2126 		__btrfs_return_cluster_to_free_space(block_group, cluster);
2127 		if (need_resched()) {
2128 			spin_unlock(&ctl->tree_lock);
2129 			cond_resched();
2130 			spin_lock(&ctl->tree_lock);
2131 		}
2132 	}
2133 	__btrfs_remove_free_space_cache_locked(ctl);
2134 	spin_unlock(&ctl->tree_lock);
2135 
2136 }
2137 
2138 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2139 			       u64 offset, u64 bytes, u64 empty_size,
2140 			       u64 *max_extent_size)
2141 {
2142 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2143 	struct btrfs_free_space *entry = NULL;
2144 	u64 bytes_search = bytes + empty_size;
2145 	u64 ret = 0;
2146 	u64 align_gap = 0;
2147 	u64 align_gap_len = 0;
2148 
2149 	spin_lock(&ctl->tree_lock);
2150 	entry = find_free_space(ctl, &offset, &bytes_search,
2151 				block_group->full_stripe_len, max_extent_size);
2152 	if (!entry)
2153 		goto out;
2154 
2155 	ret = offset;
2156 	if (entry->bitmap) {
2157 		bitmap_clear_bits(ctl, entry, offset, bytes);
2158 		if (!entry->bytes)
2159 			free_bitmap(ctl, entry);
2160 	} else {
2161 		unlink_free_space(ctl, entry);
2162 		align_gap_len = offset - entry->offset;
2163 		align_gap = entry->offset;
2164 
2165 		entry->offset = offset + bytes;
2166 		WARN_ON(entry->bytes < bytes + align_gap_len);
2167 
2168 		entry->bytes -= bytes + align_gap_len;
2169 		if (!entry->bytes)
2170 			kmem_cache_free(btrfs_free_space_cachep, entry);
2171 		else
2172 			link_free_space(ctl, entry);
2173 	}
2174 out:
2175 	spin_unlock(&ctl->tree_lock);
2176 
2177 	if (align_gap_len)
2178 		__btrfs_add_free_space(ctl, align_gap, align_gap_len);
2179 	return ret;
2180 }
2181 
2182 /*
2183  * given a cluster, put all of its extents back into the free space
2184  * cache.  If a block group is passed, this function will only free
2185  * a cluster that belongs to the passed block group.
2186  *
2187  * Otherwise, it'll get a reference on the block group pointed to by the
2188  * cluster and remove the cluster from it.
2189  */
2190 int btrfs_return_cluster_to_free_space(
2191 			       struct btrfs_block_group_cache *block_group,
2192 			       struct btrfs_free_cluster *cluster)
2193 {
2194 	struct btrfs_free_space_ctl *ctl;
2195 	int ret;
2196 
2197 	/* first, get a safe pointer to the block group */
2198 	spin_lock(&cluster->lock);
2199 	if (!block_group) {
2200 		block_group = cluster->block_group;
2201 		if (!block_group) {
2202 			spin_unlock(&cluster->lock);
2203 			return 0;
2204 		}
2205 	} else if (cluster->block_group != block_group) {
2206 		/* someone else has already freed it don't redo their work */
2207 		spin_unlock(&cluster->lock);
2208 		return 0;
2209 	}
2210 	atomic_inc(&block_group->count);
2211 	spin_unlock(&cluster->lock);
2212 
2213 	ctl = block_group->free_space_ctl;
2214 
2215 	/* now return any extents the cluster had on it */
2216 	spin_lock(&ctl->tree_lock);
2217 	ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2218 	spin_unlock(&ctl->tree_lock);
2219 
2220 	/* finally drop our ref */
2221 	btrfs_put_block_group(block_group);
2222 	return ret;
2223 }
2224 
2225 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2226 				   struct btrfs_free_cluster *cluster,
2227 				   struct btrfs_free_space *entry,
2228 				   u64 bytes, u64 min_start,
2229 				   u64 *max_extent_size)
2230 {
2231 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2232 	int err;
2233 	u64 search_start = cluster->window_start;
2234 	u64 search_bytes = bytes;
2235 	u64 ret = 0;
2236 
2237 	search_start = min_start;
2238 	search_bytes = bytes;
2239 
2240 	err = search_bitmap(ctl, entry, &search_start, &search_bytes);
2241 	if (err) {
2242 		if (search_bytes > *max_extent_size)
2243 			*max_extent_size = search_bytes;
2244 		return 0;
2245 	}
2246 
2247 	ret = search_start;
2248 	__bitmap_clear_bits(ctl, entry, ret, bytes);
2249 
2250 	return ret;
2251 }
2252 
2253 /*
2254  * given a cluster, try to allocate 'bytes' from it, returns 0
2255  * if it couldn't find anything suitably large, or a logical disk offset
2256  * if things worked out
2257  */
2258 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2259 			     struct btrfs_free_cluster *cluster, u64 bytes,
2260 			     u64 min_start, u64 *max_extent_size)
2261 {
2262 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2263 	struct btrfs_free_space *entry = NULL;
2264 	struct rb_node *node;
2265 	u64 ret = 0;
2266 
2267 	spin_lock(&cluster->lock);
2268 	if (bytes > cluster->max_size)
2269 		goto out;
2270 
2271 	if (cluster->block_group != block_group)
2272 		goto out;
2273 
2274 	node = rb_first(&cluster->root);
2275 	if (!node)
2276 		goto out;
2277 
2278 	entry = rb_entry(node, struct btrfs_free_space, offset_index);
2279 	while(1) {
2280 		if (entry->bytes < bytes && entry->bytes > *max_extent_size)
2281 			*max_extent_size = entry->bytes;
2282 
2283 		if (entry->bytes < bytes ||
2284 		    (!entry->bitmap && entry->offset < min_start)) {
2285 			node = rb_next(&entry->offset_index);
2286 			if (!node)
2287 				break;
2288 			entry = rb_entry(node, struct btrfs_free_space,
2289 					 offset_index);
2290 			continue;
2291 		}
2292 
2293 		if (entry->bitmap) {
2294 			ret = btrfs_alloc_from_bitmap(block_group,
2295 						      cluster, entry, bytes,
2296 						      cluster->window_start,
2297 						      max_extent_size);
2298 			if (ret == 0) {
2299 				node = rb_next(&entry->offset_index);
2300 				if (!node)
2301 					break;
2302 				entry = rb_entry(node, struct btrfs_free_space,
2303 						 offset_index);
2304 				continue;
2305 			}
2306 			cluster->window_start += bytes;
2307 		} else {
2308 			ret = entry->offset;
2309 
2310 			entry->offset += bytes;
2311 			entry->bytes -= bytes;
2312 		}
2313 
2314 		if (entry->bytes == 0)
2315 			rb_erase(&entry->offset_index, &cluster->root);
2316 		break;
2317 	}
2318 out:
2319 	spin_unlock(&cluster->lock);
2320 
2321 	if (!ret)
2322 		return 0;
2323 
2324 	spin_lock(&ctl->tree_lock);
2325 
2326 	ctl->free_space -= bytes;
2327 	if (entry->bytes == 0) {
2328 		ctl->free_extents--;
2329 		if (entry->bitmap) {
2330 			kfree(entry->bitmap);
2331 			ctl->total_bitmaps--;
2332 			ctl->op->recalc_thresholds(ctl);
2333 		}
2334 		kmem_cache_free(btrfs_free_space_cachep, entry);
2335 	}
2336 
2337 	spin_unlock(&ctl->tree_lock);
2338 
2339 	return ret;
2340 }
2341 
2342 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2343 				struct btrfs_free_space *entry,
2344 				struct btrfs_free_cluster *cluster,
2345 				u64 offset, u64 bytes,
2346 				u64 cont1_bytes, u64 min_bytes)
2347 {
2348 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2349 	unsigned long next_zero;
2350 	unsigned long i;
2351 	unsigned long want_bits;
2352 	unsigned long min_bits;
2353 	unsigned long found_bits;
2354 	unsigned long start = 0;
2355 	unsigned long total_found = 0;
2356 	int ret;
2357 
2358 	i = offset_to_bit(entry->offset, ctl->unit,
2359 			  max_t(u64, offset, entry->offset));
2360 	want_bits = bytes_to_bits(bytes, ctl->unit);
2361 	min_bits = bytes_to_bits(min_bytes, ctl->unit);
2362 
2363 again:
2364 	found_bits = 0;
2365 	for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2366 		next_zero = find_next_zero_bit(entry->bitmap,
2367 					       BITS_PER_BITMAP, i);
2368 		if (next_zero - i >= min_bits) {
2369 			found_bits = next_zero - i;
2370 			break;
2371 		}
2372 		i = next_zero;
2373 	}
2374 
2375 	if (!found_bits)
2376 		return -ENOSPC;
2377 
2378 	if (!total_found) {
2379 		start = i;
2380 		cluster->max_size = 0;
2381 	}
2382 
2383 	total_found += found_bits;
2384 
2385 	if (cluster->max_size < found_bits * ctl->unit)
2386 		cluster->max_size = found_bits * ctl->unit;
2387 
2388 	if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2389 		i = next_zero + 1;
2390 		goto again;
2391 	}
2392 
2393 	cluster->window_start = start * ctl->unit + entry->offset;
2394 	rb_erase(&entry->offset_index, &ctl->free_space_offset);
2395 	ret = tree_insert_offset(&cluster->root, entry->offset,
2396 				 &entry->offset_index, 1);
2397 	ASSERT(!ret); /* -EEXIST; Logic error */
2398 
2399 	trace_btrfs_setup_cluster(block_group, cluster,
2400 				  total_found * ctl->unit, 1);
2401 	return 0;
2402 }
2403 
2404 /*
2405  * This searches the block group for just extents to fill the cluster with.
2406  * Try to find a cluster with at least bytes total bytes, at least one
2407  * extent of cont1_bytes, and other clusters of at least min_bytes.
2408  */
2409 static noinline int
2410 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2411 			struct btrfs_free_cluster *cluster,
2412 			struct list_head *bitmaps, u64 offset, u64 bytes,
2413 			u64 cont1_bytes, u64 min_bytes)
2414 {
2415 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2416 	struct btrfs_free_space *first = NULL;
2417 	struct btrfs_free_space *entry = NULL;
2418 	struct btrfs_free_space *last;
2419 	struct rb_node *node;
2420 	u64 window_start;
2421 	u64 window_free;
2422 	u64 max_extent;
2423 	u64 total_size = 0;
2424 
2425 	entry = tree_search_offset(ctl, offset, 0, 1);
2426 	if (!entry)
2427 		return -ENOSPC;
2428 
2429 	/*
2430 	 * We don't want bitmaps, so just move along until we find a normal
2431 	 * extent entry.
2432 	 */
2433 	while (entry->bitmap || entry->bytes < min_bytes) {
2434 		if (entry->bitmap && list_empty(&entry->list))
2435 			list_add_tail(&entry->list, bitmaps);
2436 		node = rb_next(&entry->offset_index);
2437 		if (!node)
2438 			return -ENOSPC;
2439 		entry = rb_entry(node, struct btrfs_free_space, offset_index);
2440 	}
2441 
2442 	window_start = entry->offset;
2443 	window_free = entry->bytes;
2444 	max_extent = entry->bytes;
2445 	first = entry;
2446 	last = entry;
2447 
2448 	for (node = rb_next(&entry->offset_index); node;
2449 	     node = rb_next(&entry->offset_index)) {
2450 		entry = rb_entry(node, struct btrfs_free_space, offset_index);
2451 
2452 		if (entry->bitmap) {
2453 			if (list_empty(&entry->list))
2454 				list_add_tail(&entry->list, bitmaps);
2455 			continue;
2456 		}
2457 
2458 		if (entry->bytes < min_bytes)
2459 			continue;
2460 
2461 		last = entry;
2462 		window_free += entry->bytes;
2463 		if (entry->bytes > max_extent)
2464 			max_extent = entry->bytes;
2465 	}
2466 
2467 	if (window_free < bytes || max_extent < cont1_bytes)
2468 		return -ENOSPC;
2469 
2470 	cluster->window_start = first->offset;
2471 
2472 	node = &first->offset_index;
2473 
2474 	/*
2475 	 * now we've found our entries, pull them out of the free space
2476 	 * cache and put them into the cluster rbtree
2477 	 */
2478 	do {
2479 		int ret;
2480 
2481 		entry = rb_entry(node, struct btrfs_free_space, offset_index);
2482 		node = rb_next(&entry->offset_index);
2483 		if (entry->bitmap || entry->bytes < min_bytes)
2484 			continue;
2485 
2486 		rb_erase(&entry->offset_index, &ctl->free_space_offset);
2487 		ret = tree_insert_offset(&cluster->root, entry->offset,
2488 					 &entry->offset_index, 0);
2489 		total_size += entry->bytes;
2490 		ASSERT(!ret); /* -EEXIST; Logic error */
2491 	} while (node && entry != last);
2492 
2493 	cluster->max_size = max_extent;
2494 	trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2495 	return 0;
2496 }
2497 
2498 /*
2499  * This specifically looks for bitmaps that may work in the cluster, we assume
2500  * that we have already failed to find extents that will work.
2501  */
2502 static noinline int
2503 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2504 		     struct btrfs_free_cluster *cluster,
2505 		     struct list_head *bitmaps, u64 offset, u64 bytes,
2506 		     u64 cont1_bytes, u64 min_bytes)
2507 {
2508 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2509 	struct btrfs_free_space *entry;
2510 	int ret = -ENOSPC;
2511 	u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2512 
2513 	if (ctl->total_bitmaps == 0)
2514 		return -ENOSPC;
2515 
2516 	/*
2517 	 * The bitmap that covers offset won't be in the list unless offset
2518 	 * is just its start offset.
2519 	 */
2520 	entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2521 	if (entry->offset != bitmap_offset) {
2522 		entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2523 		if (entry && list_empty(&entry->list))
2524 			list_add(&entry->list, bitmaps);
2525 	}
2526 
2527 	list_for_each_entry(entry, bitmaps, list) {
2528 		if (entry->bytes < bytes)
2529 			continue;
2530 		ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2531 					   bytes, cont1_bytes, min_bytes);
2532 		if (!ret)
2533 			return 0;
2534 	}
2535 
2536 	/*
2537 	 * The bitmaps list has all the bitmaps that record free space
2538 	 * starting after offset, so no more search is required.
2539 	 */
2540 	return -ENOSPC;
2541 }
2542 
2543 /*
2544  * here we try to find a cluster of blocks in a block group.  The goal
2545  * is to find at least bytes+empty_size.
2546  * We might not find them all in one contiguous area.
2547  *
2548  * returns zero and sets up cluster if things worked out, otherwise
2549  * it returns -enospc
2550  */
2551 int btrfs_find_space_cluster(struct btrfs_root *root,
2552 			     struct btrfs_block_group_cache *block_group,
2553 			     struct btrfs_free_cluster *cluster,
2554 			     u64 offset, u64 bytes, u64 empty_size)
2555 {
2556 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2557 	struct btrfs_free_space *entry, *tmp;
2558 	LIST_HEAD(bitmaps);
2559 	u64 min_bytes;
2560 	u64 cont1_bytes;
2561 	int ret;
2562 
2563 	/*
2564 	 * Choose the minimum extent size we'll require for this
2565 	 * cluster.  For SSD_SPREAD, don't allow any fragmentation.
2566 	 * For metadata, allow allocates with smaller extents.  For
2567 	 * data, keep it dense.
2568 	 */
2569 	if (btrfs_test_opt(root, SSD_SPREAD)) {
2570 		cont1_bytes = min_bytes = bytes + empty_size;
2571 	} else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2572 		cont1_bytes = bytes;
2573 		min_bytes = block_group->sectorsize;
2574 	} else {
2575 		cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
2576 		min_bytes = block_group->sectorsize;
2577 	}
2578 
2579 	spin_lock(&ctl->tree_lock);
2580 
2581 	/*
2582 	 * If we know we don't have enough space to make a cluster don't even
2583 	 * bother doing all the work to try and find one.
2584 	 */
2585 	if (ctl->free_space < bytes) {
2586 		spin_unlock(&ctl->tree_lock);
2587 		return -ENOSPC;
2588 	}
2589 
2590 	spin_lock(&cluster->lock);
2591 
2592 	/* someone already found a cluster, hooray */
2593 	if (cluster->block_group) {
2594 		ret = 0;
2595 		goto out;
2596 	}
2597 
2598 	trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
2599 				 min_bytes);
2600 
2601 	INIT_LIST_HEAD(&bitmaps);
2602 	ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
2603 				      bytes + empty_size,
2604 				      cont1_bytes, min_bytes);
2605 	if (ret)
2606 		ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
2607 					   offset, bytes + empty_size,
2608 					   cont1_bytes, min_bytes);
2609 
2610 	/* Clear our temporary list */
2611 	list_for_each_entry_safe(entry, tmp, &bitmaps, list)
2612 		list_del_init(&entry->list);
2613 
2614 	if (!ret) {
2615 		atomic_inc(&block_group->count);
2616 		list_add_tail(&cluster->block_group_list,
2617 			      &block_group->cluster_list);
2618 		cluster->block_group = block_group;
2619 	} else {
2620 		trace_btrfs_failed_cluster_setup(block_group);
2621 	}
2622 out:
2623 	spin_unlock(&cluster->lock);
2624 	spin_unlock(&ctl->tree_lock);
2625 
2626 	return ret;
2627 }
2628 
2629 /*
2630  * simple code to zero out a cluster
2631  */
2632 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2633 {
2634 	spin_lock_init(&cluster->lock);
2635 	spin_lock_init(&cluster->refill_lock);
2636 	cluster->root = RB_ROOT;
2637 	cluster->max_size = 0;
2638 	INIT_LIST_HEAD(&cluster->block_group_list);
2639 	cluster->block_group = NULL;
2640 }
2641 
2642 static int do_trimming(struct btrfs_block_group_cache *block_group,
2643 		       u64 *total_trimmed, u64 start, u64 bytes,
2644 		       u64 reserved_start, u64 reserved_bytes)
2645 {
2646 	struct btrfs_space_info *space_info = block_group->space_info;
2647 	struct btrfs_fs_info *fs_info = block_group->fs_info;
2648 	int ret;
2649 	int update = 0;
2650 	u64 trimmed = 0;
2651 
2652 	spin_lock(&space_info->lock);
2653 	spin_lock(&block_group->lock);
2654 	if (!block_group->ro) {
2655 		block_group->reserved += reserved_bytes;
2656 		space_info->bytes_reserved += reserved_bytes;
2657 		update = 1;
2658 	}
2659 	spin_unlock(&block_group->lock);
2660 	spin_unlock(&space_info->lock);
2661 
2662 	ret = btrfs_error_discard_extent(fs_info->extent_root,
2663 					 start, bytes, &trimmed);
2664 	if (!ret)
2665 		*total_trimmed += trimmed;
2666 
2667 	btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
2668 
2669 	if (update) {
2670 		spin_lock(&space_info->lock);
2671 		spin_lock(&block_group->lock);
2672 		if (block_group->ro)
2673 			space_info->bytes_readonly += reserved_bytes;
2674 		block_group->reserved -= reserved_bytes;
2675 		space_info->bytes_reserved -= reserved_bytes;
2676 		spin_unlock(&space_info->lock);
2677 		spin_unlock(&block_group->lock);
2678 	}
2679 
2680 	return ret;
2681 }
2682 
2683 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
2684 			  u64 *total_trimmed, u64 start, u64 end, u64 minlen)
2685 {
2686 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2687 	struct btrfs_free_space *entry;
2688 	struct rb_node *node;
2689 	int ret = 0;
2690 	u64 extent_start;
2691 	u64 extent_bytes;
2692 	u64 bytes;
2693 
2694 	while (start < end) {
2695 		spin_lock(&ctl->tree_lock);
2696 
2697 		if (ctl->free_space < minlen) {
2698 			spin_unlock(&ctl->tree_lock);
2699 			break;
2700 		}
2701 
2702 		entry = tree_search_offset(ctl, start, 0, 1);
2703 		if (!entry) {
2704 			spin_unlock(&ctl->tree_lock);
2705 			break;
2706 		}
2707 
2708 		/* skip bitmaps */
2709 		while (entry->bitmap) {
2710 			node = rb_next(&entry->offset_index);
2711 			if (!node) {
2712 				spin_unlock(&ctl->tree_lock);
2713 				goto out;
2714 			}
2715 			entry = rb_entry(node, struct btrfs_free_space,
2716 					 offset_index);
2717 		}
2718 
2719 		if (entry->offset >= end) {
2720 			spin_unlock(&ctl->tree_lock);
2721 			break;
2722 		}
2723 
2724 		extent_start = entry->offset;
2725 		extent_bytes = entry->bytes;
2726 		start = max(start, extent_start);
2727 		bytes = min(extent_start + extent_bytes, end) - start;
2728 		if (bytes < minlen) {
2729 			spin_unlock(&ctl->tree_lock);
2730 			goto next;
2731 		}
2732 
2733 		unlink_free_space(ctl, entry);
2734 		kmem_cache_free(btrfs_free_space_cachep, entry);
2735 
2736 		spin_unlock(&ctl->tree_lock);
2737 
2738 		ret = do_trimming(block_group, total_trimmed, start, bytes,
2739 				  extent_start, extent_bytes);
2740 		if (ret)
2741 			break;
2742 next:
2743 		start += bytes;
2744 
2745 		if (fatal_signal_pending(current)) {
2746 			ret = -ERESTARTSYS;
2747 			break;
2748 		}
2749 
2750 		cond_resched();
2751 	}
2752 out:
2753 	return ret;
2754 }
2755 
2756 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
2757 			u64 *total_trimmed, u64 start, u64 end, u64 minlen)
2758 {
2759 	struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2760 	struct btrfs_free_space *entry;
2761 	int ret = 0;
2762 	int ret2;
2763 	u64 bytes;
2764 	u64 offset = offset_to_bitmap(ctl, start);
2765 
2766 	while (offset < end) {
2767 		bool next_bitmap = false;
2768 
2769 		spin_lock(&ctl->tree_lock);
2770 
2771 		if (ctl->free_space < minlen) {
2772 			spin_unlock(&ctl->tree_lock);
2773 			break;
2774 		}
2775 
2776 		entry = tree_search_offset(ctl, offset, 1, 0);
2777 		if (!entry) {
2778 			spin_unlock(&ctl->tree_lock);
2779 			next_bitmap = true;
2780 			goto next;
2781 		}
2782 
2783 		bytes = minlen;
2784 		ret2 = search_bitmap(ctl, entry, &start, &bytes);
2785 		if (ret2 || start >= end) {
2786 			spin_unlock(&ctl->tree_lock);
2787 			next_bitmap = true;
2788 			goto next;
2789 		}
2790 
2791 		bytes = min(bytes, end - start);
2792 		if (bytes < minlen) {
2793 			spin_unlock(&ctl->tree_lock);
2794 			goto next;
2795 		}
2796 
2797 		bitmap_clear_bits(ctl, entry, start, bytes);
2798 		if (entry->bytes == 0)
2799 			free_bitmap(ctl, entry);
2800 
2801 		spin_unlock(&ctl->tree_lock);
2802 
2803 		ret = do_trimming(block_group, total_trimmed, start, bytes,
2804 				  start, bytes);
2805 		if (ret)
2806 			break;
2807 next:
2808 		if (next_bitmap) {
2809 			offset += BITS_PER_BITMAP * ctl->unit;
2810 		} else {
2811 			start += bytes;
2812 			if (start >= offset + BITS_PER_BITMAP * ctl->unit)
2813 				offset += BITS_PER_BITMAP * ctl->unit;
2814 		}
2815 
2816 		if (fatal_signal_pending(current)) {
2817 			ret = -ERESTARTSYS;
2818 			break;
2819 		}
2820 
2821 		cond_resched();
2822 	}
2823 
2824 	return ret;
2825 }
2826 
2827 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
2828 			   u64 *trimmed, u64 start, u64 end, u64 minlen)
2829 {
2830 	int ret;
2831 
2832 	*trimmed = 0;
2833 
2834 	ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
2835 	if (ret)
2836 		return ret;
2837 
2838 	ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
2839 
2840 	return ret;
2841 }
2842 
2843 /*
2844  * Find the left-most item in the cache tree, and then return the
2845  * smallest inode number in the item.
2846  *
2847  * Note: the returned inode number may not be the smallest one in
2848  * the tree, if the left-most item is a bitmap.
2849  */
2850 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
2851 {
2852 	struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
2853 	struct btrfs_free_space *entry = NULL;
2854 	u64 ino = 0;
2855 
2856 	spin_lock(&ctl->tree_lock);
2857 
2858 	if (RB_EMPTY_ROOT(&ctl->free_space_offset))
2859 		goto out;
2860 
2861 	entry = rb_entry(rb_first(&ctl->free_space_offset),
2862 			 struct btrfs_free_space, offset_index);
2863 
2864 	if (!entry->bitmap) {
2865 		ino = entry->offset;
2866 
2867 		unlink_free_space(ctl, entry);
2868 		entry->offset++;
2869 		entry->bytes--;
2870 		if (!entry->bytes)
2871 			kmem_cache_free(btrfs_free_space_cachep, entry);
2872 		else
2873 			link_free_space(ctl, entry);
2874 	} else {
2875 		u64 offset = 0;
2876 		u64 count = 1;
2877 		int ret;
2878 
2879 		ret = search_bitmap(ctl, entry, &offset, &count);
2880 		/* Logic error; Should be empty if it can't find anything */
2881 		ASSERT(!ret);
2882 
2883 		ino = offset;
2884 		bitmap_clear_bits(ctl, entry, offset, 1);
2885 		if (entry->bytes == 0)
2886 			free_bitmap(ctl, entry);
2887 	}
2888 out:
2889 	spin_unlock(&ctl->tree_lock);
2890 
2891 	return ino;
2892 }
2893 
2894 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
2895 				    struct btrfs_path *path)
2896 {
2897 	struct inode *inode = NULL;
2898 
2899 	spin_lock(&root->cache_lock);
2900 	if (root->cache_inode)
2901 		inode = igrab(root->cache_inode);
2902 	spin_unlock(&root->cache_lock);
2903 	if (inode)
2904 		return inode;
2905 
2906 	inode = __lookup_free_space_inode(root, path, 0);
2907 	if (IS_ERR(inode))
2908 		return inode;
2909 
2910 	spin_lock(&root->cache_lock);
2911 	if (!btrfs_fs_closing(root->fs_info))
2912 		root->cache_inode = igrab(inode);
2913 	spin_unlock(&root->cache_lock);
2914 
2915 	return inode;
2916 }
2917 
2918 int create_free_ino_inode(struct btrfs_root *root,
2919 			  struct btrfs_trans_handle *trans,
2920 			  struct btrfs_path *path)
2921 {
2922 	return __create_free_space_inode(root, trans, path,
2923 					 BTRFS_FREE_INO_OBJECTID, 0);
2924 }
2925 
2926 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
2927 {
2928 	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2929 	struct btrfs_path *path;
2930 	struct inode *inode;
2931 	int ret = 0;
2932 	u64 root_gen = btrfs_root_generation(&root->root_item);
2933 
2934 	if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2935 		return 0;
2936 
2937 	/*
2938 	 * If we're unmounting then just return, since this does a search on the
2939 	 * normal root and not the commit root and we could deadlock.
2940 	 */
2941 	if (btrfs_fs_closing(fs_info))
2942 		return 0;
2943 
2944 	path = btrfs_alloc_path();
2945 	if (!path)
2946 		return 0;
2947 
2948 	inode = lookup_free_ino_inode(root, path);
2949 	if (IS_ERR(inode))
2950 		goto out;
2951 
2952 	if (root_gen != BTRFS_I(inode)->generation)
2953 		goto out_put;
2954 
2955 	ret = __load_free_space_cache(root, inode, ctl, path, 0);
2956 
2957 	if (ret < 0)
2958 		btrfs_err(fs_info,
2959 			"failed to load free ino cache for root %llu",
2960 			root->root_key.objectid);
2961 out_put:
2962 	iput(inode);
2963 out:
2964 	btrfs_free_path(path);
2965 	return ret;
2966 }
2967 
2968 int btrfs_write_out_ino_cache(struct btrfs_root *root,
2969 			      struct btrfs_trans_handle *trans,
2970 			      struct btrfs_path *path)
2971 {
2972 	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2973 	struct inode *inode;
2974 	int ret;
2975 
2976 	if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2977 		return 0;
2978 
2979 	inode = lookup_free_ino_inode(root, path);
2980 	if (IS_ERR(inode))
2981 		return 0;
2982 
2983 	ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
2984 	if (ret) {
2985 		btrfs_delalloc_release_metadata(inode, inode->i_size);
2986 #ifdef DEBUG
2987 		btrfs_err(root->fs_info,
2988 			"failed to write free ino cache for root %llu",
2989 			root->root_key.objectid);
2990 #endif
2991 	}
2992 
2993 	iput(inode);
2994 	return ret;
2995 }
2996 
2997 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
2998 /*
2999  * Use this if you need to make a bitmap or extent entry specifically, it
3000  * doesn't do any of the merging that add_free_space does, this acts a lot like
3001  * how the free space cache loading stuff works, so you can get really weird
3002  * configurations.
3003  */
3004 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
3005 			      u64 offset, u64 bytes, bool bitmap)
3006 {
3007 	struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3008 	struct btrfs_free_space *info = NULL, *bitmap_info;
3009 	void *map = NULL;
3010 	u64 bytes_added;
3011 	int ret;
3012 
3013 again:
3014 	if (!info) {
3015 		info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3016 		if (!info)
3017 			return -ENOMEM;
3018 	}
3019 
3020 	if (!bitmap) {
3021 		spin_lock(&ctl->tree_lock);
3022 		info->offset = offset;
3023 		info->bytes = bytes;
3024 		ret = link_free_space(ctl, info);
3025 		spin_unlock(&ctl->tree_lock);
3026 		if (ret)
3027 			kmem_cache_free(btrfs_free_space_cachep, info);
3028 		return ret;
3029 	}
3030 
3031 	if (!map) {
3032 		map = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
3033 		if (!map) {
3034 			kmem_cache_free(btrfs_free_space_cachep, info);
3035 			return -ENOMEM;
3036 		}
3037 	}
3038 
3039 	spin_lock(&ctl->tree_lock);
3040 	bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3041 					 1, 0);
3042 	if (!bitmap_info) {
3043 		info->bitmap = map;
3044 		map = NULL;
3045 		add_new_bitmap(ctl, info, offset);
3046 		bitmap_info = info;
3047 	}
3048 
3049 	bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3050 	bytes -= bytes_added;
3051 	offset += bytes_added;
3052 	spin_unlock(&ctl->tree_lock);
3053 
3054 	if (bytes)
3055 		goto again;
3056 
3057 	if (map)
3058 		kfree(map);
3059 	return 0;
3060 }
3061 
3062 /*
3063  * Checks to see if the given range is in the free space cache.  This is really
3064  * just used to check the absence of space, so if there is free space in the
3065  * range at all we will return 1.
3066  */
3067 int test_check_exists(struct btrfs_block_group_cache *cache,
3068 		      u64 offset, u64 bytes)
3069 {
3070 	struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3071 	struct btrfs_free_space *info;
3072 	int ret = 0;
3073 
3074 	spin_lock(&ctl->tree_lock);
3075 	info = tree_search_offset(ctl, offset, 0, 0);
3076 	if (!info) {
3077 		info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3078 					  1, 0);
3079 		if (!info)
3080 			goto out;
3081 	}
3082 
3083 have_info:
3084 	if (info->bitmap) {
3085 		u64 bit_off, bit_bytes;
3086 		struct rb_node *n;
3087 		struct btrfs_free_space *tmp;
3088 
3089 		bit_off = offset;
3090 		bit_bytes = ctl->unit;
3091 		ret = search_bitmap(ctl, info, &bit_off, &bit_bytes);
3092 		if (!ret) {
3093 			if (bit_off == offset) {
3094 				ret = 1;
3095 				goto out;
3096 			} else if (bit_off > offset &&
3097 				   offset + bytes > bit_off) {
3098 				ret = 1;
3099 				goto out;
3100 			}
3101 		}
3102 
3103 		n = rb_prev(&info->offset_index);
3104 		while (n) {
3105 			tmp = rb_entry(n, struct btrfs_free_space,
3106 				       offset_index);
3107 			if (tmp->offset + tmp->bytes < offset)
3108 				break;
3109 			if (offset + bytes < tmp->offset) {
3110 				n = rb_prev(&info->offset_index);
3111 				continue;
3112 			}
3113 			info = tmp;
3114 			goto have_info;
3115 		}
3116 
3117 		n = rb_next(&info->offset_index);
3118 		while (n) {
3119 			tmp = rb_entry(n, struct btrfs_free_space,
3120 				       offset_index);
3121 			if (offset + bytes < tmp->offset)
3122 				break;
3123 			if (tmp->offset + tmp->bytes < offset) {
3124 				n = rb_next(&info->offset_index);
3125 				continue;
3126 			}
3127 			info = tmp;
3128 			goto have_info;
3129 		}
3130 
3131 		goto out;
3132 	}
3133 
3134 	if (info->offset == offset) {
3135 		ret = 1;
3136 		goto out;
3137 	}
3138 
3139 	if (offset > info->offset && offset < info->offset + info->bytes)
3140 		ret = 1;
3141 out:
3142 	spin_unlock(&ctl->tree_lock);
3143 	return ret;
3144 }
3145 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */
3146