xref: /openbmc/linux/fs/btrfs/free-space-cache.c (revision b627b4ed)
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/sched.h>
20 #include "ctree.h"
21 #include "free-space-cache.h"
22 #include "transaction.h"
23 
24 struct btrfs_free_space {
25 	struct rb_node bytes_index;
26 	struct rb_node offset_index;
27 	u64 offset;
28 	u64 bytes;
29 };
30 
31 static int tree_insert_offset(struct rb_root *root, u64 offset,
32 			      struct rb_node *node)
33 {
34 	struct rb_node **p = &root->rb_node;
35 	struct rb_node *parent = NULL;
36 	struct btrfs_free_space *info;
37 
38 	while (*p) {
39 		parent = *p;
40 		info = rb_entry(parent, struct btrfs_free_space, offset_index);
41 
42 		if (offset < info->offset)
43 			p = &(*p)->rb_left;
44 		else if (offset > info->offset)
45 			p = &(*p)->rb_right;
46 		else
47 			return -EEXIST;
48 	}
49 
50 	rb_link_node(node, parent, p);
51 	rb_insert_color(node, root);
52 
53 	return 0;
54 }
55 
56 static int tree_insert_bytes(struct rb_root *root, u64 bytes,
57 			     struct rb_node *node)
58 {
59 	struct rb_node **p = &root->rb_node;
60 	struct rb_node *parent = NULL;
61 	struct btrfs_free_space *info;
62 
63 	while (*p) {
64 		parent = *p;
65 		info = rb_entry(parent, struct btrfs_free_space, bytes_index);
66 
67 		if (bytes < info->bytes)
68 			p = &(*p)->rb_left;
69 		else
70 			p = &(*p)->rb_right;
71 	}
72 
73 	rb_link_node(node, parent, p);
74 	rb_insert_color(node, root);
75 
76 	return 0;
77 }
78 
79 /*
80  * searches the tree for the given offset.
81  *
82  * fuzzy == 1: this is used for allocations where we are given a hint of where
83  * to look for free space.  Because the hint may not be completely on an offset
84  * mark, or the hint may no longer point to free space we need to fudge our
85  * results a bit.  So we look for free space starting at or after offset with at
86  * least bytes size.  We prefer to find as close to the given offset as we can.
87  * Also if the offset is within a free space range, then we will return the free
88  * space that contains the given offset, which means we can return a free space
89  * chunk with an offset before the provided offset.
90  *
91  * fuzzy == 0: this is just a normal tree search.  Give us the free space that
92  * starts at the given offset which is at least bytes size, and if its not there
93  * return NULL.
94  */
95 static struct btrfs_free_space *tree_search_offset(struct rb_root *root,
96 						   u64 offset, u64 bytes,
97 						   int fuzzy)
98 {
99 	struct rb_node *n = root->rb_node;
100 	struct btrfs_free_space *entry, *ret = NULL;
101 
102 	while (n) {
103 		entry = rb_entry(n, struct btrfs_free_space, offset_index);
104 
105 		if (offset < entry->offset) {
106 			if (fuzzy &&
107 			    (!ret || entry->offset < ret->offset) &&
108 			    (bytes <= entry->bytes))
109 				ret = entry;
110 			n = n->rb_left;
111 		} else if (offset > entry->offset) {
112 			if (fuzzy &&
113 			    (entry->offset + entry->bytes - 1) >= offset &&
114 			    bytes <= entry->bytes) {
115 				ret = entry;
116 				break;
117 			}
118 			n = n->rb_right;
119 		} else {
120 			if (bytes > entry->bytes) {
121 				n = n->rb_right;
122 				continue;
123 			}
124 			ret = entry;
125 			break;
126 		}
127 	}
128 
129 	return ret;
130 }
131 
132 /*
133  * return a chunk at least bytes size, as close to offset that we can get.
134  */
135 static struct btrfs_free_space *tree_search_bytes(struct rb_root *root,
136 						  u64 offset, u64 bytes)
137 {
138 	struct rb_node *n = root->rb_node;
139 	struct btrfs_free_space *entry, *ret = NULL;
140 
141 	while (n) {
142 		entry = rb_entry(n, struct btrfs_free_space, bytes_index);
143 
144 		if (bytes < entry->bytes) {
145 			/*
146 			 * We prefer to get a hole size as close to the size we
147 			 * are asking for so we don't take small slivers out of
148 			 * huge holes, but we also want to get as close to the
149 			 * offset as possible so we don't have a whole lot of
150 			 * fragmentation.
151 			 */
152 			if (offset <= entry->offset) {
153 				if (!ret)
154 					ret = entry;
155 				else if (entry->bytes < ret->bytes)
156 					ret = entry;
157 				else if (entry->offset < ret->offset)
158 					ret = entry;
159 			}
160 			n = n->rb_left;
161 		} else if (bytes > entry->bytes) {
162 			n = n->rb_right;
163 		} else {
164 			/*
165 			 * Ok we may have multiple chunks of the wanted size,
166 			 * so we don't want to take the first one we find, we
167 			 * want to take the one closest to our given offset, so
168 			 * keep searching just in case theres a better match.
169 			 */
170 			n = n->rb_right;
171 			if (offset > entry->offset)
172 				continue;
173 			else if (!ret || entry->offset < ret->offset)
174 				ret = entry;
175 		}
176 	}
177 
178 	return ret;
179 }
180 
181 static void unlink_free_space(struct btrfs_block_group_cache *block_group,
182 			      struct btrfs_free_space *info)
183 {
184 	rb_erase(&info->offset_index, &block_group->free_space_offset);
185 	rb_erase(&info->bytes_index, &block_group->free_space_bytes);
186 }
187 
188 static int link_free_space(struct btrfs_block_group_cache *block_group,
189 			   struct btrfs_free_space *info)
190 {
191 	int ret = 0;
192 
193 
194 	BUG_ON(!info->bytes);
195 	ret = tree_insert_offset(&block_group->free_space_offset, info->offset,
196 				 &info->offset_index);
197 	if (ret)
198 		return ret;
199 
200 	ret = tree_insert_bytes(&block_group->free_space_bytes, info->bytes,
201 				&info->bytes_index);
202 	if (ret)
203 		return ret;
204 
205 	return ret;
206 }
207 
208 int btrfs_add_free_space(struct btrfs_block_group_cache *block_group,
209 			 u64 offset, u64 bytes)
210 {
211 	struct btrfs_free_space *right_info;
212 	struct btrfs_free_space *left_info;
213 	struct btrfs_free_space *info = NULL;
214 	int ret = 0;
215 
216 	info = kzalloc(sizeof(struct btrfs_free_space), GFP_NOFS);
217 	if (!info)
218 		return -ENOMEM;
219 
220 	info->offset = offset;
221 	info->bytes = bytes;
222 
223 	spin_lock(&block_group->tree_lock);
224 
225 	/*
226 	 * first we want to see if there is free space adjacent to the range we
227 	 * are adding, if there is remove that struct and add a new one to
228 	 * cover the entire range
229 	 */
230 	right_info = tree_search_offset(&block_group->free_space_offset,
231 					offset+bytes, 0, 0);
232 	left_info = tree_search_offset(&block_group->free_space_offset,
233 				       offset-1, 0, 1);
234 
235 	if (right_info) {
236 		unlink_free_space(block_group, right_info);
237 		info->bytes += right_info->bytes;
238 		kfree(right_info);
239 	}
240 
241 	if (left_info && left_info->offset + left_info->bytes == offset) {
242 		unlink_free_space(block_group, left_info);
243 		info->offset = left_info->offset;
244 		info->bytes += left_info->bytes;
245 		kfree(left_info);
246 	}
247 
248 	ret = link_free_space(block_group, info);
249 	if (ret)
250 		kfree(info);
251 
252 	spin_unlock(&block_group->tree_lock);
253 
254 	if (ret) {
255 		printk(KERN_ERR "btrfs: unable to add free space :%d\n", ret);
256 		BUG_ON(ret == -EEXIST);
257 	}
258 
259 	return ret;
260 }
261 
262 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
263 			    u64 offset, u64 bytes)
264 {
265 	struct btrfs_free_space *info;
266 	int ret = 0;
267 
268 	spin_lock(&block_group->tree_lock);
269 
270 	info = tree_search_offset(&block_group->free_space_offset, offset, 0,
271 				  1);
272 	if (info && info->offset == offset) {
273 		if (info->bytes < bytes) {
274 			printk(KERN_ERR "Found free space at %llu, size %llu,"
275 			       "trying to use %llu\n",
276 			       (unsigned long long)info->offset,
277 			       (unsigned long long)info->bytes,
278 			       (unsigned long long)bytes);
279 			WARN_ON(1);
280 			ret = -EINVAL;
281 			spin_unlock(&block_group->tree_lock);
282 			goto out;
283 		}
284 		unlink_free_space(block_group, info);
285 
286 		if (info->bytes == bytes) {
287 			kfree(info);
288 			spin_unlock(&block_group->tree_lock);
289 			goto out;
290 		}
291 
292 		info->offset += bytes;
293 		info->bytes -= bytes;
294 
295 		ret = link_free_space(block_group, info);
296 		spin_unlock(&block_group->tree_lock);
297 		BUG_ON(ret);
298 	} else if (info && info->offset < offset &&
299 		   info->offset + info->bytes >= offset + bytes) {
300 		u64 old_start = info->offset;
301 		/*
302 		 * we're freeing space in the middle of the info,
303 		 * this can happen during tree log replay
304 		 *
305 		 * first unlink the old info and then
306 		 * insert it again after the hole we're creating
307 		 */
308 		unlink_free_space(block_group, info);
309 		if (offset + bytes < info->offset + info->bytes) {
310 			u64 old_end = info->offset + info->bytes;
311 
312 			info->offset = offset + bytes;
313 			info->bytes = old_end - info->offset;
314 			ret = link_free_space(block_group, info);
315 			BUG_ON(ret);
316 		} else {
317 			/* the hole we're creating ends at the end
318 			 * of the info struct, just free the info
319 			 */
320 			kfree(info);
321 		}
322 		spin_unlock(&block_group->tree_lock);
323 		/* step two, insert a new info struct to cover anything
324 		 * before the hole
325 		 */
326 		ret = btrfs_add_free_space(block_group, old_start,
327 					   offset - old_start);
328 		BUG_ON(ret);
329 	} else {
330 		spin_unlock(&block_group->tree_lock);
331 		if (!info) {
332 			printk(KERN_ERR "couldn't find space %llu to free\n",
333 			       (unsigned long long)offset);
334 			printk(KERN_ERR "cached is %d, offset %llu bytes %llu\n",
335 			       block_group->cached, block_group->key.objectid,
336 			       block_group->key.offset);
337 			btrfs_dump_free_space(block_group, bytes);
338 		} else if (info) {
339 			printk(KERN_ERR "hmm, found offset=%llu bytes=%llu, "
340 			       "but wanted offset=%llu bytes=%llu\n",
341 			       info->offset, info->bytes, offset, bytes);
342 		}
343 		WARN_ON(1);
344 	}
345 out:
346 	return ret;
347 }
348 
349 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
350 			   u64 bytes)
351 {
352 	struct btrfs_free_space *info;
353 	struct rb_node *n;
354 	int count = 0;
355 
356 	for (n = rb_first(&block_group->free_space_offset); n; n = rb_next(n)) {
357 		info = rb_entry(n, struct btrfs_free_space, offset_index);
358 		if (info->bytes >= bytes)
359 			count++;
360 		printk(KERN_ERR "entry offset %llu, bytes %llu\n", info->offset,
361 		       info->bytes);
362 	}
363 	printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
364 	       "\n", count);
365 }
366 
367 u64 btrfs_block_group_free_space(struct btrfs_block_group_cache *block_group)
368 {
369 	struct btrfs_free_space *info;
370 	struct rb_node *n;
371 	u64 ret = 0;
372 
373 	for (n = rb_first(&block_group->free_space_offset); n;
374 	     n = rb_next(n)) {
375 		info = rb_entry(n, struct btrfs_free_space, offset_index);
376 		ret += info->bytes;
377 	}
378 
379 	return ret;
380 }
381 
382 /*
383  * for a given cluster, put all of its extents back into the free
384  * space cache.  If the block group passed doesn't match the block group
385  * pointed to by the cluster, someone else raced in and freed the
386  * cluster already.  In that case, we just return without changing anything
387  */
388 static int
389 __btrfs_return_cluster_to_free_space(
390 			     struct btrfs_block_group_cache *block_group,
391 			     struct btrfs_free_cluster *cluster)
392 {
393 	struct btrfs_free_space *entry;
394 	struct rb_node *node;
395 
396 	spin_lock(&cluster->lock);
397 	if (cluster->block_group != block_group)
398 		goto out;
399 
400 	cluster->window_start = 0;
401 	node = rb_first(&cluster->root);
402 	while(node) {
403 		entry = rb_entry(node, struct btrfs_free_space, offset_index);
404 		node = rb_next(&entry->offset_index);
405 		rb_erase(&entry->offset_index, &cluster->root);
406 		link_free_space(block_group, entry);
407 	}
408 	list_del_init(&cluster->block_group_list);
409 
410 	btrfs_put_block_group(cluster->block_group);
411 	cluster->block_group = NULL;
412 	cluster->root.rb_node = NULL;
413 out:
414 	spin_unlock(&cluster->lock);
415 	return 0;
416 }
417 
418 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
419 {
420 	struct btrfs_free_space *info;
421 	struct rb_node *node;
422 	struct btrfs_free_cluster *cluster;
423 	struct btrfs_free_cluster *safe;
424 
425 	spin_lock(&block_group->tree_lock);
426 
427 	list_for_each_entry_safe(cluster, safe, &block_group->cluster_list,
428 				 block_group_list) {
429 
430 		WARN_ON(cluster->block_group != block_group);
431 		__btrfs_return_cluster_to_free_space(block_group, cluster);
432 	}
433 
434 	while ((node = rb_last(&block_group->free_space_bytes)) != NULL) {
435 		info = rb_entry(node, struct btrfs_free_space, bytes_index);
436 		unlink_free_space(block_group, info);
437 		kfree(info);
438 		if (need_resched()) {
439 			spin_unlock(&block_group->tree_lock);
440 			cond_resched();
441 			spin_lock(&block_group->tree_lock);
442 		}
443 	}
444 	spin_unlock(&block_group->tree_lock);
445 }
446 
447 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
448 			       u64 offset, u64 bytes, u64 empty_size)
449 {
450 	struct btrfs_free_space *entry = NULL;
451 	u64 ret = 0;
452 
453 	spin_lock(&block_group->tree_lock);
454 	entry = tree_search_offset(&block_group->free_space_offset, offset,
455 				   bytes + empty_size, 1);
456 	if (!entry)
457 		entry = tree_search_bytes(&block_group->free_space_bytes,
458 					  offset, bytes + empty_size);
459 	if (entry) {
460 		unlink_free_space(block_group, entry);
461 		ret = entry->offset;
462 		entry->offset += bytes;
463 		entry->bytes -= bytes;
464 
465 		if (!entry->bytes)
466 			kfree(entry);
467 		else
468 			link_free_space(block_group, entry);
469 	}
470 	spin_unlock(&block_group->tree_lock);
471 
472 	return ret;
473 }
474 
475 /*
476  * given a cluster, put all of its extents back into the free space
477  * cache.  If a block group is passed, this function will only free
478  * a cluster that belongs to the passed block group.
479  *
480  * Otherwise, it'll get a reference on the block group pointed to by the
481  * cluster and remove the cluster from it.
482  */
483 int btrfs_return_cluster_to_free_space(
484 			       struct btrfs_block_group_cache *block_group,
485 			       struct btrfs_free_cluster *cluster)
486 {
487 	int ret;
488 
489 	/* first, get a safe pointer to the block group */
490 	spin_lock(&cluster->lock);
491 	if (!block_group) {
492 		block_group = cluster->block_group;
493 		if (!block_group) {
494 			spin_unlock(&cluster->lock);
495 			return 0;
496 		}
497 	} else if (cluster->block_group != block_group) {
498 		/* someone else has already freed it don't redo their work */
499 		spin_unlock(&cluster->lock);
500 		return 0;
501 	}
502 	atomic_inc(&block_group->count);
503 	spin_unlock(&cluster->lock);
504 
505 	/* now return any extents the cluster had on it */
506 	spin_lock(&block_group->tree_lock);
507 	ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
508 	spin_unlock(&block_group->tree_lock);
509 
510 	/* finally drop our ref */
511 	btrfs_put_block_group(block_group);
512 	return ret;
513 }
514 
515 /*
516  * given a cluster, try to allocate 'bytes' from it, returns 0
517  * if it couldn't find anything suitably large, or a logical disk offset
518  * if things worked out
519  */
520 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
521 			     struct btrfs_free_cluster *cluster, u64 bytes,
522 			     u64 min_start)
523 {
524 	struct btrfs_free_space *entry = NULL;
525 	struct rb_node *node;
526 	u64 ret = 0;
527 
528 	spin_lock(&cluster->lock);
529 	if (bytes > cluster->max_size)
530 		goto out;
531 
532 	if (cluster->block_group != block_group)
533 		goto out;
534 
535 	node = rb_first(&cluster->root);
536 	if (!node)
537 		goto out;
538 
539 	entry = rb_entry(node, struct btrfs_free_space, offset_index);
540 
541 	while(1) {
542 		if (entry->bytes < bytes || entry->offset < min_start) {
543 			struct rb_node *node;
544 
545 			node = rb_next(&entry->offset_index);
546 			if (!node)
547 				break;
548 			entry = rb_entry(node, struct btrfs_free_space,
549 					 offset_index);
550 			continue;
551 		}
552 		ret = entry->offset;
553 
554 		entry->offset += bytes;
555 		entry->bytes -= bytes;
556 
557 		if (entry->bytes == 0) {
558 			rb_erase(&entry->offset_index, &cluster->root);
559 			kfree(entry);
560 		}
561 		break;
562 	}
563 out:
564 	spin_unlock(&cluster->lock);
565 	return ret;
566 }
567 
568 /*
569  * here we try to find a cluster of blocks in a block group.  The goal
570  * is to find at least bytes free and up to empty_size + bytes free.
571  * We might not find them all in one contiguous area.
572  *
573  * returns zero and sets up cluster if things worked out, otherwise
574  * it returns -enospc
575  */
576 int btrfs_find_space_cluster(struct btrfs_trans_handle *trans,
577 			     struct btrfs_block_group_cache *block_group,
578 			     struct btrfs_free_cluster *cluster,
579 			     u64 offset, u64 bytes, u64 empty_size)
580 {
581 	struct btrfs_free_space *entry = NULL;
582 	struct rb_node *node;
583 	struct btrfs_free_space *next;
584 	struct btrfs_free_space *last;
585 	u64 min_bytes;
586 	u64 window_start;
587 	u64 window_free;
588 	u64 max_extent = 0;
589 	int total_retries = 0;
590 	int ret;
591 
592 	/* for metadata, allow allocates with more holes */
593 	if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
594 		/*
595 		 * we want to do larger allocations when we are
596 		 * flushing out the delayed refs, it helps prevent
597 		 * making more work as we go along.
598 		 */
599 		if (trans->transaction->delayed_refs.flushing)
600 			min_bytes = max(bytes, (bytes + empty_size) >> 1);
601 		else
602 			min_bytes = max(bytes, (bytes + empty_size) >> 4);
603 	} else
604 		min_bytes = max(bytes, (bytes + empty_size) >> 2);
605 
606 	spin_lock(&block_group->tree_lock);
607 	spin_lock(&cluster->lock);
608 
609 	/* someone already found a cluster, hooray */
610 	if (cluster->block_group) {
611 		ret = 0;
612 		goto out;
613 	}
614 again:
615 	min_bytes = min(min_bytes, bytes + empty_size);
616 	entry = tree_search_bytes(&block_group->free_space_bytes,
617 				  offset, min_bytes);
618 	if (!entry) {
619 		ret = -ENOSPC;
620 		goto out;
621 	}
622 	window_start = entry->offset;
623 	window_free = entry->bytes;
624 	last = entry;
625 	max_extent = entry->bytes;
626 
627 	while(1) {
628 		/* out window is just right, lets fill it */
629 		if (window_free >= bytes + empty_size)
630 			break;
631 
632 		node = rb_next(&last->offset_index);
633 		if (!node) {
634 			ret = -ENOSPC;
635 			goto out;
636 		}
637 		next = rb_entry(node, struct btrfs_free_space, offset_index);
638 
639 		/*
640 		 * we haven't filled the empty size and the window is
641 		 * very large.  reset and try again
642 		 */
643 		if (next->offset - window_start > (bytes + empty_size) * 2) {
644 			entry = next;
645 			window_start = entry->offset;
646 			window_free = entry->bytes;
647 			last = entry;
648 			max_extent = 0;
649 			total_retries++;
650 			if (total_retries % 256 == 0) {
651 				if (min_bytes >= (bytes + empty_size)) {
652 					ret = -ENOSPC;
653 					goto out;
654 				}
655 				/*
656 				 * grow our allocation a bit, we're not having
657 				 * much luck
658 				 */
659 				min_bytes *= 2;
660 				goto again;
661 			}
662 		} else {
663 			last = next;
664 			window_free += next->bytes;
665 			if (entry->bytes > max_extent)
666 				max_extent = entry->bytes;
667 		}
668 	}
669 
670 	cluster->window_start = entry->offset;
671 
672 	/*
673 	 * now we've found our entries, pull them out of the free space
674 	 * cache and put them into the cluster rbtree
675 	 *
676 	 * The cluster includes an rbtree, but only uses the offset index
677 	 * of each free space cache entry.
678 	 */
679 	while(1) {
680 		node = rb_next(&entry->offset_index);
681 		unlink_free_space(block_group, entry);
682 		ret = tree_insert_offset(&cluster->root, entry->offset,
683 					 &entry->offset_index);
684 		BUG_ON(ret);
685 
686 		if (!node || entry == last)
687 			break;
688 
689 		entry = rb_entry(node, struct btrfs_free_space, offset_index);
690 	}
691 	ret = 0;
692 	cluster->max_size = max_extent;
693 	atomic_inc(&block_group->count);
694 	list_add_tail(&cluster->block_group_list, &block_group->cluster_list);
695 	cluster->block_group = block_group;
696 out:
697 	spin_unlock(&cluster->lock);
698 	spin_unlock(&block_group->tree_lock);
699 
700 	return ret;
701 }
702 
703 /*
704  * simple code to zero out a cluster
705  */
706 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
707 {
708 	spin_lock_init(&cluster->lock);
709 	spin_lock_init(&cluster->refill_lock);
710 	cluster->root.rb_node = NULL;
711 	cluster->max_size = 0;
712 	INIT_LIST_HEAD(&cluster->block_group_list);
713 	cluster->block_group = NULL;
714 }
715 
716