• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2007 Oracle.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public
6  * License v2 as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful,
9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
11  * General Public License for more details.
12  *
13  * You should have received a copy of the GNU General Public
14  * License along with this program; if not, write to the
15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16  * Boston, MA 021110-1307, USA.
17  */
18 
19 #include <linux/delay.h>
20 #include <linux/kthread.h>
21 #include <linux/pagemap.h>
22 
23 #include "ctree.h"
24 #include "disk-io.h"
25 #include "free-space-cache.h"
26 #include "inode-map.h"
27 #include "transaction.h"
28 
caching_kthread(void * data)29 static int caching_kthread(void *data)
30 {
31 	struct btrfs_root *root = data;
32 	struct btrfs_fs_info *fs_info = root->fs_info;
33 	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
34 	struct btrfs_key key;
35 	struct btrfs_path *path;
36 	struct extent_buffer *leaf;
37 	u64 last = (u64)-1;
38 	int slot;
39 	int ret;
40 
41 	if (!btrfs_test_opt(root, INODE_MAP_CACHE))
42 		return 0;
43 
44 	path = btrfs_alloc_path();
45 	if (!path)
46 		return -ENOMEM;
47 
48 	/* Since the commit root is read-only, we can safely skip locking. */
49 	path->skip_locking = 1;
50 	path->search_commit_root = 1;
51 	path->reada = 2;
52 
53 	key.objectid = BTRFS_FIRST_FREE_OBJECTID;
54 	key.offset = 0;
55 	key.type = BTRFS_INODE_ITEM_KEY;
56 again:
57 	/* need to make sure the commit_root doesn't disappear */
58 	down_read(&fs_info->commit_root_sem);
59 
60 	ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
61 	if (ret < 0)
62 		goto out;
63 
64 	while (1) {
65 		if (btrfs_fs_closing(fs_info))
66 			goto out;
67 
68 		leaf = path->nodes[0];
69 		slot = path->slots[0];
70 		if (slot >= btrfs_header_nritems(leaf)) {
71 			ret = btrfs_next_leaf(root, path);
72 			if (ret < 0)
73 				goto out;
74 			else if (ret > 0)
75 				break;
76 
77 			if (need_resched() ||
78 			    btrfs_transaction_in_commit(fs_info)) {
79 				leaf = path->nodes[0];
80 
81 				if (WARN_ON(btrfs_header_nritems(leaf) == 0))
82 					break;
83 
84 				/*
85 				 * Save the key so we can advances forward
86 				 * in the next search.
87 				 */
88 				btrfs_item_key_to_cpu(leaf, &key, 0);
89 				btrfs_release_path(path);
90 				root->ino_cache_progress = last;
91 				up_read(&fs_info->commit_root_sem);
92 				schedule_timeout(1);
93 				goto again;
94 			} else
95 				continue;
96 		}
97 
98 		btrfs_item_key_to_cpu(leaf, &key, slot);
99 
100 		if (key.type != BTRFS_INODE_ITEM_KEY)
101 			goto next;
102 
103 		if (key.objectid >= root->highest_objectid)
104 			break;
105 
106 		if (last != (u64)-1 && last + 1 != key.objectid) {
107 			__btrfs_add_free_space(ctl, last + 1,
108 					       key.objectid - last - 1);
109 			wake_up(&root->ino_cache_wait);
110 		}
111 
112 		last = key.objectid;
113 next:
114 		path->slots[0]++;
115 	}
116 
117 	if (last < root->highest_objectid - 1) {
118 		__btrfs_add_free_space(ctl, last + 1,
119 				       root->highest_objectid - last - 1);
120 	}
121 
122 	spin_lock(&root->ino_cache_lock);
123 	root->ino_cache_state = BTRFS_CACHE_FINISHED;
124 	spin_unlock(&root->ino_cache_lock);
125 
126 	root->ino_cache_progress = (u64)-1;
127 	btrfs_unpin_free_ino(root);
128 out:
129 	wake_up(&root->ino_cache_wait);
130 	up_read(&fs_info->commit_root_sem);
131 
132 	btrfs_free_path(path);
133 
134 	return ret;
135 }
136 
start_caching(struct btrfs_root * root)137 static void start_caching(struct btrfs_root *root)
138 {
139 	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
140 	struct task_struct *tsk;
141 	int ret;
142 	u64 objectid;
143 
144 	if (!btrfs_test_opt(root, INODE_MAP_CACHE))
145 		return;
146 
147 	spin_lock(&root->ino_cache_lock);
148 	if (root->ino_cache_state != BTRFS_CACHE_NO) {
149 		spin_unlock(&root->ino_cache_lock);
150 		return;
151 	}
152 
153 	root->ino_cache_state = BTRFS_CACHE_STARTED;
154 	spin_unlock(&root->ino_cache_lock);
155 
156 	ret = load_free_ino_cache(root->fs_info, root);
157 	if (ret == 1) {
158 		spin_lock(&root->ino_cache_lock);
159 		root->ino_cache_state = BTRFS_CACHE_FINISHED;
160 		spin_unlock(&root->ino_cache_lock);
161 		wake_up(&root->ino_cache_wait);
162 		return;
163 	}
164 
165 	/*
166 	 * It can be quite time-consuming to fill the cache by searching
167 	 * through the extent tree, and this can keep ino allocation path
168 	 * waiting. Therefore at start we quickly find out the highest
169 	 * inode number and we know we can use inode numbers which fall in
170 	 * [highest_ino + 1, BTRFS_LAST_FREE_OBJECTID].
171 	 */
172 	ret = btrfs_find_free_objectid(root, &objectid);
173 	if (!ret && objectid <= BTRFS_LAST_FREE_OBJECTID) {
174 		__btrfs_add_free_space(ctl, objectid,
175 				       BTRFS_LAST_FREE_OBJECTID - objectid + 1);
176 	}
177 
178 	tsk = kthread_run(caching_kthread, root, "btrfs-ino-cache-%llu",
179 			  root->root_key.objectid);
180 	if (IS_ERR(tsk)) {
181 		btrfs_warn(root->fs_info, "failed to start inode caching task");
182 		btrfs_clear_pending_and_info(root->fs_info, INODE_MAP_CACHE,
183 				"disabling inode map caching");
184 	}
185 }
186 
btrfs_find_free_ino(struct btrfs_root * root,u64 * objectid)187 int btrfs_find_free_ino(struct btrfs_root *root, u64 *objectid)
188 {
189 	if (!btrfs_test_opt(root, INODE_MAP_CACHE))
190 		return btrfs_find_free_objectid(root, objectid);
191 
192 again:
193 	*objectid = btrfs_find_ino_for_alloc(root);
194 
195 	if (*objectid != 0)
196 		return 0;
197 
198 	start_caching(root);
199 
200 	wait_event(root->ino_cache_wait,
201 		   root->ino_cache_state == BTRFS_CACHE_FINISHED ||
202 		   root->free_ino_ctl->free_space > 0);
203 
204 	if (root->ino_cache_state == BTRFS_CACHE_FINISHED &&
205 	    root->free_ino_ctl->free_space == 0)
206 		return -ENOSPC;
207 	else
208 		goto again;
209 }
210 
btrfs_return_ino(struct btrfs_root * root,u64 objectid)211 void btrfs_return_ino(struct btrfs_root *root, u64 objectid)
212 {
213 	struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
214 
215 	if (!btrfs_test_opt(root, INODE_MAP_CACHE))
216 		return;
217 again:
218 	if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
219 		__btrfs_add_free_space(pinned, objectid, 1);
220 	} else {
221 		down_write(&root->fs_info->commit_root_sem);
222 		spin_lock(&root->ino_cache_lock);
223 		if (root->ino_cache_state == BTRFS_CACHE_FINISHED) {
224 			spin_unlock(&root->ino_cache_lock);
225 			up_write(&root->fs_info->commit_root_sem);
226 			goto again;
227 		}
228 		spin_unlock(&root->ino_cache_lock);
229 
230 		start_caching(root);
231 
232 		__btrfs_add_free_space(pinned, objectid, 1);
233 
234 		up_write(&root->fs_info->commit_root_sem);
235 	}
236 }
237 
238 /*
239  * When a transaction is committed, we'll move those inode numbers which are
240  * smaller than root->ino_cache_progress from pinned tree to free_ino tree, and
241  * others will just be dropped, because the commit root we were searching has
242  * changed.
243  *
244  * Must be called with root->fs_info->commit_root_sem held
245  */
btrfs_unpin_free_ino(struct btrfs_root * root)246 void btrfs_unpin_free_ino(struct btrfs_root *root)
247 {
248 	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
249 	struct rb_root *rbroot = &root->free_ino_pinned->free_space_offset;
250 	spinlock_t *rbroot_lock = &root->free_ino_pinned->tree_lock;
251 	struct btrfs_free_space *info;
252 	struct rb_node *n;
253 	u64 count;
254 
255 	if (!btrfs_test_opt(root, INODE_MAP_CACHE))
256 		return;
257 
258 	while (1) {
259 		bool add_to_ctl = true;
260 
261 		spin_lock(rbroot_lock);
262 		n = rb_first(rbroot);
263 		if (!n) {
264 			spin_unlock(rbroot_lock);
265 			break;
266 		}
267 
268 		info = rb_entry(n, struct btrfs_free_space, offset_index);
269 		BUG_ON(info->bitmap); /* Logic error */
270 
271 		if (info->offset > root->ino_cache_progress)
272 			add_to_ctl = false;
273 		else if (info->offset + info->bytes > root->ino_cache_progress)
274 			count = root->ino_cache_progress - info->offset + 1;
275 		else
276 			count = info->bytes;
277 
278 		rb_erase(&info->offset_index, rbroot);
279 		spin_unlock(rbroot_lock);
280 		if (add_to_ctl)
281 			__btrfs_add_free_space(ctl, info->offset, count);
282 		kmem_cache_free(btrfs_free_space_cachep, info);
283 	}
284 }
285 
286 #define INIT_THRESHOLD	(((1024 * 32) / 2) / sizeof(struct btrfs_free_space))
287 #define INODES_PER_BITMAP (PAGE_CACHE_SIZE * 8)
288 
289 /*
290  * The goal is to keep the memory used by the free_ino tree won't
291  * exceed the memory if we use bitmaps only.
292  */
recalculate_thresholds(struct btrfs_free_space_ctl * ctl)293 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
294 {
295 	struct btrfs_free_space *info;
296 	struct rb_node *n;
297 	int max_ino;
298 	int max_bitmaps;
299 
300 	n = rb_last(&ctl->free_space_offset);
301 	if (!n) {
302 		ctl->extents_thresh = INIT_THRESHOLD;
303 		return;
304 	}
305 	info = rb_entry(n, struct btrfs_free_space, offset_index);
306 
307 	/*
308 	 * Find the maximum inode number in the filesystem. Note we
309 	 * ignore the fact that this can be a bitmap, because we are
310 	 * not doing precise calculation.
311 	 */
312 	max_ino = info->bytes - 1;
313 
314 	max_bitmaps = ALIGN(max_ino, INODES_PER_BITMAP) / INODES_PER_BITMAP;
315 	if (max_bitmaps <= ctl->total_bitmaps) {
316 		ctl->extents_thresh = 0;
317 		return;
318 	}
319 
320 	ctl->extents_thresh = (max_bitmaps - ctl->total_bitmaps) *
321 				PAGE_CACHE_SIZE / sizeof(*info);
322 }
323 
324 /*
325  * We don't fall back to bitmap, if we are below the extents threshold
326  * or this chunk of inode numbers is a big one.
327  */
use_bitmap(struct btrfs_free_space_ctl * ctl,struct btrfs_free_space * info)328 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
329 		       struct btrfs_free_space *info)
330 {
331 	if (ctl->free_extents < ctl->extents_thresh ||
332 	    info->bytes > INODES_PER_BITMAP / 10)
333 		return false;
334 
335 	return true;
336 }
337 
338 static struct btrfs_free_space_op free_ino_op = {
339 	.recalc_thresholds	= recalculate_thresholds,
340 	.use_bitmap		= use_bitmap,
341 };
342 
pinned_recalc_thresholds(struct btrfs_free_space_ctl * ctl)343 static void pinned_recalc_thresholds(struct btrfs_free_space_ctl *ctl)
344 {
345 }
346 
pinned_use_bitmap(struct btrfs_free_space_ctl * ctl,struct btrfs_free_space * info)347 static bool pinned_use_bitmap(struct btrfs_free_space_ctl *ctl,
348 			      struct btrfs_free_space *info)
349 {
350 	/*
351 	 * We always use extents for two reasons:
352 	 *
353 	 * - The pinned tree is only used during the process of caching
354 	 *   work.
355 	 * - Make code simpler. See btrfs_unpin_free_ino().
356 	 */
357 	return false;
358 }
359 
360 static struct btrfs_free_space_op pinned_free_ino_op = {
361 	.recalc_thresholds	= pinned_recalc_thresholds,
362 	.use_bitmap		= pinned_use_bitmap,
363 };
364 
btrfs_init_free_ino_ctl(struct btrfs_root * root)365 void btrfs_init_free_ino_ctl(struct btrfs_root *root)
366 {
367 	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
368 	struct btrfs_free_space_ctl *pinned = root->free_ino_pinned;
369 
370 	spin_lock_init(&ctl->tree_lock);
371 	ctl->unit = 1;
372 	ctl->start = 0;
373 	ctl->private = NULL;
374 	ctl->op = &free_ino_op;
375 	INIT_LIST_HEAD(&ctl->trimming_ranges);
376 	mutex_init(&ctl->cache_writeout_mutex);
377 
378 	/*
379 	 * Initially we allow to use 16K of ram to cache chunks of
380 	 * inode numbers before we resort to bitmaps. This is somewhat
381 	 * arbitrary, but it will be adjusted in runtime.
382 	 */
383 	ctl->extents_thresh = INIT_THRESHOLD;
384 
385 	spin_lock_init(&pinned->tree_lock);
386 	pinned->unit = 1;
387 	pinned->start = 0;
388 	pinned->private = NULL;
389 	pinned->extents_thresh = 0;
390 	pinned->op = &pinned_free_ino_op;
391 }
392 
btrfs_save_ino_cache(struct btrfs_root * root,struct btrfs_trans_handle * trans)393 int btrfs_save_ino_cache(struct btrfs_root *root,
394 			 struct btrfs_trans_handle *trans)
395 {
396 	struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
397 	struct btrfs_path *path;
398 	struct inode *inode;
399 	struct btrfs_block_rsv *rsv;
400 	u64 num_bytes;
401 	u64 alloc_hint = 0;
402 	int ret;
403 	int prealloc;
404 	bool retry = false;
405 
406 	/* only fs tree and subvol/snap needs ino cache */
407 	if (root->root_key.objectid != BTRFS_FS_TREE_OBJECTID &&
408 	    (root->root_key.objectid < BTRFS_FIRST_FREE_OBJECTID ||
409 	     root->root_key.objectid > BTRFS_LAST_FREE_OBJECTID))
410 		return 0;
411 
412 	/* Don't save inode cache if we are deleting this root */
413 	if (btrfs_root_refs(&root->root_item) == 0)
414 		return 0;
415 
416 	if (!btrfs_test_opt(root, INODE_MAP_CACHE))
417 		return 0;
418 
419 	path = btrfs_alloc_path();
420 	if (!path)
421 		return -ENOMEM;
422 
423 	rsv = trans->block_rsv;
424 	trans->block_rsv = &root->fs_info->trans_block_rsv;
425 
426 	num_bytes = trans->bytes_reserved;
427 	/*
428 	 * 1 item for inode item insertion if need
429 	 * 4 items for inode item update (in the worst case)
430 	 * 1 items for slack space if we need do truncation
431 	 * 1 item for free space object
432 	 * 3 items for pre-allocation
433 	 */
434 	trans->bytes_reserved = btrfs_calc_trans_metadata_size(root, 10);
435 	ret = btrfs_block_rsv_add(root, trans->block_rsv,
436 				  trans->bytes_reserved,
437 				  BTRFS_RESERVE_NO_FLUSH);
438 	if (ret)
439 		goto out;
440 	trace_btrfs_space_reservation(root->fs_info, "ino_cache",
441 				      trans->transid, trans->bytes_reserved, 1);
442 again:
443 	inode = lookup_free_ino_inode(root, path);
444 	if (IS_ERR(inode) && (PTR_ERR(inode) != -ENOENT || retry)) {
445 		ret = PTR_ERR(inode);
446 		goto out_release;
447 	}
448 
449 	if (IS_ERR(inode)) {
450 		BUG_ON(retry); /* Logic error */
451 		retry = true;
452 
453 		ret = create_free_ino_inode(root, trans, path);
454 		if (ret)
455 			goto out_release;
456 		goto again;
457 	}
458 
459 	BTRFS_I(inode)->generation = 0;
460 	ret = btrfs_update_inode(trans, root, inode);
461 	if (ret) {
462 		btrfs_abort_transaction(trans, root, ret);
463 		goto out_put;
464 	}
465 
466 	if (i_size_read(inode) > 0) {
467 		ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
468 		if (ret) {
469 			if (ret != -ENOSPC)
470 				btrfs_abort_transaction(trans, root, ret);
471 			goto out_put;
472 		}
473 	}
474 
475 	spin_lock(&root->ino_cache_lock);
476 	if (root->ino_cache_state != BTRFS_CACHE_FINISHED) {
477 		ret = -1;
478 		spin_unlock(&root->ino_cache_lock);
479 		goto out_put;
480 	}
481 	spin_unlock(&root->ino_cache_lock);
482 
483 	spin_lock(&ctl->tree_lock);
484 	prealloc = sizeof(struct btrfs_free_space) * ctl->free_extents;
485 	prealloc = ALIGN(prealloc, PAGE_CACHE_SIZE);
486 	prealloc += ctl->total_bitmaps * PAGE_CACHE_SIZE;
487 	spin_unlock(&ctl->tree_lock);
488 
489 	/* Just to make sure we have enough space */
490 	prealloc += 8 * PAGE_CACHE_SIZE;
491 
492 	ret = btrfs_delalloc_reserve_space(inode, 0, prealloc);
493 	if (ret)
494 		goto out_put;
495 
496 	ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, prealloc,
497 					      prealloc, prealloc, &alloc_hint);
498 	if (ret) {
499 		btrfs_delalloc_release_space(inode, 0, prealloc);
500 		goto out_put;
501 	}
502 	btrfs_free_reserved_data_space(inode, 0, prealloc);
503 
504 	ret = btrfs_write_out_ino_cache(root, trans, path, inode);
505 out_put:
506 	iput(inode);
507 out_release:
508 	trace_btrfs_space_reservation(root->fs_info, "ino_cache",
509 				      trans->transid, trans->bytes_reserved, 0);
510 	btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
511 out:
512 	trans->block_rsv = rsv;
513 	trans->bytes_reserved = num_bytes;
514 
515 	btrfs_free_path(path);
516 	return ret;
517 }
518 
btrfs_find_highest_objectid(struct btrfs_root * root,u64 * objectid)519 int btrfs_find_highest_objectid(struct btrfs_root *root, u64 *objectid)
520 {
521 	struct btrfs_path *path;
522 	int ret;
523 	struct extent_buffer *l;
524 	struct btrfs_key search_key;
525 	struct btrfs_key found_key;
526 	int slot;
527 
528 	path = btrfs_alloc_path();
529 	if (!path)
530 		return -ENOMEM;
531 
532 	search_key.objectid = BTRFS_LAST_FREE_OBJECTID;
533 	search_key.type = -1;
534 	search_key.offset = (u64)-1;
535 	ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
536 	if (ret < 0)
537 		goto error;
538 	BUG_ON(ret == 0); /* Corruption */
539 	if (path->slots[0] > 0) {
540 		slot = path->slots[0] - 1;
541 		l = path->nodes[0];
542 		btrfs_item_key_to_cpu(l, &found_key, slot);
543 		*objectid = max_t(u64, found_key.objectid,
544 				  BTRFS_FIRST_FREE_OBJECTID - 1);
545 	} else {
546 		*objectid = BTRFS_FIRST_FREE_OBJECTID - 1;
547 	}
548 	ret = 0;
549 error:
550 	btrfs_free_path(path);
551 	return ret;
552 }
553 
btrfs_find_free_objectid(struct btrfs_root * root,u64 * objectid)554 int btrfs_find_free_objectid(struct btrfs_root *root, u64 *objectid)
555 {
556 	int ret;
557 	mutex_lock(&root->objectid_mutex);
558 
559 	if (unlikely(root->highest_objectid >= BTRFS_LAST_FREE_OBJECTID)) {
560 		ret = -ENOSPC;
561 		goto out;
562 	}
563 
564 	*objectid = ++root->highest_objectid;
565 	ret = 0;
566 out:
567 	mutex_unlock(&root->objectid_mutex);
568 	return ret;
569 }
570