• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2007 Oracle.  All rights reserved.
4  */
5 
6 #include <linux/sched.h>
7 #include "ctree.h"
8 #include "disk-io.h"
9 #include "print-tree.h"
10 #include "transaction.h"
11 #include "locking.h"
12 
13 static struct kmem_cache *btrfs_inode_defrag_cachep;
14 
15 /*
16  * When auto defrag is enabled we queue up these defrag structs to remember
17  * which inodes need defragging passes.
18  */
19 struct inode_defrag {
20 	struct rb_node rb_node;
21 	/* Inode number */
22 	u64 ino;
23 	/*
24 	 * Transid where the defrag was added, we search for extents newer than
25 	 * this.
26 	 */
27 	u64 transid;
28 
29 	/* Root objectid */
30 	u64 root;
31 
32 	/*
33 	 * The extent size threshold for autodefrag.
34 	 *
35 	 * This value is different for compressed/non-compressed extents, thus
36 	 * needs to be passed from higher layer.
37 	 * (aka, inode_should_defrag())
38 	 */
39 	u32 extent_thresh;
40 };
41 
__compare_inode_defrag(struct inode_defrag * defrag1,struct inode_defrag * defrag2)42 static int __compare_inode_defrag(struct inode_defrag *defrag1,
43 				  struct inode_defrag *defrag2)
44 {
45 	if (defrag1->root > defrag2->root)
46 		return 1;
47 	else if (defrag1->root < defrag2->root)
48 		return -1;
49 	else if (defrag1->ino > defrag2->ino)
50 		return 1;
51 	else if (defrag1->ino < defrag2->ino)
52 		return -1;
53 	else
54 		return 0;
55 }
56 
57 /*
58  * Pop a record for an inode into the defrag tree.  The lock must be held
59  * already.
60  *
61  * If you're inserting a record for an older transid than an existing record,
62  * the transid already in the tree is lowered.
63  *
64  * If an existing record is found the defrag item you pass in is freed.
65  */
__btrfs_add_inode_defrag(struct btrfs_inode * inode,struct inode_defrag * defrag)66 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
67 				    struct inode_defrag *defrag)
68 {
69 	struct btrfs_fs_info *fs_info = inode->root->fs_info;
70 	struct inode_defrag *entry;
71 	struct rb_node **p;
72 	struct rb_node *parent = NULL;
73 	int ret;
74 
75 	p = &fs_info->defrag_inodes.rb_node;
76 	while (*p) {
77 		parent = *p;
78 		entry = rb_entry(parent, struct inode_defrag, rb_node);
79 
80 		ret = __compare_inode_defrag(defrag, entry);
81 		if (ret < 0)
82 			p = &parent->rb_left;
83 		else if (ret > 0)
84 			p = &parent->rb_right;
85 		else {
86 			/*
87 			 * If we're reinserting an entry for an old defrag run,
88 			 * make sure to lower the transid of our existing
89 			 * record.
90 			 */
91 			if (defrag->transid < entry->transid)
92 				entry->transid = defrag->transid;
93 			entry->extent_thresh = min(defrag->extent_thresh,
94 						   entry->extent_thresh);
95 			return -EEXIST;
96 		}
97 	}
98 	set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
99 	rb_link_node(&defrag->rb_node, parent, p);
100 	rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
101 	return 0;
102 }
103 
__need_auto_defrag(struct btrfs_fs_info * fs_info)104 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
105 {
106 	if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
107 		return 0;
108 
109 	if (btrfs_fs_closing(fs_info))
110 		return 0;
111 
112 	return 1;
113 }
114 
115 /*
116  * Insert a defrag record for this inode if auto defrag is enabled.
117  */
btrfs_add_inode_defrag(struct btrfs_trans_handle * trans,struct btrfs_inode * inode,u32 extent_thresh)118 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
119 			   struct btrfs_inode *inode, u32 extent_thresh)
120 {
121 	struct btrfs_root *root = inode->root;
122 	struct btrfs_fs_info *fs_info = root->fs_info;
123 	struct inode_defrag *defrag;
124 	u64 transid;
125 	int ret;
126 
127 	if (!__need_auto_defrag(fs_info))
128 		return 0;
129 
130 	if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
131 		return 0;
132 
133 	if (trans)
134 		transid = trans->transid;
135 	else
136 		transid = inode->root->last_trans;
137 
138 	defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
139 	if (!defrag)
140 		return -ENOMEM;
141 
142 	defrag->ino = btrfs_ino(inode);
143 	defrag->transid = transid;
144 	defrag->root = root->root_key.objectid;
145 	defrag->extent_thresh = extent_thresh;
146 
147 	spin_lock(&fs_info->defrag_inodes_lock);
148 	if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
149 		/*
150 		 * If we set IN_DEFRAG flag and evict the inode from memory,
151 		 * and then re-read this inode, this new inode doesn't have
152 		 * IN_DEFRAG flag. At the case, we may find the existed defrag.
153 		 */
154 		ret = __btrfs_add_inode_defrag(inode, defrag);
155 		if (ret)
156 			kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
157 	} else {
158 		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
159 	}
160 	spin_unlock(&fs_info->defrag_inodes_lock);
161 	return 0;
162 }
163 
164 /*
165  * Pick the defragable inode that we want, if it doesn't exist, we will get the
166  * next one.
167  */
btrfs_pick_defrag_inode(struct btrfs_fs_info * fs_info,u64 root,u64 ino)168 static struct inode_defrag *btrfs_pick_defrag_inode(
169 			struct btrfs_fs_info *fs_info, u64 root, u64 ino)
170 {
171 	struct inode_defrag *entry = NULL;
172 	struct inode_defrag tmp;
173 	struct rb_node *p;
174 	struct rb_node *parent = NULL;
175 	int ret;
176 
177 	tmp.ino = ino;
178 	tmp.root = root;
179 
180 	spin_lock(&fs_info->defrag_inodes_lock);
181 	p = fs_info->defrag_inodes.rb_node;
182 	while (p) {
183 		parent = p;
184 		entry = rb_entry(parent, struct inode_defrag, rb_node);
185 
186 		ret = __compare_inode_defrag(&tmp, entry);
187 		if (ret < 0)
188 			p = parent->rb_left;
189 		else if (ret > 0)
190 			p = parent->rb_right;
191 		else
192 			goto out;
193 	}
194 
195 	if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
196 		parent = rb_next(parent);
197 		if (parent)
198 			entry = rb_entry(parent, struct inode_defrag, rb_node);
199 		else
200 			entry = NULL;
201 	}
202 out:
203 	if (entry)
204 		rb_erase(parent, &fs_info->defrag_inodes);
205 	spin_unlock(&fs_info->defrag_inodes_lock);
206 	return entry;
207 }
208 
btrfs_cleanup_defrag_inodes(struct btrfs_fs_info * fs_info)209 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
210 {
211 	struct inode_defrag *defrag;
212 	struct rb_node *node;
213 
214 	spin_lock(&fs_info->defrag_inodes_lock);
215 	node = rb_first(&fs_info->defrag_inodes);
216 	while (node) {
217 		rb_erase(node, &fs_info->defrag_inodes);
218 		defrag = rb_entry(node, struct inode_defrag, rb_node);
219 		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
220 
221 		cond_resched_lock(&fs_info->defrag_inodes_lock);
222 
223 		node = rb_first(&fs_info->defrag_inodes);
224 	}
225 	spin_unlock(&fs_info->defrag_inodes_lock);
226 }
227 
228 #define BTRFS_DEFRAG_BATCH	1024
229 
__btrfs_run_defrag_inode(struct btrfs_fs_info * fs_info,struct inode_defrag * defrag)230 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
231 				    struct inode_defrag *defrag)
232 {
233 	struct btrfs_root *inode_root;
234 	struct inode *inode;
235 	struct btrfs_ioctl_defrag_range_args range;
236 	int ret = 0;
237 	u64 cur = 0;
238 
239 again:
240 	if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
241 		goto cleanup;
242 	if (!__need_auto_defrag(fs_info))
243 		goto cleanup;
244 
245 	/* Get the inode */
246 	inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
247 	if (IS_ERR(inode_root)) {
248 		ret = PTR_ERR(inode_root);
249 		goto cleanup;
250 	}
251 
252 	inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
253 	btrfs_put_root(inode_root);
254 	if (IS_ERR(inode)) {
255 		ret = PTR_ERR(inode);
256 		goto cleanup;
257 	}
258 
259 	if (cur >= i_size_read(inode)) {
260 		iput(inode);
261 		goto cleanup;
262 	}
263 
264 	/* Do a chunk of defrag */
265 	clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
266 	memset(&range, 0, sizeof(range));
267 	range.len = (u64)-1;
268 	range.start = cur;
269 	range.extent_thresh = defrag->extent_thresh;
270 
271 	sb_start_write(fs_info->sb);
272 	ret = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
273 				       BTRFS_DEFRAG_BATCH);
274 	sb_end_write(fs_info->sb);
275 	iput(inode);
276 
277 	if (ret < 0)
278 		goto cleanup;
279 
280 	cur = max(cur + fs_info->sectorsize, range.start);
281 	goto again;
282 
283 cleanup:
284 	kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
285 	return ret;
286 }
287 
288 /*
289  * Run through the list of inodes in the FS that need defragging.
290  */
btrfs_run_defrag_inodes(struct btrfs_fs_info * fs_info)291 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
292 {
293 	struct inode_defrag *defrag;
294 	u64 first_ino = 0;
295 	u64 root_objectid = 0;
296 
297 	atomic_inc(&fs_info->defrag_running);
298 	while (1) {
299 		/* Pause the auto defragger. */
300 		if (test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state))
301 			break;
302 
303 		if (!__need_auto_defrag(fs_info))
304 			break;
305 
306 		/* find an inode to defrag */
307 		defrag = btrfs_pick_defrag_inode(fs_info, root_objectid, first_ino);
308 		if (!defrag) {
309 			if (root_objectid || first_ino) {
310 				root_objectid = 0;
311 				first_ino = 0;
312 				continue;
313 			} else {
314 				break;
315 			}
316 		}
317 
318 		first_ino = defrag->ino + 1;
319 		root_objectid = defrag->root;
320 
321 		__btrfs_run_defrag_inode(fs_info, defrag);
322 	}
323 	atomic_dec(&fs_info->defrag_running);
324 
325 	/*
326 	 * During unmount, we use the transaction_wait queue to wait for the
327 	 * defragger to stop.
328 	 */
329 	wake_up(&fs_info->transaction_wait);
330 	return 0;
331 }
332 
333 /*
334  * Defrag all the leaves in a given btree.
335  * Read all the leaves and try to get key order to
336  * better reflect disk order
337  */
338 
btrfs_defrag_leaves(struct btrfs_trans_handle * trans,struct btrfs_root * root)339 int btrfs_defrag_leaves(struct btrfs_trans_handle *trans,
340 			struct btrfs_root *root)
341 {
342 	struct btrfs_path *path = NULL;
343 	struct btrfs_key key;
344 	int ret = 0;
345 	int wret;
346 	int level;
347 	int next_key_ret = 0;
348 	u64 last_ret = 0;
349 
350 	if (!test_bit(BTRFS_ROOT_SHAREABLE, &root->state))
351 		goto out;
352 
353 	path = btrfs_alloc_path();
354 	if (!path) {
355 		ret = -ENOMEM;
356 		goto out;
357 	}
358 
359 	level = btrfs_header_level(root->node);
360 
361 	if (level == 0)
362 		goto out;
363 
364 	if (root->defrag_progress.objectid == 0) {
365 		struct extent_buffer *root_node;
366 		u32 nritems;
367 
368 		root_node = btrfs_lock_root_node(root);
369 		nritems = btrfs_header_nritems(root_node);
370 		root->defrag_max.objectid = 0;
371 		/* from above we know this is not a leaf */
372 		btrfs_node_key_to_cpu(root_node, &root->defrag_max,
373 				      nritems - 1);
374 		btrfs_tree_unlock(root_node);
375 		free_extent_buffer(root_node);
376 		memset(&key, 0, sizeof(key));
377 	} else {
378 		memcpy(&key, &root->defrag_progress, sizeof(key));
379 	}
380 
381 	path->keep_locks = 1;
382 
383 	ret = btrfs_search_forward(root, &key, path, BTRFS_OLDEST_GENERATION);
384 	if (ret < 0)
385 		goto out;
386 	if (ret > 0) {
387 		ret = 0;
388 		goto out;
389 	}
390 	btrfs_release_path(path);
391 	/*
392 	 * We don't need a lock on a leaf. btrfs_realloc_node() will lock all
393 	 * leafs from path->nodes[1], so set lowest_level to 1 to avoid later
394 	 * a deadlock (attempting to write lock an already write locked leaf).
395 	 */
396 	path->lowest_level = 1;
397 	wret = btrfs_search_slot(trans, root, &key, path, 0, 1);
398 
399 	if (wret < 0) {
400 		ret = wret;
401 		goto out;
402 	}
403 	if (!path->nodes[1]) {
404 		ret = 0;
405 		goto out;
406 	}
407 	/*
408 	 * The node at level 1 must always be locked when our path has
409 	 * keep_locks set and lowest_level is 1, regardless of the value of
410 	 * path->slots[1].
411 	 */
412 	BUG_ON(path->locks[1] == 0);
413 	ret = btrfs_realloc_node(trans, root,
414 				 path->nodes[1], 0,
415 				 &last_ret,
416 				 &root->defrag_progress);
417 	if (ret) {
418 		WARN_ON(ret == -EAGAIN);
419 		goto out;
420 	}
421 	/*
422 	 * Now that we reallocated the node we can find the next key. Note that
423 	 * btrfs_find_next_key() can release our path and do another search
424 	 * without COWing, this is because even with path->keep_locks = 1,
425 	 * btrfs_search_slot() / ctree.c:unlock_up() does not keeps a lock on a
426 	 * node when path->slots[node_level - 1] does not point to the last
427 	 * item or a slot beyond the last item (ctree.c:unlock_up()). Therefore
428 	 * we search for the next key after reallocating our node.
429 	 */
430 	path->slots[1] = btrfs_header_nritems(path->nodes[1]);
431 	next_key_ret = btrfs_find_next_key(root, path, &key, 1,
432 					   BTRFS_OLDEST_GENERATION);
433 	if (next_key_ret == 0) {
434 		memcpy(&root->defrag_progress, &key, sizeof(key));
435 		ret = -EAGAIN;
436 	}
437 out:
438 	btrfs_free_path(path);
439 	if (ret == -EAGAIN) {
440 		if (root->defrag_max.objectid > root->defrag_progress.objectid)
441 			goto done;
442 		if (root->defrag_max.type > root->defrag_progress.type)
443 			goto done;
444 		if (root->defrag_max.offset > root->defrag_progress.offset)
445 			goto done;
446 		ret = 0;
447 	}
448 done:
449 	if (ret != -EAGAIN)
450 		memset(&root->defrag_progress, 0,
451 		       sizeof(root->defrag_progress));
452 
453 	return ret;
454 }
455 
btrfs_auto_defrag_exit(void)456 void __cold btrfs_auto_defrag_exit(void)
457 {
458 	kmem_cache_destroy(btrfs_inode_defrag_cachep);
459 }
460 
btrfs_auto_defrag_init(void)461 int __init btrfs_auto_defrag_init(void)
462 {
463 	btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
464 					sizeof(struct inode_defrag), 0,
465 					SLAB_MEM_SPREAD,
466 					NULL);
467 	if (!btrfs_inode_defrag_cachep)
468 		return -ENOMEM;
469 
470 	return 0;
471 }
472