xref: /linux/fs/ocfs2/uptodate.c (revision 2da68a77)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * uptodate.c
4  *
5  * Tracking the up-to-date-ness of a local buffer_head with respect to
6  * the cluster.
7  *
8  * Copyright (C) 2002, 2004, 2005 Oracle.  All rights reserved.
9  *
10  * Standard buffer head caching flags (uptodate, etc) are insufficient
11  * in a clustered environment - a buffer may be marked up to date on
12  * our local node but could have been modified by another cluster
13  * member. As a result an additional (and performant) caching scheme
14  * is required. A further requirement is that we consume as little
15  * memory as possible - we never pin buffer_head structures in order
16  * to cache them.
17  *
18  * We track the existence of up to date buffers on the inodes which
19  * are associated with them. Because we don't want to pin
20  * buffer_heads, this is only a (strong) hint and several other checks
21  * are made in the I/O path to ensure that we don't use a stale or
22  * invalid buffer without going to disk:
23  *	- buffer_jbd is used liberally - if a bh is in the journal on
24  *	  this node then it *must* be up to date.
25  *	- the standard buffer_uptodate() macro is used to detect buffers
26  *	  which may be invalid (even if we have an up to date tracking
27  * 	  item for them)
28  *
29  * For a full understanding of how this code works together, one
30  * should read the callers in dlmglue.c, the I/O functions in
31  * buffer_head_io.c and ocfs2_journal_access in journal.c
32  */
33 
34 #include <linux/fs.h>
35 #include <linux/types.h>
36 #include <linux/slab.h>
37 #include <linux/highmem.h>
38 #include <linux/buffer_head.h>
39 #include <linux/rbtree.h>
40 
41 #include <cluster/masklog.h>
42 
43 #include "ocfs2.h"
44 
45 #include "inode.h"
46 #include "uptodate.h"
47 #include "ocfs2_trace.h"
48 
49 struct ocfs2_meta_cache_item {
50 	struct rb_node	c_node;
51 	sector_t	c_block;
52 };
53 
54 static struct kmem_cache *ocfs2_uptodate_cachep;
55 
56 u64 ocfs2_metadata_cache_owner(struct ocfs2_caching_info *ci)
57 {
58 	BUG_ON(!ci || !ci->ci_ops);
59 
60 	return ci->ci_ops->co_owner(ci);
61 }
62 
63 struct super_block *ocfs2_metadata_cache_get_super(struct ocfs2_caching_info *ci)
64 {
65 	BUG_ON(!ci || !ci->ci_ops);
66 
67 	return ci->ci_ops->co_get_super(ci);
68 }
69 
70 static void ocfs2_metadata_cache_lock(struct ocfs2_caching_info *ci)
71 {
72 	BUG_ON(!ci || !ci->ci_ops);
73 
74 	ci->ci_ops->co_cache_lock(ci);
75 }
76 
77 static void ocfs2_metadata_cache_unlock(struct ocfs2_caching_info *ci)
78 {
79 	BUG_ON(!ci || !ci->ci_ops);
80 
81 	ci->ci_ops->co_cache_unlock(ci);
82 }
83 
84 void ocfs2_metadata_cache_io_lock(struct ocfs2_caching_info *ci)
85 {
86 	BUG_ON(!ci || !ci->ci_ops);
87 
88 	ci->ci_ops->co_io_lock(ci);
89 }
90 
91 void ocfs2_metadata_cache_io_unlock(struct ocfs2_caching_info *ci)
92 {
93 	BUG_ON(!ci || !ci->ci_ops);
94 
95 	ci->ci_ops->co_io_unlock(ci);
96 }
97 
98 
99 static void ocfs2_metadata_cache_reset(struct ocfs2_caching_info *ci,
100 				       int clear)
101 {
102 	ci->ci_flags |= OCFS2_CACHE_FL_INLINE;
103 	ci->ci_num_cached = 0;
104 
105 	if (clear) {
106 		ci->ci_created_trans = 0;
107 		ci->ci_last_trans = 0;
108 	}
109 }
110 
111 void ocfs2_metadata_cache_init(struct ocfs2_caching_info *ci,
112 			       const struct ocfs2_caching_operations *ops)
113 {
114 	BUG_ON(!ops);
115 
116 	ci->ci_ops = ops;
117 	ocfs2_metadata_cache_reset(ci, 1);
118 }
119 
120 void ocfs2_metadata_cache_exit(struct ocfs2_caching_info *ci)
121 {
122 	ocfs2_metadata_cache_purge(ci);
123 	ocfs2_metadata_cache_reset(ci, 1);
124 }
125 
126 
127 /* No lock taken here as 'root' is not expected to be visible to other
128  * processes. */
129 static unsigned int ocfs2_purge_copied_metadata_tree(struct rb_root *root)
130 {
131 	unsigned int purged = 0;
132 	struct rb_node *node;
133 	struct ocfs2_meta_cache_item *item;
134 
135 	while ((node = rb_last(root)) != NULL) {
136 		item = rb_entry(node, struct ocfs2_meta_cache_item, c_node);
137 
138 		trace_ocfs2_purge_copied_metadata_tree(
139 					(unsigned long long) item->c_block);
140 
141 		rb_erase(&item->c_node, root);
142 		kmem_cache_free(ocfs2_uptodate_cachep, item);
143 
144 		purged++;
145 	}
146 	return purged;
147 }
148 
149 /* Called from locking and called from ocfs2_clear_inode. Dump the
150  * cache for a given inode.
151  *
152  * This function is a few more lines longer than necessary due to some
153  * accounting done here, but I think it's worth tracking down those
154  * bugs sooner -- Mark */
155 void ocfs2_metadata_cache_purge(struct ocfs2_caching_info *ci)
156 {
157 	unsigned int tree, to_purge, purged;
158 	struct rb_root root = RB_ROOT;
159 
160 	BUG_ON(!ci || !ci->ci_ops);
161 
162 	ocfs2_metadata_cache_lock(ci);
163 	tree = !(ci->ci_flags & OCFS2_CACHE_FL_INLINE);
164 	to_purge = ci->ci_num_cached;
165 
166 	trace_ocfs2_metadata_cache_purge(
167 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
168 		to_purge, tree);
169 
170 	/* If we're a tree, save off the root so that we can safely
171 	 * initialize the cache. We do the work to free tree members
172 	 * without the spinlock. */
173 	if (tree)
174 		root = ci->ci_cache.ci_tree;
175 
176 	ocfs2_metadata_cache_reset(ci, 0);
177 	ocfs2_metadata_cache_unlock(ci);
178 
179 	purged = ocfs2_purge_copied_metadata_tree(&root);
180 	/* If possible, track the number wiped so that we can more
181 	 * easily detect counting errors. Unfortunately, this is only
182 	 * meaningful for trees. */
183 	if (tree && purged != to_purge)
184 		mlog(ML_ERROR, "Owner %llu, count = %u, purged = %u\n",
185 		     (unsigned long long)ocfs2_metadata_cache_owner(ci),
186 		     to_purge, purged);
187 }
188 
189 /* Returns the index in the cache array, -1 if not found.
190  * Requires ip_lock. */
191 static int ocfs2_search_cache_array(struct ocfs2_caching_info *ci,
192 				    sector_t item)
193 {
194 	int i;
195 
196 	for (i = 0; i < ci->ci_num_cached; i++) {
197 		if (item == ci->ci_cache.ci_array[i])
198 			return i;
199 	}
200 
201 	return -1;
202 }
203 
204 /* Returns the cache item if found, otherwise NULL.
205  * Requires ip_lock. */
206 static struct ocfs2_meta_cache_item *
207 ocfs2_search_cache_tree(struct ocfs2_caching_info *ci,
208 			sector_t block)
209 {
210 	struct rb_node * n = ci->ci_cache.ci_tree.rb_node;
211 	struct ocfs2_meta_cache_item *item = NULL;
212 
213 	while (n) {
214 		item = rb_entry(n, struct ocfs2_meta_cache_item, c_node);
215 
216 		if (block < item->c_block)
217 			n = n->rb_left;
218 		else if (block > item->c_block)
219 			n = n->rb_right;
220 		else
221 			return item;
222 	}
223 
224 	return NULL;
225 }
226 
227 static int ocfs2_buffer_cached(struct ocfs2_caching_info *ci,
228 			       struct buffer_head *bh)
229 {
230 	int index = -1;
231 	struct ocfs2_meta_cache_item *item = NULL;
232 
233 	ocfs2_metadata_cache_lock(ci);
234 
235 	trace_ocfs2_buffer_cached_begin(
236 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
237 		(unsigned long long) bh->b_blocknr,
238 		!!(ci->ci_flags & OCFS2_CACHE_FL_INLINE));
239 
240 	if (ci->ci_flags & OCFS2_CACHE_FL_INLINE)
241 		index = ocfs2_search_cache_array(ci, bh->b_blocknr);
242 	else
243 		item = ocfs2_search_cache_tree(ci, bh->b_blocknr);
244 
245 	ocfs2_metadata_cache_unlock(ci);
246 
247 	trace_ocfs2_buffer_cached_end(index, item);
248 
249 	return (index != -1) || (item != NULL);
250 }
251 
252 /* Warning: even if it returns true, this does *not* guarantee that
253  * the block is stored in our inode metadata cache.
254  *
255  * This can be called under lock_buffer()
256  */
257 int ocfs2_buffer_uptodate(struct ocfs2_caching_info *ci,
258 			  struct buffer_head *bh)
259 {
260 	/* Doesn't matter if the bh is in our cache or not -- if it's
261 	 * not marked uptodate then we know it can't have correct
262 	 * data. */
263 	if (!buffer_uptodate(bh))
264 		return 0;
265 
266 	/* OCFS2 does not allow multiple nodes to be changing the same
267 	 * block at the same time. */
268 	if (buffer_jbd(bh))
269 		return 1;
270 
271 	/* Ok, locally the buffer is marked as up to date, now search
272 	 * our cache to see if we can trust that. */
273 	return ocfs2_buffer_cached(ci, bh);
274 }
275 
276 /*
277  * Determine whether a buffer is currently out on a read-ahead request.
278  * ci_io_sem should be held to serialize submitters with the logic here.
279  */
280 int ocfs2_buffer_read_ahead(struct ocfs2_caching_info *ci,
281 			    struct buffer_head *bh)
282 {
283 	return buffer_locked(bh) && ocfs2_buffer_cached(ci, bh);
284 }
285 
286 /* Requires ip_lock */
287 static void ocfs2_append_cache_array(struct ocfs2_caching_info *ci,
288 				     sector_t block)
289 {
290 	BUG_ON(ci->ci_num_cached >= OCFS2_CACHE_INFO_MAX_ARRAY);
291 
292 	trace_ocfs2_append_cache_array(
293 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
294 		(unsigned long long)block, ci->ci_num_cached);
295 
296 	ci->ci_cache.ci_array[ci->ci_num_cached] = block;
297 	ci->ci_num_cached++;
298 }
299 
300 /* By now the caller should have checked that the item does *not*
301  * exist in the tree.
302  * Requires ip_lock. */
303 static void __ocfs2_insert_cache_tree(struct ocfs2_caching_info *ci,
304 				      struct ocfs2_meta_cache_item *new)
305 {
306 	sector_t block = new->c_block;
307 	struct rb_node *parent = NULL;
308 	struct rb_node **p = &ci->ci_cache.ci_tree.rb_node;
309 	struct ocfs2_meta_cache_item *tmp;
310 
311 	trace_ocfs2_insert_cache_tree(
312 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
313 		(unsigned long long)block, ci->ci_num_cached);
314 
315 	while(*p) {
316 		parent = *p;
317 
318 		tmp = rb_entry(parent, struct ocfs2_meta_cache_item, c_node);
319 
320 		if (block < tmp->c_block)
321 			p = &(*p)->rb_left;
322 		else if (block > tmp->c_block)
323 			p = &(*p)->rb_right;
324 		else {
325 			/* This should never happen! */
326 			mlog(ML_ERROR, "Duplicate block %llu cached!\n",
327 			     (unsigned long long) block);
328 			BUG();
329 		}
330 	}
331 
332 	rb_link_node(&new->c_node, parent, p);
333 	rb_insert_color(&new->c_node, &ci->ci_cache.ci_tree);
334 	ci->ci_num_cached++;
335 }
336 
337 /* co_cache_lock() must be held */
338 static inline int ocfs2_insert_can_use_array(struct ocfs2_caching_info *ci)
339 {
340 	return (ci->ci_flags & OCFS2_CACHE_FL_INLINE) &&
341 		(ci->ci_num_cached < OCFS2_CACHE_INFO_MAX_ARRAY);
342 }
343 
344 /* tree should be exactly OCFS2_CACHE_INFO_MAX_ARRAY wide. NULL the
345  * pointers in tree after we use them - this allows caller to detect
346  * when to free in case of error.
347  *
348  * The co_cache_lock() must be held. */
349 static void ocfs2_expand_cache(struct ocfs2_caching_info *ci,
350 			       struct ocfs2_meta_cache_item **tree)
351 {
352 	int i;
353 
354 	mlog_bug_on_msg(ci->ci_num_cached != OCFS2_CACHE_INFO_MAX_ARRAY,
355 			"Owner %llu, num cached = %u, should be %u\n",
356 			(unsigned long long)ocfs2_metadata_cache_owner(ci),
357 			ci->ci_num_cached, OCFS2_CACHE_INFO_MAX_ARRAY);
358 	mlog_bug_on_msg(!(ci->ci_flags & OCFS2_CACHE_FL_INLINE),
359 			"Owner %llu not marked as inline anymore!\n",
360 			(unsigned long long)ocfs2_metadata_cache_owner(ci));
361 
362 	/* Be careful to initialize the tree members *first* because
363 	 * once the ci_tree is used, the array is junk... */
364 	for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++)
365 		tree[i]->c_block = ci->ci_cache.ci_array[i];
366 
367 	ci->ci_flags &= ~OCFS2_CACHE_FL_INLINE;
368 	ci->ci_cache.ci_tree = RB_ROOT;
369 	/* this will be set again by __ocfs2_insert_cache_tree */
370 	ci->ci_num_cached = 0;
371 
372 	for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) {
373 		__ocfs2_insert_cache_tree(ci, tree[i]);
374 		tree[i] = NULL;
375 	}
376 
377 	trace_ocfs2_expand_cache(
378 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
379 		ci->ci_flags, ci->ci_num_cached);
380 }
381 
382 /* Slow path function - memory allocation is necessary. See the
383  * comment above ocfs2_set_buffer_uptodate for more information. */
384 static void __ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
385 					sector_t block,
386 					int expand_tree)
387 {
388 	int i;
389 	struct ocfs2_meta_cache_item *new = NULL;
390 	struct ocfs2_meta_cache_item *tree[OCFS2_CACHE_INFO_MAX_ARRAY] =
391 		{ NULL, };
392 
393 	trace_ocfs2_set_buffer_uptodate(
394 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
395 		(unsigned long long)block, expand_tree);
396 
397 	new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_NOFS);
398 	if (!new) {
399 		mlog_errno(-ENOMEM);
400 		return;
401 	}
402 	new->c_block = block;
403 
404 	if (expand_tree) {
405 		/* Do *not* allocate an array here - the removal code
406 		 * has no way of tracking that. */
407 		for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) {
408 			tree[i] = kmem_cache_alloc(ocfs2_uptodate_cachep,
409 						   GFP_NOFS);
410 			if (!tree[i]) {
411 				mlog_errno(-ENOMEM);
412 				goto out_free;
413 			}
414 
415 			/* These are initialized in ocfs2_expand_cache! */
416 		}
417 	}
418 
419 	ocfs2_metadata_cache_lock(ci);
420 	if (ocfs2_insert_can_use_array(ci)) {
421 		/* Ok, items were removed from the cache in between
422 		 * locks. Detect this and revert back to the fast path */
423 		ocfs2_append_cache_array(ci, block);
424 		ocfs2_metadata_cache_unlock(ci);
425 		goto out_free;
426 	}
427 
428 	if (expand_tree)
429 		ocfs2_expand_cache(ci, tree);
430 
431 	__ocfs2_insert_cache_tree(ci, new);
432 	ocfs2_metadata_cache_unlock(ci);
433 
434 	new = NULL;
435 out_free:
436 	if (new)
437 		kmem_cache_free(ocfs2_uptodate_cachep, new);
438 
439 	/* If these were used, then ocfs2_expand_cache re-set them to
440 	 * NULL for us. */
441 	if (tree[0]) {
442 		for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++)
443 			if (tree[i])
444 				kmem_cache_free(ocfs2_uptodate_cachep,
445 						tree[i]);
446 	}
447 }
448 
449 /* Item insertion is guarded by co_io_lock(), so the insertion path takes
450  * advantage of this by not rechecking for a duplicate insert during
451  * the slow case. Additionally, if the cache needs to be bumped up to
452  * a tree, the code will not recheck after acquiring the lock --
453  * multiple paths cannot be expanding to a tree at the same time.
454  *
455  * The slow path takes into account that items can be removed
456  * (including the whole tree wiped and reset) when this process it out
457  * allocating memory. In those cases, it reverts back to the fast
458  * path.
459  *
460  * Note that this function may actually fail to insert the block if
461  * memory cannot be allocated. This is not fatal however (but may
462  * result in a performance penalty)
463  *
464  * Readahead buffers can be passed in here before the I/O request is
465  * completed.
466  */
467 void ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
468 			       struct buffer_head *bh)
469 {
470 	int expand;
471 
472 	/* The block may very well exist in our cache already, so avoid
473 	 * doing any more work in that case. */
474 	if (ocfs2_buffer_cached(ci, bh))
475 		return;
476 
477 	trace_ocfs2_set_buffer_uptodate_begin(
478 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
479 		(unsigned long long)bh->b_blocknr);
480 
481 	/* No need to recheck under spinlock - insertion is guarded by
482 	 * co_io_lock() */
483 	ocfs2_metadata_cache_lock(ci);
484 	if (ocfs2_insert_can_use_array(ci)) {
485 		/* Fast case - it's an array and there's a free
486 		 * spot. */
487 		ocfs2_append_cache_array(ci, bh->b_blocknr);
488 		ocfs2_metadata_cache_unlock(ci);
489 		return;
490 	}
491 
492 	expand = 0;
493 	if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
494 		/* We need to bump things up to a tree. */
495 		expand = 1;
496 	}
497 	ocfs2_metadata_cache_unlock(ci);
498 
499 	__ocfs2_set_buffer_uptodate(ci, bh->b_blocknr, expand);
500 }
501 
502 /* Called against a newly allocated buffer. Most likely nobody should
503  * be able to read this sort of metadata while it's still being
504  * allocated, but this is careful to take co_io_lock() anyway. */
505 void ocfs2_set_new_buffer_uptodate(struct ocfs2_caching_info *ci,
506 				   struct buffer_head *bh)
507 {
508 	/* This should definitely *not* exist in our cache */
509 	BUG_ON(ocfs2_buffer_cached(ci, bh));
510 
511 	set_buffer_uptodate(bh);
512 
513 	ocfs2_metadata_cache_io_lock(ci);
514 	ocfs2_set_buffer_uptodate(ci, bh);
515 	ocfs2_metadata_cache_io_unlock(ci);
516 }
517 
518 /* Requires ip_lock. */
519 static void ocfs2_remove_metadata_array(struct ocfs2_caching_info *ci,
520 					int index)
521 {
522 	sector_t *array = ci->ci_cache.ci_array;
523 	int bytes;
524 
525 	BUG_ON(index < 0 || index >= OCFS2_CACHE_INFO_MAX_ARRAY);
526 	BUG_ON(index >= ci->ci_num_cached);
527 	BUG_ON(!ci->ci_num_cached);
528 
529 	trace_ocfs2_remove_metadata_array(
530 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
531 		index, ci->ci_num_cached);
532 
533 	ci->ci_num_cached--;
534 
535 	/* don't need to copy if the array is now empty, or if we
536 	 * removed at the tail */
537 	if (ci->ci_num_cached && index < ci->ci_num_cached) {
538 		bytes = sizeof(sector_t) * (ci->ci_num_cached - index);
539 		memmove(&array[index], &array[index + 1], bytes);
540 	}
541 }
542 
543 /* Requires ip_lock. */
544 static void ocfs2_remove_metadata_tree(struct ocfs2_caching_info *ci,
545 				       struct ocfs2_meta_cache_item *item)
546 {
547 	trace_ocfs2_remove_metadata_tree(
548 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
549 		(unsigned long long)item->c_block);
550 
551 	rb_erase(&item->c_node, &ci->ci_cache.ci_tree);
552 	ci->ci_num_cached--;
553 }
554 
555 static void ocfs2_remove_block_from_cache(struct ocfs2_caching_info *ci,
556 					  sector_t block)
557 {
558 	int index;
559 	struct ocfs2_meta_cache_item *item = NULL;
560 
561 	ocfs2_metadata_cache_lock(ci);
562 	trace_ocfs2_remove_block_from_cache(
563 		(unsigned long long)ocfs2_metadata_cache_owner(ci),
564 		(unsigned long long) block, ci->ci_num_cached,
565 		ci->ci_flags);
566 
567 	if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
568 		index = ocfs2_search_cache_array(ci, block);
569 		if (index != -1)
570 			ocfs2_remove_metadata_array(ci, index);
571 	} else {
572 		item = ocfs2_search_cache_tree(ci, block);
573 		if (item)
574 			ocfs2_remove_metadata_tree(ci, item);
575 	}
576 	ocfs2_metadata_cache_unlock(ci);
577 
578 	if (item)
579 		kmem_cache_free(ocfs2_uptodate_cachep, item);
580 }
581 
582 /*
583  * Called when we remove a chunk of metadata from an inode. We don't
584  * bother reverting things to an inlined array in the case of a remove
585  * which moves us back under the limit.
586  */
587 void ocfs2_remove_from_cache(struct ocfs2_caching_info *ci,
588 			     struct buffer_head *bh)
589 {
590 	sector_t block = bh->b_blocknr;
591 
592 	ocfs2_remove_block_from_cache(ci, block);
593 }
594 
595 /* Called when we remove xattr clusters from an inode. */
596 void ocfs2_remove_xattr_clusters_from_cache(struct ocfs2_caching_info *ci,
597 					    sector_t block,
598 					    u32 c_len)
599 {
600 	struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
601 	unsigned int i, b_len = ocfs2_clusters_to_blocks(sb, 1) * c_len;
602 
603 	for (i = 0; i < b_len; i++, block++)
604 		ocfs2_remove_block_from_cache(ci, block);
605 }
606 
607 int __init init_ocfs2_uptodate_cache(void)
608 {
609 	ocfs2_uptodate_cachep = kmem_cache_create("ocfs2_uptodate",
610 				  sizeof(struct ocfs2_meta_cache_item),
611 				  0, SLAB_HWCACHE_ALIGN, NULL);
612 	if (!ocfs2_uptodate_cachep)
613 		return -ENOMEM;
614 
615 	return 0;
616 }
617 
618 void exit_ocfs2_uptodate_cache(void)
619 {
620 	kmem_cache_destroy(ocfs2_uptodate_cachep);
621 }
622