1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Tracking the up-to-date-ness of a local buffer_head with respect to
9 * Copyright (C) 2002, 2004, 2005 Oracle. All rights reserved.
11 * This program is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU General Public
13 * License as published by the Free Software Foundation; either
14 * version 2 of the License, or (at your option) any later version.
16 * This program is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * General Public License for more details.
21 * You should have received a copy of the GNU General Public
22 * License along with this program; if not, write to the
23 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
24 * Boston, MA 021110-1307, USA.
26 * Standard buffer head caching flags (uptodate, etc) are insufficient
27 * in a clustered environment - a buffer may be marked up to date on
28 * our local node but could have been modified by another cluster
29 * member. As a result an additional (and performant) caching scheme
30 * is required. A further requirement is that we consume as little
31 * memory as possible - we never pin buffer_head structures in order
34 * We track the existence of up to date buffers on the inodes which
35 * are associated with them. Because we don't want to pin
36 * buffer_heads, this is only a (strong) hint and several other checks
37 * are made in the I/O path to ensure that we don't use a stale or
38 * invalid buffer without going to disk:
39 * - buffer_jbd is used liberally - if a bh is in the journal on
40 * this node then it *must* be up to date.
41 * - the standard buffer_uptodate() macro is used to detect buffers
42 * which may be invalid (even if we have an up to date tracking
45 * For a full understanding of how this code works together, one
46 * should read the callers in dlmglue.c, the I/O functions in
47 * buffer_head_io.c and ocfs2_journal_access in journal.c
51 #include <linux/types.h>
52 #include <linux/slab.h>
53 #include <linux/highmem.h>
54 #include <linux/buffer_head.h>
55 #include <linux/rbtree.h>
57 #define MLOG_MASK_PREFIX ML_UPTODATE
59 #include <cluster/masklog.h>
66 struct ocfs2_meta_cache_item {
67 struct rb_node c_node;
71 static struct kmem_cache *ocfs2_uptodate_cachep = NULL;
73 u64 ocfs2_metadata_cache_owner(struct ocfs2_caching_info *ci)
75 BUG_ON(!ci || !ci->ci_ops);
77 return ci->ci_ops->co_owner(ci);
80 struct super_block *ocfs2_metadata_cache_get_super(struct ocfs2_caching_info *ci)
82 BUG_ON(!ci || !ci->ci_ops);
84 return ci->ci_ops->co_get_super(ci);
87 static void ocfs2_metadata_cache_lock(struct ocfs2_caching_info *ci)
89 BUG_ON(!ci || !ci->ci_ops);
91 ci->ci_ops->co_cache_lock(ci);
94 static void ocfs2_metadata_cache_unlock(struct ocfs2_caching_info *ci)
96 BUG_ON(!ci || !ci->ci_ops);
98 ci->ci_ops->co_cache_unlock(ci);
101 void ocfs2_metadata_cache_io_lock(struct ocfs2_caching_info *ci)
103 BUG_ON(!ci || !ci->ci_ops);
105 ci->ci_ops->co_io_lock(ci);
108 void ocfs2_metadata_cache_io_unlock(struct ocfs2_caching_info *ci)
110 BUG_ON(!ci || !ci->ci_ops);
112 ci->ci_ops->co_io_unlock(ci);
116 static void ocfs2_metadata_cache_reset(struct ocfs2_caching_info *ci,
119 ci->ci_flags |= OCFS2_CACHE_FL_INLINE;
120 ci->ci_num_cached = 0;
123 ci->ci_created_trans = 0;
124 ci->ci_last_trans = 0;
128 void ocfs2_metadata_cache_init(struct ocfs2_caching_info *ci,
129 const struct ocfs2_caching_operations *ops)
134 ocfs2_metadata_cache_reset(ci, 1);
137 void ocfs2_metadata_cache_exit(struct ocfs2_caching_info *ci)
139 ocfs2_metadata_cache_purge(ci);
140 ocfs2_metadata_cache_reset(ci, 1);
144 /* No lock taken here as 'root' is not expected to be visible to other
146 static unsigned int ocfs2_purge_copied_metadata_tree(struct rb_root *root)
148 unsigned int purged = 0;
149 struct rb_node *node;
150 struct ocfs2_meta_cache_item *item;
152 while ((node = rb_last(root)) != NULL) {
153 item = rb_entry(node, struct ocfs2_meta_cache_item, c_node);
155 mlog(0, "Purge item %llu\n",
156 (unsigned long long) item->c_block);
158 rb_erase(&item->c_node, root);
159 kmem_cache_free(ocfs2_uptodate_cachep, item);
166 /* Called from locking and called from ocfs2_clear_inode. Dump the
167 * cache for a given inode.
169 * This function is a few more lines longer than necessary due to some
170 * accounting done here, but I think it's worth tracking down those
171 * bugs sooner -- Mark */
172 void ocfs2_metadata_cache_purge(struct ocfs2_caching_info *ci)
174 unsigned int tree, to_purge, purged;
175 struct rb_root root = RB_ROOT;
177 BUG_ON(!ci || !ci->ci_ops);
179 ocfs2_metadata_cache_lock(ci);
180 tree = !(ci->ci_flags & OCFS2_CACHE_FL_INLINE);
181 to_purge = ci->ci_num_cached;
183 mlog(0, "Purge %u %s items from Owner %llu\n", to_purge,
184 tree ? "array" : "tree",
185 (unsigned long long)ocfs2_metadata_cache_owner(ci));
187 /* If we're a tree, save off the root so that we can safely
188 * initialize the cache. We do the work to free tree members
189 * without the spinlock. */
191 root = ci->ci_cache.ci_tree;
193 ocfs2_metadata_cache_reset(ci, 0);
194 ocfs2_metadata_cache_unlock(ci);
196 purged = ocfs2_purge_copied_metadata_tree(&root);
197 /* If possible, track the number wiped so that we can more
198 * easily detect counting errors. Unfortunately, this is only
199 * meaningful for trees. */
200 if (tree && purged != to_purge)
201 mlog(ML_ERROR, "Owner %llu, count = %u, purged = %u\n",
202 (unsigned long long)ocfs2_metadata_cache_owner(ci),
206 /* Returns the index in the cache array, -1 if not found.
207 * Requires ip_lock. */
208 static int ocfs2_search_cache_array(struct ocfs2_caching_info *ci,
213 for (i = 0; i < ci->ci_num_cached; i++) {
214 if (item == ci->ci_cache.ci_array[i])
221 /* Returns the cache item if found, otherwise NULL.
222 * Requires ip_lock. */
223 static struct ocfs2_meta_cache_item *
224 ocfs2_search_cache_tree(struct ocfs2_caching_info *ci,
227 struct rb_node * n = ci->ci_cache.ci_tree.rb_node;
228 struct ocfs2_meta_cache_item *item = NULL;
231 item = rb_entry(n, struct ocfs2_meta_cache_item, c_node);
233 if (block < item->c_block)
235 else if (block > item->c_block)
244 static int ocfs2_buffer_cached(struct ocfs2_caching_info *ci,
245 struct buffer_head *bh)
248 struct ocfs2_meta_cache_item *item = NULL;
250 ocfs2_metadata_cache_lock(ci);
252 mlog(0, "Owner %llu, query block %llu (inline = %u)\n",
253 (unsigned long long)ocfs2_metadata_cache_owner(ci),
254 (unsigned long long) bh->b_blocknr,
255 !!(ci->ci_flags & OCFS2_CACHE_FL_INLINE));
257 if (ci->ci_flags & OCFS2_CACHE_FL_INLINE)
258 index = ocfs2_search_cache_array(ci, bh->b_blocknr);
260 item = ocfs2_search_cache_tree(ci, bh->b_blocknr);
262 ocfs2_metadata_cache_unlock(ci);
264 mlog(0, "index = %d, item = %p\n", index, item);
266 return (index != -1) || (item != NULL);
269 /* Warning: even if it returns true, this does *not* guarantee that
270 * the block is stored in our inode metadata cache.
272 * This can be called under lock_buffer()
274 int ocfs2_buffer_uptodate(struct ocfs2_caching_info *ci,
275 struct buffer_head *bh)
277 /* Doesn't matter if the bh is in our cache or not -- if it's
278 * not marked uptodate then we know it can't have correct
280 if (!buffer_uptodate(bh))
283 /* OCFS2 does not allow multiple nodes to be changing the same
284 * block at the same time. */
288 /* Ok, locally the buffer is marked as up to date, now search
289 * our cache to see if we can trust that. */
290 return ocfs2_buffer_cached(ci, bh);
294 * Determine whether a buffer is currently out on a read-ahead request.
295 * ci_io_sem should be held to serialize submitters with the logic here.
297 int ocfs2_buffer_read_ahead(struct ocfs2_caching_info *ci,
298 struct buffer_head *bh)
300 return buffer_locked(bh) && ocfs2_buffer_cached(ci, bh);
303 /* Requires ip_lock */
304 static void ocfs2_append_cache_array(struct ocfs2_caching_info *ci,
307 BUG_ON(ci->ci_num_cached >= OCFS2_CACHE_INFO_MAX_ARRAY);
309 mlog(0, "block %llu takes position %u\n", (unsigned long long) block,
312 ci->ci_cache.ci_array[ci->ci_num_cached] = block;
316 /* By now the caller should have checked that the item does *not*
318 * Requires ip_lock. */
319 static void __ocfs2_insert_cache_tree(struct ocfs2_caching_info *ci,
320 struct ocfs2_meta_cache_item *new)
322 sector_t block = new->c_block;
323 struct rb_node *parent = NULL;
324 struct rb_node **p = &ci->ci_cache.ci_tree.rb_node;
325 struct ocfs2_meta_cache_item *tmp;
327 mlog(0, "Insert block %llu num = %u\n", (unsigned long long) block,
333 tmp = rb_entry(parent, struct ocfs2_meta_cache_item, c_node);
335 if (block < tmp->c_block)
337 else if (block > tmp->c_block)
340 /* This should never happen! */
341 mlog(ML_ERROR, "Duplicate block %llu cached!\n",
342 (unsigned long long) block);
347 rb_link_node(&new->c_node, parent, p);
348 rb_insert_color(&new->c_node, &ci->ci_cache.ci_tree);
352 /* co_cache_lock() must be held */
353 static inline int ocfs2_insert_can_use_array(struct ocfs2_caching_info *ci)
355 return (ci->ci_flags & OCFS2_CACHE_FL_INLINE) &&
356 (ci->ci_num_cached < OCFS2_CACHE_INFO_MAX_ARRAY);
359 /* tree should be exactly OCFS2_CACHE_INFO_MAX_ARRAY wide. NULL the
360 * pointers in tree after we use them - this allows caller to detect
361 * when to free in case of error.
363 * The co_cache_lock() must be held. */
364 static void ocfs2_expand_cache(struct ocfs2_caching_info *ci,
365 struct ocfs2_meta_cache_item **tree)
369 mlog_bug_on_msg(ci->ci_num_cached != OCFS2_CACHE_INFO_MAX_ARRAY,
370 "Owner %llu, num cached = %u, should be %u\n",
371 (unsigned long long)ocfs2_metadata_cache_owner(ci),
372 ci->ci_num_cached, OCFS2_CACHE_INFO_MAX_ARRAY);
373 mlog_bug_on_msg(!(ci->ci_flags & OCFS2_CACHE_FL_INLINE),
374 "Owner %llu not marked as inline anymore!\n",
375 (unsigned long long)ocfs2_metadata_cache_owner(ci));
377 /* Be careful to initialize the tree members *first* because
378 * once the ci_tree is used, the array is junk... */
379 for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++)
380 tree[i]->c_block = ci->ci_cache.ci_array[i];
382 ci->ci_flags &= ~OCFS2_CACHE_FL_INLINE;
383 ci->ci_cache.ci_tree = RB_ROOT;
384 /* this will be set again by __ocfs2_insert_cache_tree */
385 ci->ci_num_cached = 0;
387 for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) {
388 __ocfs2_insert_cache_tree(ci, tree[i]);
392 mlog(0, "Expanded %llu to a tree cache: flags 0x%x, num = %u\n",
393 (unsigned long long)ocfs2_metadata_cache_owner(ci),
394 ci->ci_flags, ci->ci_num_cached);
397 /* Slow path function - memory allocation is necessary. See the
398 * comment above ocfs2_set_buffer_uptodate for more information. */
399 static void __ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
404 struct ocfs2_meta_cache_item *new = NULL;
405 struct ocfs2_meta_cache_item *tree[OCFS2_CACHE_INFO_MAX_ARRAY] =
408 mlog(0, "Owner %llu, block %llu, expand = %d\n",
409 (unsigned long long)ocfs2_metadata_cache_owner(ci),
410 (unsigned long long)block, expand_tree);
412 new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_NOFS);
417 new->c_block = block;
420 /* Do *not* allocate an array here - the removal code
421 * has no way of tracking that. */
422 for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) {
423 tree[i] = kmem_cache_alloc(ocfs2_uptodate_cachep,
430 /* These are initialized in ocfs2_expand_cache! */
434 ocfs2_metadata_cache_lock(ci);
435 if (ocfs2_insert_can_use_array(ci)) {
436 mlog(0, "Someone cleared the tree underneath us\n");
437 /* Ok, items were removed from the cache in between
438 * locks. Detect this and revert back to the fast path */
439 ocfs2_append_cache_array(ci, block);
440 ocfs2_metadata_cache_unlock(ci);
445 ocfs2_expand_cache(ci, tree);
447 __ocfs2_insert_cache_tree(ci, new);
448 ocfs2_metadata_cache_unlock(ci);
453 kmem_cache_free(ocfs2_uptodate_cachep, new);
455 /* If these were used, then ocfs2_expand_cache re-set them to
458 for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++)
460 kmem_cache_free(ocfs2_uptodate_cachep,
465 /* Item insertion is guarded by co_io_lock(), so the insertion path takes
466 * advantage of this by not rechecking for a duplicate insert during
467 * the slow case. Additionally, if the cache needs to be bumped up to
468 * a tree, the code will not recheck after acquiring the lock --
469 * multiple paths cannot be expanding to a tree at the same time.
471 * The slow path takes into account that items can be removed
472 * (including the whole tree wiped and reset) when this process it out
473 * allocating memory. In those cases, it reverts back to the fast
476 * Note that this function may actually fail to insert the block if
477 * memory cannot be allocated. This is not fatal however (but may
478 * result in a performance penalty)
480 * Readahead buffers can be passed in here before the I/O request is
483 void ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
484 struct buffer_head *bh)
488 /* The block may very well exist in our cache already, so avoid
489 * doing any more work in that case. */
490 if (ocfs2_buffer_cached(ci, bh))
493 mlog(0, "Owner %llu, inserting block %llu\n",
494 (unsigned long long)ocfs2_metadata_cache_owner(ci),
495 (unsigned long long)bh->b_blocknr);
497 /* No need to recheck under spinlock - insertion is guarded by
499 ocfs2_metadata_cache_lock(ci);
500 if (ocfs2_insert_can_use_array(ci)) {
501 /* Fast case - it's an array and there's a free
503 ocfs2_append_cache_array(ci, bh->b_blocknr);
504 ocfs2_metadata_cache_unlock(ci);
509 if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
510 /* We need to bump things up to a tree. */
513 ocfs2_metadata_cache_unlock(ci);
515 __ocfs2_set_buffer_uptodate(ci, bh->b_blocknr, expand);
518 /* Called against a newly allocated buffer. Most likely nobody should
519 * be able to read this sort of metadata while it's still being
520 * allocated, but this is careful to take co_io_lock() anyway. */
521 void ocfs2_set_new_buffer_uptodate(struct ocfs2_caching_info *ci,
522 struct buffer_head *bh)
524 /* This should definitely *not* exist in our cache */
525 BUG_ON(ocfs2_buffer_cached(ci, bh));
527 set_buffer_uptodate(bh);
529 ocfs2_metadata_cache_io_lock(ci);
530 ocfs2_set_buffer_uptodate(ci, bh);
531 ocfs2_metadata_cache_io_unlock(ci);
534 /* Requires ip_lock. */
535 static void ocfs2_remove_metadata_array(struct ocfs2_caching_info *ci,
538 sector_t *array = ci->ci_cache.ci_array;
541 BUG_ON(index < 0 || index >= OCFS2_CACHE_INFO_MAX_ARRAY);
542 BUG_ON(index >= ci->ci_num_cached);
543 BUG_ON(!ci->ci_num_cached);
545 mlog(0, "remove index %d (num_cached = %u\n", index,
550 /* don't need to copy if the array is now empty, or if we
551 * removed at the tail */
552 if (ci->ci_num_cached && index < ci->ci_num_cached) {
553 bytes = sizeof(sector_t) * (ci->ci_num_cached - index);
554 memmove(&array[index], &array[index + 1], bytes);
558 /* Requires ip_lock. */
559 static void ocfs2_remove_metadata_tree(struct ocfs2_caching_info *ci,
560 struct ocfs2_meta_cache_item *item)
562 mlog(0, "remove block %llu from tree\n",
563 (unsigned long long) item->c_block);
565 rb_erase(&item->c_node, &ci->ci_cache.ci_tree);
569 static void ocfs2_remove_block_from_cache(struct ocfs2_caching_info *ci,
573 struct ocfs2_meta_cache_item *item = NULL;
575 ocfs2_metadata_cache_lock(ci);
576 mlog(0, "Owner %llu, remove %llu, items = %u, array = %u\n",
577 (unsigned long long)ocfs2_metadata_cache_owner(ci),
578 (unsigned long long) block, ci->ci_num_cached,
579 ci->ci_flags & OCFS2_CACHE_FL_INLINE);
581 if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
582 index = ocfs2_search_cache_array(ci, block);
584 ocfs2_remove_metadata_array(ci, index);
586 item = ocfs2_search_cache_tree(ci, block);
588 ocfs2_remove_metadata_tree(ci, item);
590 ocfs2_metadata_cache_unlock(ci);
593 kmem_cache_free(ocfs2_uptodate_cachep, item);
597 * Called when we remove a chunk of metadata from an inode. We don't
598 * bother reverting things to an inlined array in the case of a remove
599 * which moves us back under the limit.
601 void ocfs2_remove_from_cache(struct ocfs2_caching_info *ci,
602 struct buffer_head *bh)
604 sector_t block = bh->b_blocknr;
606 ocfs2_remove_block_from_cache(ci, block);
609 /* Called when we remove xattr clusters from an inode. */
610 void ocfs2_remove_xattr_clusters_from_cache(struct ocfs2_caching_info *ci,
614 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
615 unsigned int i, b_len = ocfs2_clusters_to_blocks(sb, 1) * c_len;
617 for (i = 0; i < b_len; i++, block++)
618 ocfs2_remove_block_from_cache(ci, block);
621 int __init init_ocfs2_uptodate_cache(void)
623 ocfs2_uptodate_cachep = kmem_cache_create("ocfs2_uptodate",
624 sizeof(struct ocfs2_meta_cache_item),
625 0, SLAB_HWCACHE_ALIGN, NULL);
626 if (!ocfs2_uptodate_cachep)
629 mlog(0, "%u inlined cache items per inode.\n",
630 OCFS2_CACHE_INFO_MAX_ARRAY);
635 void exit_ocfs2_uptodate_cache(void)
637 if (ocfs2_uptodate_cachep)
638 kmem_cache_destroy(ocfs2_uptodate_cachep);