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>
56 #ifndef CONFIG_OCFS2_COMPAT_JBD
57 # include <linux/jbd2.h>
59 # include <linux/jbd.h>
62 #define MLOG_MASK_PREFIX ML_UPTODATE
64 #include <cluster/masklog.h>
71 struct ocfs2_meta_cache_item {
72 struct rb_node c_node;
76 static struct kmem_cache *ocfs2_uptodate_cachep = NULL;
78 u64 ocfs2_metadata_cache_owner(struct ocfs2_caching_info *ci)
80 BUG_ON(!ci || !ci->ci_ops);
82 return ci->ci_ops->co_owner(ci);
85 struct super_block *ocfs2_metadata_cache_get_super(struct ocfs2_caching_info *ci)
87 BUG_ON(!ci || !ci->ci_ops);
89 return ci->ci_ops->co_get_super(ci);
92 static void ocfs2_metadata_cache_lock(struct ocfs2_caching_info *ci)
94 BUG_ON(!ci || !ci->ci_ops);
96 ci->ci_ops->co_cache_lock(ci);
99 static void ocfs2_metadata_cache_unlock(struct ocfs2_caching_info *ci)
101 BUG_ON(!ci || !ci->ci_ops);
103 ci->ci_ops->co_cache_unlock(ci);
106 void ocfs2_metadata_cache_io_lock(struct ocfs2_caching_info *ci)
108 BUG_ON(!ci || !ci->ci_ops);
110 ci->ci_ops->co_io_lock(ci);
113 void ocfs2_metadata_cache_io_unlock(struct ocfs2_caching_info *ci)
115 BUG_ON(!ci || !ci->ci_ops);
117 ci->ci_ops->co_io_unlock(ci);
121 static void ocfs2_metadata_cache_reset(struct ocfs2_caching_info *ci,
124 ci->ci_flags |= OCFS2_CACHE_FL_INLINE;
125 ci->ci_num_cached = 0;
128 ci->ci_last_trans = 0;
131 void ocfs2_metadata_cache_init(struct ocfs2_caching_info *ci,
132 const struct ocfs2_caching_operations *ops)
137 ocfs2_metadata_cache_reset(ci, 1);
140 void ocfs2_metadata_cache_exit(struct ocfs2_caching_info *ci)
142 ocfs2_metadata_cache_purge(ci);
143 ocfs2_metadata_cache_reset(ci, 1);
147 /* No lock taken here as 'root' is not expected to be visible to other
149 static unsigned int ocfs2_purge_copied_metadata_tree(struct rb_root *root)
151 unsigned int purged = 0;
152 struct rb_node *node;
153 struct ocfs2_meta_cache_item *item;
155 while ((node = rb_last(root)) != NULL) {
156 item = rb_entry(node, struct ocfs2_meta_cache_item, c_node);
158 mlog(0, "Purge item %llu\n",
159 (unsigned long long) item->c_block);
161 rb_erase(&item->c_node, root);
162 kmem_cache_free(ocfs2_uptodate_cachep, item);
169 /* Called from locking and called from ocfs2_clear_inode. Dump the
170 * cache for a given inode.
172 * This function is a few more lines longer than necessary due to some
173 * accounting done here, but I think it's worth tracking down those
174 * bugs sooner -- Mark */
175 void ocfs2_metadata_cache_purge(struct ocfs2_caching_info *ci)
177 unsigned int tree, to_purge, purged;
178 struct rb_root root = RB_ROOT;
180 BUG_ON(!ci || !ci->ci_ops);
182 ocfs2_metadata_cache_lock(ci);
183 tree = !(ci->ci_flags & OCFS2_CACHE_FL_INLINE);
184 to_purge = ci->ci_num_cached;
186 mlog(0, "Purge %u %s items from Owner %llu\n", to_purge,
187 tree ? "array" : "tree",
188 (unsigned long long)ocfs2_metadata_cache_owner(ci));
190 /* If we're a tree, save off the root so that we can safely
191 * initialize the cache. We do the work to free tree members
192 * without the spinlock. */
194 root = ci->ci_cache.ci_tree;
196 ocfs2_metadata_cache_reset(ci, 0);
197 ocfs2_metadata_cache_unlock(ci);
199 purged = ocfs2_purge_copied_metadata_tree(&root);
200 /* If possible, track the number wiped so that we can more
201 * easily detect counting errors. Unfortunately, this is only
202 * meaningful for trees. */
203 if (tree && purged != to_purge)
204 mlog(ML_ERROR, "Owner %llu, count = %u, purged = %u\n",
205 (unsigned long long)ocfs2_metadata_cache_owner(ci),
209 /* Returns the index in the cache array, -1 if not found.
210 * Requires ip_lock. */
211 static int ocfs2_search_cache_array(struct ocfs2_caching_info *ci,
216 for (i = 0; i < ci->ci_num_cached; i++) {
217 if (item == ci->ci_cache.ci_array[i])
224 /* Returns the cache item if found, otherwise NULL.
225 * Requires ip_lock. */
226 static struct ocfs2_meta_cache_item *
227 ocfs2_search_cache_tree(struct ocfs2_caching_info *ci,
230 struct rb_node * n = ci->ci_cache.ci_tree.rb_node;
231 struct ocfs2_meta_cache_item *item = NULL;
234 item = rb_entry(n, struct ocfs2_meta_cache_item, c_node);
236 if (block < item->c_block)
238 else if (block > item->c_block)
247 static int ocfs2_buffer_cached(struct ocfs2_caching_info *ci,
248 struct buffer_head *bh)
251 struct ocfs2_meta_cache_item *item = NULL;
253 ocfs2_metadata_cache_lock(ci);
255 mlog(0, "Owner %llu, query block %llu (inline = %u)\n",
256 (unsigned long long)ocfs2_metadata_cache_owner(ci),
257 (unsigned long long) bh->b_blocknr,
258 !!(ci->ci_flags & OCFS2_CACHE_FL_INLINE));
260 if (ci->ci_flags & OCFS2_CACHE_FL_INLINE)
261 index = ocfs2_search_cache_array(ci, bh->b_blocknr);
263 item = ocfs2_search_cache_tree(ci, bh->b_blocknr);
265 ocfs2_metadata_cache_unlock(ci);
267 mlog(0, "index = %d, item = %p\n", index, item);
269 return (index != -1) || (item != NULL);
272 /* Warning: even if it returns true, this does *not* guarantee that
273 * the block is stored in our inode metadata cache.
275 * This can be called under lock_buffer()
277 int ocfs2_buffer_uptodate(struct ocfs2_caching_info *ci,
278 struct buffer_head *bh)
280 /* Doesn't matter if the bh is in our cache or not -- if it's
281 * not marked uptodate then we know it can't have correct
283 if (!buffer_uptodate(bh))
286 /* OCFS2 does not allow multiple nodes to be changing the same
287 * block at the same time. */
291 /* Ok, locally the buffer is marked as up to date, now search
292 * our cache to see if we can trust that. */
293 return ocfs2_buffer_cached(ci, bh);
297 * Determine whether a buffer is currently out on a read-ahead request.
298 * ci_io_sem should be held to serialize submitters with the logic here.
300 int ocfs2_buffer_read_ahead(struct ocfs2_caching_info *ci,
301 struct buffer_head *bh)
303 return buffer_locked(bh) && ocfs2_buffer_cached(ci, bh);
306 /* Requires ip_lock */
307 static void ocfs2_append_cache_array(struct ocfs2_caching_info *ci,
310 BUG_ON(ci->ci_num_cached >= OCFS2_CACHE_INFO_MAX_ARRAY);
312 mlog(0, "block %llu takes position %u\n", (unsigned long long) block,
315 ci->ci_cache.ci_array[ci->ci_num_cached] = block;
319 /* By now the caller should have checked that the item does *not*
321 * Requires ip_lock. */
322 static void __ocfs2_insert_cache_tree(struct ocfs2_caching_info *ci,
323 struct ocfs2_meta_cache_item *new)
325 sector_t block = new->c_block;
326 struct rb_node *parent = NULL;
327 struct rb_node **p = &ci->ci_cache.ci_tree.rb_node;
328 struct ocfs2_meta_cache_item *tmp;
330 mlog(0, "Insert block %llu num = %u\n", (unsigned long long) block,
336 tmp = rb_entry(parent, struct ocfs2_meta_cache_item, c_node);
338 if (block < tmp->c_block)
340 else if (block > tmp->c_block)
343 /* This should never happen! */
344 mlog(ML_ERROR, "Duplicate block %llu cached!\n",
345 (unsigned long long) block);
350 rb_link_node(&new->c_node, parent, p);
351 rb_insert_color(&new->c_node, &ci->ci_cache.ci_tree);
355 /* co_cache_lock() must be held */
356 static inline int ocfs2_insert_can_use_array(struct ocfs2_caching_info *ci)
358 return (ci->ci_flags & OCFS2_CACHE_FL_INLINE) &&
359 (ci->ci_num_cached < OCFS2_CACHE_INFO_MAX_ARRAY);
362 /* tree should be exactly OCFS2_CACHE_INFO_MAX_ARRAY wide. NULL the
363 * pointers in tree after we use them - this allows caller to detect
364 * when to free in case of error.
366 * The co_cache_lock() must be held. */
367 static void ocfs2_expand_cache(struct ocfs2_caching_info *ci,
368 struct ocfs2_meta_cache_item **tree)
372 mlog_bug_on_msg(ci->ci_num_cached != OCFS2_CACHE_INFO_MAX_ARRAY,
373 "Owner %llu, num cached = %u, should be %u\n",
374 (unsigned long long)ocfs2_metadata_cache_owner(ci),
375 ci->ci_num_cached, OCFS2_CACHE_INFO_MAX_ARRAY);
376 mlog_bug_on_msg(!(ci->ci_flags & OCFS2_CACHE_FL_INLINE),
377 "Owner %llu not marked as inline anymore!\n",
378 (unsigned long long)ocfs2_metadata_cache_owner(ci));
380 /* Be careful to initialize the tree members *first* because
381 * once the ci_tree is used, the array is junk... */
382 for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++)
383 tree[i]->c_block = ci->ci_cache.ci_array[i];
385 ci->ci_flags &= ~OCFS2_CACHE_FL_INLINE;
386 ci->ci_cache.ci_tree = RB_ROOT;
387 /* this will be set again by __ocfs2_insert_cache_tree */
388 ci->ci_num_cached = 0;
390 for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) {
391 __ocfs2_insert_cache_tree(ci, tree[i]);
395 mlog(0, "Expanded %llu to a tree cache: flags 0x%x, num = %u\n",
396 (unsigned long long)ocfs2_metadata_cache_owner(ci),
397 ci->ci_flags, ci->ci_num_cached);
400 /* Slow path function - memory allocation is necessary. See the
401 * comment above ocfs2_set_buffer_uptodate for more information. */
402 static void __ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
407 struct ocfs2_meta_cache_item *new = NULL;
408 struct ocfs2_meta_cache_item *tree[OCFS2_CACHE_INFO_MAX_ARRAY] =
411 mlog(0, "Owner %llu, block %llu, expand = %d\n",
412 (unsigned long long)ocfs2_metadata_cache_owner(ci),
413 (unsigned long long)block, expand_tree);
415 new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_NOFS);
420 new->c_block = block;
423 /* Do *not* allocate an array here - the removal code
424 * has no way of tracking that. */
425 for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) {
426 tree[i] = kmem_cache_alloc(ocfs2_uptodate_cachep,
433 /* These are initialized in ocfs2_expand_cache! */
437 ocfs2_metadata_cache_lock(ci);
438 if (ocfs2_insert_can_use_array(ci)) {
439 mlog(0, "Someone cleared the tree underneath us\n");
440 /* Ok, items were removed from the cache in between
441 * locks. Detect this and revert back to the fast path */
442 ocfs2_append_cache_array(ci, block);
443 ocfs2_metadata_cache_unlock(ci);
448 ocfs2_expand_cache(ci, tree);
450 __ocfs2_insert_cache_tree(ci, new);
451 ocfs2_metadata_cache_unlock(ci);
456 kmem_cache_free(ocfs2_uptodate_cachep, new);
458 /* If these were used, then ocfs2_expand_cache re-set them to
461 for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++)
463 kmem_cache_free(ocfs2_uptodate_cachep,
468 /* Item insertion is guarded by co_io_lock(), so the insertion path takes
469 * advantage of this by not rechecking for a duplicate insert during
470 * the slow case. Additionally, if the cache needs to be bumped up to
471 * a tree, the code will not recheck after acquiring the lock --
472 * multiple paths cannot be expanding to a tree at the same time.
474 * The slow path takes into account that items can be removed
475 * (including the whole tree wiped and reset) when this process it out
476 * allocating memory. In those cases, it reverts back to the fast
479 * Note that this function may actually fail to insert the block if
480 * memory cannot be allocated. This is not fatal however (but may
481 * result in a performance penalty)
483 * Readahead buffers can be passed in here before the I/O request is
486 void ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
487 struct buffer_head *bh)
491 /* The block may very well exist in our cache already, so avoid
492 * doing any more work in that case. */
493 if (ocfs2_buffer_cached(ci, bh))
496 mlog(0, "Owner %llu, inserting block %llu\n",
497 (unsigned long long)ocfs2_metadata_cache_owner(ci),
498 (unsigned long long)bh->b_blocknr);
500 /* No need to recheck under spinlock - insertion is guarded by
502 ocfs2_metadata_cache_lock(ci);
503 if (ocfs2_insert_can_use_array(ci)) {
504 /* Fast case - it's an array and there's a free
506 ocfs2_append_cache_array(ci, bh->b_blocknr);
507 ocfs2_metadata_cache_unlock(ci);
512 if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
513 /* We need to bump things up to a tree. */
516 ocfs2_metadata_cache_unlock(ci);
518 __ocfs2_set_buffer_uptodate(ci, bh->b_blocknr, expand);
521 /* Called against a newly allocated buffer. Most likely nobody should
522 * be able to read this sort of metadata while it's still being
523 * allocated, but this is careful to take co_io_lock() anyway. */
524 void ocfs2_set_new_buffer_uptodate(struct ocfs2_caching_info *ci,
525 struct buffer_head *bh)
527 /* This should definitely *not* exist in our cache */
528 BUG_ON(ocfs2_buffer_cached(ci, bh));
530 set_buffer_uptodate(bh);
532 ocfs2_metadata_cache_io_lock(ci);
533 ocfs2_set_buffer_uptodate(ci, bh);
534 ocfs2_metadata_cache_io_unlock(ci);
537 /* Requires ip_lock. */
538 static void ocfs2_remove_metadata_array(struct ocfs2_caching_info *ci,
541 sector_t *array = ci->ci_cache.ci_array;
544 BUG_ON(index < 0 || index >= OCFS2_CACHE_INFO_MAX_ARRAY);
545 BUG_ON(index >= ci->ci_num_cached);
546 BUG_ON(!ci->ci_num_cached);
548 mlog(0, "remove index %d (num_cached = %u\n", index,
553 /* don't need to copy if the array is now empty, or if we
554 * removed at the tail */
555 if (ci->ci_num_cached && index < ci->ci_num_cached) {
556 bytes = sizeof(sector_t) * (ci->ci_num_cached - index);
557 memmove(&array[index], &array[index + 1], bytes);
561 /* Requires ip_lock. */
562 static void ocfs2_remove_metadata_tree(struct ocfs2_caching_info *ci,
563 struct ocfs2_meta_cache_item *item)
565 mlog(0, "remove block %llu from tree\n",
566 (unsigned long long) item->c_block);
568 rb_erase(&item->c_node, &ci->ci_cache.ci_tree);
572 static void ocfs2_remove_block_from_cache(struct ocfs2_caching_info *ci,
576 struct ocfs2_meta_cache_item *item = NULL;
578 ocfs2_metadata_cache_lock(ci);
579 mlog(0, "Owner %llu, remove %llu, items = %u, array = %u\n",
580 (unsigned long long)ocfs2_metadata_cache_owner(ci),
581 (unsigned long long) block, ci->ci_num_cached,
582 ci->ci_flags & OCFS2_CACHE_FL_INLINE);
584 if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
585 index = ocfs2_search_cache_array(ci, block);
587 ocfs2_remove_metadata_array(ci, index);
589 item = ocfs2_search_cache_tree(ci, block);
591 ocfs2_remove_metadata_tree(ci, item);
593 ocfs2_metadata_cache_unlock(ci);
596 kmem_cache_free(ocfs2_uptodate_cachep, item);
600 * Called when we remove a chunk of metadata from an inode. We don't
601 * bother reverting things to an inlined array in the case of a remove
602 * which moves us back under the limit.
604 void ocfs2_remove_from_cache(struct ocfs2_caching_info *ci,
605 struct buffer_head *bh)
607 sector_t block = bh->b_blocknr;
609 ocfs2_remove_block_from_cache(ci, block);
612 /* Called when we remove xattr clusters from an inode. */
613 void ocfs2_remove_xattr_clusters_from_cache(struct ocfs2_caching_info *ci,
617 struct super_block *sb = ocfs2_metadata_cache_get_super(ci);
618 unsigned int i, b_len = ocfs2_clusters_to_blocks(sb, 1) * c_len;
620 for (i = 0; i < b_len; i++, block++)
621 ocfs2_remove_block_from_cache(ci, block);
624 int __init init_ocfs2_uptodate_cache(void)
626 ocfs2_uptodate_cachep = kmem_cache_create("ocfs2_uptodate",
627 sizeof(struct ocfs2_meta_cache_item),
628 0, SLAB_HWCACHE_ALIGN, NULL);
629 if (!ocfs2_uptodate_cachep)
632 mlog(0, "%u inlined cache items per inode.\n",
633 OCFS2_CACHE_INFO_MAX_ARRAY);
638 void exit_ocfs2_uptodate_cache(void)
640 if (ocfs2_uptodate_cachep)
641 kmem_cache_destroy(ocfs2_uptodate_cachep);