sparc64: Fix bootup with mcount in some configs.
[linux-2.6/mini2440.git] / fs / ocfs2 / uptodate.c
blob187b99ff03688992fc0f141390ca6bec667fa027
1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
4 * uptodate.c
6 * Tracking the up-to-date-ness of a local buffer_head with respect to
7 * the cluster.
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
32 * to cache them.
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
43 * item for them)
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
50 #include <linux/fs.h>
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>
58 #else
59 # include <linux/jbd.h>
60 #endif
62 #define MLOG_MASK_PREFIX ML_UPTODATE
64 #include <cluster/masklog.h>
66 #include "ocfs2.h"
68 #include "inode.h"
69 #include "uptodate.h"
71 struct ocfs2_meta_cache_item {
72 struct rb_node c_node;
73 sector_t c_block;
76 static struct kmem_cache *ocfs2_uptodate_cachep = NULL;
78 void ocfs2_metadata_cache_init(struct inode *inode)
80 struct ocfs2_inode_info *oi = OCFS2_I(inode);
81 struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;
83 oi->ip_flags |= OCFS2_INODE_CACHE_INLINE;
84 ci->ci_num_cached = 0;
87 /* No lock taken here as 'root' is not expected to be visible to other
88 * processes. */
89 static unsigned int ocfs2_purge_copied_metadata_tree(struct rb_root *root)
91 unsigned int purged = 0;
92 struct rb_node *node;
93 struct ocfs2_meta_cache_item *item;
95 while ((node = rb_last(root)) != NULL) {
96 item = rb_entry(node, struct ocfs2_meta_cache_item, c_node);
98 mlog(0, "Purge item %llu\n",
99 (unsigned long long) item->c_block);
101 rb_erase(&item->c_node, root);
102 kmem_cache_free(ocfs2_uptodate_cachep, item);
104 purged++;
106 return purged;
109 /* Called from locking and called from ocfs2_clear_inode. Dump the
110 * cache for a given inode.
112 * This function is a few more lines longer than necessary due to some
113 * accounting done here, but I think it's worth tracking down those
114 * bugs sooner -- Mark */
115 void ocfs2_metadata_cache_purge(struct inode *inode)
117 struct ocfs2_inode_info *oi = OCFS2_I(inode);
118 unsigned int tree, to_purge, purged;
119 struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;
120 struct rb_root root = RB_ROOT;
122 spin_lock(&oi->ip_lock);
123 tree = !(oi->ip_flags & OCFS2_INODE_CACHE_INLINE);
124 to_purge = ci->ci_num_cached;
126 mlog(0, "Purge %u %s items from Inode %llu\n", to_purge,
127 tree ? "array" : "tree", (unsigned long long)oi->ip_blkno);
129 /* If we're a tree, save off the root so that we can safely
130 * initialize the cache. We do the work to free tree members
131 * without the spinlock. */
132 if (tree)
133 root = ci->ci_cache.ci_tree;
135 ocfs2_metadata_cache_init(inode);
136 spin_unlock(&oi->ip_lock);
138 purged = ocfs2_purge_copied_metadata_tree(&root);
139 /* If possible, track the number wiped so that we can more
140 * easily detect counting errors. Unfortunately, this is only
141 * meaningful for trees. */
142 if (tree && purged != to_purge)
143 mlog(ML_ERROR, "Inode %llu, count = %u, purged = %u\n",
144 (unsigned long long)oi->ip_blkno, to_purge, purged);
147 /* Returns the index in the cache array, -1 if not found.
148 * Requires ip_lock. */
149 static int ocfs2_search_cache_array(struct ocfs2_caching_info *ci,
150 sector_t item)
152 int i;
154 for (i = 0; i < ci->ci_num_cached; i++) {
155 if (item == ci->ci_cache.ci_array[i])
156 return i;
159 return -1;
162 /* Returns the cache item if found, otherwise NULL.
163 * Requires ip_lock. */
164 static struct ocfs2_meta_cache_item *
165 ocfs2_search_cache_tree(struct ocfs2_caching_info *ci,
166 sector_t block)
168 struct rb_node * n = ci->ci_cache.ci_tree.rb_node;
169 struct ocfs2_meta_cache_item *item = NULL;
171 while (n) {
172 item = rb_entry(n, struct ocfs2_meta_cache_item, c_node);
174 if (block < item->c_block)
175 n = n->rb_left;
176 else if (block > item->c_block)
177 n = n->rb_right;
178 else
179 return item;
182 return NULL;
185 static int ocfs2_buffer_cached(struct ocfs2_inode_info *oi,
186 struct buffer_head *bh)
188 int index = -1;
189 struct ocfs2_meta_cache_item *item = NULL;
191 spin_lock(&oi->ip_lock);
193 mlog(0, "Inode %llu, query block %llu (inline = %u)\n",
194 (unsigned long long)oi->ip_blkno,
195 (unsigned long long) bh->b_blocknr,
196 !!(oi->ip_flags & OCFS2_INODE_CACHE_INLINE));
198 if (oi->ip_flags & OCFS2_INODE_CACHE_INLINE)
199 index = ocfs2_search_cache_array(&oi->ip_metadata_cache,
200 bh->b_blocknr);
201 else
202 item = ocfs2_search_cache_tree(&oi->ip_metadata_cache,
203 bh->b_blocknr);
205 spin_unlock(&oi->ip_lock);
207 mlog(0, "index = %d, item = %p\n", index, item);
209 return (index != -1) || (item != NULL);
212 /* Warning: even if it returns true, this does *not* guarantee that
213 * the block is stored in our inode metadata cache.
215 * This can be called under lock_buffer()
217 int ocfs2_buffer_uptodate(struct inode *inode,
218 struct buffer_head *bh)
220 /* Doesn't matter if the bh is in our cache or not -- if it's
221 * not marked uptodate then we know it can't have correct
222 * data. */
223 if (!buffer_uptodate(bh))
224 return 0;
226 /* OCFS2 does not allow multiple nodes to be changing the same
227 * block at the same time. */
228 if (buffer_jbd(bh))
229 return 1;
231 /* Ok, locally the buffer is marked as up to date, now search
232 * our cache to see if we can trust that. */
233 return ocfs2_buffer_cached(OCFS2_I(inode), bh);
237 * Determine whether a buffer is currently out on a read-ahead request.
238 * ip_io_sem should be held to serialize submitters with the logic here.
240 int ocfs2_buffer_read_ahead(struct inode *inode,
241 struct buffer_head *bh)
243 return buffer_locked(bh) && ocfs2_buffer_cached(OCFS2_I(inode), bh);
246 /* Requires ip_lock */
247 static void ocfs2_append_cache_array(struct ocfs2_caching_info *ci,
248 sector_t block)
250 BUG_ON(ci->ci_num_cached >= OCFS2_INODE_MAX_CACHE_ARRAY);
252 mlog(0, "block %llu takes position %u\n", (unsigned long long) block,
253 ci->ci_num_cached);
255 ci->ci_cache.ci_array[ci->ci_num_cached] = block;
256 ci->ci_num_cached++;
259 /* By now the caller should have checked that the item does *not*
260 * exist in the tree.
261 * Requires ip_lock. */
262 static void __ocfs2_insert_cache_tree(struct ocfs2_caching_info *ci,
263 struct ocfs2_meta_cache_item *new)
265 sector_t block = new->c_block;
266 struct rb_node *parent = NULL;
267 struct rb_node **p = &ci->ci_cache.ci_tree.rb_node;
268 struct ocfs2_meta_cache_item *tmp;
270 mlog(0, "Insert block %llu num = %u\n", (unsigned long long) block,
271 ci->ci_num_cached);
273 while(*p) {
274 parent = *p;
276 tmp = rb_entry(parent, struct ocfs2_meta_cache_item, c_node);
278 if (block < tmp->c_block)
279 p = &(*p)->rb_left;
280 else if (block > tmp->c_block)
281 p = &(*p)->rb_right;
282 else {
283 /* This should never happen! */
284 mlog(ML_ERROR, "Duplicate block %llu cached!\n",
285 (unsigned long long) block);
286 BUG();
290 rb_link_node(&new->c_node, parent, p);
291 rb_insert_color(&new->c_node, &ci->ci_cache.ci_tree);
292 ci->ci_num_cached++;
295 static inline int ocfs2_insert_can_use_array(struct ocfs2_inode_info *oi,
296 struct ocfs2_caching_info *ci)
298 assert_spin_locked(&oi->ip_lock);
300 return (oi->ip_flags & OCFS2_INODE_CACHE_INLINE) &&
301 (ci->ci_num_cached < OCFS2_INODE_MAX_CACHE_ARRAY);
304 /* tree should be exactly OCFS2_INODE_MAX_CACHE_ARRAY wide. NULL the
305 * pointers in tree after we use them - this allows caller to detect
306 * when to free in case of error. */
307 static void ocfs2_expand_cache(struct ocfs2_inode_info *oi,
308 struct ocfs2_meta_cache_item **tree)
310 int i;
311 struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;
313 mlog_bug_on_msg(ci->ci_num_cached != OCFS2_INODE_MAX_CACHE_ARRAY,
314 "Inode %llu, num cached = %u, should be %u\n",
315 (unsigned long long)oi->ip_blkno, ci->ci_num_cached,
316 OCFS2_INODE_MAX_CACHE_ARRAY);
317 mlog_bug_on_msg(!(oi->ip_flags & OCFS2_INODE_CACHE_INLINE),
318 "Inode %llu not marked as inline anymore!\n",
319 (unsigned long long)oi->ip_blkno);
320 assert_spin_locked(&oi->ip_lock);
322 /* Be careful to initialize the tree members *first* because
323 * once the ci_tree is used, the array is junk... */
324 for(i = 0; i < OCFS2_INODE_MAX_CACHE_ARRAY; i++)
325 tree[i]->c_block = ci->ci_cache.ci_array[i];
327 oi->ip_flags &= ~OCFS2_INODE_CACHE_INLINE;
328 ci->ci_cache.ci_tree = RB_ROOT;
329 /* this will be set again by __ocfs2_insert_cache_tree */
330 ci->ci_num_cached = 0;
332 for(i = 0; i < OCFS2_INODE_MAX_CACHE_ARRAY; i++) {
333 __ocfs2_insert_cache_tree(ci, tree[i]);
334 tree[i] = NULL;
337 mlog(0, "Expanded %llu to a tree cache: flags 0x%x, num = %u\n",
338 (unsigned long long)oi->ip_blkno, oi->ip_flags, ci->ci_num_cached);
341 /* Slow path function - memory allocation is necessary. See the
342 * comment above ocfs2_set_buffer_uptodate for more information. */
343 static void __ocfs2_set_buffer_uptodate(struct ocfs2_inode_info *oi,
344 sector_t block,
345 int expand_tree)
347 int i;
348 struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;
349 struct ocfs2_meta_cache_item *new = NULL;
350 struct ocfs2_meta_cache_item *tree[OCFS2_INODE_MAX_CACHE_ARRAY] =
351 { NULL, };
353 mlog(0, "Inode %llu, block %llu, expand = %d\n",
354 (unsigned long long)oi->ip_blkno,
355 (unsigned long long)block, expand_tree);
357 new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_NOFS);
358 if (!new) {
359 mlog_errno(-ENOMEM);
360 return;
362 new->c_block = block;
364 if (expand_tree) {
365 /* Do *not* allocate an array here - the removal code
366 * has no way of tracking that. */
367 for(i = 0; i < OCFS2_INODE_MAX_CACHE_ARRAY; i++) {
368 tree[i] = kmem_cache_alloc(ocfs2_uptodate_cachep,
369 GFP_NOFS);
370 if (!tree[i]) {
371 mlog_errno(-ENOMEM);
372 goto out_free;
375 /* These are initialized in ocfs2_expand_cache! */
379 spin_lock(&oi->ip_lock);
380 if (ocfs2_insert_can_use_array(oi, ci)) {
381 mlog(0, "Someone cleared the tree underneath us\n");
382 /* Ok, items were removed from the cache in between
383 * locks. Detect this and revert back to the fast path */
384 ocfs2_append_cache_array(ci, block);
385 spin_unlock(&oi->ip_lock);
386 goto out_free;
389 if (expand_tree)
390 ocfs2_expand_cache(oi, tree);
392 __ocfs2_insert_cache_tree(ci, new);
393 spin_unlock(&oi->ip_lock);
395 new = NULL;
396 out_free:
397 if (new)
398 kmem_cache_free(ocfs2_uptodate_cachep, new);
400 /* If these were used, then ocfs2_expand_cache re-set them to
401 * NULL for us. */
402 if (tree[0]) {
403 for(i = 0; i < OCFS2_INODE_MAX_CACHE_ARRAY; i++)
404 if (tree[i])
405 kmem_cache_free(ocfs2_uptodate_cachep,
406 tree[i]);
410 /* Item insertion is guarded by ip_io_mutex, so the insertion path takes
411 * advantage of this by not rechecking for a duplicate insert during
412 * the slow case. Additionally, if the cache needs to be bumped up to
413 * a tree, the code will not recheck after acquiring the lock --
414 * multiple paths cannot be expanding to a tree at the same time.
416 * The slow path takes into account that items can be removed
417 * (including the whole tree wiped and reset) when this process it out
418 * allocating memory. In those cases, it reverts back to the fast
419 * path.
421 * Note that this function may actually fail to insert the block if
422 * memory cannot be allocated. This is not fatal however (but may
423 * result in a performance penalty)
425 * Readahead buffers can be passed in here before the I/O request is
426 * completed.
428 void ocfs2_set_buffer_uptodate(struct inode *inode,
429 struct buffer_head *bh)
431 int expand;
432 struct ocfs2_inode_info *oi = OCFS2_I(inode);
433 struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;
435 /* The block may very well exist in our cache already, so avoid
436 * doing any more work in that case. */
437 if (ocfs2_buffer_cached(oi, bh))
438 return;
440 mlog(0, "Inode %llu, inserting block %llu\n",
441 (unsigned long long)oi->ip_blkno,
442 (unsigned long long)bh->b_blocknr);
444 /* No need to recheck under spinlock - insertion is guarded by
445 * ip_io_mutex */
446 spin_lock(&oi->ip_lock);
447 if (ocfs2_insert_can_use_array(oi, ci)) {
448 /* Fast case - it's an array and there's a free
449 * spot. */
450 ocfs2_append_cache_array(ci, bh->b_blocknr);
451 spin_unlock(&oi->ip_lock);
452 return;
455 expand = 0;
456 if (oi->ip_flags & OCFS2_INODE_CACHE_INLINE) {
457 /* We need to bump things up to a tree. */
458 expand = 1;
460 spin_unlock(&oi->ip_lock);
462 __ocfs2_set_buffer_uptodate(oi, bh->b_blocknr, expand);
465 /* Called against a newly allocated buffer. Most likely nobody should
466 * be able to read this sort of metadata while it's still being
467 * allocated, but this is careful to take ip_io_mutex anyway. */
468 void ocfs2_set_new_buffer_uptodate(struct inode *inode,
469 struct buffer_head *bh)
471 struct ocfs2_inode_info *oi = OCFS2_I(inode);
473 /* This should definitely *not* exist in our cache */
474 BUG_ON(ocfs2_buffer_cached(oi, bh));
476 set_buffer_uptodate(bh);
478 mutex_lock(&oi->ip_io_mutex);
479 ocfs2_set_buffer_uptodate(inode, bh);
480 mutex_unlock(&oi->ip_io_mutex);
483 /* Requires ip_lock. */
484 static void ocfs2_remove_metadata_array(struct ocfs2_caching_info *ci,
485 int index)
487 sector_t *array = ci->ci_cache.ci_array;
488 int bytes;
490 BUG_ON(index < 0 || index >= OCFS2_INODE_MAX_CACHE_ARRAY);
491 BUG_ON(index >= ci->ci_num_cached);
492 BUG_ON(!ci->ci_num_cached);
494 mlog(0, "remove index %d (num_cached = %u\n", index,
495 ci->ci_num_cached);
497 ci->ci_num_cached--;
499 /* don't need to copy if the array is now empty, or if we
500 * removed at the tail */
501 if (ci->ci_num_cached && index < ci->ci_num_cached) {
502 bytes = sizeof(sector_t) * (ci->ci_num_cached - index);
503 memmove(&array[index], &array[index + 1], bytes);
507 /* Requires ip_lock. */
508 static void ocfs2_remove_metadata_tree(struct ocfs2_caching_info *ci,
509 struct ocfs2_meta_cache_item *item)
511 mlog(0, "remove block %llu from tree\n",
512 (unsigned long long) item->c_block);
514 rb_erase(&item->c_node, &ci->ci_cache.ci_tree);
515 ci->ci_num_cached--;
518 static void ocfs2_remove_block_from_cache(struct inode *inode,
519 sector_t block)
521 int index;
522 struct ocfs2_meta_cache_item *item = NULL;
523 struct ocfs2_inode_info *oi = OCFS2_I(inode);
524 struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;
526 spin_lock(&oi->ip_lock);
527 mlog(0, "Inode %llu, remove %llu, items = %u, array = %u\n",
528 (unsigned long long)oi->ip_blkno,
529 (unsigned long long) block, ci->ci_num_cached,
530 oi->ip_flags & OCFS2_INODE_CACHE_INLINE);
532 if (oi->ip_flags & OCFS2_INODE_CACHE_INLINE) {
533 index = ocfs2_search_cache_array(ci, block);
534 if (index != -1)
535 ocfs2_remove_metadata_array(ci, index);
536 } else {
537 item = ocfs2_search_cache_tree(ci, block);
538 if (item)
539 ocfs2_remove_metadata_tree(ci, item);
541 spin_unlock(&oi->ip_lock);
543 if (item)
544 kmem_cache_free(ocfs2_uptodate_cachep, item);
548 * Called when we remove a chunk of metadata from an inode. We don't
549 * bother reverting things to an inlined array in the case of a remove
550 * which moves us back under the limit.
552 void ocfs2_remove_from_cache(struct inode *inode,
553 struct buffer_head *bh)
555 sector_t block = bh->b_blocknr;
557 ocfs2_remove_block_from_cache(inode, block);
560 /* Called when we remove xattr clusters from an inode. */
561 void ocfs2_remove_xattr_clusters_from_cache(struct inode *inode,
562 sector_t block,
563 u32 c_len)
565 unsigned int i, b_len = ocfs2_clusters_to_blocks(inode->i_sb, 1) * c_len;
567 for (i = 0; i < b_len; i++, block++)
568 ocfs2_remove_block_from_cache(inode, block);
571 int __init init_ocfs2_uptodate_cache(void)
573 ocfs2_uptodate_cachep = kmem_cache_create("ocfs2_uptodate",
574 sizeof(struct ocfs2_meta_cache_item),
575 0, SLAB_HWCACHE_ALIGN, NULL);
576 if (!ocfs2_uptodate_cachep)
577 return -ENOMEM;
579 mlog(0, "%u inlined cache items per inode.\n",
580 OCFS2_INODE_MAX_CACHE_ARRAY);
582 return 0;
585 void exit_ocfs2_uptodate_cache(void)
587 if (ocfs2_uptodate_cachep)
588 kmem_cache_destroy(ocfs2_uptodate_cachep);