Root/fs/ocfs2/uptodate.c

1/* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
3 *
4 * uptodate.c
5 *
6 * Tracking the up-to-date-ness of a local buffer_head with respect to
7 * the cluster.
8 *
9 * Copyright (C) 2002, 2004, 2005 Oracle. All rights reserved.
10 *
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.
15 *
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.
20 *
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.
25 *
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.
33 *
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)
44 *
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
48 */
49
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
57#define MLOG_MASK_PREFIX ML_UPTODATE
58
59#include <cluster/masklog.h>
60
61#include "ocfs2.h"
62
63#include "inode.h"
64#include "uptodate.h"
65
66struct ocfs2_meta_cache_item {
67    struct rb_node c_node;
68    sector_t c_block;
69};
70
71static struct kmem_cache *ocfs2_uptodate_cachep = NULL;
72
73u64 ocfs2_metadata_cache_owner(struct ocfs2_caching_info *ci)
74{
75    BUG_ON(!ci || !ci->ci_ops);
76
77    return ci->ci_ops->co_owner(ci);
78}
79
80struct super_block *ocfs2_metadata_cache_get_super(struct ocfs2_caching_info *ci)
81{
82    BUG_ON(!ci || !ci->ci_ops);
83
84    return ci->ci_ops->co_get_super(ci);
85}
86
87static void ocfs2_metadata_cache_lock(struct ocfs2_caching_info *ci)
88{
89    BUG_ON(!ci || !ci->ci_ops);
90
91    ci->ci_ops->co_cache_lock(ci);
92}
93
94static void ocfs2_metadata_cache_unlock(struct ocfs2_caching_info *ci)
95{
96    BUG_ON(!ci || !ci->ci_ops);
97
98    ci->ci_ops->co_cache_unlock(ci);
99}
100
101void ocfs2_metadata_cache_io_lock(struct ocfs2_caching_info *ci)
102{
103    BUG_ON(!ci || !ci->ci_ops);
104
105    ci->ci_ops->co_io_lock(ci);
106}
107
108void ocfs2_metadata_cache_io_unlock(struct ocfs2_caching_info *ci)
109{
110    BUG_ON(!ci || !ci->ci_ops);
111
112    ci->ci_ops->co_io_unlock(ci);
113}
114
115
116static void ocfs2_metadata_cache_reset(struct ocfs2_caching_info *ci,
117                       int clear)
118{
119    ci->ci_flags |= OCFS2_CACHE_FL_INLINE;
120    ci->ci_num_cached = 0;
121
122    if (clear) {
123        ci->ci_created_trans = 0;
124        ci->ci_last_trans = 0;
125    }
126}
127
128void ocfs2_metadata_cache_init(struct ocfs2_caching_info *ci,
129                   const struct ocfs2_caching_operations *ops)
130{
131    BUG_ON(!ops);
132
133    ci->ci_ops = ops;
134    ocfs2_metadata_cache_reset(ci, 1);
135}
136
137void ocfs2_metadata_cache_exit(struct ocfs2_caching_info *ci)
138{
139    ocfs2_metadata_cache_purge(ci);
140    ocfs2_metadata_cache_reset(ci, 1);
141}
142
143
144/* No lock taken here as 'root' is not expected to be visible to other
145 * processes. */
146static unsigned int ocfs2_purge_copied_metadata_tree(struct rb_root *root)
147{
148    unsigned int purged = 0;
149    struct rb_node *node;
150    struct ocfs2_meta_cache_item *item;
151
152    while ((node = rb_last(root)) != NULL) {
153        item = rb_entry(node, struct ocfs2_meta_cache_item, c_node);
154
155        mlog(0, "Purge item %llu\n",
156             (unsigned long long) item->c_block);
157
158        rb_erase(&item->c_node, root);
159        kmem_cache_free(ocfs2_uptodate_cachep, item);
160
161        purged++;
162    }
163    return purged;
164}
165
166/* Called from locking and called from ocfs2_clear_inode. Dump the
167 * cache for a given inode.
168 *
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 */
172void ocfs2_metadata_cache_purge(struct ocfs2_caching_info *ci)
173{
174    unsigned int tree, to_purge, purged;
175    struct rb_root root = RB_ROOT;
176
177    BUG_ON(!ci || !ci->ci_ops);
178
179    ocfs2_metadata_cache_lock(ci);
180    tree = !(ci->ci_flags & OCFS2_CACHE_FL_INLINE);
181    to_purge = ci->ci_num_cached;
182
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));
186
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. */
190    if (tree)
191        root = ci->ci_cache.ci_tree;
192
193    ocfs2_metadata_cache_reset(ci, 0);
194    ocfs2_metadata_cache_unlock(ci);
195
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),
203             to_purge, purged);
204}
205
206/* Returns the index in the cache array, -1 if not found.
207 * Requires ip_lock. */
208static int ocfs2_search_cache_array(struct ocfs2_caching_info *ci,
209                    sector_t item)
210{
211    int i;
212
213    for (i = 0; i < ci->ci_num_cached; i++) {
214        if (item == ci->ci_cache.ci_array[i])
215            return i;
216    }
217
218    return -1;
219}
220
221/* Returns the cache item if found, otherwise NULL.
222 * Requires ip_lock. */
223static struct ocfs2_meta_cache_item *
224ocfs2_search_cache_tree(struct ocfs2_caching_info *ci,
225            sector_t block)
226{
227    struct rb_node * n = ci->ci_cache.ci_tree.rb_node;
228    struct ocfs2_meta_cache_item *item = NULL;
229
230    while (n) {
231        item = rb_entry(n, struct ocfs2_meta_cache_item, c_node);
232
233        if (block < item->c_block)
234            n = n->rb_left;
235        else if (block > item->c_block)
236            n = n->rb_right;
237        else
238            return item;
239    }
240
241    return NULL;
242}
243
244static int ocfs2_buffer_cached(struct ocfs2_caching_info *ci,
245                   struct buffer_head *bh)
246{
247    int index = -1;
248    struct ocfs2_meta_cache_item *item = NULL;
249
250    ocfs2_metadata_cache_lock(ci);
251
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));
256
257    if (ci->ci_flags & OCFS2_CACHE_FL_INLINE)
258        index = ocfs2_search_cache_array(ci, bh->b_blocknr);
259    else
260        item = ocfs2_search_cache_tree(ci, bh->b_blocknr);
261
262    ocfs2_metadata_cache_unlock(ci);
263
264    mlog(0, "index = %d, item = %p\n", index, item);
265
266    return (index != -1) || (item != NULL);
267}
268
269/* Warning: even if it returns true, this does *not* guarantee that
270 * the block is stored in our inode metadata cache.
271 *
272 * This can be called under lock_buffer()
273 */
274int ocfs2_buffer_uptodate(struct ocfs2_caching_info *ci,
275              struct buffer_head *bh)
276{
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
279     * data. */
280    if (!buffer_uptodate(bh))
281        return 0;
282
283    /* OCFS2 does not allow multiple nodes to be changing the same
284     * block at the same time. */
285    if (buffer_jbd(bh))
286        return 1;
287
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);
291}
292
293/*
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.
296 */
297int ocfs2_buffer_read_ahead(struct ocfs2_caching_info *ci,
298                struct buffer_head *bh)
299{
300    return buffer_locked(bh) && ocfs2_buffer_cached(ci, bh);
301}
302
303/* Requires ip_lock */
304static void ocfs2_append_cache_array(struct ocfs2_caching_info *ci,
305                     sector_t block)
306{
307    BUG_ON(ci->ci_num_cached >= OCFS2_CACHE_INFO_MAX_ARRAY);
308
309    mlog(0, "block %llu takes position %u\n", (unsigned long long) block,
310         ci->ci_num_cached);
311
312    ci->ci_cache.ci_array[ci->ci_num_cached] = block;
313    ci->ci_num_cached++;
314}
315
316/* By now the caller should have checked that the item does *not*
317 * exist in the tree.
318 * Requires ip_lock. */
319static void __ocfs2_insert_cache_tree(struct ocfs2_caching_info *ci,
320                      struct ocfs2_meta_cache_item *new)
321{
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;
326
327    mlog(0, "Insert block %llu num = %u\n", (unsigned long long) block,
328         ci->ci_num_cached);
329
330    while(*p) {
331        parent = *p;
332
333        tmp = rb_entry(parent, struct ocfs2_meta_cache_item, c_node);
334
335        if (block < tmp->c_block)
336            p = &(*p)->rb_left;
337        else if (block > tmp->c_block)
338            p = &(*p)->rb_right;
339        else {
340            /* This should never happen! */
341            mlog(ML_ERROR, "Duplicate block %llu cached!\n",
342                 (unsigned long long) block);
343            BUG();
344        }
345    }
346
347    rb_link_node(&new->c_node, parent, p);
348    rb_insert_color(&new->c_node, &ci->ci_cache.ci_tree);
349    ci->ci_num_cached++;
350}
351
352/* co_cache_lock() must be held */
353static inline int ocfs2_insert_can_use_array(struct ocfs2_caching_info *ci)
354{
355    return (ci->ci_flags & OCFS2_CACHE_FL_INLINE) &&
356        (ci->ci_num_cached < OCFS2_CACHE_INFO_MAX_ARRAY);
357}
358
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.
362 *
363 * The co_cache_lock() must be held. */
364static void ocfs2_expand_cache(struct ocfs2_caching_info *ci,
365                   struct ocfs2_meta_cache_item **tree)
366{
367    int i;
368
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));
376
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];
381
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;
386
387    for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++) {
388        __ocfs2_insert_cache_tree(ci, tree[i]);
389        tree[i] = NULL;
390    }
391
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);
395}
396
397/* Slow path function - memory allocation is necessary. See the
398 * comment above ocfs2_set_buffer_uptodate for more information. */
399static void __ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
400                    sector_t block,
401                    int expand_tree)
402{
403    int i;
404    struct ocfs2_meta_cache_item *new = NULL;
405    struct ocfs2_meta_cache_item *tree[OCFS2_CACHE_INFO_MAX_ARRAY] =
406        { NULL, };
407
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);
411
412    new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_NOFS);
413    if (!new) {
414        mlog_errno(-ENOMEM);
415        return;
416    }
417    new->c_block = block;
418
419    if (expand_tree) {
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,
424                           GFP_NOFS);
425            if (!tree[i]) {
426                mlog_errno(-ENOMEM);
427                goto out_free;
428            }
429
430            /* These are initialized in ocfs2_expand_cache! */
431        }
432    }
433
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);
441        goto out_free;
442    }
443
444    if (expand_tree)
445        ocfs2_expand_cache(ci, tree);
446
447    __ocfs2_insert_cache_tree(ci, new);
448    ocfs2_metadata_cache_unlock(ci);
449
450    new = NULL;
451out_free:
452    if (new)
453        kmem_cache_free(ocfs2_uptodate_cachep, new);
454
455    /* If these were used, then ocfs2_expand_cache re-set them to
456     * NULL for us. */
457    if (tree[0]) {
458        for (i = 0; i < OCFS2_CACHE_INFO_MAX_ARRAY; i++)
459            if (tree[i])
460                kmem_cache_free(ocfs2_uptodate_cachep,
461                        tree[i]);
462    }
463}
464
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.
470 *
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
474 * path.
475 *
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)
479 *
480 * Readahead buffers can be passed in here before the I/O request is
481 * completed.
482 */
483void ocfs2_set_buffer_uptodate(struct ocfs2_caching_info *ci,
484                   struct buffer_head *bh)
485{
486    int expand;
487
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))
491        return;
492
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);
496
497    /* No need to recheck under spinlock - insertion is guarded by
498     * co_io_lock() */
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
502         * spot. */
503        ocfs2_append_cache_array(ci, bh->b_blocknr);
504        ocfs2_metadata_cache_unlock(ci);
505        return;
506    }
507
508    expand = 0;
509    if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
510        /* We need to bump things up to a tree. */
511        expand = 1;
512    }
513    ocfs2_metadata_cache_unlock(ci);
514
515    __ocfs2_set_buffer_uptodate(ci, bh->b_blocknr, expand);
516}
517
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. */
521void ocfs2_set_new_buffer_uptodate(struct ocfs2_caching_info *ci,
522                   struct buffer_head *bh)
523{
524    /* This should definitely *not* exist in our cache */
525    BUG_ON(ocfs2_buffer_cached(ci, bh));
526
527    set_buffer_uptodate(bh);
528
529    ocfs2_metadata_cache_io_lock(ci);
530    ocfs2_set_buffer_uptodate(ci, bh);
531    ocfs2_metadata_cache_io_unlock(ci);
532}
533
534/* Requires ip_lock. */
535static void ocfs2_remove_metadata_array(struct ocfs2_caching_info *ci,
536                    int index)
537{
538    sector_t *array = ci->ci_cache.ci_array;
539    int bytes;
540
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);
544
545    mlog(0, "remove index %d (num_cached = %u\n", index,
546         ci->ci_num_cached);
547
548    ci->ci_num_cached--;
549
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);
555    }
556}
557
558/* Requires ip_lock. */
559static void ocfs2_remove_metadata_tree(struct ocfs2_caching_info *ci,
560                       struct ocfs2_meta_cache_item *item)
561{
562    mlog(0, "remove block %llu from tree\n",
563         (unsigned long long) item->c_block);
564
565    rb_erase(&item->c_node, &ci->ci_cache.ci_tree);
566    ci->ci_num_cached--;
567}
568
569static void ocfs2_remove_block_from_cache(struct ocfs2_caching_info *ci,
570                      sector_t block)
571{
572    int index;
573    struct ocfs2_meta_cache_item *item = NULL;
574
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);
580
581    if (ci->ci_flags & OCFS2_CACHE_FL_INLINE) {
582        index = ocfs2_search_cache_array(ci, block);
583        if (index != -1)
584            ocfs2_remove_metadata_array(ci, index);
585    } else {
586        item = ocfs2_search_cache_tree(ci, block);
587        if (item)
588            ocfs2_remove_metadata_tree(ci, item);
589    }
590    ocfs2_metadata_cache_unlock(ci);
591
592    if (item)
593        kmem_cache_free(ocfs2_uptodate_cachep, item);
594}
595
596/*
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.
600 */
601void ocfs2_remove_from_cache(struct ocfs2_caching_info *ci,
602                 struct buffer_head *bh)
603{
604    sector_t block = bh->b_blocknr;
605
606    ocfs2_remove_block_from_cache(ci, block);
607}
608
609/* Called when we remove xattr clusters from an inode. */
610void ocfs2_remove_xattr_clusters_from_cache(struct ocfs2_caching_info *ci,
611                        sector_t block,
612                        u32 c_len)
613{
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;
616
617    for (i = 0; i < b_len; i++, block++)
618        ocfs2_remove_block_from_cache(ci, block);
619}
620
621int __init init_ocfs2_uptodate_cache(void)
622{
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)
627        return -ENOMEM;
628
629    mlog(0, "%u inlined cache items per inode.\n",
630         OCFS2_CACHE_INFO_MAX_ARRAY);
631
632    return 0;
633}
634
635void exit_ocfs2_uptodate_cache(void)
636{
637    if (ocfs2_uptodate_cachep)
638        kmem_cache_destroy(ocfs2_uptodate_cachep);
639}
640

Archive Download this file



interactive