uptodate.c

linux 内核源代码

	}

	rb_link_node(&new->c_node, parent, p);
	rb_insert_color(&new->c_node, &ci->ci_cache.ci_tree);
	ci->ci_num_cached++;
}

static inline int ocfs2_insert_can_use_array(struct ocfs2_inode_info *oi,
					     struct ocfs2_caching_info *ci)
{
	assert_spin_locked(&oi->ip_lock);

	return (oi->ip_flags & OCFS2_INODE_CACHE_INLINE) &&
		(ci->ci_num_cached < OCFS2_INODE_MAX_CACHE_ARRAY);
}

/* tree should be exactly OCFS2_INODE_MAX_CACHE_ARRAY wide. NULL the
 * pointers in tree after we use them - this allows caller to detect
 * when to free in case of error. */
static void ocfs2_expand_cache(struct ocfs2_inode_info *oi,
			       struct ocfs2_meta_cache_item **tree)
{
	int i;
	struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;

	mlog_bug_on_msg(ci->ci_num_cached != OCFS2_INODE_MAX_CACHE_ARRAY,
			"Inode %llu, num cached = %u, should be %u\n",
			(unsigned long long)oi->ip_blkno, ci->ci_num_cached,
			OCFS2_INODE_MAX_CACHE_ARRAY);
	mlog_bug_on_msg(!(oi->ip_flags & OCFS2_INODE_CACHE_INLINE),
			"Inode %llu not marked as inline anymore!\n",
			(unsigned long long)oi->ip_blkno);
	assert_spin_locked(&oi->ip_lock);

	/* Be careful to initialize the tree members *first* because
	 * once the ci_tree is used, the array is junk... */
	for(i = 0; i < OCFS2_INODE_MAX_CACHE_ARRAY; i++)
		tree[i]->c_block = ci->ci_cache.ci_array[i];

	oi->ip_flags &= ~OCFS2_INODE_CACHE_INLINE;
	ci->ci_cache.ci_tree = RB_ROOT;
	/* this will be set again by __ocfs2_insert_cache_tree */
	ci->ci_num_cached = 0;

	for(i = 0; i < OCFS2_INODE_MAX_CACHE_ARRAY; i++) {
		__ocfs2_insert_cache_tree(ci, tree[i]);
		tree[i] = NULL;
	}

	mlog(0, "Expanded %llu to a tree cache: flags 0x%x, num = %u\n",
	     (unsigned long long)oi->ip_blkno, oi->ip_flags, ci->ci_num_cached);
}

/* Slow path function - memory allocation is necessary. See the
 * comment above ocfs2_set_buffer_uptodate for more information. */
static void __ocfs2_set_buffer_uptodate(struct ocfs2_inode_info *oi,
					sector_t block,
					int expand_tree)
{
	int i;
	struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;
	struct ocfs2_meta_cache_item *new = NULL;
	struct ocfs2_meta_cache_item *tree[OCFS2_INODE_MAX_CACHE_ARRAY] =
		{ NULL, };

	mlog(0, "Inode %llu, block %llu, expand = %d\n",
	     (unsigned long long)oi->ip_blkno,
	     (unsigned long long)block, expand_tree);

	new = kmem_cache_alloc(ocfs2_uptodate_cachep, GFP_NOFS);
	if (!new) {
		mlog_errno(-ENOMEM);
		return;
	}
	new->c_block = block;

	if (expand_tree) {
		/* Do *not* allocate an array here - the removal code
		 * has no way of tracking that. */
		for(i = 0; i < OCFS2_INODE_MAX_CACHE_ARRAY; i++) {
			tree[i] = kmem_cache_alloc(ocfs2_uptodate_cachep,
						   GFP_NOFS);
			if (!tree[i]) {
				mlog_errno(-ENOMEM);
				goto out_free;
			}

			/* These are initialized in ocfs2_expand_cache! */
		}
	}

	spin_lock(&oi->ip_lock);
	if (ocfs2_insert_can_use_array(oi, ci)) {
		mlog(0, "Someone cleared the tree underneath us\n");
		/* Ok, items were removed from the cache in between
		 * locks. Detect this and revert back to the fast path */
		ocfs2_append_cache_array(ci, block);
		spin_unlock(&oi->ip_lock);
		goto out_free;
	}

	if (expand_tree)
		ocfs2_expand_cache(oi, tree);

	__ocfs2_insert_cache_tree(ci, new);
	spin_unlock(&oi->ip_lock);

	new = NULL;
out_free:
	if (new)
		kmem_cache_free(ocfs2_uptodate_cachep, new);

	/* If these were used, then ocfs2_expand_cache re-set them to
	 * NULL for us. */
	if (tree[0]) {
		for(i = 0; i < OCFS2_INODE_MAX_CACHE_ARRAY; i++)
			if (tree[i])
				kmem_cache_free(ocfs2_uptodate_cachep,
						tree[i]);
	}
}

/* Item insertion is guarded by ip_io_mutex, so the insertion path takes
 * advantage of this by not rechecking for a duplicate insert during
 * the slow case. Additionally, if the cache needs to be bumped up to
 * a tree, the code will not recheck after acquiring the lock --
 * multiple paths cannot be expanding to a tree at the same time.
 *
 * The slow path takes into account that items can be removed
 * (including the whole tree wiped and reset) when this process it out
 * allocating memory. In those cases, it reverts back to the fast
 * path.
 *
 * Note that this function may actually fail to insert the block if
 * memory cannot be allocated. This is not fatal however (but may
 * result in a performance penalty)
 *
 * Readahead buffers can be passed in here before the I/O request is
 * completed.
 */
void ocfs2_set_buffer_uptodate(struct inode *inode,
			       struct buffer_head *bh)
{
	int expand;
	struct ocfs2_inode_info *oi = OCFS2_I(inode);
	struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;

	/* The block may very well exist in our cache already, so avoid
	 * doing any more work in that case. */
	if (ocfs2_buffer_cached(oi, bh))
		return;

	mlog(0, "Inode %llu, inserting block %llu\n",
	     (unsigned long long)oi->ip_blkno,
	     (unsigned long long)bh->b_blocknr);

	/* No need to recheck under spinlock - insertion is guarded by
	 * ip_io_mutex */
	spin_lock(&oi->ip_lock);
	if (ocfs2_insert_can_use_array(oi, ci)) {
		/* Fast case - it's an array and there's a free
		 * spot. */
		ocfs2_append_cache_array(ci, bh->b_blocknr);
		spin_unlock(&oi->ip_lock);
		return;
	}

	expand = 0;
	if (oi->ip_flags & OCFS2_INODE_CACHE_INLINE) {
		/* We need to bump things up to a tree. */
		expand = 1;
	}
	spin_unlock(&oi->ip_lock);

	__ocfs2_set_buffer_uptodate(oi, bh->b_blocknr, expand);
}

/* Called against a newly allocated buffer. Most likely nobody should
 * be able to read this sort of metadata while it's still being
 * allocated, but this is careful to take ip_io_mutex anyway. */
void ocfs2_set_new_buffer_uptodate(struct inode *inode,
				   struct buffer_head *bh)
{
	struct ocfs2_inode_info *oi = OCFS2_I(inode);

	/* This should definitely *not* exist in our cache */
	BUG_ON(ocfs2_buffer_cached(oi, bh));

	set_buffer_uptodate(bh);

	mutex_lock(&oi->ip_io_mutex);
	ocfs2_set_buffer_uptodate(inode, bh);
	mutex_unlock(&oi->ip_io_mutex);
}

/* Requires ip_lock. */
static void ocfs2_remove_metadata_array(struct ocfs2_caching_info *ci,
					int index)
{
	sector_t *array = ci->ci_cache.ci_array;
	int bytes;

	BUG_ON(index < 0 || index >= OCFS2_INODE_MAX_CACHE_ARRAY);
	BUG_ON(index >= ci->ci_num_cached);
	BUG_ON(!ci->ci_num_cached);

	mlog(0, "remove index %d (num_cached = %u\n", index,
	     ci->ci_num_cached);

	ci->ci_num_cached--;

	/* don't need to copy if the array is now empty, or if we
	 * removed at the tail */
	if (ci->ci_num_cached && index < ci->ci_num_cached) {
		bytes = sizeof(sector_t) * (ci->ci_num_cached - index);
		memmove(&array[index], &array[index + 1], bytes);
	}
}

/* Requires ip_lock. */
static void ocfs2_remove_metadata_tree(struct ocfs2_caching_info *ci,
				       struct ocfs2_meta_cache_item *item)
{
	mlog(0, "remove block %llu from tree\n",
	     (unsigned long long) item->c_block);

	rb_erase(&item->c_node, &ci->ci_cache.ci_tree);
	ci->ci_num_cached--;
}

/* Called when we remove a chunk of metadata from an inode. We don't
 * bother reverting things to an inlined array in the case of a remove
 * which moves us back under the limit. */
void ocfs2_remove_from_cache(struct inode *inode,
			     struct buffer_head *bh)
{
	int index;
	sector_t block = bh->b_blocknr;
	struct ocfs2_meta_cache_item *item = NULL;
	struct ocfs2_inode_info *oi = OCFS2_I(inode);
	struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;

	spin_lock(&oi->ip_lock);
	mlog(0, "Inode %llu, remove %llu, items = %u, array = %u\n",
	     (unsigned long long)oi->ip_blkno,
	     (unsigned long long) block, ci->ci_num_cached,
	     oi->ip_flags & OCFS2_INODE_CACHE_INLINE);

	if (oi->ip_flags & OCFS2_INODE_CACHE_INLINE) {
		index = ocfs2_search_cache_array(ci, block);
		if (index != -1)
			ocfs2_remove_metadata_array(ci, index);
	} else {
		item = ocfs2_search_cache_tree(ci, block);
		if (item)
			ocfs2_remove_metadata_tree(ci, item);
	}
	spin_unlock(&oi->ip_lock);

	if (item)
		kmem_cache_free(ocfs2_uptodate_cachep, item);
}

int __init init_ocfs2_uptodate_cache(void)
{
	ocfs2_uptodate_cachep = kmem_cache_create("ocfs2_uptodate",
				  sizeof(struct ocfs2_meta_cache_item),
				  0, SLAB_HWCACHE_ALIGN, NULL);
	if (!ocfs2_uptodate_cachep)
		return -ENOMEM;

	mlog(0, "%u inlined cache items per inode.\n",
	     OCFS2_INODE_MAX_CACHE_ARRAY);

	return 0;
}

void exit_ocfs2_uptodate_cache(void)
{
	if (ocfs2_uptodate_cachep)
		kmem_cache_destroy(ocfs2_uptodate_cachep);
}