uptodate.c

linux 内核源代码

/* -*- mode: c; c-basic-offset: 8; -*-
 * vim: noexpandtab sw=8 ts=8 sts=0:
 *
 * uptodate.c
 *
 * Tracking the up-to-date-ness of a local buffer_head with respect to
 * the cluster.
 *
 * Copyright (C) 2002, 2004, 2005 Oracle.  All rights reserved.
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License for more details.
 *
 * You should have received a copy of the GNU General Public
 * License along with this program; if not, write to the
 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
 * Boston, MA 021110-1307, USA.
 *
 * Standard buffer head caching flags (uptodate, etc) are insufficient
 * in a clustered environment - a buffer may be marked up to date on
 * our local node but could have been modified by another cluster
 * member. As a result an additional (and performant) caching scheme
 * is required. A further requirement is that we consume as little
 * memory as possible - we never pin buffer_head structures in order
 * to cache them.
 *
 * We track the existence of up to date buffers on the inodes which
 * are associated with them. Because we don't want to pin
 * buffer_heads, this is only a (strong) hint and several other checks
 * are made in the I/O path to ensure that we don't use a stale or
 * invalid buffer without going to disk:
 *	- buffer_jbd is used liberally - if a bh is in the journal on
 *	  this node then it *must* be up to date.
 *	- the standard buffer_uptodate() macro is used to detect buffers
 *	  which may be invalid (even if we have an up to date tracking
 * 	  item for them)
 *
 * For a full understanding of how this code works together, one
 * should read the callers in dlmglue.c, the I/O functions in
 * buffer_head_io.c and ocfs2_journal_access in journal.c
 */

#include <linux/fs.h>
#include <linux/types.h>
#include <linux/slab.h>
#include <linux/highmem.h>
#include <linux/buffer_head.h>
#include <linux/rbtree.h>
#include <linux/jbd.h>

#define MLOG_MASK_PREFIX ML_UPTODATE

#include <cluster/masklog.h>

#include "ocfs2.h"

#include "inode.h"
#include "uptodate.h"

struct ocfs2_meta_cache_item {
	struct rb_node	c_node;
	sector_t	c_block;
};

static struct kmem_cache *ocfs2_uptodate_cachep = NULL;

void ocfs2_metadata_cache_init(struct inode *inode)
{
	struct ocfs2_inode_info *oi = OCFS2_I(inode);
	struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;

	oi->ip_flags |= OCFS2_INODE_CACHE_INLINE;
	ci->ci_num_cached = 0;
}

/* No lock taken here as 'root' is not expected to be visible to other
 * processes. */
static unsigned int ocfs2_purge_copied_metadata_tree(struct rb_root *root)
{
	unsigned int purged = 0;
	struct rb_node *node;
	struct ocfs2_meta_cache_item *item;

	while ((node = rb_last(root)) != NULL) {
		item = rb_entry(node, struct ocfs2_meta_cache_item, c_node);

		mlog(0, "Purge item %llu\n",
		     (unsigned long long) item->c_block);

		rb_erase(&item->c_node, root);
		kmem_cache_free(ocfs2_uptodate_cachep, item);

		purged++;
	}
	return purged;
}

/* Called from locking and called from ocfs2_clear_inode. Dump the
 * cache for a given inode.
 *
 * This function is a few more lines longer than necessary due to some
 * accounting done here, but I think it's worth tracking down those
 * bugs sooner -- Mark */
void ocfs2_metadata_cache_purge(struct inode *inode)
{
	struct ocfs2_inode_info *oi = OCFS2_I(inode);
	unsigned int tree, to_purge, purged;
	struct ocfs2_caching_info *ci = &oi->ip_metadata_cache;
	struct rb_root root = RB_ROOT;

	spin_lock(&oi->ip_lock);
	tree = !(oi->ip_flags & OCFS2_INODE_CACHE_INLINE);
	to_purge = ci->ci_num_cached;

	mlog(0, "Purge %u %s items from Inode %llu\n", to_purge,
	     tree ? "array" : "tree", (unsigned long long)oi->ip_blkno);

	/* If we're a tree, save off the root so that we can safely
	 * initialize the cache. We do the work to free tree members
	 * without the spinlock. */
	if (tree)
		root = ci->ci_cache.ci_tree;

	ocfs2_metadata_cache_init(inode);
	spin_unlock(&oi->ip_lock);

	purged = ocfs2_purge_copied_metadata_tree(&root);
	/* If possible, track the number wiped so that we can more
	 * easily detect counting errors. Unfortunately, this is only
	 * meaningful for trees. */
	if (tree && purged != to_purge)
		mlog(ML_ERROR, "Inode %llu, count = %u, purged = %u\n",
		     (unsigned long long)oi->ip_blkno, to_purge, purged);
}

/* Returns the index in the cache array, -1 if not found.
 * Requires ip_lock. */
static int ocfs2_search_cache_array(struct ocfs2_caching_info *ci,
				    sector_t item)
{
	int i;

	for (i = 0; i < ci->ci_num_cached; i++) {
		if (item == ci->ci_cache.ci_array[i])
			return i;
	}

	return -1;
}

/* Returns the cache item if found, otherwise NULL.
 * Requires ip_lock. */
static struct ocfs2_meta_cache_item *
ocfs2_search_cache_tree(struct ocfs2_caching_info *ci,
			sector_t block)
{
	struct rb_node * n = ci->ci_cache.ci_tree.rb_node;
	struct ocfs2_meta_cache_item *item = NULL;

	while (n) {
		item = rb_entry(n, struct ocfs2_meta_cache_item, c_node);

		if (block < item->c_block)
			n = n->rb_left;
		else if (block > item->c_block)
			n = n->rb_right;
		else
			return item;
	}

	return NULL;
}

static int ocfs2_buffer_cached(struct ocfs2_inode_info *oi,
			       struct buffer_head *bh)
{
	int index = -1;
	struct ocfs2_meta_cache_item *item = NULL;

	spin_lock(&oi->ip_lock);

	mlog(0, "Inode %llu, query block %llu (inline = %u)\n",
	     (unsigned long long)oi->ip_blkno,
	     (unsigned long long) bh->b_blocknr,
	     !!(oi->ip_flags & OCFS2_INODE_CACHE_INLINE));

	if (oi->ip_flags & OCFS2_INODE_CACHE_INLINE)
		index = ocfs2_search_cache_array(&oi->ip_metadata_cache,
						 bh->b_blocknr);
	else
		item = ocfs2_search_cache_tree(&oi->ip_metadata_cache,
					       bh->b_blocknr);

	spin_unlock(&oi->ip_lock);

	mlog(0, "index = %d, item = %p\n", index, item);

	return (index != -1) || (item != NULL);
}

/* Warning: even if it returns true, this does *not* guarantee that
 * the block is stored in our inode metadata cache. 
 * 
 * This can be called under lock_buffer()
 */
int ocfs2_buffer_uptodate(struct inode *inode,
			  struct buffer_head *bh)
{
	/* Doesn't matter if the bh is in our cache or not -- if it's
	 * not marked uptodate then we know it can't have correct
	 * data. */
	if (!buffer_uptodate(bh))
		return 0;

	/* OCFS2 does not allow multiple nodes to be changing the same
	 * block at the same time. */
	if (buffer_jbd(bh))
		return 1;

	/* Ok, locally the buffer is marked as up to date, now search
	 * our cache to see if we can trust that. */
	return ocfs2_buffer_cached(OCFS2_I(inode), bh);
}

/* 
 * Determine whether a buffer is currently out on a read-ahead request.
 * ip_io_sem should be held to serialize submitters with the logic here.
 */
int ocfs2_buffer_read_ahead(struct inode *inode,
			    struct buffer_head *bh)
{
	return buffer_locked(bh) && ocfs2_buffer_cached(OCFS2_I(inode), bh);
}

/* Requires ip_lock */
static void ocfs2_append_cache_array(struct ocfs2_caching_info *ci,
				     sector_t block)
{
	BUG_ON(ci->ci_num_cached >= OCFS2_INODE_MAX_CACHE_ARRAY);

	mlog(0, "block %llu takes position %u\n", (unsigned long long) block,
	     ci->ci_num_cached);

	ci->ci_cache.ci_array[ci->ci_num_cached] = block;
	ci->ci_num_cached++;
}

/* By now the caller should have checked that the item does *not*
 * exist in the tree.
 * Requires ip_lock. */
static void __ocfs2_insert_cache_tree(struct ocfs2_caching_info *ci,
				      struct ocfs2_meta_cache_item *new)
{
	sector_t block = new->c_block;
	struct rb_node *parent = NULL;
	struct rb_node **p = &ci->ci_cache.ci_tree.rb_node;
	struct ocfs2_meta_cache_item *tmp;

	mlog(0, "Insert block %llu num = %u\n", (unsigned long long) block,
	     ci->ci_num_cached);

	while(*p) {
		parent = *p;

		tmp = rb_entry(parent, struct ocfs2_meta_cache_item, c_node);

		if (block < tmp->c_block)
			p = &(*p)->rb_left;
		else if (block > tmp->c_block)
			p = &(*p)->rb_right;
		else {
			/* This should never happen! */
			mlog(ML_ERROR, "Duplicate block %llu cached!\n",
			     (unsigned long long) block);
			BUG();
		}