path: root/fs/ocfs2/uptodate.c

// SPDX-License-Identifier: GPL-2.0-or-later
/* -*- 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.
 *
 * 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 <cluster/masklog.h>

#include "ocfs2.h"

#include "inode.h"
#include "uptodate.h"
#include "ocfs2_trace.h"

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

static struct kmem_cache *ocfs2_uptodate_cachep;

u64 ocfs2_metadata_cache_owner(struct ocfs2_caching_info *ci)
{
	BUG_ON(!ci || !ci->ci_ops);

	return ci->ci_ops->co_owner(ci);
}

struct super_block *ocfs2_metadata_cache_get_super(struct ocfs2_caching_info *ci)
{
	BUG_ON(!ci || !ci->ci_ops);

	return ci->ci_ops->co_get_super(ci);
}

static void ocfs2_metadata_cache_lock(struct ocfs2_caching_info *ci)
{
	BUG_ON(!ci || !ci->ci_ops);

	ci->ci_ops->co_cache_lock(ci);
}

static void ocfs2_metadata_cache_unlock(struct ocfs2_caching_info *ci)
{
	BUG_ON(!ci || !ci->ci_ops);

	ci->ci_ops->co_cache_unlock(ci);
}

void ocfs2_metadata_cache_io_lock(struct ocfs2_caching_info *ci)
{
	BUG_ON(!ci || !ci->ci_ops);

	ci->ci_ops->co_io_lock(ci);
}

void ocfs2_metadata_cache_io_unlock(struct ocfs2_caching_info *ci)
{
	BUG_ON(!ci || !ci->ci_ops);

	ci->ci_ops->co_io_unlock(ci);
}


static void ocfs2_metadata_cache_reset(struct ocfs2_caching_info *ci,
				       int clear)
{
	ci->ci_flags |= OCFS2_CACHE_FL_INLINE;
	ci->ci_num_cached = 0;

	if (clear) {
		ci->ci_created_trans = 0;
		ci->ci_last_trans = 0;
	}
}

void ocfs2_metadata_cache_init(struct ocfs2_caching_info *ci,
			       const struct ocfs2_caching_operations *ops)
{
	BUG_ON(!ops);

	ci->ci_ops = ops;
	ocfs2_metadata_cache_reset(ci, 1);
}

void ocfs2_metadata_cache_exit(struct ocfs2_caching_info *ci)
{
	ocfs2_metadata_cache_purge(ci);
	ocfs2_metadata_cache_reset(ci, 1);
}


/* 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);

		trace_ocfs2_purge_copied_metadata_tree(
					(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 ocfs2_caching_info *ci)
{
	unsigned int tree, to_purge, purged;
	struct rb_root root = RB_ROOT;

	BUG_ON(!ci || !ci->ci_ops);

	ocfs2_metadata_cache_lock(ci);
	tree = !(ci->ci_flags & OCFS2_CACHE_FL_INLINE);
	to_purge = ci->ci_num_cached;

	trace_ocfs2_metadata_cache_purge(
		(unsigned long long)ocfs2_metadata_cache_owner(ci),
		to_purge, tree);

	/* 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_reset(ci, 0);
	ocfs2_metadata_cache_unlock(ci);

	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, "Owner %llu, count = %u, purged = %u\n",
		     (unsigned long long)ocfs2_metadata_cache_owner(ci),
		     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_caching_info *ci,
			       struct buffer_head *bh)
{
	int index = -1;
	struct ocfs2_meta_cache_item *item = NULL;

	ocfs2_metadata_cache_lock(ci);

	trace_ocfs2_buffer_cached_begin(
		(unsigned long long)ocfs2_metadata_cache_owner(ci),
		(unsigned long long) bh->b_blocknr,
		!!(ci->ci_flags & OCFS2_CACHE_FL_INLINE));

	if (ci->ci_flags & OCFS2_CACHE_FL_INLINE)
		index = ocfs2_search_cache_array(ci, bh->b_blocknr);
	else
		item = ocfs2_search_cache_tree(ci, bh->b_blocknr);

	ocfs2_metadata_cache_unlock(ci);

	trace_ocfs2_buffer_cached_end(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 ocfs2_caching_info *ci,
			  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