extent_map.c

ocfs1.2.7 源码

	}

	rb_erase(&old_ent->e_node, &em->em_extents);
	/* Now that he's erased, set him up for deletion */
	ctxt->old_ent = old_ent;

	if (ctxt->need_left) {
		ret = ocfs2_extent_map_insert_entry(em,
						    ctxt->left_ent);
		if (ret)
			goto out_unlock;
		ctxt->left_ent = NULL;
	}

	if (ctxt->need_right) {
		ret = ocfs2_extent_map_insert_entry(em,
						    ctxt->right_ent);
		if (ret)
			goto out_unlock;
		ctxt->right_ent = NULL;
	}

	ret = ocfs2_extent_map_insert_entry(em, ctxt->new_ent);

	if (!ret)
		ctxt->new_ent = NULL;

out_unlock:
	spin_unlock(&OCFS2_I(inode)->ip_lock);

	return ret;
}


int ocfs2_extent_map_insert(struct inode *inode, struct ocfs2_extent_rec *rec,
			    int tree_depth)
{
	int ret;
	struct ocfs2_em_insert_context ctxt = {0, };

	if ((le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters)) >
	    OCFS2_I(inode)->ip_map.em_clusters) {
		ret = -EBADR;
		mlog_errno(ret);
		return ret;
	}

	/* Zero e_clusters means a truncated tail record.  It better be EOF */
	if (!rec->e_clusters) {
		if ((le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters)) !=
		    OCFS2_I(inode)->ip_map.em_clusters) {
			ret = -EBADR;
			mlog_errno(ret);
			ocfs2_error(inode->i_sb,
				    "Zero e_clusters on non-tail extent record at e_blkno %"MLFu64" on inode %"MLFu64"\n",
				    le64_to_cpu(rec->e_blkno),
				    OCFS2_I(inode)->ip_blkno);
			return ret;
		}

		/* Ignore the truncated tail */
		return 0;
	}

	ret = -ENOMEM;
	ctxt.new_ent = kmem_cache_alloc(ocfs2_em_ent_cachep,
					GFP_NOFS);
	if (!ctxt.new_ent) {
		mlog_errno(ret);
		return ret;
	}

	ctxt.new_ent->e_rec = *rec;
	ctxt.new_ent->e_tree_depth = tree_depth;

	do {
		ret = -ENOMEM;
		if (ctxt.need_left && !ctxt.left_ent) {
			ctxt.left_ent =
				kmem_cache_alloc(ocfs2_em_ent_cachep,
						 GFP_NOFS);
			if (!ctxt.left_ent)
				break;
		}
		if (ctxt.need_right && !ctxt.right_ent) {
			ctxt.right_ent =
				kmem_cache_alloc(ocfs2_em_ent_cachep,
						 GFP_NOFS);
			if (!ctxt.right_ent)
				break;
		}

		ret = ocfs2_extent_map_try_insert(inode, rec,
						  tree_depth, &ctxt);
	} while (ret == -EAGAIN);

	if ((ret < 0) && (ret != -EEXIST))
		mlog_errno(ret);

	if (ctxt.left_ent)
		kmem_cache_free(ocfs2_em_ent_cachep, ctxt.left_ent);
	if (ctxt.right_ent)
		kmem_cache_free(ocfs2_em_ent_cachep, ctxt.right_ent);
	if (ctxt.old_ent)
		kmem_cache_free(ocfs2_em_ent_cachep, ctxt.old_ent);
	if (ctxt.new_ent)
		kmem_cache_free(ocfs2_em_ent_cachep, ctxt.new_ent);

	return ret;
}

/*
 * Append this record to the tail of the extent map.  It must be
 * tree_depth 0.  The record might be an extension of an existing
 * record, and as such that needs to be handled.  eg:
 *
 * Existing record in the extent map:
 *
 *	cpos = 10, len = 10
 *	|---------|
 *
 * New Record:
 *
 *	cpos = 10, len = 20
 *	|------------------|
 *
 * The passed record is the new on-disk record.  The new_clusters value
 * is how many clusters were added to the file.  If the append is a
 * contiguous append, the new_clusters has been added to
 * rec->e_clusters.  If the append is an entirely new extent, then
 * rec->e_clusters is == new_clusters.
 */
int ocfs2_extent_map_append(struct inode *inode,
			    struct ocfs2_extent_rec *rec,
			    u32 new_clusters)
{
	int ret;
	struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
	struct ocfs2_extent_map_entry *ent;
	struct ocfs2_extent_rec *old;

	BUG_ON(!new_clusters);
	BUG_ON(le32_to_cpu(rec->e_clusters) < new_clusters);

	if (em->em_clusters < OCFS2_I(inode)->ip_clusters) {
		/*
		 * Size changed underneath us on disk.  Drop any
		 * straddling records and update our idea of
		 * i_clusters
		 */
		ocfs2_extent_map_drop(inode, em->em_clusters - 1);
		em->em_clusters = OCFS2_I(inode)->ip_clusters;
	}

	mlog_bug_on_msg((le32_to_cpu(rec->e_cpos) +
			 le32_to_cpu(rec->e_clusters)) !=
			(em->em_clusters + new_clusters),
			"Inode %"MLFu64":\n"
			"rec->e_cpos = %u + rec->e_clusters = %u = %u\n"
			"em->em_clusters = %u + new_clusters = %u = %u\n",
			OCFS2_I(inode)->ip_blkno,
			le32_to_cpu(rec->e_cpos), le32_to_cpu(rec->e_clusters),
			le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters),
			em->em_clusters, new_clusters,
			em->em_clusters + new_clusters);

	em->em_clusters += new_clusters;

	ret = -ENOENT;
	if (le32_to_cpu(rec->e_clusters) > new_clusters) {
		/* This is a contiguous append */
		ent = ocfs2_extent_map_lookup(em, le32_to_cpu(rec->e_cpos), 1,
					      NULL, NULL);
		if (ent) {
			old = &ent->e_rec;
			BUG_ON((le32_to_cpu(rec->e_cpos) +
				le32_to_cpu(rec->e_clusters)) !=
				 (le32_to_cpu(old->e_cpos) +
				  le32_to_cpu(old->e_clusters) +
				  new_clusters));
			if (ent->e_tree_depth == 0) {
				BUG_ON(le32_to_cpu(old->e_cpos) !=
				       le32_to_cpu(rec->e_cpos));
				BUG_ON(le64_to_cpu(old->e_blkno) !=
				       le64_to_cpu(rec->e_blkno));
				ret = 0;
			}
			/*
			 * Let non-leafs fall through as -ENOENT to
			 * force insertion of the new leaf.
			 */
			le32_add_cpu(&old->e_clusters, new_clusters);
		}
	}

	if (ret == -ENOENT)
		ret = ocfs2_extent_map_insert(inode, rec, 0);
	if (ret < 0)
		mlog_errno(ret);
	return ret;
}

/*
 * Look up the record containing this cluster offset.  This record is
 * part of the extent map.  Do not free it.  Any changes you make to
 * it will reflect in the extent map.  So, if your last extent
 * is (cpos = 10, clusters = 10) and you truncate the file by 5
 * clusters, you can do:
 *
 * ret = ocfs2_extent_map_get_rec(em, orig_size - 5, &rec);
 * rec->e_clusters -= 5;
 *
 * The lookup does not read from disk.  If the map isn't filled in for
 * an entry, you won't find it.
 *
 * Also note that the returned record is valid until alloc_sem is
 * dropped.  After that, truncate and extend can happen.  Caveat Emptor.
 */
int ocfs2_extent_map_get_rec(struct inode *inode, u32 cpos,
			     struct ocfs2_extent_rec **rec,
			     int *tree_depth)
{
	int ret = -ENOENT;
	struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
	struct ocfs2_extent_map_entry *ent;

	*rec = NULL;

	if (cpos >= OCFS2_I(inode)->ip_clusters)
		return -EINVAL;

	if (cpos >= em->em_clusters) {
		/*
		 * Size changed underneath us on disk.  Drop any
		 * straddling records and update our idea of
		 * i_clusters
		 */
		ocfs2_extent_map_drop(inode, em->em_clusters - 1);
		em->em_clusters = OCFS2_I(inode)->ip_clusters ;
	}

	ent = ocfs2_extent_map_lookup(&OCFS2_I(inode)->ip_map, cpos, 1,
				      NULL, NULL);

	if (ent) {
		*rec = &ent->e_rec;
		if (tree_depth)
			*tree_depth = ent->e_tree_depth;
		ret = 0;
	}

	return ret;
}

int ocfs2_extent_map_get_clusters(struct inode *inode,
				  u32 v_cpos, int count,
				  u32 *p_cpos, int *ret_count)
{
	int ret;
	u32 coff, ccount;
	struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
	struct ocfs2_extent_map_entry *ent = NULL;

	*p_cpos = ccount = 0;

	if ((v_cpos + count) > OCFS2_I(inode)->ip_clusters)
		return -EINVAL;

	if ((v_cpos + count) > em->em_clusters) {
		/*
		 * Size changed underneath us on disk.  Drop any
		 * straddling records and update our idea of
		 * i_clusters
		 */
		ocfs2_extent_map_drop(inode, em->em_clusters - 1);
		em->em_clusters = OCFS2_I(inode)->ip_clusters;
	}


	ret = ocfs2_extent_map_lookup_read(inode, v_cpos, count, &ent);
	if (ret)
		return ret;

	if (ent) {
		/* We should never find ourselves straddling an interval */
		if (!ocfs2_extent_rec_contains_clusters(&ent->e_rec,
							v_cpos,
							count))
			return -ESRCH;

		coff = v_cpos - le32_to_cpu(ent->e_rec.e_cpos);
		*p_cpos = ocfs2_blocks_to_clusters(inode->i_sb,
				le64_to_cpu(ent->e_rec.e_blkno)) +
			  coff;

		if (ret_count)
			*ret_count = le32_to_cpu(ent->e_rec.e_clusters) - coff;

		return 0;
	}


	return -ENOENT;
}

int ocfs2_extent_map_get_blocks(struct inode *inode,
				u64 v_blkno, int count,
				u64 *p_blkno, int *ret_count)
{
	int ret;
	u64 boff;
	u32 cpos, clusters;
	int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
	struct ocfs2_extent_map_entry *ent = NULL;
	struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
	struct ocfs2_extent_rec *rec;

	*p_blkno = 0;

	cpos = ocfs2_blocks_to_clusters(inode->i_sb, v_blkno);
	clusters = ocfs2_blocks_to_clusters(inode->i_sb,
					    (u64)count + bpc - 1);
	if ((cpos + clusters) > OCFS2_I(inode)->ip_clusters) {
		ret = -EINVAL;
		mlog_errno(ret);
		return ret;
	}

	if ((cpos + clusters) > em->em_clusters) {
		/*
		 * Size changed underneath us on disk.  Drop any
		 * straddling records and update our idea of
		 * i_clusters
		 */
		ocfs2_extent_map_drop(inode, em->em_clusters - 1);
		em->em_clusters = OCFS2_I(inode)->ip_clusters;
	}

	ret = ocfs2_extent_map_lookup_read(inode, cpos, clusters, &ent);
	if (ret) {
		mlog_errno(ret);
		return ret;
	}

	if (ent)
	{
		rec = &ent->e_rec;

		/* We should never find ourselves straddling an interval */
		if (!ocfs2_extent_rec_contains_clusters(rec, cpos, clusters)) {
			ret = -ESRCH;
			mlog_errno(ret);
			return ret;
		}

		boff = ocfs2_clusters_to_blocks(inode->i_sb, cpos -
						le32_to_cpu(rec->e_cpos));
		boff += (v_blkno & (u64)(bpc - 1));
		*p_blkno = le64_to_cpu(rec->e_blkno) + boff;

		if (ret_count) {
			*ret_count = ocfs2_clusters_to_blocks(inode->i_sb,
					le32_to_cpu(rec->e_clusters)) - boff;
		}

		return 0;
	}

	return -ENOENT;
}

int ocfs2_extent_map_init(struct inode *inode)
{
	struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;

	em->em_extents = RB_ROOT;
	em->em_clusters = 0;

	return 0;
}

/* Needs the lock */
static void __ocfs2_extent_map_drop(struct inode *inode,
				    u32 new_clusters,
				    struct rb_node **free_head,
				    struct ocfs2_extent_map_entry **tail_ent)
{
	struct rb_node *node, *next;
	struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
	struct ocfs2_extent_map_entry *ent;

	*free_head = NULL;

	ent = NULL;
	node = rb_last(&em->em_extents);
	while (node)
	{
		next = rb_prev(node);

		ent = rb_entry(node, struct ocfs2_extent_map_entry,
			       e_node);
		if (le32_to_cpu(ent->e_rec.e_cpos) < new_clusters)
			break;

		rb_erase(&ent->e_node, &em->em_extents);

		node->rb_right = *free_head;
		*free_head = node;

		ent = NULL;
		node = next;
	}

	/* Do we have an entry straddling new_clusters? */
	if (tail_ent) {
		if (ent &&
		    ((le32_to_cpu(ent->e_rec.e_cpos) +
		      le32_to_cpu(ent->e_rec.e_clusters)) > new_clusters))
			*tail_ent = ent;
		else
			*tail_ent = NULL;
	}
}

static void __ocfs2_extent_map_drop_cleanup(struct rb_node *free_head)
{
	struct rb_node *node;
	struct ocfs2_extent_map_entry *ent;

	while (free_head) {
		node = free_head;
		free_head = node->rb_right;

		ent = rb_entry(node, struct ocfs2_extent_map_entry,
			       e_node);
		kmem_cache_free(ocfs2_em_ent_cachep, ent);
	}
}

/*
 * Remove all entries past new_clusters, inclusive of an entry that
 * contains new_clusters.  This is effectively a cache forget.
 *
 * If you want to also clip the last extent by some number of clusters,
 * you need to call ocfs2_extent_map_trunc().
 * This code does not check or modify ip_clusters.
 */
int ocfs2_extent_map_drop(struct inode *inode, u32 new_clusters)
{
	struct rb_node *free_head = NULL;
	struct ocfs2_extent_map *em = &OCFS2_I(inode)->ip_map;
	struct ocfs2_extent_map_entry *ent;

	spin_lock(&OCFS2_I(inode)->ip_lock);

	__ocfs2_extent_map_drop(inode, new_clusters, &free_head, &ent);

	if (ent) {
		rb_erase(&ent->e_node, &em->em_extents);
		ent->e_node.rb_right = free_head;
		free_head = &ent->e_node;
	}

	spin_unlock(&OCFS2_I(inode)->ip_lock);

	if (free_head)
		__ocfs2_extent_map_drop_cleanup(free_head);

	return 0;
}

/*
 * Remove all entries past new_clusters and also clip any extent
 * straddling new_clusters, if there is one.  This does not check
 * or modify ip_clusters
 */
int ocfs2_extent_map_trunc(struct inode *inode, u32 new_clusters)
{
	struct rb_node *free_head = NULL;
	struct ocfs2_extent_map_entry *ent = NULL;

	spin_lock(&OCFS2_I(inode)->ip_lock);

	__ocfs2_extent_map_drop(inode, new_clusters, &free_head, &ent);

	if (ent)
		ent->e_rec.e_clusters = cpu_to_le32(new_clusters -
					       le32_to_cpu(ent->e_rec.e_cpos));

	OCFS2_I(inode)->ip_map.em_clusters = new_clusters;

	spin_unlock(&OCFS2_I(inode)->ip_lock);

	if (free_head)
		__ocfs2_extent_map_drop_cleanup(free_head);

	return 0;
}

int __init init_ocfs2_extent_maps(void)
{
	ocfs2_em_ent_cachep =
		kmem_cache_create("ocfs2_em_ent",
				  sizeof(struct ocfs2_extent_map_entry),
				  0, SLAB_HWCACHE_ALIGN, NULL, NULL);
	if (!ocfs2_em_ent_cachep)
		return -ENOMEM;

	return 0;
}

void __exit exit_ocfs2_extent_maps(void)
{
	kmem_cache_destroy(ocfs2_em_ent_cachep);
}