fs/ocfs2/alloc.c

// SPDX-License-Identifier: GPL-2.0-or-later
/*
 * alloc.c
 *
 * Extent allocs and frees
 *
 * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
 */

#include <linux/fs.h>
#include <linux/types.h>
#include <linux/slab.h>
#include <linux/highmem.h>
#include <linux/swap.h>
#include <linux/quotaops.h>
#include <linux/blkdev.h>
#include <linux/sched/signal.h>

#include <cluster/masklog.h>

#include "ocfs2.h"

#include "alloc.h"
#include "aops.h"
#include "blockcheck.h"
#include "dlmglue.h"
#include "extent_map.h"
#include "inode.h"
#include "journal.h"
#include "localalloc.h"
#include "suballoc.h"
#include "sysfile.h"
#include "file.h"
#include "super.h"
#include "uptodate.h"
#include "xattr.h"
#include "refcounttree.h"
#include "ocfs2_trace.h"

#include "buffer_head_io.h"

enum ocfs2_contig_type {
	CONTIG_NONE = 0,
	CONTIG_LEFT,
	CONTIG_RIGHT,
	CONTIG_LEFTRIGHT,
};

static enum ocfs2_contig_type
	ocfs2_extent_rec_contig(struct super_block *sb,
				struct ocfs2_extent_rec *ext,
				struct ocfs2_extent_rec *insert_rec);
/*
 * Operations for a specific extent tree type.
 *
 * To implement an on-disk btree (extent tree) type in ocfs2, add
 * an ocfs2_extent_tree_operations structure and the matching
 * ocfs2_init_<thingy>_extent_tree() function.  That's pretty much it
 * for the allocation portion of the extent tree.
 */
struct ocfs2_extent_tree_operations {
	/*
	 * last_eb_blk is the block number of the right most leaf extent
	 * block.  Most on-disk structures containing an extent tree store
	 * this value for fast access.  The ->eo_set_last_eb_blk() and
	 * ->eo_get_last_eb_blk() operations access this value.  They are
	 *  both required.
	 */
	void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et,
				   u64 blkno);
	u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et);

	/*
	 * The on-disk structure usually keeps track of how many total
	 * clusters are stored in this extent tree.  This function updates
	 * that value.  new_clusters is the delta, and must be
	 * added to the total.  Required.
	 */
	void (*eo_update_clusters)(struct ocfs2_extent_tree *et,
				   u32 new_clusters);

	/*
	 * If this extent tree is supported by an extent map, insert
	 * a record into the map.
	 */
	void (*eo_extent_map_insert)(struct ocfs2_extent_tree *et,
				     struct ocfs2_extent_rec *rec);

	/*
	 * If this extent tree is supported by an extent map, truncate the
	 * map to clusters,
	 */
	void (*eo_extent_map_truncate)(struct ocfs2_extent_tree *et,
				       u32 clusters);

	/*
	 * If ->eo_insert_check() exists, it is called before rec is
	 * inserted into the extent tree.  It is optional.
	 */
	int (*eo_insert_check)(struct ocfs2_extent_tree *et,
			       struct ocfs2_extent_rec *rec);
	int (*eo_sanity_check)(struct ocfs2_extent_tree *et);

	/*
	 * --------------------------------------------------------------
	 * The remaining are internal to ocfs2_extent_tree and don't have
	 * accessor functions
	 */

	/*
	 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el.
	 * It is required.
	 */
	void (*eo_fill_root_el)(struct ocfs2_extent_tree *et);

	/*
	 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if
	 * it exists.  If it does not, et->et_max_leaf_clusters is set
	 * to 0 (unlimited).  Optional.
	 */
	void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et);

	/*
	 * ->eo_extent_contig test whether the 2 ocfs2_extent_rec
	 * are contiguous or not. Optional. Don't need to set it if use
	 * ocfs2_extent_rec as the tree leaf.
	 */
	enum ocfs2_contig_type
		(*eo_extent_contig)(struct ocfs2_extent_tree *et,
				    struct ocfs2_extent_rec *ext,
				    struct ocfs2_extent_rec *insert_rec);
};


/*
 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check
 * in the methods.
 */
static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et);
static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
					 u64 blkno);
static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
					 u32 clusters);
static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
					   struct ocfs2_extent_rec *rec);
static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
					     u32 clusters);
static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
				     struct ocfs2_extent_rec *rec);
static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et);
static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et);

static int ocfs2_reuse_blk_from_dealloc(handle_t *handle,
					struct ocfs2_extent_tree *et,
					struct buffer_head **new_eb_bh,
					int blk_wanted, int *blk_given);
static int ocfs2_is_dealloc_empty(struct ocfs2_extent_tree *et);

static const struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = {
	.eo_set_last_eb_blk	= ocfs2_dinode_set_last_eb_blk,
	.eo_get_last_eb_blk	= ocfs2_dinode_get_last_eb_blk,
	.eo_update_clusters	= ocfs2_dinode_update_clusters,
	.eo_extent_map_insert	= ocfs2_dinode_extent_map_insert,
	.eo_extent_map_truncate	= ocfs2_dinode_extent_map_truncate,
	.eo_insert_check	= ocfs2_dinode_insert_check,
	.eo_sanity_check	= ocfs2_dinode_sanity_check,
	.eo_fill_root_el	= ocfs2_dinode_fill_root_el,
};

static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et,
					 u64 blkno)
{
	struct ocfs2_dinode *di = et->et_object;

	BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
	di->i_last_eb_blk = cpu_to_le64(blkno);
}

static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et)
{
	struct ocfs2_dinode *di = et->et_object;

	BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
	return le64_to_cpu(di->i_last_eb_blk);
}

static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et,
					 u32 clusters)
{
	struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
	struct ocfs2_dinode *di = et->et_object;

	le32_add_cpu(&di->i_clusters, clusters);
	spin_lock(&oi->ip_lock);
	oi->ip_clusters = le32_to_cpu(di->i_clusters);
	spin_unlock(&oi->ip_lock);
}

static void ocfs2_dinode_extent_map_insert(struct ocfs2_extent_tree *et,
					   struct ocfs2_extent_rec *rec)
{
	struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;

	ocfs2_extent_map_insert_rec(inode, rec);
}

static void ocfs2_dinode_extent_map_truncate(struct ocfs2_extent_tree *et,
					     u32 clusters)
{
	struct inode *inode = &cache_info_to_inode(et->et_ci)->vfs_inode;

	ocfs2_extent_map_trunc(inode, clusters);
}

static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et,
				     struct ocfs2_extent_rec *rec)
{
	struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci);
	struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb);

	BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL);
	mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
			(oi->ip_clusters != le32_to_cpu(rec->e_cpos)),
			"Device %s, asking for sparse allocation: inode %llu, "
			"cpos %u, clusters %u\n",
			osb->dev_str,
			(unsigned long long)oi->ip_blkno,
			rec->e_cpos, oi->ip_clusters);

	return 0;
}

static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et)
{
	struct ocfs2_dinode *di = et->et_object;

	BUG_ON(et->et_ops != &ocfs2_dinode_et_ops);
	BUG_ON(!OCFS2_IS_VALID_DINODE(di));

	return 0;
}

static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et)
{
	struct ocfs2_dinode *di = et->et_object;

	et->et_root_el = &di->id2.i_list;
}


static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et)
{
	struct ocfs2_xattr_value_buf *vb = et->et_object;

	et->et_root_el = &vb->vb_xv->xr_list;
}

static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et,
					      u64 blkno)
{
	struct ocfs2_xattr_value_buf *vb = et->et_object;

	vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno);
}

static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et)
{
	struct ocfs2_xattr_value_buf *vb = et->et_object;

	return le64_to_cpu(vb->vb_xv->xr_last_eb_blk);
}

static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et,
					      u32 clusters)
{
	struct ocfs2_xattr_value_buf *vb = et->et_object;

	le32_add_cpu(&vb->vb_xv->xr_clusters, clusters);
}

static const struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = {
	.eo_set_last_eb_blk	= ocfs2_xattr_value_set_last_eb_blk,
	.eo_get_last_eb_blk	= ocfs2_xattr_value_get_last_eb_blk,
	.eo_update_clusters	= ocfs2_xattr_value_update_clusters,
	.eo_fill_root_el	= ocfs2_xattr_value_fill_root_el,
};

static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et)
{
	struct ocfs2_xattr_block *xb = et->et_object;

	et->et_root_el = &xb->xb_attrs.xb_root.xt_list;
}

static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et)
{
	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);
	et->et_max_leaf_clusters =
		ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE);
}

static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
					     u64 blkno)
{
	struct ocfs2_xattr_block *xb = et->et_object;
	struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;

	xt->xt_last_eb_blk = cpu_to_le64(blkno);
}

static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
{
	struct ocfs2_xattr_block *xb = et->et_object;
	struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root;

	return le64_to_cpu(xt->xt_last_eb_blk);
}

static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et,
					     u32 clusters)
{
	struct ocfs2_xattr_block *xb = et->et_object;

	le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters);
}

static const struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = {
	.eo_set_last_eb_blk	= ocfs2_xattr_tree_set_last_eb_blk,
	.eo_get_last_eb_blk	= ocfs2_xattr_tree_get_last_eb_blk,
	.eo_update_clusters	= ocfs2_xattr_tree_update_clusters,
	.eo_fill_root_el	= ocfs2_xattr_tree_fill_root_el,
	.eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters,
};

static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et,
					  u64 blkno)
{
	struct ocfs2_dx_root_block *dx_root = et->et_object;

	dx_root->dr_last_eb_blk = cpu_to_le64(blkno);
}

static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et)
{
	struct ocfs2_dx_root_block *dx_root = et->et_object;

	return le64_to_cpu(dx_root->dr_last_eb_blk);
}

static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et,
					  u32 clusters)
{
	struct ocfs2_dx_root_block *dx_root = et->et_object;

	le32_add_cpu(&dx_root->dr_clusters, clusters);
}

static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et)
{
	struct ocfs2_dx_root_block *dx_root = et->et_object;

	BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root));

	return 0;
}

static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et)
{
	struct ocfs2_dx_root_block *dx_root = et->et_object;

	et->et_root_el = &dx_root->dr_list;
}

static const struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = {
	.eo_set_last_eb_blk	= ocfs2_dx_root_set_last_eb_blk,
	.eo_get_last_eb_blk	= ocfs2_dx_root_get_last_eb_blk,
	.eo_update_clusters	= ocfs2_dx_root_update_clusters,
	.eo_sanity_check	= ocfs2_dx_root_sanity_check,
	.eo_fill_root_el	= ocfs2_dx_root_fill_root_el,
};

static void ocfs2_refcount_tree_fill_root_el(struct ocfs2_extent_tree *et)
{
	struct ocfs2_refcount_block *rb = et->et_object;

	et->et_root_el = &rb->rf_list;
}

static void ocfs2_refcount_tree_set_last_eb_blk(struct ocfs2_extent_tree *et,
						u64 blkno)
{
	struct ocfs2_refcount_block *rb = et->et_object;

	rb->rf_last_eb_blk = cpu_to_le64(blkno);
}

static u64 ocfs2_refcount_tree_get_last_eb_blk(struct ocfs2_extent_tree *et)
{
	struct ocfs2_refcount_block *rb = et->et_object;

	return le64_to_cpu(rb->rf_last_eb_blk);
}

static void ocfs2_refcount_tree_update_clusters(struct ocfs2_extent_tree *et,
						u32 clusters)
{
	struct ocfs2_refcount_block *rb = et->et_object;

	le32_add_cpu(&rb->rf_clusters, clusters);
}

static enum ocfs2_contig_type
ocfs2_refcount_tree_extent_contig(struct ocfs2_extent_tree *et,
				  struct ocfs2_extent_rec *ext,
				  struct ocfs2_extent_rec *insert_rec)
{
	return CONTIG_NONE;
}

static const struct ocfs2_extent_tree_operations ocfs2_refcount_tree_et_ops = {
	.eo_set_last_eb_blk	= ocfs2_refcount_tree_set_last_eb_blk,
	.eo_get_last_eb_blk	= ocfs2_refcount_tree_get_last_eb_blk,
	.eo_update_clusters	= ocfs2_refcount_tree_update_clusters,
	.eo_fill_root_el	= ocfs2_refcount_tree_fill_root_el,
	.eo_extent_contig	= ocfs2_refcount_tree_extent_contig,
};

static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et,
				     struct ocfs2_caching_info *ci,
				     struct buffer_head *bh,
				     ocfs2_journal_access_func access,
				     void *obj,
				     const struct ocfs2_extent_tree_operations *ops)
{
	et->et_ops = ops;
	et->et_root_bh = bh;
	et->et_ci = ci;
	et->et_root_journal_access = access;
	if (!obj)
		obj = (void *)bh->b_data;
	et->et_object = obj;
	et->et_dealloc = NULL;

	et->et_ops->eo_fill_root_el(et);
	if (!et->et_ops->eo_fill_max_leaf_clusters)
		et->et_max_leaf_clusters = 0;
	else
		et->et_ops->eo_fill_max_leaf_clusters(et);
}

void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et,
				   struct ocfs2_caching_info *ci,
				   struct buffer_head *bh)
{
	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_di,
				 NULL, &ocfs2_dinode_et_ops);
}

void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et,
				       struct ocfs2_caching_info *ci,
				       struct buffer_head *bh)
{
	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_xb,
				 NULL, &ocfs2_xattr_tree_et_ops);
}

void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et,
					struct ocfs2_caching_info *ci,
					struct ocfs2_xattr_value_buf *vb)
{
	__ocfs2_init_extent_tree(et, ci, vb->vb_bh, vb->vb_access, vb,
				 &ocfs2_xattr_value_et_ops);
}

void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et,
				    struct ocfs2_caching_info *ci,
				    struct buffer_head *bh)
{
	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_dr,
				 NULL, &ocfs2_dx_root_et_ops);
}

void ocfs2_init_refcount_extent_tree(struct ocfs2_extent_tree *et,
				     struct ocfs2_caching_info *ci,
				     struct buffer_head *bh)
{
	__ocfs2_init_extent_tree(et, ci, bh, ocfs2_journal_access_rb,
				 NULL, &ocfs2_refcount_tree_et_ops);
}

static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et,
					    u64 new_last_eb_blk)
{
	et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk);
}

static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et)
{
	return et->et_ops->eo_get_last_eb_blk(et);
}

static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et,
					    u32 clusters)
{
	et->et_ops->eo_update_clusters(et, clusters);
}

static inline void ocfs2_et_extent_map_insert(struct ocfs2_extent_tree *et,
					      struct ocfs2_extent_rec *rec)
{
	if (et->et_ops->eo_extent_map_insert)
		et->et_ops->eo_extent_map_insert(et, rec);
}

static inline void ocfs2_et_extent_map_truncate(struct ocfs2_extent_tree *et,
						u32 clusters)
{
	if (et->et_ops->eo_extent_map_truncate)
		et->et_ops->eo_extent_map_truncate(et, clusters);
}

static inline int ocfs2_et_root_journal_access(handle_t *handle,
					       struct ocfs2_extent_tree *et,
					       int type)
{
	return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh,
					  type);
}

static inline enum ocfs2_contig_type
	ocfs2_et_extent_contig(struct ocfs2_extent_tree *et,
			       struct ocfs2_extent_rec *rec,
			       struct ocfs2_extent_rec *insert_rec)
{
	if (et->et_ops->eo_extent_contig)
		return et->et_ops->eo_extent_contig(et, rec, insert_rec);

	return ocfs2_extent_rec_contig(
				ocfs2_metadata_cache_get_super(et->et_ci),
				rec, insert_rec);
}

static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et,
					struct ocfs2_extent_rec *rec)
{
	int ret = 0;

	if (et->et_ops->eo_insert_check)
		ret = et->et_ops->eo_insert_check(et, rec);
	return ret;
}

static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et)
{
	int ret = 0;

	if (et->et_ops->eo_sanity_check)
		ret = et->et_ops->eo_sanity_check(et);
	return ret;
}

static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt,
					 struct ocfs2_extent_block *eb);
static void ocfs2_adjust_rightmost_records(handle_t *handle,
					   struct ocfs2_extent_tree *et,
					   struct ocfs2_path *path,
					   struct ocfs2_extent_rec *insert_rec);
/*
 * Reset the actual path elements so that we can re-use the structure
 * to build another path. Generally, this involves freeing the buffer
 * heads.
 */
void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root)
{
	int i, start = 0, depth = 0;
	struct ocfs2_path_item *node;

	if (keep_root)
		start = 1;

	for(i = start; i < path_num_items(path); i++) {
		node = &path->p_node[i];

		brelse(node->bh);
		node->bh = NULL;
		node->el = NULL;
	}

	/*
	 * Tree depth may change during truncate, or insert. If we're
	 * keeping the root extent list, then make sure that our path
	 * structure reflects the proper depth.
	 */
	if (keep_root)
		depth = le16_to_cpu(path_root_el(path)->l_tree_depth);
	else
		path_root_access(path) = NULL;

	path->p_tree_depth = depth;
}

void ocfs2_free_path(struct ocfs2_path *path)
{
	if (path) {
		ocfs2_reinit_path(path, 0);
		kfree(path);
	}
}

/*
 * All the elements of src into dest. After this call, src could be freed
 * without affecting dest.
 *
 * Both paths should have the same root. Any non-root elements of dest
 * will be freed.
 */
static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src)
{
	int i;

	BUG_ON(path_root_bh(dest) != path_root_bh(src));
	BUG_ON(path_root_el(dest) != path_root_el(src));
	BUG_ON(path_root_access(dest) != path_root_access(src));

	ocfs2_reinit_path(dest, 1);

	for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
		dest->p_node[i].bh = src->p_node[i].bh;
		dest->p_node[i].el = src->p_node[i].el;

		if (dest->p_node[i].bh)
			get_bh(dest->p_node[i].bh);
	}
}

/*
 * Make the *dest path the same as src and re-initialize src path to
 * have a root only.
 */
static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src)
{
	int i;

	BUG_ON(path_root_bh(dest) != path_root_bh(src));
	BUG_ON(path_root_access(dest) != path_root_access(src));

	for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) {
		brelse(dest->p_node[i].bh);

		dest->p_node[i].bh = src->p_node[i].bh;
		dest->p_node[i].el = src->p_node[i].el;

		src->p_node[i].bh = NULL;
		src->p_node[i].el = NULL;
	}
}

/*
 * Insert an extent block at given index.
 *
 * This will not take an additional reference on eb_bh.
 */
static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index,
					struct buffer_head *eb_bh)
{
	struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data;

	/*
	 * Right now, no root bh is an extent block, so this helps
	 * catch code errors with dinode trees. The assertion can be
	 * safely removed if we ever need to insert extent block
	 * structures at the root.
	 */
	BUG_ON(index == 0);

	path->p_node[index].bh = eb_bh;
	path->p_node[index].el = &eb->h_list;
}

static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh,
					 struct ocfs2_extent_list *root_el,
					 ocfs2_journal_access_func access)
{
	struct ocfs2_path *path;

	BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH);

	path = kzalloc(sizeof(*path), GFP_NOFS);
	if (path) {
		path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth);
		get_bh(root_bh);
		path_root_bh(path) = root_bh;
		path_root_el(path) = root_el;
		path_root_access(path) = access;
	}

	return path;
}

struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path)
{
	return ocfs2_new_path(path_root_bh(path), path_root_el(path),
			      path_root_access(path));
}

struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et)
{
	return ocfs2_new_path(et->et_root_bh, et->et_root_el,
			      et->et_root_journal_access);
}

/*
 * Journal the buffer at depth idx.  All idx>0 are extent_blocks,
 * otherwise it's the root_access function.
 *
 * I don't like the way this function's name looks next to
 * ocfs2_journal_access_path(), but I don't have a better one.
 */
int ocfs2_path_bh_journal_access(handle_t *handle,
				 struct ocfs2_caching_info *ci,
				 struct ocfs2_path *path,
				 int idx)
{
	ocfs2_journal_access_func access = path_root_access(path);

	if (!access)
		access = ocfs2_journal_access;

	if (idx)
		access = ocfs2_journal_access_eb;

	return access(handle, ci, path->p_node[idx].bh,
		      OCFS2_JOURNAL_ACCESS_WRITE);
}

/*
 * Convenience function to journal all components in a path.
 */
int ocfs2_journal_access_path(struct ocfs2_caching_info *ci,
			      handle_t *handle,
			      struct ocfs2_path *path)
{
	int i, ret = 0;

	if (!path)
		goto out;

	for(i = 0; i < path_num_items(path); i++) {
		ret = ocfs2_path_bh_journal_access(handle, ci, path, i);
		if (ret < 0) {
			mlog_errno(ret);
			goto out;
		}
	}

out:
	return ret;
}

/*
 * Return the index of the extent record which contains cluster #v_cluster.
 * -1 is returned if it was not found.
 *
 * Should work fine on interior and exterior nodes.
 */
int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster)
{
	int ret = -1;
	int i;
	struct ocfs2_extent_rec *rec;
	u32 rec_end, rec_start, clusters;

	for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
		rec = &el->l_recs[i];

		rec_start = le32_to_cpu(rec->e_cpos);
		clusters = ocfs2_rec_clusters(el, rec);

		rec_end = rec_start + clusters;

		if (v_cluster >= rec_start && v_cluster < rec_end) {
			ret = i;
			break;
		}
	}

	return ret;
}

/*
 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and
 * ocfs2_extent_rec_contig only work properly against leaf nodes!
 */
static int ocfs2_block_extent_contig(struct super_block *sb,
				     struct ocfs2_extent_rec *ext,
				     u64 blkno)
{
	u64 blk_end = le64_to_cpu(ext->e_blkno);

	blk_end += ocfs2_clusters_to_blocks(sb,
				    le16_to_cpu(ext->e_leaf_clusters));

	return blkno == blk_end;
}

static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
				  struct ocfs2_extent_rec *right)
{
	u32 left_range;

	left_range = le32_to_cpu(left->e_cpos) +
		le16_to_cpu(left->e_leaf_clusters);

	return (left_range == le32_to_cpu(right->e_cpos));
}

static enum ocfs2_contig_type
	ocfs2_extent_rec_contig(struct super_block *sb,
				struct ocfs2_extent_rec *ext,
				struct ocfs2_extent_rec *insert_rec)
{
	u64 blkno = le64_to_cpu(insert_rec->e_blkno);

	/*
	 * Refuse to coalesce extent records with different flag
	 * fields - we don't want to mix unwritten extents with user
	 * data.
	 */
	if (ext->e_flags != insert_rec->e_flags)
		return CONTIG_NONE;

	if (ocfs2_extents_adjacent(ext, insert_rec) &&
	    ocfs2_block_extent_contig(sb, ext, blkno))
			return CONTIG_RIGHT;

	blkno = le64_to_cpu(ext->e_blkno);
	if (ocfs2_extents_adjacent(insert_rec, ext) &&
	    ocfs2_block_extent_contig(sb, insert_rec, blkno))
		return CONTIG_LEFT;

	return CONTIG_NONE;
}

/*
 * NOTE: We can have pretty much any combination of contiguousness and
 * appending.
 *
 * The usefulness of APPEND_TAIL is more in that it lets us know that
 * we'll have to update the path to that leaf.
 */
enum ocfs2_append_type {
	APPEND_NONE = 0,
	APPEND_TAIL,
};

enum ocfs2_split_type {
	SPLIT_NONE = 0,
	SPLIT_LEFT,
	SPLIT_RIGHT,
};

struct ocfs2_insert_type {
	enum ocfs2_split_type	ins_split;
	enum ocfs2_append_type	ins_appending;
	enum ocfs2_contig_type	ins_contig;
	int			ins_contig_index;
	int			ins_tree_depth;
};

struct ocfs2_merge_ctxt {
	enum ocfs2_contig_type	c_contig_type;
	int			c_has_empty_extent;
	int			c_split_covers_rec;
};

static int ocfs2_validate_extent_block(struct super_block *sb,
				       struct buffer_head *bh)
{
	int rc;
	struct ocfs2_extent_block *eb =
		(struct ocfs2_extent_block *)bh->b_data;

	trace_ocfs2_validate_extent_block((unsigned long long)bh->b_blocknr);

	BUG_ON(!buffer_uptodate(bh));

	/*
	 * If the ecc fails, we return the error but otherwise
	 * leave the filesystem running.  We know any error is
	 * local to this block.
	 */
	rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check);
	if (rc) {
		mlog(ML_ERROR, "Checksum failed for extent block %llu\n",
		     (unsigned long long)bh->b_blocknr);
		return rc;
	}

	/*
	 * Errors after here are fatal.
	 */

	if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
		rc = ocfs2_error(sb,
				 "Extent block #%llu has bad signature %.*s\n",
				 (unsigned long long)bh->b_blocknr, 7,
				 eb->h_signature);
		goto bail;
	}

	if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) {
		rc = ocfs2_error(sb,
				 "Extent block #%llu has an invalid h_blkno of %llu\n",
				 (unsigned long long)bh->b_blocknr,
				 (unsigned long long)le64_to_cpu(eb->h_blkno));
		goto bail;
	}

	if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation)
		rc = ocfs2_error(sb,
				 "Extent block #%llu has an invalid h_fs_generation of #%u\n",
				 (unsigned long long)bh->b_blocknr,
				 le32_to_cpu(eb->h_fs_generation));
bail:
	return rc;
}

int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno,
			    struct buffer_head **bh)
{
	int rc;
	struct buffer_head *tmp = *bh;

	rc = ocfs2_read_block(ci, eb_blkno, &tmp,
			      ocfs2_validate_extent_block);

	/* If ocfs2_read_block() got us a new bh, pass it up. */
	if (!rc && !*bh)
		*bh = tmp;

	return rc;
}


/*
 * How many free extents have we got before we need more meta data?
 */
int ocfs2_num_free_extents(struct ocfs2_extent_tree *et)
{
	int retval;
	struct ocfs2_extent_list *el = NULL;
	struct ocfs2_extent_block *eb;
	struct buffer_head *eb_bh = NULL;
	u64 last_eb_blk = 0;

	el = et->et_root_el;
	last_eb_blk = ocfs2_et_get_last_eb_blk(et);

	if (last_eb_blk) {
		retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk,
						 &eb_bh);
		if (retval < 0) {
			mlog_errno(retval);
			goto bail;
		}
		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
		el = &eb->h_list;
	}

	BUG_ON(el->l_tree_depth != 0);

	retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec);
bail:
	brelse(eb_bh);

	trace_ocfs2_num_free_extents(retval);
	return retval;
}

/* expects array to already be allocated
 *
 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and
 * l_count for you
 */
static int ocfs2_create_new_meta_bhs(handle_t *handle,
				     struct ocfs2_extent_tree *et,
				     int wanted,
				     struct ocfs2_alloc_context *meta_ac,
				     struct buffer_head *bhs[])
{
	int count, status, i;
	u16 suballoc_bit_start;
	u32 num_got;
	u64 suballoc_loc, first_blkno;
	struct ocfs2_super *osb =
		OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci));
	struct ocfs2_extent_block *eb;

	count = 0;
	while (count < wanted) {
		status = ocfs2_claim_metadata(handle,
					      meta_ac,
					      wanted - count,
					      &suballoc_loc,
					      &suballoc_bit_start,
					      &num_got,
					      &first_blkno);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}

		for(i = count;  i < (num_got + count); i++) {
			bhs[i] = sb_getblk(osb->sb, first_blkno);
			if (bhs[i] == NULL) {
				status = -ENOMEM;
				mlog_errno(status);
				goto bail;
			}
			ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]);

			status = ocfs2_journal_access_eb(handle, et->et_ci,
							 bhs[i],
							 OCFS2_JOURNAL_ACCESS_CREATE);
			if (status < 0) {
				mlog_errno(status);
				goto bail;
			}

			memset(bhs[i]->b_data, 0, osb->sb->s_blocksize);
			eb = (struct ocfs2_extent_block *) bhs[i]->b_data;
			/* Ok, setup the minimal stuff here. */
			strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE);
			eb->h_blkno = cpu_to_le64(first_blkno);
			eb->h_fs_generation = cpu_to_le32(osb->fs_generation);
			eb->h_suballoc_slot =
				cpu_to_le16(meta_ac->ac_alloc_slot);
			eb->h_suballoc_loc = cpu_to_le64(suballoc_loc);
			eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start);
			eb->h_list.l_count =
				cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb));

			suballoc_bit_start++;
			first_blkno++;

			/* We'll also be dirtied by the caller, so
			 * this isn't absolutely necessary. */
			ocfs2_journal_dirty(handle, bhs[i]);
		}

		count += num_got;
	}

	status = 0;
bail:
	if (status < 0) {
		for(i = 0; i < wanted; i++) {
			brelse(bhs[i]);
			bhs[i] = NULL;
		}
	}
	return status;
}

/*
 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth().
 *
 * Returns the sum of the rightmost extent rec logical offset and
 * cluster count.
 *
 * ocfs2_add_branch() uses this to determine what logical cluster
 * value should be populated into the leftmost new branch records.
 *
 * ocfs2_shift_tree_depth() uses this to determine the # clusters
 * value for the new topmost tree record.
 */
static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list  *el)
{
	int i;

	i = le16_to_cpu(el->l_next_free_rec) - 1;

	return le32_to_cpu(el->l_recs[i].e_cpos) +
		ocfs2_rec_clusters(el, &el->l_recs[i]);
}

/*
 * Change range of the branches in the right most path according to the leaf
 * extent block's rightmost record.
 */
static int ocfs2_adjust_rightmost_branch(handle_t *handle,
					 struct ocfs2_extent_tree *et)
{
	int status;
	struct ocfs2_path *path = NULL;
	struct ocfs2_extent_list *el;
	struct ocfs2_extent_rec *rec;

	path = ocfs2_new_path_from_et(et);
	if (!path) {
		status = -ENOMEM;
		return status;
	}

	status = ocfs2_find_path(et->et_ci, path, UINT_MAX);
	if (status < 0) {
		mlog_errno(status);
		goto out;
	}

	status = ocfs2_extend_trans(handle, path_num_items(path));
	if (status < 0) {
		mlog_errno(status);
		goto out;
	}

	status = ocfs2_journal_access_path(et->et_ci, handle, path);
	if (status < 0) {
		mlog_errno(status);
		goto out;
	}

	el = path_leaf_el(path);
	rec = &el->l_recs[le16_to_cpu(el->l_next_free_rec) - 1];

	ocfs2_adjust_rightmost_records(handle, et, path, rec);

out:
	ocfs2_free_path(path);
	return status;
}

/*
 * Add an entire tree branch to our inode. eb_bh is the extent block
 * to start at, if we don't want to start the branch at the root
 * structure.
 *
 * last_eb_bh is required as we have to update it's next_leaf pointer
 * for the new last extent block.
 *
 * the new branch will be 'empty' in the sense that every block will
 * contain a single record with cluster count == 0.
 */
static int ocfs2_add_branch(handle_t *handle,
			    struct ocfs2_extent_tree *et,
			    struct buffer_head *eb_bh,
			    struct buffer_head **last_eb_bh,
			    struct ocfs2_alloc_context *meta_ac)
{
	int status, new_blocks, i, block_given = 0;
	u64 next_blkno, new_last_eb_blk;
	struct buffer_head *bh;
	struct buffer_head **new_eb_bhs = NULL;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list  *eb_el;
	struct ocfs2_extent_list  *el;
	u32 new_cpos, root_end;

	BUG_ON(!last_eb_bh || !*last_eb_bh);

	if (eb_bh) {
		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
		el = &eb->h_list;
	} else
		el = et->et_root_el;

	/* we never add a branch to a leaf. */
	BUG_ON(!el->l_tree_depth);

	new_blocks = le16_to_cpu(el->l_tree_depth);

	eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data;
	new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list);
	root_end = ocfs2_sum_rightmost_rec(et->et_root_el);

	/*
	 * If there is a gap before the root end and the real end
	 * of the righmost leaf block, we need to remove the gap
	 * between new_cpos and root_end first so that the tree
	 * is consistent after we add a new branch(it will start
	 * from new_cpos).
	 */
	if (root_end > new_cpos) {
		trace_ocfs2_adjust_rightmost_branch(
			(unsigned long long)
			ocfs2_metadata_cache_owner(et->et_ci),
			root_end, new_cpos);

		status = ocfs2_adjust_rightmost_branch(handle, et);
		if (status) {
			mlog_errno(status);
			goto bail;
		}
	}

	/* allocate the number of new eb blocks we need */
	new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *),
			     GFP_KERNEL);
	if (!new_eb_bhs) {
		status = -ENOMEM;
		mlog_errno(status);
		goto bail;
	}

	/* Firstyly, try to reuse dealloc since we have already estimated how
	 * many extent blocks we may use.
	 */
	if (!ocfs2_is_dealloc_empty(et)) {
		status = ocfs2_reuse_blk_from_dealloc(handle, et,
						      new_eb_bhs, new_blocks,
						      &block_given);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}
	}

	BUG_ON(block_given > new_blocks);

	if (block_given < new_blocks) {
		BUG_ON(!meta_ac);
		status = ocfs2_create_new_meta_bhs(handle, et,
						   new_blocks - block_given,
						   meta_ac,
						   &new_eb_bhs[block_given]);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}
	}

	/* Note: new_eb_bhs[new_blocks - 1] is the guy which will be
	 * linked with the rest of the tree.
	 * conversly, new_eb_bhs[0] is the new bottommost leaf.
	 *
	 * when we leave the loop, new_last_eb_blk will point to the
	 * newest leaf, and next_blkno will point to the topmost extent
	 * block. */
	next_blkno = new_last_eb_blk = 0;
	for(i = 0; i < new_blocks; i++) {
		bh = new_eb_bhs[i];
		eb = (struct ocfs2_extent_block *) bh->b_data;
		/* ocfs2_create_new_meta_bhs() should create it right! */
		BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));
		eb_el = &eb->h_list;

		status = ocfs2_journal_access_eb(handle, et->et_ci, bh,
						 OCFS2_JOURNAL_ACCESS_CREATE);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}

		eb->h_next_leaf_blk = 0;
		eb_el->l_tree_depth = cpu_to_le16(i);
		eb_el->l_next_free_rec = cpu_to_le16(1);
		/*
		 * This actually counts as an empty extent as
		 * c_clusters == 0
		 */
		eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos);
		eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno);
		/*
		 * eb_el isn't always an interior node, but even leaf
		 * nodes want a zero'd flags and reserved field so
		 * this gets the whole 32 bits regardless of use.
		 */
		eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0);
		if (!eb_el->l_tree_depth)
			new_last_eb_blk = le64_to_cpu(eb->h_blkno);

		ocfs2_journal_dirty(handle, bh);
		next_blkno = le64_to_cpu(eb->h_blkno);
	}

	/* This is a bit hairy. We want to update up to three blocks
	 * here without leaving any of them in an inconsistent state
	 * in case of error. We don't have to worry about
	 * journal_dirty erroring as it won't unless we've aborted the
	 * handle (in which case we would never be here) so reserving
	 * the write with journal_access is all we need to do. */
	status = ocfs2_journal_access_eb(handle, et->et_ci, *last_eb_bh,
					 OCFS2_JOURNAL_ACCESS_WRITE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}
	status = ocfs2_et_root_journal_access(handle, et,
					      OCFS2_JOURNAL_ACCESS_WRITE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}
	if (eb_bh) {
		status = ocfs2_journal_access_eb(handle, et->et_ci, eb_bh,
						 OCFS2_JOURNAL_ACCESS_WRITE);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}
	}

	/* Link the new branch into the rest of the tree (el will
	 * either be on the root_bh, or the extent block passed in. */
	i = le16_to_cpu(el->l_next_free_rec);
	el->l_recs[i].e_blkno = cpu_to_le64(next_blkno);
	el->l_recs[i].e_cpos = cpu_to_le32(new_cpos);
	el->l_recs[i].e_int_clusters = 0;
	le16_add_cpu(&el->l_next_free_rec, 1);

	/* fe needs a new last extent block pointer, as does the
	 * next_leaf on the previously last-extent-block. */
	ocfs2_et_set_last_eb_blk(et, new_last_eb_blk);

	eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data;
	eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk);

	ocfs2_journal_dirty(handle, *last_eb_bh);
	ocfs2_journal_dirty(handle, et->et_root_bh);
	if (eb_bh)
		ocfs2_journal_dirty(handle, eb_bh);

	/*
	 * Some callers want to track the rightmost leaf so pass it
	 * back here.
	 */
	brelse(*last_eb_bh);
	get_bh(new_eb_bhs[0]);
	*last_eb_bh = new_eb_bhs[0];

	status = 0;
bail:
	if (new_eb_bhs) {
		for (i = 0; i < new_blocks; i++)
			brelse(new_eb_bhs[i]);
		kfree(new_eb_bhs);
	}

	return status;
}

/*
 * adds another level to the allocation tree.
 * returns back the new extent block so you can add a branch to it
 * after this call.
 */
static int ocfs2_shift_tree_depth(handle_t *handle,
				  struct ocfs2_extent_tree *et,
				  struct ocfs2_alloc_context *meta_ac,
				  struct buffer_head **ret_new_eb_bh)
{
	int status, i, block_given = 0;
	u32 new_clusters;
	struct buffer_head *new_eb_bh = NULL;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list  *root_el;
	struct ocfs2_extent_list  *eb_el;

	if (!ocfs2_is_dealloc_empty(et)) {
		status = ocfs2_reuse_blk_from_dealloc(handle, et,
						      &new_eb_bh, 1,
						      &block_given);
	} else if (meta_ac) {
		status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac,
						   &new_eb_bh);

	} else {
		BUG();
	}

	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
	/* ocfs2_create_new_meta_bhs() should create it right! */
	BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb));

	eb_el = &eb->h_list;
	root_el = et->et_root_el;

	status = ocfs2_journal_access_eb(handle, et->et_ci, new_eb_bh,
					 OCFS2_JOURNAL_ACCESS_CREATE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	/* copy the root extent list data into the new extent block */
	eb_el->l_tree_depth = root_el->l_tree_depth;
	eb_el->l_next_free_rec = root_el->l_next_free_rec;
	for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++)
		eb_el->l_recs[i] = root_el->l_recs[i];

	ocfs2_journal_dirty(handle, new_eb_bh);

	status = ocfs2_et_root_journal_access(handle, et,
					      OCFS2_JOURNAL_ACCESS_WRITE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	new_clusters = ocfs2_sum_rightmost_rec(eb_el);

	/* update root_bh now */
	le16_add_cpu(&root_el->l_tree_depth, 1);
	root_el->l_recs[0].e_cpos = 0;
	root_el->l_recs[0].e_blkno = eb->h_blkno;
	root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters);
	for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
		memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
	root_el->l_next_free_rec = cpu_to_le16(1);

	/* If this is our 1st tree depth shift, then last_eb_blk
	 * becomes the allocated extent block */
	if (root_el->l_tree_depth == cpu_to_le16(1))
		ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));

	ocfs2_journal_dirty(handle, et->et_root_bh);

	*ret_new_eb_bh = new_eb_bh;
	new_eb_bh = NULL;
	status = 0;
bail:
	brelse(new_eb_bh);

	return status;
}

/*
 * Should only be called when there is no space left in any of the
 * leaf nodes. What we want to do is find the lowest tree depth
 * non-leaf extent block with room for new records. There are three
 * valid results of this search:
 *
 * 1) a lowest extent block is found, then we pass it back in
 *    *lowest_eb_bh and return '0'
 *
 * 2) the search fails to find anything, but the root_el has room. We
 *    pass NULL back in *lowest_eb_bh, but still return '0'
 *
 * 3) the search fails to find anything AND the root_el is full, in
 *    which case we return > 0
 *
 * return status < 0 indicates an error.
 */
static int ocfs2_find_branch_target(struct ocfs2_extent_tree *et,
				    struct buffer_head **target_bh)
{
	int status = 0, i;
	u64 blkno;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list  *el;
	struct buffer_head *bh = NULL;
	struct buffer_head *lowest_bh = NULL;

	*target_bh = NULL;

	el = et->et_root_el;

	while(le16_to_cpu(el->l_tree_depth) > 1) {
		if (le16_to_cpu(el->l_next_free_rec) == 0) {
			status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
					"Owner %llu has empty extent list (next_free_rec == 0)\n",
					(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci));
			goto bail;
		}
		i = le16_to_cpu(el->l_next_free_rec) - 1;
		blkno = le64_to_cpu(el->l_recs[i].e_blkno);
		if (!blkno) {
			status = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
					"Owner %llu has extent list where extent # %d has no physical block start\n",
					(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i);
			goto bail;
		}

		brelse(bh);
		bh = NULL;

		status = ocfs2_read_extent_block(et->et_ci, blkno, &bh);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}

		eb = (struct ocfs2_extent_block *) bh->b_data;
		el = &eb->h_list;

		if (le16_to_cpu(el->l_next_free_rec) <
		    le16_to_cpu(el->l_count)) {
			brelse(lowest_bh);
			lowest_bh = bh;
			get_bh(lowest_bh);
		}
	}

	/* If we didn't find one and the fe doesn't have any room,
	 * then return '1' */
	el = et->et_root_el;
	if (!lowest_bh && (el->l_next_free_rec == el->l_count))
		status = 1;

	*target_bh = lowest_bh;
bail:
	brelse(bh);

	return status;
}

/*
 * Grow a b-tree so that it has more records.
 *
 * We might shift the tree depth in which case existing paths should
 * be considered invalid.
 *
 * Tree depth after the grow is returned via *final_depth.
 *
 * *last_eb_bh will be updated by ocfs2_add_branch().
 */
static int ocfs2_grow_tree(handle_t *handle, struct ocfs2_extent_tree *et,
			   int *final_depth, struct buffer_head **last_eb_bh,
			   struct ocfs2_alloc_context *meta_ac)
{
	int ret, shift;
	struct ocfs2_extent_list *el = et->et_root_el;
	int depth = le16_to_cpu(el->l_tree_depth);
	struct buffer_head *bh = NULL;

	BUG_ON(meta_ac == NULL && ocfs2_is_dealloc_empty(et));

	shift = ocfs2_find_branch_target(et, &bh);
	if (shift < 0) {
		ret = shift;
		mlog_errno(ret);
		goto out;
	}

	/* We traveled all the way to the bottom of the allocation tree
	 * and didn't find room for any more extents - we need to add
	 * another tree level */
	if (shift) {
		BUG_ON(bh);
		trace_ocfs2_grow_tree(
			(unsigned long long)
			ocfs2_metadata_cache_owner(et->et_ci),
			depth);

		/* ocfs2_shift_tree_depth will return us a buffer with
		 * the new extent block (so we can pass that to
		 * ocfs2_add_branch). */
		ret = ocfs2_shift_tree_depth(handle, et, meta_ac, &bh);
		if (ret < 0) {
			mlog_errno(ret);
			goto out;
		}
		depth++;
		if (depth == 1) {
			/*
			 * Special case: we have room now if we shifted from
			 * tree_depth 0, so no more work needs to be done.
			 *
			 * We won't be calling add_branch, so pass
			 * back *last_eb_bh as the new leaf. At depth
			 * zero, it should always be null so there's
			 * no reason to brelse.
			 */
			BUG_ON(*last_eb_bh);
			get_bh(bh);
			*last_eb_bh = bh;
			goto out;
		}
	}

	/* call ocfs2_add_branch to add the final part of the tree with
	 * the new data. */
	ret = ocfs2_add_branch(handle, et, bh, last_eb_bh,
			       meta_ac);
	if (ret < 0)
		mlog_errno(ret);

out:
	if (final_depth)
		*final_depth = depth;
	brelse(bh);
	return ret;
}

/*
 * This function will discard the rightmost extent record.
 */
static void ocfs2_shift_records_right(struct ocfs2_extent_list *el)
{
	int next_free = le16_to_cpu(el->l_next_free_rec);
	int count = le16_to_cpu(el->l_count);
	unsigned int num_bytes;

	BUG_ON(!next_free);
	/* This will cause us to go off the end of our extent list. */
	BUG_ON(next_free >= count);

	num_bytes = sizeof(struct ocfs2_extent_rec) * next_free;

	memmove(&el->l_recs[1], &el->l_recs[0], num_bytes);
}

static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el,
			      struct ocfs2_extent_rec *insert_rec)
{
	int i, insert_index, next_free, has_empty, num_bytes;
	u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos);
	struct ocfs2_extent_rec *rec;

	next_free = le16_to_cpu(el->l_next_free_rec);
	has_empty = ocfs2_is_empty_extent(&el->l_recs[0]);

	BUG_ON(!next_free);

	/* The tree code before us didn't allow enough room in the leaf. */
	BUG_ON(el->l_next_free_rec == el->l_count && !has_empty);

	/*
	 * The easiest way to approach this is to just remove the
	 * empty extent and temporarily decrement next_free.
	 */
	if (has_empty) {
		/*
		 * If next_free was 1 (only an empty extent), this
		 * loop won't execute, which is fine. We still want
		 * the decrement above to happen.
		 */
		for(i = 0; i < (next_free - 1); i++)
			el->l_recs[i] = el->l_recs[i+1];

		next_free--;
	}

	/*
	 * Figure out what the new record index should be.
	 */
	for(i = 0; i < next_free; i++) {
		rec = &el->l_recs[i];

		if (insert_cpos < le32_to_cpu(rec->e_cpos))
			break;
	}
	insert_index = i;

	trace_ocfs2_rotate_leaf(insert_cpos, insert_index,
				has_empty, next_free,
				le16_to_cpu(el->l_count));

	BUG_ON(insert_index < 0);
	BUG_ON(insert_index >= le16_to_cpu(el->l_count));
	BUG_ON(insert_index > next_free);

	/*
	 * No need to memmove if we're just adding to the tail.
	 */
	if (insert_index != next_free) {
		BUG_ON(next_free >= le16_to_cpu(el->l_count));

		num_bytes = next_free - insert_index;
		num_bytes *= sizeof(struct ocfs2_extent_rec);
		memmove(&el->l_recs[insert_index + 1],
			&el->l_recs[insert_index],
			num_bytes);
	}

	/*
	 * Either we had an empty extent, and need to re-increment or
	 * there was no empty extent on a non full rightmost leaf node,
	 * in which case we still need to increment.
	 */
	next_free++;
	el->l_next_free_rec = cpu_to_le16(next_free);
	/*
	 * Make sure none of the math above just messed up our tree.
	 */
	BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count));

	el->l_recs[insert_index] = *insert_rec;

}

static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el)
{
	int size, num_recs = le16_to_cpu(el->l_next_free_rec);

	BUG_ON(num_recs == 0);

	if (ocfs2_is_empty_extent(&el->l_recs[0])) {
		num_recs--;
		size = num_recs * sizeof(struct ocfs2_extent_rec);
		memmove(&el->l_recs[0], &el->l_recs[1], size);
		memset(&el->l_recs[num_recs], 0,
		       sizeof(struct ocfs2_extent_rec));
		el->l_next_free_rec = cpu_to_le16(num_recs);
	}
}

/*
 * Create an empty extent record .
 *
 * l_next_free_rec may be updated.
 *
 * If an empty extent already exists do nothing.
 */
static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el)
{
	int next_free = le16_to_cpu(el->l_next_free_rec);

	BUG_ON(le16_to_cpu(el->l_tree_depth) != 0);

	if (next_free == 0)
		goto set_and_inc;

	if (ocfs2_is_empty_extent(&el->l_recs[0]))
		return;

	mlog_bug_on_msg(el->l_count == el->l_next_free_rec,
			"Asked to create an empty extent in a full list:\n"
			"count = %u, tree depth = %u",
			le16_to_cpu(el->l_count),
			le16_to_cpu(el->l_tree_depth));

	ocfs2_shift_records_right(el);

set_and_inc:
	le16_add_cpu(&el->l_next_free_rec, 1);
	memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
}

/*
 * For a rotation which involves two leaf nodes, the "root node" is
 * the lowest level tree node which contains a path to both leafs. This
 * resulting set of information can be used to form a complete "subtree"
 *
 * This function is passed two full paths from the dinode down to a
 * pair of adjacent leaves. It's task is to figure out which path
 * index contains the subtree root - this can be the root index itself
 * in a worst-case rotation.
 *
 * The array index of the subtree root is passed back.
 */
int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et,
			    struct ocfs2_path *left,
			    struct ocfs2_path *right)
{
	int i = 0;

	/*
	 * Check that the caller passed in two paths from the same tree.
	 */
	BUG_ON(path_root_bh(left) != path_root_bh(right));

	do {
		i++;

		/*
		 * The caller didn't pass two adjacent paths.
		 */
		mlog_bug_on_msg(i > left->p_tree_depth,
				"Owner %llu, left depth %u, right depth %u\n"
				"left leaf blk %llu, right leaf blk %llu\n",
				(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
				left->p_tree_depth, right->p_tree_depth,
				(unsigned long long)path_leaf_bh(left)->b_blocknr,
				(unsigned long long)path_leaf_bh(right)->b_blocknr);
	} while (left->p_node[i].bh->b_blocknr ==
		 right->p_node[i].bh->b_blocknr);

	return i - 1;
}

typedef void (path_insert_t)(void *, struct buffer_head *);

/*
 * Traverse a btree path in search of cpos, starting at root_el.
 *
 * This code can be called with a cpos larger than the tree, in which
 * case it will return the rightmost path.
 */
static int __ocfs2_find_path(struct ocfs2_caching_info *ci,
			     struct ocfs2_extent_list *root_el, u32 cpos,
			     path_insert_t *func, void *data)
{
	int i, ret = 0;
	u32 range;
	u64 blkno;
	struct buffer_head *bh = NULL;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list *el;
	struct ocfs2_extent_rec *rec;

	el = root_el;
	while (el->l_tree_depth) {
		if (le16_to_cpu(el->l_next_free_rec) == 0) {
			ocfs2_error(ocfs2_metadata_cache_get_super(ci),
				    "Owner %llu has empty extent list at depth %u\n",
				    (unsigned long long)ocfs2_metadata_cache_owner(ci),
				    le16_to_cpu(el->l_tree_depth));
			ret = -EROFS;
			goto out;

		}

		for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) {
			rec = &el->l_recs[i];

			/*
			 * In the case that cpos is off the allocation
			 * tree, this should just wind up returning the
			 * rightmost record.
			 */
			range = le32_to_cpu(rec->e_cpos) +
				ocfs2_rec_clusters(el, rec);
			if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
			    break;
		}

		blkno = le64_to_cpu(el->l_recs[i].e_blkno);
		if (blkno == 0) {
			ocfs2_error(ocfs2_metadata_cache_get_super(ci),
				    "Owner %llu has bad blkno in extent list at depth %u (index %d)\n",
				    (unsigned long long)ocfs2_metadata_cache_owner(ci),
				    le16_to_cpu(el->l_tree_depth), i);
			ret = -EROFS;
			goto out;
		}

		brelse(bh);
		bh = NULL;
		ret = ocfs2_read_extent_block(ci, blkno, &bh);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		eb = (struct ocfs2_extent_block *) bh->b_data;
		el = &eb->h_list;

		if (le16_to_cpu(el->l_next_free_rec) >
		    le16_to_cpu(el->l_count)) {
			ocfs2_error(ocfs2_metadata_cache_get_super(ci),
				    "Owner %llu has bad count in extent list at block %llu (next free=%u, count=%u)\n",
				    (unsigned long long)ocfs2_metadata_cache_owner(ci),
				    (unsigned long long)bh->b_blocknr,
				    le16_to_cpu(el->l_next_free_rec),
				    le16_to_cpu(el->l_count));
			ret = -EROFS;
			goto out;
		}

		if (func)
			func(data, bh);
	}

out:
	/*
	 * Catch any trailing bh that the loop didn't handle.
	 */
	brelse(bh);

	return ret;
}

/*
 * Given an initialized path (that is, it has a valid root extent
 * list), this function will traverse the btree in search of the path
 * which would contain cpos.
 *
 * The path traveled is recorded in the path structure.
 *
 * Note that this will not do any comparisons on leaf node extent
 * records, so it will work fine in the case that we just added a tree
 * branch.
 */
struct find_path_data {
	int index;
	struct ocfs2_path *path;
};
static void find_path_ins(void *data, struct buffer_head *bh)
{
	struct find_path_data *fp = data;

	get_bh(bh);
	ocfs2_path_insert_eb(fp->path, fp->index, bh);
	fp->index++;
}
int ocfs2_find_path(struct ocfs2_caching_info *ci,
		    struct ocfs2_path *path, u32 cpos)
{
	struct find_path_data data;

	data.index = 1;
	data.path = path;
	return __ocfs2_find_path(ci, path_root_el(path), cpos,
				 find_path_ins, &data);
}

static void find_leaf_ins(void *data, struct buffer_head *bh)
{
	struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data;
	struct ocfs2_extent_list *el = &eb->h_list;
	struct buffer_head **ret = data;

	/* We want to retain only the leaf block. */
	if (le16_to_cpu(el->l_tree_depth) == 0) {
		get_bh(bh);
		*ret = bh;
	}
}
/*
 * Find the leaf block in the tree which would contain cpos. No
 * checking of the actual leaf is done.
 *
 * Some paths want to call this instead of allocating a path structure
 * and calling ocfs2_find_path().
 *
 * This function doesn't handle non btree extent lists.
 */
int ocfs2_find_leaf(struct ocfs2_caching_info *ci,
		    struct ocfs2_extent_list *root_el, u32 cpos,
		    struct buffer_head **leaf_bh)
{
	int ret;
	struct buffer_head *bh = NULL;

	ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh);
	if (ret) {
		mlog_errno(ret);
		goto out;
	}

	*leaf_bh = bh;
out:
	return ret;
}

/*
 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation.
 *
 * Basically, we've moved stuff around at the bottom of the tree and
 * we need to fix up the extent records above the changes to reflect
 * the new changes.
 *
 * left_rec: the record on the left.
 * right_rec: the record to the right of left_rec
 * right_child_el: is the child list pointed to by right_rec
 *
 * By definition, this only works on interior nodes.
 */
static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec,
				  struct ocfs2_extent_rec *right_rec,
				  struct ocfs2_extent_list *right_child_el)
{
	u32 left_clusters, right_end;

	/*
	 * Interior nodes never have holes. Their cpos is the cpos of
	 * the leftmost record in their child list. Their cluster
	 * count covers the full theoretical range of their child list
	 * - the range between their cpos and the cpos of the record
	 * immediately to their right.
	 */
	left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos);
	if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) {
		BUG_ON(right_child_el->l_tree_depth);
		BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1);
		left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos);
	}
	left_clusters -= le32_to_cpu(left_rec->e_cpos);
	left_rec->e_int_clusters = cpu_to_le32(left_clusters);

	/*
	 * Calculate the rightmost cluster count boundary before
	 * moving cpos - we will need to adjust clusters after
	 * updating e_cpos to keep the same highest cluster count.
	 */
	right_end = le32_to_cpu(right_rec->e_cpos);
	right_end += le32_to_cpu(right_rec->e_int_clusters);

	right_rec->e_cpos = left_rec->e_cpos;
	le32_add_cpu(&right_rec->e_cpos, left_clusters);

	right_end -= le32_to_cpu(right_rec->e_cpos);
	right_rec->e_int_clusters = cpu_to_le32(right_end);
}

/*
 * Adjust the adjacent root node records involved in a
 * rotation. left_el_blkno is passed in as a key so that we can easily
 * find it's index in the root list.
 */
static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el,
				      struct ocfs2_extent_list *left_el,
				      struct ocfs2_extent_list *right_el,
				      u64 left_el_blkno)
{
	int i;

	BUG_ON(le16_to_cpu(root_el->l_tree_depth) <=
	       le16_to_cpu(left_el->l_tree_depth));

	for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) {
		if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno)
			break;
	}

	/*
	 * The path walking code should have never returned a root and
	 * two paths which are not adjacent.
	 */
	BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1));

	ocfs2_adjust_adjacent_records(&root_el->l_recs[i],
				      &root_el->l_recs[i + 1], right_el);
}

/*
 * We've changed a leaf block (in right_path) and need to reflect that
 * change back up the subtree.
 *
 * This happens in multiple places:
 *   - When we've moved an extent record from the left path leaf to the right
 *     path leaf to make room for an empty extent in the left path leaf.
 *   - When our insert into the right path leaf is at the leftmost edge
 *     and requires an update of the path immediately to it's left. This
 *     can occur at the end of some types of rotation and appending inserts.
 *   - When we've adjusted the last extent record in the left path leaf and the
 *     1st extent record in the right path leaf during cross extent block merge.
 */
static void ocfs2_complete_edge_insert(handle_t *handle,
				       struct ocfs2_path *left_path,
				       struct ocfs2_path *right_path,
				       int subtree_index)
{
	int i, idx;
	struct ocfs2_extent_list *el, *left_el, *right_el;
	struct ocfs2_extent_rec *left_rec, *right_rec;
	struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;

	/*
	 * Update the counts and position values within all the
	 * interior nodes to reflect the leaf rotation we just did.
	 *
	 * The root node is handled below the loop.
	 *
	 * We begin the loop with right_el and left_el pointing to the
	 * leaf lists and work our way up.
	 *
	 * NOTE: within this loop, left_el and right_el always refer
	 * to the *child* lists.
	 */
	left_el = path_leaf_el(left_path);
	right_el = path_leaf_el(right_path);
	for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) {
		trace_ocfs2_complete_edge_insert(i);

		/*
		 * One nice property of knowing that all of these
		 * nodes are below the root is that we only deal with
		 * the leftmost right node record and the rightmost
		 * left node record.
		 */
		el = left_path->p_node[i].el;
		idx = le16_to_cpu(left_el->l_next_free_rec) - 1;
		left_rec = &el->l_recs[idx];

		el = right_path->p_node[i].el;
		right_rec = &el->l_recs[0];

		ocfs2_adjust_adjacent_records(left_rec, right_rec, right_el);

		ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
		ocfs2_journal_dirty(handle, right_path->p_node[i].bh);

		/*
		 * Setup our list pointers now so that the current
		 * parents become children in the next iteration.
		 */
		left_el = left_path->p_node[i].el;
		right_el = right_path->p_node[i].el;
	}

	/*
	 * At the root node, adjust the two adjacent records which
	 * begin our path to the leaves.
	 */

	el = left_path->p_node[subtree_index].el;
	left_el = left_path->p_node[subtree_index + 1].el;
	right_el = right_path->p_node[subtree_index + 1].el;

	ocfs2_adjust_root_records(el, left_el, right_el,
				  left_path->p_node[subtree_index + 1].bh->b_blocknr);

	root_bh = left_path->p_node[subtree_index].bh;

	ocfs2_journal_dirty(handle, root_bh);
}

static int ocfs2_rotate_subtree_right(handle_t *handle,
				      struct ocfs2_extent_tree *et,
				      struct ocfs2_path *left_path,
				      struct ocfs2_path *right_path,
				      int subtree_index)
{
	int ret, i;
	struct buffer_head *right_leaf_bh;
	struct buffer_head *left_leaf_bh = NULL;
	struct buffer_head *root_bh;
	struct ocfs2_extent_list *right_el, *left_el;
	struct ocfs2_extent_rec move_rec;

	left_leaf_bh = path_leaf_bh(left_path);
	left_el = path_leaf_el(left_path);

	if (left_el->l_next_free_rec != left_el->l_count) {
		ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
			    "Inode %llu has non-full interior leaf node %llu (next free = %u)\n",
			    (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
			    (unsigned long long)left_leaf_bh->b_blocknr,
			    le16_to_cpu(left_el->l_next_free_rec));
		return -EROFS;
	}

	/*
	 * This extent block may already have an empty record, so we
	 * return early if so.
	 */
	if (ocfs2_is_empty_extent(&left_el->l_recs[0]))
		return 0;

	root_bh = left_path->p_node[subtree_index].bh;
	BUG_ON(root_bh != right_path->p_node[subtree_index].bh);

	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
					   subtree_index);
	if (ret) {
		mlog_errno(ret);
		goto out;
	}

	for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
						   right_path, i);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
						   left_path, i);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}
	}

	right_leaf_bh = path_leaf_bh(right_path);
	right_el = path_leaf_el(right_path);

	/* This is a code error, not a disk corruption. */
	mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails "
			"because rightmost leaf block %llu is empty\n",
			(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
			(unsigned long long)right_leaf_bh->b_blocknr);

	ocfs2_create_empty_extent(right_el);

	ocfs2_journal_dirty(handle, right_leaf_bh);

	/* Do the copy now. */
	i = le16_to_cpu(left_el->l_next_free_rec) - 1;
	move_rec = left_el->l_recs[i];
	right_el->l_recs[0] = move_rec;

	/*
	 * Clear out the record we just copied and shift everything
	 * over, leaving an empty extent in the left leaf.
	 *
	 * We temporarily subtract from next_free_rec so that the
	 * shift will lose the tail record (which is now defunct).
	 */
	le16_add_cpu(&left_el->l_next_free_rec, -1);
	ocfs2_shift_records_right(left_el);
	memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));
	le16_add_cpu(&left_el->l_next_free_rec, 1);

	ocfs2_journal_dirty(handle, left_leaf_bh);

	ocfs2_complete_edge_insert(handle, left_path, right_path,
				   subtree_index);

out:
	return ret;
}

/*
 * Given a full path, determine what cpos value would return us a path
 * containing the leaf immediately to the left of the current one.
 *
 * Will return zero if the path passed in is already the leftmost path.
 */
int ocfs2_find_cpos_for_left_leaf(struct super_block *sb,
				  struct ocfs2_path *path, u32 *cpos)
{
	int i, j, ret = 0;
	u64 blkno;
	struct ocfs2_extent_list *el;

	BUG_ON(path->p_tree_depth == 0);

	*cpos = 0;

	blkno = path_leaf_bh(path)->b_blocknr;

	/* Start at the tree node just above the leaf and work our way up. */
	i = path->p_tree_depth - 1;
	while (i >= 0) {
		el = path->p_node[i].el;

		/*
		 * Find the extent record just before the one in our
		 * path.
		 */
		for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
			if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
				if (j == 0) {
					if (i == 0) {
						/*
						 * We've determined that the
						 * path specified is already
						 * the leftmost one - return a
						 * cpos of zero.
						 */
						goto out;
					}
					/*
					 * The leftmost record points to our
					 * leaf - we need to travel up the
					 * tree one level.
					 */
					goto next_node;
				}

				*cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos);
				*cpos = *cpos + ocfs2_rec_clusters(el,
							   &el->l_recs[j - 1]);
				*cpos = *cpos - 1;
				goto out;
			}
		}

		/*
		 * If we got here, we never found a valid node where
		 * the tree indicated one should be.
		 */
		ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
			    (unsigned long long)blkno);
		ret = -EROFS;
		goto out;

next_node:
		blkno = path->p_node[i].bh->b_blocknr;
		i--;
	}

out:
	return ret;
}

/*
 * Extend the transaction by enough credits to complete the rotation,
 * and still leave at least the original number of credits allocated
 * to this transaction.
 */
static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
					   int op_credits,
					   struct ocfs2_path *path)
{
	int ret = 0;
	int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits;

	if (jbd2_handle_buffer_credits(handle) < credits)
		ret = ocfs2_extend_trans(handle,
				credits - jbd2_handle_buffer_credits(handle));

	return ret;
}

/*
 * Trap the case where we're inserting into the theoretical range past
 * the _actual_ left leaf range. Otherwise, we'll rotate a record
 * whose cpos is less than ours into the right leaf.
 *
 * It's only necessary to look at the rightmost record of the left
 * leaf because the logic that calls us should ensure that the
 * theoretical ranges in the path components above the leaves are
 * correct.
 */
static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path,
						 u32 insert_cpos)
{
	struct ocfs2_extent_list *left_el;
	struct ocfs2_extent_rec *rec;
	int next_free;

	left_el = path_leaf_el(left_path);
	next_free = le16_to_cpu(left_el->l_next_free_rec);
	rec = &left_el->l_recs[next_free - 1];

	if (insert_cpos > le32_to_cpu(rec->e_cpos))
		return 1;
	return 0;
}

static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos)
{
	int next_free = le16_to_cpu(el->l_next_free_rec);
	unsigned int range;
	struct ocfs2_extent_rec *rec;

	if (next_free == 0)
		return 0;

	rec = &el->l_recs[0];
	if (ocfs2_is_empty_extent(rec)) {
		/* Empty list. */
		if (next_free == 1)
			return 0;
		rec = &el->l_recs[1];
	}

	range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);
	if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range)
		return 1;
	return 0;
}

/*
 * Rotate all the records in a btree right one record, starting at insert_cpos.
 *
 * The path to the rightmost leaf should be passed in.
 *
 * The array is assumed to be large enough to hold an entire path (tree depth).
 *
 * Upon successful return from this function:
 *
 * - The 'right_path' array will contain a path to the leaf block
 *   whose range contains e_cpos.
 * - That leaf block will have a single empty extent in list index 0.
 * - In the case that the rotation requires a post-insert update,
 *   *ret_left_path will contain a valid path which can be passed to
 *   ocfs2_insert_path().
 */
static int ocfs2_rotate_tree_right(handle_t *handle,
				   struct ocfs2_extent_tree *et,
				   enum ocfs2_split_type split,
				   u32 insert_cpos,
				   struct ocfs2_path *right_path,
				   struct ocfs2_path **ret_left_path)
{
	int ret, start, orig_credits = jbd2_handle_buffer_credits(handle);
	u32 cpos;
	struct ocfs2_path *left_path = NULL;
	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);

	*ret_left_path = NULL;

	left_path = ocfs2_new_path_from_path(right_path);
	if (!left_path) {
		ret = -ENOMEM;
		mlog_errno(ret);
		goto out;
	}

	ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
	if (ret) {
		mlog_errno(ret);
		goto out;
	}

	trace_ocfs2_rotate_tree_right(
		(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
		insert_cpos, cpos);

	/*
	 * What we want to do here is:
	 *
	 * 1) Start with the rightmost path.
	 *
	 * 2) Determine a path to the leaf block directly to the left
	 *    of that leaf.
	 *
	 * 3) Determine the 'subtree root' - the lowest level tree node
	 *    which contains a path to both leaves.
	 *
	 * 4) Rotate the subtree.
	 *
	 * 5) Find the next subtree by considering the left path to be
	 *    the new right path.
	 *
	 * The check at the top of this while loop also accepts
	 * insert_cpos == cpos because cpos is only a _theoretical_
	 * value to get us the left path - insert_cpos might very well
	 * be filling that hole.
	 *
	 * Stop at a cpos of '0' because we either started at the
	 * leftmost branch (i.e., a tree with one branch and a
	 * rotation inside of it), or we've gone as far as we can in
	 * rotating subtrees.
	 */
	while (cpos && insert_cpos <= cpos) {
		trace_ocfs2_rotate_tree_right(
			(unsigned long long)
			ocfs2_metadata_cache_owner(et->et_ci),
			insert_cpos, cpos);

		ret = ocfs2_find_path(et->et_ci, left_path, cpos);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		mlog_bug_on_msg(path_leaf_bh(left_path) ==
				path_leaf_bh(right_path),
				"Owner %llu: error during insert of %u "
				"(left path cpos %u) results in two identical "
				"paths ending at %llu\n",
				(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
				insert_cpos, cpos,
				(unsigned long long)
				path_leaf_bh(left_path)->b_blocknr);

		if (split == SPLIT_NONE &&
		    ocfs2_rotate_requires_path_adjustment(left_path,
							  insert_cpos)) {

			/*
			 * We've rotated the tree as much as we
			 * should. The rest is up to
			 * ocfs2_insert_path() to complete, after the
			 * record insertion. We indicate this
			 * situation by returning the left path.
			 *
			 * The reason we don't adjust the records here
			 * before the record insert is that an error
			 * later might break the rule where a parent
			 * record e_cpos will reflect the actual
			 * e_cpos of the 1st nonempty record of the
			 * child list.
			 */
			*ret_left_path = left_path;
			goto out_ret_path;
		}

		start = ocfs2_find_subtree_root(et, left_path, right_path);

		trace_ocfs2_rotate_subtree(start,
			(unsigned long long)
			right_path->p_node[start].bh->b_blocknr,
			right_path->p_tree_depth);

		ret = ocfs2_extend_rotate_transaction(handle, start,
						      orig_credits, right_path);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		ret = ocfs2_rotate_subtree_right(handle, et, left_path,
						 right_path, start);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		if (split != SPLIT_NONE &&
		    ocfs2_leftmost_rec_contains(path_leaf_el(right_path),
						insert_cpos)) {
			/*
			 * A rotate moves the rightmost left leaf
			 * record over to the leftmost right leaf
			 * slot. If we're doing an extent split
			 * instead of a real insert, then we have to
			 * check that the extent to be split wasn't
			 * just moved over. If it was, then we can
			 * exit here, passing left_path back -
			 * ocfs2_split_extent() is smart enough to
			 * search both leaves.
			 */
			*ret_left_path = left_path;
			goto out_ret_path;
		}

		/*
		 * There is no need to re-read the next right path
		 * as we know that it'll be our current left
		 * path. Optimize by copying values instead.
		 */
		ocfs2_mv_path(right_path, left_path);

		ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}
	}

out:
	ocfs2_free_path(left_path);

out_ret_path:
	return ret;
}

static int ocfs2_update_edge_lengths(handle_t *handle,
				     struct ocfs2_extent_tree *et,
				     struct ocfs2_path *path)
{
	int i, idx, ret;
	struct ocfs2_extent_rec *rec;
	struct ocfs2_extent_list *el;
	struct ocfs2_extent_block *eb;
	u32 range;

	ret = ocfs2_journal_access_path(et->et_ci, handle, path);
	if (ret) {
		mlog_errno(ret);
		goto out;
	}

	/* Path should always be rightmost. */
	eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
	BUG_ON(eb->h_next_leaf_blk != 0ULL);

	el = &eb->h_list;
	BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
	idx = le16_to_cpu(el->l_next_free_rec) - 1;
	rec = &el->l_recs[idx];
	range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec);

	for (i = 0; i < path->p_tree_depth; i++) {
		el = path->p_node[i].el;
		idx = le16_to_cpu(el->l_next_free_rec) - 1;
		rec = &el->l_recs[idx];

		rec->e_int_clusters = cpu_to_le32(range);
		le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos));

		ocfs2_journal_dirty(handle, path->p_node[i].bh);
	}
out:
	return ret;
}

static void ocfs2_unlink_path(handle_t *handle,
			      struct ocfs2_extent_tree *et,
			      struct ocfs2_cached_dealloc_ctxt *dealloc,
			      struct ocfs2_path *path, int unlink_start)
{
	int ret, i;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list *el;
	struct buffer_head *bh;

	for(i = unlink_start; i < path_num_items(path); i++) {
		bh = path->p_node[i].bh;

		eb = (struct ocfs2_extent_block *)bh->b_data;
		/*
		 * Not all nodes might have had their final count
		 * decremented by the caller - handle this here.
		 */
		el = &eb->h_list;
		if (le16_to_cpu(el->l_next_free_rec) > 1) {
			mlog(ML_ERROR,
			     "Inode %llu, attempted to remove extent block "
			     "%llu with %u records\n",
			     (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
			     (unsigned long long)le64_to_cpu(eb->h_blkno),
			     le16_to_cpu(el->l_next_free_rec));

			ocfs2_journal_dirty(handle, bh);
			ocfs2_remove_from_cache(et->et_ci, bh);
			continue;
		}

		el->l_next_free_rec = 0;
		memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));

		ocfs2_journal_dirty(handle, bh);

		ret = ocfs2_cache_extent_block_free(dealloc, eb);
		if (ret)
			mlog_errno(ret);

		ocfs2_remove_from_cache(et->et_ci, bh);
	}
}

static void ocfs2_unlink_subtree(handle_t *handle,
				 struct ocfs2_extent_tree *et,
				 struct ocfs2_path *left_path,
				 struct ocfs2_path *right_path,
				 int subtree_index,
				 struct ocfs2_cached_dealloc_ctxt *dealloc)
{
	int i;
	struct buffer_head *root_bh = left_path->p_node[subtree_index].bh;
	struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el;
	struct ocfs2_extent_block *eb;

	eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data;

	for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++)
		if (root_el->l_recs[i].e_blkno == eb->h_blkno)
			break;

	BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec));

	memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec));
	le16_add_cpu(&root_el->l_next_free_rec, -1);

	eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
	eb->h_next_leaf_blk = 0;

	ocfs2_journal_dirty(handle, root_bh);
	ocfs2_journal_dirty(handle, path_leaf_bh(left_path));

	ocfs2_unlink_path(handle, et, dealloc, right_path,
			  subtree_index + 1);
}

static int ocfs2_rotate_subtree_left(handle_t *handle,
				     struct ocfs2_extent_tree *et,
				     struct ocfs2_path *left_path,
				     struct ocfs2_path *right_path,
				     int subtree_index,
				     struct ocfs2_cached_dealloc_ctxt *dealloc,
				     int *deleted)
{
	int ret, i, del_right_subtree = 0, right_has_empty = 0;
	struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path);
	struct ocfs2_extent_list *right_leaf_el, *left_leaf_el;
	struct ocfs2_extent_block *eb;

	*deleted = 0;

	right_leaf_el = path_leaf_el(right_path);
	left_leaf_el = path_leaf_el(left_path);
	root_bh = left_path->p_node[subtree_index].bh;
	BUG_ON(root_bh != right_path->p_node[subtree_index].bh);

	if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0]))
		return 0;

	eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data;
	if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) {
		/*
		 * It's legal for us to proceed if the right leaf is
		 * the rightmost one and it has an empty extent. There
		 * are two cases to handle - whether the leaf will be
		 * empty after removal or not. If the leaf isn't empty
		 * then just remove the empty extent up front. The
		 * next block will handle empty leaves by flagging
		 * them for unlink.
		 *
		 * Non rightmost leaves will throw -EAGAIN and the
		 * caller can manually move the subtree and retry.
		 */

		if (eb->h_next_leaf_blk != 0ULL)
			return -EAGAIN;

		if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) {
			ret = ocfs2_journal_access_eb(handle, et->et_ci,
						      path_leaf_bh(right_path),
						      OCFS2_JOURNAL_ACCESS_WRITE);
			if (ret) {
				mlog_errno(ret);
				goto out;
			}

			ocfs2_remove_empty_extent(right_leaf_el);
		} else
			right_has_empty = 1;
	}

	if (eb->h_next_leaf_blk == 0ULL &&
	    le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) {
		/*
		 * We have to update i_last_eb_blk during the meta
		 * data delete.
		 */
		ret = ocfs2_et_root_journal_access(handle, et,
						   OCFS2_JOURNAL_ACCESS_WRITE);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		del_right_subtree = 1;
	}

	/*
	 * Getting here with an empty extent in the right path implies
	 * that it's the rightmost path and will be deleted.
	 */
	BUG_ON(right_has_empty && !del_right_subtree);

	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path,
					   subtree_index);
	if (ret) {
		mlog_errno(ret);
		goto out;
	}

	for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
						   right_path, i);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
						   left_path, i);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}
	}

	if (!right_has_empty) {
		/*
		 * Only do this if we're moving a real
		 * record. Otherwise, the action is delayed until
		 * after removal of the right path in which case we
		 * can do a simple shift to remove the empty extent.
		 */
		ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]);
		memset(&right_leaf_el->l_recs[0], 0,
		       sizeof(struct ocfs2_extent_rec));
	}
	if (eb->h_next_leaf_blk == 0ULL) {
		/*
		 * Move recs over to get rid of empty extent, decrease
		 * next_free. This is allowed to remove the last
		 * extent in our leaf (setting l_next_free_rec to
		 * zero) - the delete code below won't care.
		 */
		ocfs2_remove_empty_extent(right_leaf_el);
	}

	ocfs2_journal_dirty(handle, path_leaf_bh(left_path));
	ocfs2_journal_dirty(handle, path_leaf_bh(right_path));

	if (del_right_subtree) {
		ocfs2_unlink_subtree(handle, et, left_path, right_path,
				     subtree_index, dealloc);
		ret = ocfs2_update_edge_lengths(handle, et, left_path);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
		ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));

		/*
		 * Removal of the extent in the left leaf was skipped
		 * above so we could delete the right path
		 * 1st.
		 */
		if (right_has_empty)
			ocfs2_remove_empty_extent(left_leaf_el);

		ocfs2_journal_dirty(handle, et_root_bh);

		*deleted = 1;
	} else
		ocfs2_complete_edge_insert(handle, left_path, right_path,
					   subtree_index);

out:
	return ret;
}

/*
 * Given a full path, determine what cpos value would return us a path
 * containing the leaf immediately to the right of the current one.
 *
 * Will return zero if the path passed in is already the rightmost path.
 *
 * This looks similar, but is subtly different to
 * ocfs2_find_cpos_for_left_leaf().
 */
int ocfs2_find_cpos_for_right_leaf(struct super_block *sb,
				   struct ocfs2_path *path, u32 *cpos)
{
	int i, j, ret = 0;
	u64 blkno;
	struct ocfs2_extent_list *el;

	*cpos = 0;

	if (path->p_tree_depth == 0)
		return 0;

	blkno = path_leaf_bh(path)->b_blocknr;

	/* Start at the tree node just above the leaf and work our way up. */
	i = path->p_tree_depth - 1;
	while (i >= 0) {
		int next_free;

		el = path->p_node[i].el;

		/*
		 * Find the extent record just after the one in our
		 * path.
		 */
		next_free = le16_to_cpu(el->l_next_free_rec);
		for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) {
			if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) {
				if (j == (next_free - 1)) {
					if (i == 0) {
						/*
						 * We've determined that the
						 * path specified is already
						 * the rightmost one - return a
						 * cpos of zero.
						 */
						goto out;
					}
					/*
					 * The rightmost record points to our
					 * leaf - we need to travel up the
					 * tree one level.
					 */
					goto next_node;
				}

				*cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos);
				goto out;
			}
		}

		/*
		 * If we got here, we never found a valid node where
		 * the tree indicated one should be.
		 */
		ocfs2_error(sb, "Invalid extent tree at extent block %llu\n",
			    (unsigned long long)blkno);
		ret = -EROFS;
		goto out;

next_node:
		blkno = path->p_node[i].bh->b_blocknr;
		i--;
	}

out:
	return ret;
}

static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle,
					    struct ocfs2_extent_tree *et,
					    struct ocfs2_path *path)
{
	int ret;
	struct buffer_head *bh = path_leaf_bh(path);
	struct ocfs2_extent_list *el = path_leaf_el(path);

	if (!ocfs2_is_empty_extent(&el->l_recs[0]))
		return 0;

	ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path,
					   path_num_items(path) - 1);
	if (ret) {
		mlog_errno(ret);
		goto out;
	}

	ocfs2_remove_empty_extent(el);
	ocfs2_journal_dirty(handle, bh);

out:
	return ret;
}

static int __ocfs2_rotate_tree_left(handle_t *handle,
				    struct ocfs2_extent_tree *et,
				    int orig_credits,
				    struct ocfs2_path *path,
				    struct ocfs2_cached_dealloc_ctxt *dealloc,
				    struct ocfs2_path **empty_extent_path)
{
	int ret, subtree_root, deleted;
	u32 right_cpos;
	struct ocfs2_path *left_path = NULL;
	struct ocfs2_path *right_path = NULL;
	struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci);

	if (!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0])))
		return 0;

	*empty_extent_path = NULL;

	ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos);
	if (ret) {
		mlog_errno(ret);
		goto out;
	}

	left_path = ocfs2_new_path_from_path(path);
	if (!left_path) {
		ret = -ENOMEM;
		mlog_errno(ret);
		goto out;
	}

	ocfs2_cp_path(left_path, path);

	right_path = ocfs2_new_path_from_path(path);
	if (!right_path) {
		ret = -ENOMEM;
		mlog_errno(ret);
		goto out;
	}

	while (right_cpos) {
		ret = ocfs2_find_path(et->et_ci, right_path, right_cpos);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		subtree_root = ocfs2_find_subtree_root(et, left_path,
						       right_path);

		trace_ocfs2_rotate_subtree(subtree_root,
		     (unsigned long long)
		     right_path->p_node[subtree_root].bh->b_blocknr,
		     right_path->p_tree_depth);

		ret = ocfs2_extend_rotate_transaction(handle, 0,
						      orig_credits, left_path);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		/*
		 * Caller might still want to make changes to the
		 * tree root, so re-add it to the journal here.
		 */
		ret = ocfs2_path_bh_journal_access(handle, et->et_ci,
						   left_path, 0);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		ret = ocfs2_rotate_subtree_left(handle, et, left_path,
						right_path, subtree_root,
						dealloc, &deleted);
		if (ret == -EAGAIN) {
			/*
			 * The rotation has to temporarily stop due to
			 * the right subtree having an empty
			 * extent. Pass it back to the caller for a
			 * fixup.
			 */
			*empty_extent_path = right_path;
			right_path = NULL;
			goto out;
		}
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		/*
		 * The subtree rotate might have removed records on
		 * the rightmost edge. If so, then rotation is
		 * complete.
		 */
		if (deleted)
			break;

		ocfs2_mv_path(left_path, right_path);

		ret = ocfs2_find_cpos_for_right_leaf(sb, left_path,
						     &right_cpos);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}
	}

out:
	ocfs2_free_path(right_path);
	ocfs2_free_path(left_path);

	return ret;
}

static int ocfs2_remove_rightmost_path(handle_t *handle,
				struct ocfs2_extent_tree *et,
				struct ocfs2_path *path,
				struct ocfs2_cached_dealloc_ctxt *dealloc)
{
	int ret, subtree_index;
	u32 cpos;
	struct ocfs2_path *left_path = NULL;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list *el;

	ret = ocfs2_et_sanity_check(et);
	if (ret)
		goto out;

	ret = ocfs2_journal_access_path(et->et_ci, handle, path);
	if (ret) {
		mlog_errno(ret);
		goto out;
	}

	ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci),
					    path, &cpos);
	if (ret) {
		mlog_errno(ret);
		goto out;
	}

	if (cpos) {
		/*
		 * We have a path to the left of this one - it needs
		 * an update too.
		 */
		left_path = ocfs2_new_path_from_path(path);
		if (!left_path) {
			ret = -ENOMEM;
			mlog_errno(ret);
			goto out;
		}

		ret = ocfs2_find_path(et->et_ci, left_path, cpos);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		ret = ocfs2_journal_access_path(et->et_ci, handle, left_path);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		subtree_index = ocfs2_find_subtree_root(et, left_path, path);

		ocfs2_unlink_subtree(handle, et, left_path, path,
				     subtree_index, dealloc);
		ret = ocfs2_update_edge_lengths(handle, et, left_path);
		if (ret) {
			mlog_errno(ret);
			goto out;
		}

		eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data;
		ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno));
	} else {
		/*
		 * 'path' is also the leftmost path which
		 * means it must be the only one. This gets
		 * handled differently because we want to
		 * revert the root back to having extents
		 * in-line.
		 */
		ocfs2_unlink_path(handle, et, dealloc, path, 1);

		el = et->et_root_el;
		el->l_tree_depth = 0;
		el->l_next_free_rec = 0;
		memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec));

		ocfs2_et_set_last_eb_blk(et, 0);
	}

	ocfs2_journal_dirty(handle, path_root_bh(path));

out:
	ocfs2_free_path(left_path);
	return ret;
}

static int ocfs2_remove_rightmost_empty_extent(struct ocfs2_super *osb,
				struct ocfs2_extent_tree *et,
				struct ocfs2_path *path,
				struct ocfs2_cached_dealloc_ctxt *dealloc)
{
	handle_t *handle;
	int ret;
	int credits = path->p_tree_depth * 2 + 1;

	handle = ocfs2_start_trans(osb, credits);
	if (IS_ERR(handle)) {
		ret = PTR_ERR(handle);
		mlog_errno(ret);
		return ret;
	}

	ret = ocfs2_remove_rightmost_path(handle, et, path, dealloc);
	if (ret)
		mlog_errno(ret);

	ocfs2_commit_trans(osb, handle);
	return ret;
}

/*
 * Left rotation of btree records.
 *
 * In many ways, this is (unsurprisingly) the opposite of right
 * rotation. We start at some non-rightmost path containing an empty
 * extent in the leaf block. The code works its way to the rightmost
 * path by rotating records to the left in every subtree.
 *
 * This is used by any code which reduces the number of extent records
 * in a leaf. After removal, an empty record should be placed in the
 * leftmost list position.
 *
 * This won't handle a length update of the rightmost path records if
 * the rightmost tree leaf record is removed so the caller is
 * responsible for detecting and correcting that.
 */
static int ocfs2_rotate_tree_left(handle_t *handle,
				  struct ocfs2_extent_tree *et,
				  struct ocfs2_path *path,
				  struct ocfs2_cached_dealloc_ctxt *dealloc)
{
	int ret, orig_credits = jbd2_handle_buffer_credits(handle);
	struct ocfs2_path *tmp_path = NULL, *restart_path = NULL;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list *el;

	el = path_leaf_el(path);
	if (!ocfs2_is_empty_extent(&el->l_recs[0]))
		return 0;

	if (path->p_tree_depth == 0) {
rightmost_no_delete:
		/*
		 * Inline extents. This is trivially handled, so do
		 * it up front.
		 */
		ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path);
		if (ret)
			mlog_errno(ret);
		goto out;
	}

	/*
	 * Handle rightmost branch now. There's several cases:
	 *  1) simple rotation leaving records in there. That's trivial.
	 *  2) rotation requiring a branch delete - there's no more
	 *     records left. Two cases of this:
	 *     a) There are branches to the left.
	 *     b) This is also the leftmost (the only) branch.
	 *
	 *  1) is handled via ocfs2_rotate_rightmost_leaf_left()
	 *  2a) we need the left branch so that we can update it with the unlink
	 *  2b) we need to bring the root back to inline extents.
	 */

	eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data;
	el = &eb->h_list;
	if (eb->h_next_leaf_blk == 0) {
		/*
		 * This gets a bit tricky if we're going to delete the
		 * rightmost path. Get the other cases out of the way
		 * 1st.
		 */
		if (le16_to_cpu(el->l_next_free_rec) > 1)
			goto rightmost_no_delete;

		if (le16_to_cpu(el->l_next_free_rec) == 0) {
			ret = ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci),
					"Owner %llu has empty extent block at %llu\n",
					(unsigned long long)ocfs2_metadata_cache_owner(et->et_ci),
					(unsigned long long)le64_to_cpu(eb->h_blkno));
			goto out;
		}

		/*
		 * XXX: The caller can not trust "path" any more after
		 * this as it will have been deleted. What do we do?
		 *
		 * In theory the rotate-for-merge code will never get
		 * here because it'll always ask for a rotate in a
		 * nonempty list.
		 */

		ret = ocfs2_remove_rightmost_path(handle, et, path,
						  dealloc);
		if (ret)
			mlog_errno(ret);
		goto out;
	}

	/*
	 * Now we can loop, remembering the path we get from -EAGAIN
	 * and restarting from there.
	 */
try_rotate:
	ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path,
				       dealloc, &restart_path);
	if (ret && ret != -EAGAIN) {
		mlog_errno(ret);
		goto out;
	}

	while (ret == -EAGAIN) {