alloc.c

/* -*- mode: c; c-basic-offset: 8; -*-
 * vim: noexpandtab sw=8 ts=8 sts=0:
 *
 * alloc.c
 *
 * Extent allocs and frees
 *
 * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public
 * License as published by the Free Software Foundation; either
 * version 2 of the License, or (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * General Public License for more details.
 *
 * You should have received a copy of the GNU General Public
 * License along with this program; if not, write to the
 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
 * Boston, MA 021110-1307, USA.
 */

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

#define MLOG_MASK_PREFIX ML_DISK_ALLOC
#include <cluster/masklog.h>

#include "ocfs2.h"

#include "alloc.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 "buffer_head_io.h"

static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);

/*
 * Structures which describe a path through a btree, and functions to
 * manipulate them.
 *
 * The idea here is to be as generic as possible with the tree
 * manipulation code.
 */
struct ocfs2_path_item {
	struct buffer_head		*bh;
	struct ocfs2_extent_list	*el;
};

#define OCFS2_MAX_PATH_DEPTH	5

struct ocfs2_path {
	int			p_tree_depth;
	struct ocfs2_path_item	p_node[OCFS2_MAX_PATH_DEPTH];
};

#define path_root_bh(_path) ((_path)->p_node[0].bh)
#define path_root_el(_path) ((_path)->p_node[0].el)
#define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh)
#define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el)
#define path_num_items(_path) ((_path)->p_tree_depth + 1)

/*
 * Reset the actual path elements so that we can re-use the structure
 * to build another path. Generally, this involves freeing the buffer
 * heads.
 */
static 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);

	path->p_tree_depth = depth;
}

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

/*
 * 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));

	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)
{
	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;
	}

	return path;
}

/*
 * Allocate and initialize a new path based on a disk inode tree.
 */
static struct ocfs2_path *ocfs2_new_inode_path(struct buffer_head *di_bh)
{
	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
	struct ocfs2_extent_list *el = &di->id2.i_list;

	return ocfs2_new_path(di_bh, el);
}

/*
 * Convenience function to journal all components in a path.
 */
static int ocfs2_journal_access_path(struct inode *inode, 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_journal_access(handle, inode, path->p_node[i].bh,
					   OCFS2_JOURNAL_ACCESS_WRITE);
		if (ret < 0) {
			mlog_errno(ret);
			goto out;
		}
	}

out:
	return ret;
}

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

static int ocfs2_block_extent_contig(struct super_block *sb,
				     struct ocfs2_extent_rec *ext,
				     u64 blkno)
{
	return blkno == (le64_to_cpu(ext->e_blkno) +
			 ocfs2_clusters_to_blocks(sb,
						  le32_to_cpu(ext->e_clusters)));
}

static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left,
				  struct ocfs2_extent_rec *right)
{
	return (le32_to_cpu(left->e_cpos) + le32_to_cpu(left->e_clusters) ==
		le32_to_cpu(right->e_cpos));
}

static enum ocfs2_contig_type
	ocfs2_extent_contig(struct inode *inode,
			    struct ocfs2_extent_rec *ext,
			    struct ocfs2_extent_rec *insert_rec)
{
	u64 blkno = le64_to_cpu(insert_rec->e_blkno);

	if (ocfs2_extents_adjacent(ext, insert_rec) &&
	    ocfs2_block_extent_contig(inode->i_sb, ext, blkno))
			return CONTIG_RIGHT;

	blkno = le64_to_cpu(ext->e_blkno);
	if (ocfs2_extents_adjacent(insert_rec, ext) &&
	    ocfs2_block_extent_contig(inode->i_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,
};

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

/*
 * How many free extents have we got before we need more meta data?
 */
int ocfs2_num_free_extents(struct ocfs2_super *osb,
			   struct inode *inode,
			   struct ocfs2_dinode *fe)
{
	int retval;
	struct ocfs2_extent_list *el;
	struct ocfs2_extent_block *eb;
	struct buffer_head *eb_bh = NULL;

	mlog_entry_void();

	if (!OCFS2_IS_VALID_DINODE(fe)) {
		OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
		retval = -EIO;
		goto bail;
	}

	if (fe->i_last_eb_blk) {
		retval = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk),
					  &eb_bh, OCFS2_BH_CACHED, inode);
		if (retval < 0) {
			mlog_errno(retval);
			goto bail;
		}
		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
		el = &eb->h_list;
	} else
		el = &fe->id2.i_list;

	BUG_ON(el->l_tree_depth != 0);

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

	mlog_exit(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(struct ocfs2_super *osb,
				     handle_t *handle,
				     struct inode *inode,
				     int wanted,
				     struct ocfs2_alloc_context *meta_ac,
				     struct buffer_head *bhs[])
{
	int count, status, i;
	u16 suballoc_bit_start;
	u32 num_got;
	u64 first_blkno;
	struct ocfs2_extent_block *eb;

	mlog_entry_void();

	count = 0;
	while (count < wanted) {
		status = ocfs2_claim_metadata(osb,
					      handle,
					      meta_ac,
					      wanted - count,
					      &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 = -EIO;
				mlog_errno(status);
				goto bail;
			}
			ocfs2_set_new_buffer_uptodate(inode, bhs[i]);

			status = ocfs2_journal_access(handle, inode, 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);

#ifndef OCFS2_USE_ALL_METADATA_SUBALLOCATORS
			/* we always use slot zero's suballocator */
			eb->h_suballoc_slot = 0;
#else
			eb->h_suballoc_slot = cpu_to_le16(osb->slot_num);
#endif
			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. */
			status = ocfs2_journal_dirty(handle, bhs[i]);
			if (status < 0) {
				mlog_errno(status);
				goto bail;
			}
		}

		count += num_got;
	}

	status = 0;
bail:
	if (status < 0) {
		for(i = 0; i < wanted; i++) {
			if (bhs[i])
				brelse(bhs[i]);
			bhs[i] = NULL;
		}
	}
	mlog_exit(status);
	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) +
		le32_to_cpu(el->l_recs[i].e_clusters);
}

/*
 * 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 dinode
 * 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 e_clusters == 0.
 */
static int ocfs2_add_branch(struct ocfs2_super *osb,
			    handle_t *handle,
			    struct inode *inode,
			    struct buffer_head *fe_bh,
			    struct buffer_head *eb_bh,
			    struct buffer_head *last_eb_bh,
			    struct ocfs2_alloc_context *meta_ac)
{
	int status, new_blocks, i;
	u64 next_blkno, new_last_eb_blk;
	struct buffer_head *bh;
	struct buffer_head **new_eb_bhs = NULL;
	struct ocfs2_dinode *fe;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list  *eb_el;
	struct ocfs2_extent_list  *el;
	u32 new_cpos;

	mlog_entry_void();

	BUG_ON(!last_eb_bh);

	fe = (struct ocfs2_dinode *) fe_bh->b_data;

	if (eb_bh) {
		eb = (struct ocfs2_extent_block *) eb_bh->b_data;
		el = &eb->h_list;
	} else
		el = &fe->id2.i_list;

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

	new_blocks = le16_to_cpu(el->l_tree_depth);

	/* 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;
	}

	status = ocfs2_create_new_meta_bhs(osb, handle, inode, new_blocks,
					   meta_ac, new_eb_bhs);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

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

	/* 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;
		if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
			OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
			status = -EIO;
			goto bail;
		}
		eb_el = &eb->h_list;

		status = ocfs2_journal_access(handle, inode, 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->l_recs[0].e_clusters = cpu_to_le32(0);
		if (!eb_el->l_tree_depth)
			new_last_eb_blk = le64_to_cpu(eb->h_blkno);

		status = ocfs2_journal_dirty(handle, bh);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}

		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(handle, inode, last_eb_bh,
				      OCFS2_JOURNAL_ACCESS_WRITE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}
	status = ocfs2_journal_access(handle, inode, fe_bh,
				      OCFS2_JOURNAL_ACCESS_WRITE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}
	if (eb_bh) {
		status = ocfs2_journal_access(handle, inode, 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 fe, 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_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. */
	fe->i_last_eb_blk = cpu_to_le64(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);

	status = ocfs2_journal_dirty(handle, last_eb_bh);
	if (status < 0)
		mlog_errno(status);
	status = ocfs2_journal_dirty(handle, fe_bh);
	if (status < 0)
		mlog_errno(status);
	if (eb_bh) {
		status = ocfs2_journal_dirty(handle, eb_bh);
		if (status < 0)
			mlog_errno(status);
	}

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

	mlog_exit(status);
	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(struct ocfs2_super *osb,
				  handle_t *handle,
				  struct inode *inode,
				  struct buffer_head *fe_bh,
				  struct ocfs2_alloc_context *meta_ac,
				  struct buffer_head **ret_new_eb_bh)
{
	int status, i;
	u32 new_clusters;
	struct buffer_head *new_eb_bh = NULL;
	struct ocfs2_dinode *fe;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list  *fe_el;
	struct ocfs2_extent_list  *eb_el;

	mlog_entry_void();

	status = ocfs2_create_new_meta_bhs(osb, handle, inode, 1, meta_ac,
					   &new_eb_bh);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	eb = (struct ocfs2_extent_block *) new_eb_bh->b_data;
	if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
		OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
		status = -EIO;
		goto bail;
	}

	eb_el = &eb->h_list;
	fe = (struct ocfs2_dinode *) fe_bh->b_data;
	fe_el = &fe->id2.i_list;

	status = ocfs2_journal_access(handle, inode, new_eb_bh,
				      OCFS2_JOURNAL_ACCESS_CREATE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	/* copy the fe data into the new extent block */
	eb_el->l_tree_depth = fe_el->l_tree_depth;
	eb_el->l_next_free_rec = fe_el->l_next_free_rec;
	for(i = 0; i < le16_to_cpu(fe_el->l_next_free_rec); i++) {
		eb_el->l_recs[i].e_cpos = fe_el->l_recs[i].e_cpos;
		eb_el->l_recs[i].e_clusters = fe_el->l_recs[i].e_clusters;
		eb_el->l_recs[i].e_blkno = fe_el->l_recs[i].e_blkno;
	}

	status = ocfs2_journal_dirty(handle, new_eb_bh);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	status = ocfs2_journal_access(handle, inode, fe_bh,
				      OCFS2_JOURNAL_ACCESS_WRITE);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	new_clusters = ocfs2_sum_rightmost_rec(eb_el);

	/* update fe now */
	le16_add_cpu(&fe_el->l_tree_depth, 1);
	fe_el->l_recs[0].e_cpos = 0;
	fe_el->l_recs[0].e_blkno = eb->h_blkno;
	fe_el->l_recs[0].e_clusters = cpu_to_le32(new_clusters);
	for(i = 1; i < le16_to_cpu(fe_el->l_next_free_rec); i++) {
		fe_el->l_recs[i].e_cpos = 0;
		fe_el->l_recs[i].e_clusters = 0;
		fe_el->l_recs[i].e_blkno = 0;
	}
	fe_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 (fe_el->l_tree_depth == cpu_to_le16(1))
		fe->i_last_eb_blk = eb->h_blkno;

	status = ocfs2_journal_dirty(handle, fe_bh);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

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

	mlog_exit(status);
	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 dinode has room. We
 *    pass NULL back in *lowest_eb_bh, but still return '0'
 *
 * 3) the search fails to find anything AND the dinode is full, in
 *    which case we return > 0
 *
 * return status < 0 indicates an error.
 */
static int ocfs2_find_branch_target(struct ocfs2_super *osb,
				    struct inode *inode,
				    struct buffer_head *fe_bh,
				    struct buffer_head **target_bh)
{
	int status = 0, i;
	u64 blkno;
	struct ocfs2_dinode *fe;
	struct ocfs2_extent_block *eb;
	struct ocfs2_extent_list  *el;
	struct buffer_head *bh = NULL;
	struct buffer_head *lowest_bh = NULL;

	mlog_entry_void();

	*target_bh = NULL;

	fe = (struct ocfs2_dinode *) fe_bh->b_data;
	el = &fe->id2.i_list;

	while(le16_to_cpu(el->l_tree_depth) > 1) {
		if (le16_to_cpu(el->l_next_free_rec) == 0) {
			ocfs2_error(inode->i_sb, "Dinode %llu has empty "
				    "extent list (next_free_rec == 0)",
				    (unsigned long long)OCFS2_I(inode)->ip_blkno);
			status = -EIO;
			goto bail;
		}
		i = le16_to_cpu(el->l_next_free_rec) - 1;
		blkno = le64_to_cpu(el->l_recs[i].e_blkno);
		if (!blkno) {
			ocfs2_error(inode->i_sb, "Dinode %llu has extent "
				    "list where extent # %d has no physical "
				    "block start",
				    (unsigned long long)OCFS2_I(inode)->ip_blkno, i);
			status = -EIO;
			goto bail;
		}

		if (bh) {
			brelse(bh);
			bh = NULL;
		}

		status = ocfs2_read_block(osb, blkno, &bh, OCFS2_BH_CACHED,
					  inode);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}

		eb = (struct ocfs2_extent_block *) bh->b_data;
		if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
			OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
			status = -EIO;
			goto bail;
		}
		el = &eb->h_list;

		if (le16_to_cpu(el->l_next_free_rec) <
		    le16_to_cpu(el->l_count)) {
			if (lowest_bh)
				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' */
	if (!lowest_bh
	    && (fe->id2.i_list.l_next_free_rec == fe->id2.i_list.l_count))
		status = 1;

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

	mlog_exit(status);
	return status;
}

static inline int ocfs2_is_empty_extent(struct ocfs2_extent_rec *rec)
{
	return !rec->e_clusters;
}

/*
 * 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. */
	if (el->l_next_free_rec == el->l_count && !has_empty)
		BUG();

	/*
	 * 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;

	mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n",
	     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;

}

/*
 * 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);

	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.
 */
static int ocfs2_find_subtree_root(struct inode *inode,
				   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,
				"Inode %lu, left depth %u, right depth %u\n"
				"left leaf blk %llu, right leaf blk %llu\n",
				inode->i_ino, 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 inode *inode,