ocfs2: sparse b-tree support

Introduce tree rotations into the b-tree code. This will allow ocfs2 to
support sparse files. Much of the added code is designed to be generic (in
the ocfs2 sense) so that it can later be re-used to implement large
extended attributes.

This patch only adds the rotation code and does minimal updates to callers
of the extent api.

Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>

fs/ocfs2/alloc.c

@@ -47,62 +47,230 @@
 
 #include "buffer_head_io.h"
 
-static int ocfs2_extent_contig(struct inode *inode,
-			       struct ocfs2_extent_rec *ext,
-			       u64 blkno);
+static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
 
-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[]);
+/*
+ * 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;
+};
 
-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);
+#define OCFS2_MAX_PATH_DEPTH	5
 
-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);
+struct ocfs2_path {
+	int			p_tree_depth;
+	struct ocfs2_path_item	p_node[OCFS2_MAX_PATH_DEPTH];
+};
 
-static int ocfs2_do_insert_extent(struct ocfs2_super *osb,
-				  handle_t *handle,
-				  struct inode *inode,
-				  struct buffer_head *fe_bh,
-				  u64 blkno,
-				  u32 new_clusters);
+#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)
 
-static int ocfs2_find_branch_target(struct ocfs2_super *osb,
-				    struct inode *inode,
-				    struct buffer_head *fe_bh,
-				    struct buffer_head **target_bh);
+/*
+ * 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;
 
-static int ocfs2_find_new_last_ext_blk(struct ocfs2_super *osb,
-				       struct inode *inode,
-				       struct ocfs2_dinode *fe,
-				       unsigned int new_i_clusters,
-				       struct buffer_head *old_last_eb,
-				       struct buffer_head **new_last_eb);
+	if (keep_root)
+		start = 1;
 
-static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc);
+	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;
 
-static int ocfs2_extent_contig(struct inode *inode,
-			       struct ocfs2_extent_rec *ext,
-			       u64 blkno)
+	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(inode->i_sb,
+			 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?
  */
@@ -241,6 +409,28 @@ bail:
 	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
@@ -268,6 +458,7 @@ static int ocfs2_add_branch(struct ocfs2_super *osb,
 	struct ocfs2_extent_block *eb;
 	struct ocfs2_extent_list  *eb_el;
 	struct ocfs2_extent_list  *el;
+	u32 new_cpos;
 
 	mlog_entry_void();
 
@@ -302,6 +493,9 @@ static int ocfs2_add_branch(struct ocfs2_super *osb,
 		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.
@@ -330,7 +524,11 @@ static int ocfs2_add_branch(struct ocfs2_super *osb,
 		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);
-		eb_el->l_recs[0].e_cpos = fe->i_clusters;
+		/*
+		 * 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)
@@ -376,7 +574,7 @@ static int ocfs2_add_branch(struct ocfs2_super *osb,
 	 * 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 = fe->i_clusters;
+	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);
 
@@ -425,6 +623,7 @@ static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
 				  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;
@@ -480,11 +679,13 @@ static int ocfs2_shift_tree_depth(struct ocfs2_super *osb,
 		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 = fe->i_clusters;
+	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;
@@ -514,199 +715,6 @@ bail:
 	return status;
 }
 
-/*
- * Expects the tree to already have room in the rightmost leaf for the
- * extent.  Updates all the extent blocks (and the dinode) on the way
- * down.
- */
-static int ocfs2_do_insert_extent(struct ocfs2_super *osb,
-				  handle_t *handle,
-				  struct inode *inode,
-				  struct buffer_head *fe_bh,
-				  u64 start_blk,
-				  u32 new_clusters)
-{
-	int status, i, num_bhs = 0;
-	u64 next_blkno;
-	u16 next_free;
-	struct buffer_head **eb_bhs = NULL;
-	struct ocfs2_dinode *fe;
-	struct ocfs2_extent_block *eb;
-	struct ocfs2_extent_list  *el;
-
-	mlog_entry_void();
-
-	status = ocfs2_journal_access(handle, inode, fe_bh,
-				      OCFS2_JOURNAL_ACCESS_WRITE);
-	if (status < 0) {
-		mlog_errno(status);
-		goto bail;
-	}
-
-	fe = (struct ocfs2_dinode *) fe_bh->b_data;
-	el = &fe->id2.i_list;
-	if (el->l_tree_depth) {
-		/* This is another operation where we want to be
-		 * careful about our tree updates. An error here means
-		 * none of the previous changes we made should roll
-		 * forward. As a result, we have to record the buffers
-		 * for this part of the tree in an array and reserve a
-		 * journal write to them before making any changes. */
-		num_bhs = le16_to_cpu(fe->id2.i_list.l_tree_depth);
-		eb_bhs = kcalloc(num_bhs, sizeof(struct buffer_head *),
-				 GFP_KERNEL);
-		if (!eb_bhs) {
-			status = -ENOMEM;
-			mlog_errno(status);
-			goto bail;
-		}
-
-		i = 0;
-		while(el->l_tree_depth) {
-			next_free = le16_to_cpu(el->l_next_free_rec);
-			if (next_free == 0) {
-				ocfs2_error(inode->i_sb,
-					    "Dinode %llu has a bad extent list",
-					    (unsigned long long)OCFS2_I(inode)->ip_blkno);
-				status = -EIO;
-				goto bail;
-			}
-			next_blkno = le64_to_cpu(el->l_recs[next_free - 1].e_blkno);
-
-			BUG_ON(i >= num_bhs);
-			status = ocfs2_read_block(osb, next_blkno, &eb_bhs[i],
-						  OCFS2_BH_CACHED, inode);
-			if (status < 0) {
-				mlog_errno(status);
-				goto bail;
-			}
-			eb = (struct ocfs2_extent_block *) eb_bhs[i]->b_data;
-			if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
-				OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb,
-								 eb);
-				status = -EIO;
-				goto bail;
-			}
-
-			status = ocfs2_journal_access(handle, inode, eb_bhs[i],
-						      OCFS2_JOURNAL_ACCESS_WRITE);
-			if (status < 0) {
-				mlog_errno(status);
-				goto bail;
-			}
-
-			el = &eb->h_list;
-			i++;
-			/* When we leave this loop, eb_bhs[num_bhs - 1] will
-			 * hold the bottom-most leaf extent block. */
-		}
-		BUG_ON(el->l_tree_depth);
-
-		el = &fe->id2.i_list;
-		/* If we have tree depth, then the fe update is
-		 * trivial, and we want to switch el out for the
-		 * bottom-most leaf in order to update it with the
-		 * actual extent data below. */
-		next_free = le16_to_cpu(el->l_next_free_rec);
-		if (next_free == 0) {
-			ocfs2_error(inode->i_sb,
-				    "Dinode %llu has a bad extent list",
-				    (unsigned long long)OCFS2_I(inode)->ip_blkno);
-			status = -EIO;
-			goto bail;
-		}
-		le32_add_cpu(&el->l_recs[next_free - 1].e_clusters,
-			     new_clusters);
-		/* (num_bhs - 1) to avoid the leaf */
-		for(i = 0; i < (num_bhs - 1); i++) {
-			eb = (struct ocfs2_extent_block *) eb_bhs[i]->b_data;
-			el = &eb->h_list;
-
-			/* finally, make our actual change to the
-			 * intermediate extent blocks. */
-			next_free = le16_to_cpu(el->l_next_free_rec);
-			le32_add_cpu(&el->l_recs[next_free - 1].e_clusters,
-				     new_clusters);
-
-			status = ocfs2_journal_dirty(handle, eb_bhs[i]);
-			if (status < 0)
-				mlog_errno(status);
-		}
-		BUG_ON(i != (num_bhs - 1));
-		/* note that the leaf block wasn't touched in
-		 * the loop above */
-		eb = (struct ocfs2_extent_block *) eb_bhs[num_bhs - 1]->b_data;
-		el = &eb->h_list;
-		BUG_ON(el->l_tree_depth);
-	}
-
-	/* yay, we can finally add the actual extent now! */
-	i = le16_to_cpu(el->l_next_free_rec) - 1;
-	if (le16_to_cpu(el->l_next_free_rec) &&
-	    ocfs2_extent_contig(inode, &el->l_recs[i], start_blk)) {
-		le32_add_cpu(&el->l_recs[i].e_clusters, new_clusters);
-	} else if (le16_to_cpu(el->l_next_free_rec) &&
-		   (le32_to_cpu(el->l_recs[i].e_clusters) == 0)) {
-		/* having an empty extent at eof is legal. */
-		if (el->l_recs[i].e_cpos != fe->i_clusters) {
-			ocfs2_error(inode->i_sb,
-				    "Dinode %llu trailing extent is bad: "
-				    "cpos (%u) != number of clusters (%u)",
-				    (unsigned long long)OCFS2_I(inode)->ip_blkno,
-				    le32_to_cpu(el->l_recs[i].e_cpos),
-				    le32_to_cpu(fe->i_clusters));
-			status = -EIO;
-			goto bail;
-		}
-		el->l_recs[i].e_blkno = cpu_to_le64(start_blk);
-		el->l_recs[i].e_clusters = cpu_to_le32(new_clusters);
-	} else {
-		/* No contiguous record, or no empty record at eof, so
-		 * we add a new one. */
-
-		BUG_ON(le16_to_cpu(el->l_next_free_rec) >=
-		       le16_to_cpu(el->l_count));
-		i = le16_to_cpu(el->l_next_free_rec);
-
-		el->l_recs[i].e_blkno = cpu_to_le64(start_blk);
-		el->l_recs[i].e_clusters = cpu_to_le32(new_clusters);
-		el->l_recs[i].e_cpos = fe->i_clusters;
-		le16_add_cpu(&el->l_next_free_rec, 1);
-	}
-
-	/*
-	 * extent_map errors are not fatal, so they are ignored outside
-	 * of flushing the thing.
-	 */
-	status = ocfs2_extent_map_append(inode, &el->l_recs[i],
-					 new_clusters);
-	if (status) {
-		mlog_errno(status);
-		ocfs2_extent_map_drop(inode, le32_to_cpu(fe->i_clusters));
-	}
-
-	status = ocfs2_journal_dirty(handle, fe_bh);
-	if (status < 0)
-		mlog_errno(status);
-	if (fe->id2.i_list.l_tree_depth) {
-		status = ocfs2_journal_dirty(handle, eb_bhs[num_bhs - 1]);
-		if (status < 0)
-			mlog_errno(status);
-	}
-
-	status = 0;
-bail:
-	if (eb_bhs) {
-		for (i = 0; i < num_bhs; i++)
-			if (eb_bhs[i])
-				brelse(eb_bhs[i]);
-		kfree(eb_bhs);
-	}
-
-	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
@@ -774,86 +782,1556 @@ static int ocfs2_find_branch_target(struct ocfs2_super *osb,
 			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,
+			     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;
+	struct ocfs2_inode_info *oi = OCFS2_I(inode);
+
+	el = root_el;
+	while (el->l_tree_depth) {
+		if (le16_to_cpu(el->l_next_free_rec) == 0) {
+			ocfs2_error(inode->i_sb,
+				    "Inode %llu has empty extent list at "
+				    "depth %u\n",
+				    (unsigned long long)oi->ip_blkno,
+				    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) +
+				le32_to_cpu(rec->e_clusters);
+			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(inode->i_sb,
+				    "Inode %llu has bad blkno in extent list "
+				    "at depth %u (index %d)\n",
+				    (unsigned long long)oi->ip_blkno,
+				    le16_to_cpu(el->l_tree_depth), i);
+			ret = -EROFS;
+			goto out;
+		}
+
+		brelse(bh);
+		bh = NULL;
+		ret = ocfs2_read_block(OCFS2_SB(inode->i_sb), blkno,
+				       &bh, OCFS2_BH_CACHED, inode);
+		if (ret) {
+			mlog_errno(ret);
+			goto out;
+		}
+
+		eb = (struct ocfs2_extent_block *) bh->b_data;
+		el = &eb->h_list;
+		if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
+			OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
+			ret = -EIO;
+			goto out;
+		}
+
+		if (le16_to_cpu(el->l_next_free_rec) >
+		    le16_to_cpu(el->l_count)) {
+			ocfs2_error(inode->i_sb,
+				    "Inode %llu has bad count in extent list "
+				    "at block %llu (next free=%u, count=%u)\n",
+				    (unsigned long long)oi->ip_blkno,
+				    (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++;
+}
+static int ocfs2_find_path(struct inode *inode, struct ocfs2_path *path,
+			   u32 cpos)
+{
+	struct find_path_data data;
+
+	data.index = 1;
+	data.path = path;
+	return __ocfs2_find_path(inode, 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.
+ */
+static int ocfs2_find_leaf(struct inode *inode,
+			   struct ocfs2_extent_list *root_el, u32 cpos,
+			   struct buffer_head **leaf_bh)
+{
+	int ret;
+	struct buffer_head *bh = NULL;
+
+	ret = __ocfs2_find_path(inode, 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.
+ * left_child_el: is the child list pointed to by left_rec
+ * 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_list *left_child_el,
+				  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);
+	left_clusters -= le32_to_cpu(left_rec->e_cpos);
+	left_rec->e_clusters = cpu_to_le32(left_clusters);
+
+	/*
+	 * Calculate the rightmost cluster count boundary before
+	 * moving cpos - we will need to adjust e_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_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_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], left_el,
+				      &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.
+ */
+static void ocfs2_complete_edge_insert(struct inode *inode, handle_t *handle,
+				       struct ocfs2_path *left_path,
+				       struct ocfs2_path *right_path,
+				       int subtree_index)
+{
+	int ret, 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--) {
+		mlog(0, "Adjust records at index %u\n", 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, left_el, right_rec,
+					      right_el);
+
+		ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh);
+		if (ret)
+			mlog_errno(ret);
+
+		ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh);
+		if (ret)
+			mlog_errno(ret);
+
+		/*
+		 * 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;
+
+	ret = ocfs2_journal_dirty(handle, root_bh);
+	if (ret)
+		mlog_errno(ret);
+}
+
+static int ocfs2_rotate_subtree_right(struct inode *inode,
+				      handle_t *handle,
+				      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(inode->i_sb,
+			    "Inode %llu has non-full interior leaf node %llu"
+			    "(next free = %u)",
+			    (unsigned long long)OCFS2_I(inode)->ip_blkno,
+			    (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_journal_access(handle, inode, root_bh,
+				   OCFS2_JOURNAL_ACCESS_WRITE);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
+
+	for(i = subtree_index + 1; i < path_num_items(right_path); i++) {
+		ret = ocfs2_journal_access(handle, inode,
+					   right_path->p_node[i].bh,
+					   OCFS2_JOURNAL_ACCESS_WRITE);
+		if (ret) {
+			mlog_errno(ret);
+			goto out;
+		}
+
+		ret = ocfs2_journal_access(handle, inode,
+					   left_path->p_node[i].bh,
+					   OCFS2_JOURNAL_ACCESS_WRITE);
+		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_I(inode)->ip_blkno,
+			(unsigned long long)right_leaf_bh->b_blocknr);
+
+	ocfs2_create_empty_extent(right_el);
+
+	ret = ocfs2_journal_dirty(handle, right_leaf_bh);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
+
+	/* 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);
+
+	ret = ocfs2_journal_dirty(handle, left_leaf_bh);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
+
+	ocfs2_complete_edge_insert(inode, 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.
+ */
+static 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;
+
+	*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 + le32_to_cpu(el->l_recs[j - 1].e_clusters) - 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;
+}
+
+static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth,
+					   struct ocfs2_path *path)
+{
+	int credits = (path->p_tree_depth - subtree_depth) * 2 + 1;
+
+	if (handle->h_buffer_credits < credits)
+		return ocfs2_extend_trans(handle, credits);
+
+	return 0;
+}
+
+/*
+ * 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;
+}
+
+/*
+ * 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 succesful 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(struct inode *inode,
+				   handle_t *handle,
+				   u32 insert_cpos,
+				   struct ocfs2_path *right_path,
+				   struct ocfs2_path **ret_left_path)
+{
+	int ret, start;
+	u32 cpos;
+	struct ocfs2_path *left_path = NULL;
+
+	*ret_left_path = NULL;
+
+	left_path = ocfs2_new_path(path_root_bh(right_path),
+				   path_root_el(right_path));
+	if (!left_path) {
+		ret = -ENOMEM;
+		mlog_errno(ret);
+		goto out;
+	}
+
+	ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, &cpos);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
+
+	mlog(0, "Insert: %u, first left path cpos: %u\n", 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) {
+		mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n",
+		     insert_cpos, cpos);
+
+		ret = ocfs2_find_path(inode, left_path, cpos);
+		if (ret) {
+			mlog_errno(ret);
+			goto out;
+		}
+
+		mlog_bug_on_msg(path_leaf_bh(left_path) ==
+				path_leaf_bh(right_path),
+				"Inode %lu: error during insert of %u "
+				"(left path cpos %u) results in two identical "
+				"paths ending at %llu\n",
+				inode->i_ino, insert_cpos, cpos,
+				(unsigned long long)
+				path_leaf_bh(left_path)->b_blocknr);
+
+		if (ocfs2_rotate_requires_path_adjustment(left_path,
+							  insert_cpos)) {
+			mlog(0, "Path adjustment required\n");
+
+			/*
+			 * 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(inode, left_path, right_path);
+
+		mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n",
+		     start,
+		     (unsigned long long) right_path->p_node[start].bh->b_blocknr,
+		     right_path->p_tree_depth);
+
+		ret = ocfs2_extend_rotate_transaction(handle, start,
+						      right_path);
+		if (ret) {
+			mlog_errno(ret);
+			goto out;
+		}
+
+		ret = ocfs2_rotate_subtree_right(inode, handle, left_path,
+						 right_path, start);
+		if (ret) {
+			mlog_errno(ret);
+			goto out;
+		}
+
+		/*
+		 * 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(inode->i_sb, right_path,
+						    &cpos);
+		if (ret) {
+			mlog_errno(ret);
+			goto out;
+		}
+	}
+
+out:
+	ocfs2_free_path(left_path);
+
+out_ret_path:
+	return ret;
+}
+
+/*
+ * Do the final bits of extent record insertion at the target leaf
+ * list. If this leaf is part of an allocation tree, it is assumed
+ * that the tree above has been prepared.
+ */
+static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec,
+				 struct ocfs2_extent_list *el,
+				 struct ocfs2_insert_type *insert,
+				 struct inode *inode)
+{
+	int i = insert->ins_contig_index;
+	unsigned int range;
+	struct ocfs2_extent_rec *rec;
+
+	BUG_ON(el->l_tree_depth);
+
+	/*
+	 * Contiguous insert - either left or right.
+	 */
+	if (insert->ins_contig != CONTIG_NONE) {
+		rec = &el->l_recs[i];
+		if (insert->ins_contig == CONTIG_LEFT) {
+			rec->e_blkno = insert_rec->e_blkno;
+			rec->e_cpos = insert_rec->e_cpos;
+		}
+		le32_add_cpu(&rec->e_clusters,
+			     le32_to_cpu(insert_rec->e_clusters));
+		return;
+	}
+
+	/*
+	 * Handle insert into an empty leaf.
+	 */
+	if (le16_to_cpu(el->l_next_free_rec) == 0 ||
+	    ((le16_to_cpu(el->l_next_free_rec) == 1) &&
+	     ocfs2_is_empty_extent(&el->l_recs[0]))) {
+		el->l_recs[0] = *insert_rec;
+		el->l_next_free_rec = cpu_to_le16(1);
+		return;
+	}
+
+	/*
+	 * Appending insert.
+	 */
+	if (insert->ins_appending == APPEND_TAIL) {
+		i = le16_to_cpu(el->l_next_free_rec) - 1;
+		rec = &el->l_recs[i];
+		range = le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters);
+		BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range);
+
+		mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >=
+				le16_to_cpu(el->l_count),
+				"inode %lu, depth %u, count %u, next free %u, "
+				"rec.cpos %u, rec.clusters %u, "
+				"insert.cpos %u, insert.clusters %u\n",
+				inode->i_ino,
+				le16_to_cpu(el->l_tree_depth),
+				le16_to_cpu(el->l_count),
+				le16_to_cpu(el->l_next_free_rec),
+				le32_to_cpu(el->l_recs[i].e_cpos),
+				le32_to_cpu(el->l_recs[i].e_clusters),
+				le32_to_cpu(insert_rec->e_cpos),
+				le32_to_cpu(insert_rec->e_clusters));
+		i++;
+		el->l_recs[i] = *insert_rec;
+		le16_add_cpu(&el->l_next_free_rec, 1);
+		return;
+	}
+
+	/*
+	 * Ok, we have to rotate.
+	 *
+	 * At this point, it is safe to assume that inserting into an
+	 * empty leaf and appending to a leaf have both been handled
+	 * above.
+	 *
+	 * This leaf needs to have space, either by the empty 1st
+	 * extent record, or by virtue of an l_next_rec < l_count.
+	 */
+	ocfs2_rotate_leaf(el, insert_rec);
+}
+
+static inline void ocfs2_update_dinode_clusters(struct inode *inode,
+						struct ocfs2_dinode *di,
+						u32 clusters)
+{
+	le32_add_cpu(&di->i_clusters, clusters);
+	spin_lock(&OCFS2_I(inode)->ip_lock);
+	OCFS2_I(inode)->ip_clusters = le32_to_cpu(di->i_clusters);
+	spin_unlock(&OCFS2_I(inode)->ip_lock);
+}
+
+static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle,
+				    struct ocfs2_extent_rec *insert_rec,
+				    struct ocfs2_path *right_path,
+				    struct ocfs2_path **ret_left_path)
+{
+	int ret, i, next_free;
+	struct buffer_head *bh;
+	struct ocfs2_extent_list *el;
+	struct ocfs2_path *left_path = NULL;
+
+	*ret_left_path = NULL;
+
+	/*
+	 * If our appending insert is at the leftmost edge of a leaf,
+	 * then we might need to update the rightmost records of the
+	 * neighboring path.
+	 */
+	el = path_leaf_el(right_path);
+	next_free = le16_to_cpu(el->l_next_free_rec);
+	if (next_free == 0 ||
+	    (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) {
+		u32 left_cpos;
+
+		ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path,
+						    &left_cpos);
+		if (ret) {
+			mlog_errno(ret);
+			goto out;
+		}
+
+		mlog(0, "Append may need a left path update. cpos: %u, "
+		     "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos),
+		     left_cpos);
+
+		/*
+		 * No need to worry if the append is already in the
+		 * leftmost leaf.
+		 */
+		if (left_cpos) {
+			left_path = ocfs2_new_path(path_root_bh(right_path),
+						   path_root_el(right_path));
+			if (!left_path) {
+				ret = -ENOMEM;
+				mlog_errno(ret);
+				goto out;
+			}
+
+			ret = ocfs2_find_path(inode, left_path, left_cpos);
+			if (ret) {
+				mlog_errno(ret);
+				goto out;
+			}
+
+			/*
+			 * ocfs2_insert_path() will pass the left_path to the
+			 * journal for us.
+			 */
+		}
+	}
+
+	ret = ocfs2_journal_access_path(inode, handle, right_path);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
+
+	el = path_root_el(right_path);
+	bh = path_root_bh(right_path);
+	i = 0;
+	while (1) {
+		next_free = le16_to_cpu(el->l_next_free_rec);
+		if (next_free == 0) {
+			ocfs2_error(inode->i_sb,
+				    "Dinode %llu has a bad extent list",
+				    (unsigned long long)OCFS2_I(inode)->ip_blkno);
+			ret = -EIO;
+			goto out;
+		}
+
+		el->l_recs[next_free - 1].e_clusters = insert_rec->e_cpos;
+		le32_add_cpu(&el->l_recs[next_free - 1].e_clusters,
+			     le32_to_cpu(insert_rec->e_clusters));
+		le32_add_cpu(&el->l_recs[next_free - 1].e_clusters,
+			    -le32_to_cpu(el->l_recs[next_free - 1].e_cpos));
+
+		ret = ocfs2_journal_dirty(handle, bh);
+		if (ret)
+			mlog_errno(ret);
+
+		if (++i >= right_path->p_tree_depth)
+			break;
+
+		bh = right_path->p_node[i].bh;
+		el = right_path->p_node[i].el;
+	}
+
+	*ret_left_path = left_path;
+	ret = 0;
+out:
+	if (ret != 0)
+		ocfs2_free_path(left_path);
+
+	return ret;
+}
+
+/*
+ * This function only does inserts on an allocation b-tree. For dinode
+ * lists, ocfs2_insert_at_leaf() is called directly.
+ *
+ * right_path is the path we want to do the actual insert
+ * in. left_path should only be passed in if we need to update that
+ * portion of the tree after an edge insert.
+ */
+static int ocfs2_insert_path(struct inode *inode,
+			     handle_t *handle,
+			     struct ocfs2_path *left_path,
+			     struct ocfs2_path *right_path,
+			     struct ocfs2_extent_rec *insert_rec,
+			     struct ocfs2_insert_type *insert)
+{
+	int ret, subtree_index;
+	struct buffer_head *leaf_bh = path_leaf_bh(right_path);
+	struct ocfs2_extent_list *el;
+
+	/*
+	 * Pass both paths to the journal. The majority of inserts
+	 * will be touching all components anyway.
+	 */
+	ret = ocfs2_journal_access_path(inode, handle, right_path);
+	if (ret < 0) {
+		mlog_errno(ret);
+		goto out;
+	}
+
+	if (left_path) {
+		int credits = handle->h_buffer_credits;
+
+		/*
+		 * There's a chance that left_path got passed back to
+		 * us without being accounted for in the
+		 * journal. Extend our transaction here to be sure we
+		 * can change those blocks.
+		 */
+		credits += left_path->p_tree_depth;
+
+		ret = ocfs2_extend_trans(handle, credits);
+		if (ret < 0) {
+			mlog_errno(ret);
+			goto out;
+		}
+
+		ret = ocfs2_journal_access_path(inode, handle, left_path);
+		if (ret < 0) {
+			mlog_errno(ret);
+			goto out;
+		}
+	}
+
+	el = path_leaf_el(right_path);
+
+	ocfs2_insert_at_leaf(insert_rec, el, insert, inode);
+	ret = ocfs2_journal_dirty(handle, leaf_bh);
+	if (ret)
+		mlog_errno(ret);
+
+	if (left_path) {
+		/*
+		 * The rotate code has indicated that we need to fix
+		 * up portions of the tree after the insert.
+		 *
+		 * XXX: Should we extend the transaction here?
+		 */
+		subtree_index = ocfs2_find_subtree_root(inode, left_path,
+							right_path);
+		ocfs2_complete_edge_insert(inode, handle, left_path,
+					   right_path, subtree_index);
+	}
+
+	ret = 0;
+out:
+	return ret;
+}
+
+static int ocfs2_do_insert_extent(struct inode *inode,
+				  handle_t *handle,
+				  struct buffer_head *di_bh,
+				  struct ocfs2_extent_rec *insert_rec,
+				  struct ocfs2_insert_type *type)
+{
+	int ret, rotate = 0;
+	u32 cpos;
+	struct ocfs2_path *right_path = NULL;
+	struct ocfs2_path *left_path = NULL;
+	struct ocfs2_dinode *di;
+	struct ocfs2_extent_list *el;
+
+	di = (struct ocfs2_dinode *) di_bh->b_data;
+	el = &di->id2.i_list;
+
+	ret = ocfs2_journal_access(handle, inode, di_bh,
+				   OCFS2_JOURNAL_ACCESS_WRITE);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
+
+	if (le16_to_cpu(el->l_tree_depth) == 0) {
+		ocfs2_insert_at_leaf(insert_rec, el, type, inode);
+		goto out_update_clusters;
+	}
+
+	right_path = ocfs2_new_inode_path(di_bh);
+	if (!right_path) {
+		ret = -ENOMEM;
+		mlog_errno(ret);
+		goto out;
+	}
+
+	/*
+	 * Determine the path to start with. Rotations need the
+	 * rightmost path, everything else can go directly to the
+	 * target leaf.
+	 */
+	cpos = le32_to_cpu(insert_rec->e_cpos);
+	if (type->ins_appending == APPEND_NONE &&
+	    type->ins_contig == CONTIG_NONE) {
+		rotate = 1;
+		cpos = UINT_MAX;
+	}
+
+	ret = ocfs2_find_path(inode, right_path, cpos);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
+
+	/*
+	 * Rotations and appends need special treatment - they modify
+	 * parts of the tree's above them.
+	 *
+	 * Both might pass back a path immediate to the left of the
+	 * one being inserted to. This will be cause
+	 * ocfs2_insert_path() to modify the rightmost records of
+	 * left_path to account for an edge insert.
+	 *
+	 * XXX: When modifying this code, keep in mind that an insert
+	 * can wind up skipping both of these two special cases...
+	 */
+	if (rotate) {
+		ret = ocfs2_rotate_tree_right(inode, handle,
+					      le32_to_cpu(insert_rec->e_cpos),
+					      right_path, &left_path);
+		if (ret) {
+			mlog_errno(ret);
+			goto out;
+		}
+	} else if (type->ins_appending == APPEND_TAIL
+		   && type->ins_contig != CONTIG_LEFT) {
+		ret = ocfs2_append_rec_to_path(inode, handle, insert_rec,
+					       right_path, &left_path);
+		if (ret) {
+			mlog_errno(ret);
+			goto out;
+		}
+	}
+
+	ret = ocfs2_insert_path(inode, handle, left_path, right_path,
+				insert_rec, type);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
+
+out_update_clusters:
+	ocfs2_update_dinode_clusters(inode, di,
+				     le32_to_cpu(insert_rec->e_clusters));
+
+	ret = ocfs2_journal_dirty(handle, di_bh);
+	if (ret)
+		mlog_errno(ret);
+
+out:
+	ocfs2_free_path(left_path);
+	ocfs2_free_path(right_path);
+
+	return ret;
+}
+
+static void ocfs2_figure_contig_type(struct inode *inode,
+				     struct ocfs2_insert_type *insert,
+				     struct ocfs2_extent_list *el,
+				     struct ocfs2_extent_rec *insert_rec)
+{
+	int i;
+	enum ocfs2_contig_type contig_type = CONTIG_NONE;
+
+	for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) {
+		contig_type = ocfs2_extent_contig(inode, &el->l_recs[i],
+						  insert_rec);
+		if (contig_type != CONTIG_NONE) {
+			insert->ins_contig_index = i;
+			break;
+		}
+	}
+	insert->ins_contig = contig_type;
+}
+
+/*
+ * This should only be called against the righmost leaf extent list.
+ *
+ * ocfs2_figure_appending_type() will figure out whether we'll have to
+ * insert at the tail of the rightmost leaf.
+ *
+ * This should also work against the dinode list for tree's with 0
+ * depth. If we consider the dinode list to be the rightmost leaf node
+ * then the logic here makes sense.
+ */
+static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert,
+					struct ocfs2_extent_list *el,
+					struct ocfs2_extent_rec *insert_rec)
+{
+	int i;
+	u32 cpos = le32_to_cpu(insert_rec->e_cpos);
+	struct ocfs2_extent_rec *rec;
+
+	insert->ins_appending = APPEND_NONE;
+
+	BUG_ON(el->l_tree_depth);
+
+	if (!el->l_next_free_rec)
+		goto set_tail_append;
+
+	if (ocfs2_is_empty_extent(&el->l_recs[0])) {
+		/* Were all records empty? */
+		if (le16_to_cpu(el->l_next_free_rec) == 1)
+			goto set_tail_append;
+	}
+
+	i = le16_to_cpu(el->l_next_free_rec) - 1;
+	rec = &el->l_recs[i];
+
+	if (cpos >= (le32_to_cpu(rec->e_cpos) + le32_to_cpu(rec->e_clusters)))
+		goto set_tail_append;
+
+	return;
+
+set_tail_append:
+	insert->ins_appending = APPEND_TAIL;
+}
+
+/*
+ * Helper function called at the begining of an insert.
+ *
+ * This computes a few things that are commonly used in the process of
+ * inserting into the btree:
+ *   - Whether the new extent is contiguous with an existing one.
+ *   - The current tree depth.
+ *   - Whether the insert is an appending one.
+ *   - The total # of free records in the tree.
+ *
+ * All of the information is stored on the ocfs2_insert_type
+ * structure.
+ */
+static int ocfs2_figure_insert_type(struct inode *inode,
+				    struct buffer_head *di_bh,
+				    struct buffer_head **last_eb_bh,
+				    struct ocfs2_extent_rec *insert_rec,
+				    struct ocfs2_insert_type *insert)
+{
+	int ret;
+	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
+	struct ocfs2_extent_block *eb;
+	struct ocfs2_extent_list *el;
+	struct ocfs2_path *path = NULL;
+	struct buffer_head *bh = NULL;
+
+	el = &di->id2.i_list;
+	insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth);
+
+	if (el->l_tree_depth) {
+		/*
+		 * If we have tree depth, we read in the
+		 * rightmost extent block ahead of time as
+		 * ocfs2_figure_insert_type() and ocfs2_add_branch()
+		 * may want it later.
+		 */
+		ret = ocfs2_read_block(OCFS2_SB(inode->i_sb),
+				       le64_to_cpu(di->i_last_eb_blk), &bh,
+				       OCFS2_BH_CACHED, inode);
+		if (ret) {
+			mlog_exit(ret);
+			goto out;
+		}
 		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);
-		}
+	/*
+	 * Unless we have a contiguous insert, we'll need to know if
+	 * there is room left in our allocation tree for another
+	 * extent record.
+	 *
+	 * XXX: This test is simplistic, we can search for empty
+	 * extent records too.
+	 */
+	insert->ins_free_records = le16_to_cpu(el->l_count) -
+		le16_to_cpu(el->l_next_free_rec);
+
+	if (!insert->ins_tree_depth) {
+		ocfs2_figure_contig_type(inode, insert, el, insert_rec);
+		ocfs2_figure_appending_type(insert, el, insert_rec);
+		return 0;
 	}
 
-	/* 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;
+	path = ocfs2_new_inode_path(di_bh);
+	if (!path) {
+		ret = -ENOMEM;
+		mlog_errno(ret);
+		goto out;
+	}
 
-	*target_bh = lowest_bh;
-bail:
-	if (bh)
-		brelse(bh);
+	/*
+	 * In the case that we're inserting past what the tree
+	 * currently accounts for, ocfs2_find_path() will return for
+	 * us the rightmost tree path. This is accounted for below in
+	 * the appending code.
+	 */
+	ret = ocfs2_find_path(inode, path, le32_to_cpu(insert_rec->e_cpos));
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
 
-	mlog_exit(status);
-	return status;
+	el = path_leaf_el(path);
+
+	/*
+	 * Now that we have the path, there's two things we want to determine:
+	 * 1) Contiguousness (also set contig_index if this is so)
+	 *
+	 * 2) Are we doing an append? We can trivially break this up
+         *     into two types of appends: simple record append, or a
+         *     rotate inside the tail leaf.
+	 */
+	ocfs2_figure_contig_type(inode, insert, el, insert_rec);
+
+	/*
+	 * The insert code isn't quite ready to deal with all cases of
+	 * left contiguousness. Specifically, if it's an insert into
+	 * the 1st record in a leaf, it will require the adjustment of
+	 * e_clusters on the last record of the path directly to it's
+	 * left. For now, just catch that case and fool the layers
+	 * above us. This works just fine for tree_depth == 0, which
+	 * is why we allow that above.
+	 */
+	if (insert->ins_contig == CONTIG_LEFT &&
+	    insert->ins_contig_index == 0)
+		insert->ins_contig = CONTIG_NONE;
+
+	/*
+	 * Ok, so we can simply compare against last_eb to figure out
+	 * whether the path doesn't exist. This will only happen in
+	 * the case that we're doing a tail append, so maybe we can
+	 * take advantage of that information somehow.
+	 */
+	if (le64_to_cpu(di->i_last_eb_blk) == path_leaf_bh(path)->b_blocknr) {
+		/*
+		 * Ok, ocfs2_find_path() returned us the rightmost
+		 * tree path. This might be an appending insert. There are
+		 * two cases:
+		 *    1) We're doing a true append at the tail:
+		 *	-This might even be off the end of the leaf
+		 *    2) We're "appending" by rotating in the tail
+		 */
+		ocfs2_figure_appending_type(insert, el, insert_rec);
+	}
+
+out:
+	ocfs2_free_path(path);
+
+	if (ret == 0)
+		*last_eb_bh = bh;
+	else
+		brelse(bh);
+	return ret;
 }
 
-/* the caller needs to update fe->i_clusters */
+/*
+ * Insert an extent into an inode btree.
+ *
+ * The caller needs to update fe->i_clusters
+ */
 int ocfs2_insert_extent(struct ocfs2_super *osb,
 			handle_t *handle,
 			struct inode *inode,
 			struct buffer_head *fe_bh,
+			u32 cpos,
 			u64 start_blk,
 			u32 new_clusters,
 			struct ocfs2_alloc_context *meta_ac)
 {
-	int status, i, shift;
+	int status, shift;
 	struct buffer_head *last_eb_bh = NULL;
 	struct buffer_head *bh = NULL;
-	struct ocfs2_dinode *fe;
-	struct ocfs2_extent_block *eb;
-	struct ocfs2_extent_list  *el;
-
-	mlog_entry_void();
-
-	mlog(0, "add %u clusters starting at block %llu to inode %llu\n",
-	     new_clusters, (unsigned long long)start_blk,
-	     (unsigned long long)OCFS2_I(inode)->ip_blkno);
-
-	fe = (struct ocfs2_dinode *) fe_bh->b_data;
-	el = &fe->id2.i_list;
-
-	if (el->l_tree_depth) {
-		/* jump to end of tree */
-		status = ocfs2_read_block(osb, le64_to_cpu(fe->i_last_eb_blk),
-					  &last_eb_bh, OCFS2_BH_CACHED, inode);
-		if (status < 0) {
-			mlog_exit(status);
-			goto bail;
-		}
-		eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
-		el = &eb->h_list;
+	struct ocfs2_insert_type insert = {0, };
+	struct ocfs2_extent_rec rec;
+
+	mlog(0, "add %u clusters at position %u to inode %llu\n",
+	     new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno);
+
+	mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) &&
+			(OCFS2_I(inode)->ip_clusters != cpos),
+			"Device %s, asking for sparse allocation: inode %llu, "
+			"cpos %u, clusters %u\n",
+			osb->dev_str,
+			(unsigned long long)OCFS2_I(inode)->ip_blkno, cpos,
+			OCFS2_I(inode)->ip_clusters);
+
+	rec.e_cpos = cpu_to_le32(cpos);
+	rec.e_blkno = cpu_to_le64(start_blk);
+	rec.e_clusters = cpu_to_le32(new_clusters);
+
+	status = ocfs2_figure_insert_type(inode, fe_bh, &last_eb_bh, &rec,
+					  &insert);
+	if (status < 0) {
+		mlog_errno(status);
+		goto bail;
 	}
 
-	/* Can we allocate without adding/shifting tree bits? */
-	i = le16_to_cpu(el->l_next_free_rec) - 1;
-	if (le16_to_cpu(el->l_next_free_rec) == 0
-	    || (le16_to_cpu(el->l_next_free_rec) < le16_to_cpu(el->l_count))
-	    || le32_to_cpu(el->l_recs[i].e_clusters) == 0
-	    || ocfs2_extent_contig(inode, &el->l_recs[i], start_blk))
-		goto out_add;
+	mlog(0, "Insert.appending: %u, Insert.Contig: %u, "
+	     "Insert.contig_index: %d, Insert.free_records: %d, "
+	     "Insert.tree_depth: %d\n",
+	     insert.ins_appending, insert.ins_contig, insert.ins_contig_index,
+	     insert.ins_free_records, insert.ins_tree_depth);
 
-	mlog(0, "ocfs2_allocate_extent: couldn't do a simple add, traversing "
-	     "tree now.\n");
+	/*
+	 * Avoid growing the tree unless we're out of records and the
+	 * insert type requres one.
+	 */
+	if (insert.ins_contig != CONTIG_NONE || insert.ins_free_records)
+		goto out_add;
 
 	shift = ocfs2_find_branch_target(osb, inode, fe_bh, &bh);
 	if (shift < 0) {
@@ -866,13 +2344,9 @@ int ocfs2_insert_extent(struct ocfs2_super *osb,
 	 * and didn't find room for any more extents - we need to add
 	 * another tree level */
 	if (shift) {
-		/* if we hit a leaf, we'd better be empty :) */
-		BUG_ON(le16_to_cpu(el->l_next_free_rec) !=
-		       le16_to_cpu(el->l_count));
 		BUG_ON(bh);
-		mlog(0, "ocfs2_allocate_extent: need to shift tree depth "
-		     "(current = %u)\n",
-		     le16_to_cpu(fe->id2.i_list.l_tree_depth));
+		mlog(0, "need to shift tree depth "
+		     "(current = %d)\n", insert.ins_tree_depth);
 
 		/* ocfs2_shift_tree_depth will return us a buffer with
 		 * the new extent block (so we can pass that to
@@ -883,15 +2357,16 @@ int ocfs2_insert_extent(struct ocfs2_super *osb,
 			mlog_errno(status);
 			goto bail;
 		}
+		insert.ins_tree_depth++;
 		/* Special case: we have room now if we shifted from
 		 * tree_depth 0 */
-		if (fe->id2.i_list.l_tree_depth == cpu_to_le16(1))
+		if (insert.ins_tree_depth == 1)
 			goto out_add;
 	}
 
 	/* call ocfs2_add_branch to add the final part of the tree with
 	 * the new data. */
-	mlog(0, "ocfs2_allocate_extent: add branch. bh = %p\n", bh);
+	mlog(0, "add branch. bh = %p\n", bh);
 	status = ocfs2_add_branch(osb, handle, inode, fe_bh, bh, last_eb_bh,
 				  meta_ac);
 	if (status < 0) {
@@ -900,9 +2375,8 @@ int ocfs2_insert_extent(struct ocfs2_super *osb,
 	}
 
 out_add:
-	/* Finally, we can add clusters. */
-	status = ocfs2_do_insert_extent(osb, handle, inode, fe_bh,
-					start_blk, new_clusters);
+	/* Finally, we can add clusters. This might rotate the tree for us. */
+	status = ocfs2_do_insert_extent(inode, handle, fe_bh, &rec, &insert);
 	if (status < 0)
 		mlog_errno(status);
 
@@ -1447,140 +2921,141 @@ int ocfs2_truncate_log_init(struct ocfs2_super *osb)
  * block will be deleted, and if it will, what the new last extent
  * block will be so we can update his h_next_leaf_blk field, as well
  * as the dinodes i_last_eb_blk */
-static int ocfs2_find_new_last_ext_blk(struct ocfs2_super *osb,
-				       struct inode *inode,
-				       struct ocfs2_dinode *fe,
+static int ocfs2_find_new_last_ext_blk(struct inode *inode,
 				       u32 new_i_clusters,
-				       struct buffer_head *old_last_eb,
+				       struct ocfs2_path *path,
 				       struct buffer_head **new_last_eb)
 {
-	int i, status = 0;
-	u64 block = 0;
+	int ret = 0;
+	u32 cpos;
 	struct ocfs2_extent_block *eb;
 	struct ocfs2_extent_list *el;
 	struct buffer_head *bh = NULL;
 
 	*new_last_eb = NULL;
 
-	if (!OCFS2_IS_VALID_DINODE(fe)) {
-		OCFS2_RO_ON_INVALID_DINODE(inode->i_sb, fe);
-		status = -EIO;
-		goto bail;
-	}
-
 	/* we have no tree, so of course, no last_eb. */
-	if (!fe->id2.i_list.l_tree_depth)
-		goto bail;
+	if (!path->p_tree_depth)
+		goto out;
 
 	/* trunc to zero special case - this makes tree_depth = 0
 	 * regardless of what it is.  */
 	if (!new_i_clusters)
-		goto bail;
+		goto out;
 
-	eb = (struct ocfs2_extent_block *) old_last_eb->b_data;
-	el = &(eb->h_list);
+	el = path_leaf_el(path);
 	BUG_ON(!el->l_next_free_rec);
 
 	/* Make sure that this guy will actually be empty after we
 	 * clear away the data. */
-	if (le32_to_cpu(el->l_recs[0].e_cpos) < new_i_clusters)
-		goto bail;
+	if (ocfs2_is_empty_extent(&el->l_recs[0])) {
+		if (le16_to_cpu(el->l_next_free_rec) > 1 &&
+		    le32_to_cpu(el->l_recs[1].e_cpos) < new_i_clusters)
+			goto out;
+	} else if (le32_to_cpu(el->l_recs[0].e_cpos) < new_i_clusters)
+		goto out;
 
-	/* Ok, at this point, we know that last_eb will definitely
-	 * change, so lets traverse the tree and find the second to
-	 * last extent block. */
-	el = &(fe->id2.i_list);
-	/* go down the tree, */
-	do {
-		for(i = (le16_to_cpu(el->l_next_free_rec) - 1); i >= 0; i--) {
-			if (le32_to_cpu(el->l_recs[i].e_cpos) <
-			    new_i_clusters) {
-				block = le64_to_cpu(el->l_recs[i].e_blkno);
-				break;
-			}
-		}
-		BUG_ON(i < 0);
+	ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
 
-		if (bh) {
-			brelse(bh);
-			bh = NULL;
-		}
+	ret = ocfs2_find_leaf(inode, path_root_el(path), cpos, &bh);
+	if (ret) {
+		mlog_errno(ret);
+		goto out;
+	}
 
-		status = ocfs2_read_block(osb, block, &bh, OCFS2_BH_CACHED,
-					 inode);
-		if (status < 0) {
-			mlog_errno(status);
-			goto bail;
-		}
-		eb = (struct ocfs2_extent_block *) bh->b_data;
-		el = &eb->h_list;
-		if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
-			OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
-			status = -EIO;
-			goto bail;
-		}
-	} while (el->l_tree_depth);
+	eb = (struct ocfs2_extent_block *) bh->b_data;
+	el = &eb->h_list;
+	if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) {
+		OCFS2_RO_ON_INVALID_EXTENT_BLOCK(inode->i_sb, eb);
+		ret = -EROFS;
+		goto out;
+	}
 
 	*new_last_eb = bh;
 	get_bh(*new_last_eb);
-	mlog(0, "returning block %llu\n",
-	     (unsigned long long)le64_to_cpu(eb->h_blkno));
-bail:
-	if (bh)
-		brelse(bh);
+	mlog(0, "returning block %llu, (cpos: %u)\n",
+	     (unsigned long long)le64_to_cpu(eb->h_blkno), cpos);
+out:
+	brelse(bh);
 
-	return status;
+	return ret;
 }
 
 static int ocfs2_do_truncate(struct ocfs2_super *osb,
 			     unsigned int clusters_to_del,
 			     struct inode *inode,
 			     struct buffer_head *fe_bh,
-			     struct buffer_head *old_last_eb_bh,
 			     handle_t *handle,
-			     struct ocfs2_truncate_context *tc)
+			     struct ocfs2_truncate_context *tc,
+			     struct ocfs2_path *path)
 {
-	int status, i, depth;
+	int status, i, index;
 	struct ocfs2_dinode *fe;
 	struct ocfs2_extent_block *eb;
 	struct ocfs2_extent_block *last_eb = NULL;
 	struct ocfs2_extent_list *el;
 	struct buffer_head *eb_bh = NULL;
 	struct buffer_head *last_eb_bh = NULL;
-	u64 next_eb = 0;
 	u64 delete_blk = 0;
 
 	fe = (struct ocfs2_dinode *) fe_bh->b_data;
 
-	status = ocfs2_find_new_last_ext_blk(osb,
-					     inode,
-					     fe,
+	status = ocfs2_find_new_last_ext_blk(inode,
 					     le32_to_cpu(fe->i_clusters) -
-					     		clusters_to_del,
-					     old_last_eb_bh,
-					     &last_eb_bh);
+					     clusters_to_del,
+					     path, &last_eb_bh);
 	if (status < 0) {
 		mlog_errno(status);
 		goto bail;
 	}
-	if (last_eb_bh)
+
+	/*
+	 * Each component will be touched, so we might as well journal
+	 * here to avoid having to handle errors later.
+	 */
+	for (i = 0; i < path_num_items(path); i++) {
+		status = ocfs2_journal_access(handle, inode,
+					      path->p_node[i].bh,
+					      OCFS2_JOURNAL_ACCESS_WRITE);
+		if (status < 0) {
+			mlog_errno(status);
+			goto bail;
+		}
+	}
+
+	if (last_eb_bh) {
+		status = ocfs2_journal_access(handle, inode, last_eb_bh,
+					      OCFS2_JOURNAL_ACCESS_WRITE);
+		if (status < 0) {
+			mlog_errno(status);
+			goto bail;
+		}
+
 		last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
+	}
 
-	status = ocfs2_journal_access(handle, inode, fe_bh,
-				      OCFS2_JOURNAL_ACCESS_WRITE);
-	if (status < 0) {
-		mlog_errno(status);
+	el = &(fe->id2.i_list);
+
+	/*
+	 * Lower levels depend on this never happening, but it's best
+	 * to check it up here before changing the tree.
+	 */
+	if (el->l_tree_depth && ocfs2_is_empty_extent(&el->l_recs[0])) {
+		ocfs2_error(inode->i_sb,
+			    "Inode %lu has an empty extent record, depth %u\n",
+			    inode->i_ino, le16_to_cpu(el->l_tree_depth));
 		goto bail;
 	}
-	el = &(fe->id2.i_list);
 
 	spin_lock(&OCFS2_I(inode)->ip_lock);
 	OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) -
 				      clusters_to_del;
 	spin_unlock(&OCFS2_I(inode)->ip_lock);
 	le32_add_cpu(&fe->i_clusters, -clusters_to_del);
-	fe->i_mtime = cpu_to_le64(CURRENT_TIME.tv_sec);
-	fe->i_mtime_nsec = cpu_to_le32(CURRENT_TIME.tv_nsec);
 
 	i = le16_to_cpu(el->l_next_free_rec) - 1;
 
@@ -1589,26 +3064,34 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
 	/* tree depth zero, we can just delete the clusters, otherwise
 	 * we need to record the offset of the next level extent block
 	 * as we may overwrite it. */
-	if (!el->l_tree_depth)
+	if (!el->l_tree_depth) {
 		delete_blk = le64_to_cpu(el->l_recs[i].e_blkno)
 			+ ocfs2_clusters_to_blocks(osb->sb,
 					le32_to_cpu(el->l_recs[i].e_clusters));
-	else
-		next_eb = le64_to_cpu(el->l_recs[i].e_blkno);
 
-	if (!el->l_recs[i].e_clusters) {
-		/* if we deleted the whole extent record, then clear
-		 * out the other fields and update the extent
-		 * list. For depth > 0 trees, we've already recorded
-		 * the extent block in 'next_eb' */
-		el->l_recs[i].e_cpos = 0;
-		el->l_recs[i].e_blkno = 0;
-		BUG_ON(!el->l_next_free_rec);
-		le16_add_cpu(&el->l_next_free_rec, -1);
+		if (!el->l_recs[i].e_clusters) {
+			/* if we deleted the whole extent record, then clear
+			 * out the other fields and update the extent
+			 * list.
+			 */
+			el->l_recs[i].e_cpos = 0;
+			el->l_recs[i].e_blkno = 0;
+			BUG_ON(!el->l_next_free_rec);
+			le16_add_cpu(&el->l_next_free_rec, -1);
+
+			/*
+			 * The leftmost record might be an empty extent -
+			 * delete it here too.
+			 */
+			if (i == 1 && ocfs2_is_empty_extent(&el->l_recs[0])) {
+				el->l_recs[0].e_cpos = 0;
+				el->l_recs[0].e_blkno = 0;
+				el->l_next_free_rec = 0;
+			}
+		}
 	}
 
-	depth = le16_to_cpu(el->l_tree_depth);
-	if (!fe->i_clusters) {
+	if (le32_to_cpu(fe->i_clusters) == 0) {
 		/* trunc to zero is a special case. */
 		el->l_tree_depth = 0;
 		fe->i_last_eb_blk = 0;
@@ -1625,12 +3108,6 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
 		/* If there will be a new last extent block, then by
 		 * definition, there cannot be any leaves to the right of
 		 * him. */
-		status = ocfs2_journal_access(handle, inode, last_eb_bh,
-					      OCFS2_JOURNAL_ACCESS_WRITE);
-		if (status < 0) {
-			mlog_errno(status);
-			goto bail;
-		}
 		last_eb->h_next_leaf_blk = 0;
 		status = ocfs2_journal_dirty(handle, last_eb_bh);
 		if (status < 0) {
@@ -1639,33 +3116,25 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
 		}
 	}
 
+	index = 1;
 	/* if our tree depth > 0, update all the tree blocks below us. */
-	while (depth) {
-		mlog(0, "traveling tree (depth = %d, next_eb = %llu)\n",
-		     depth,  (unsigned long long)next_eb);
-		status = ocfs2_read_block(osb, next_eb, &eb_bh,
-					  OCFS2_BH_CACHED, inode);
-		if (status < 0) {
-			mlog_errno(status);
-			goto bail;
-		}
+	while (index <= path->p_tree_depth) {
+		eb_bh = path->p_node[index].bh;
 		eb = (struct ocfs2_extent_block *)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;
-		}
-		el = &(eb->h_list);
+		el = path->p_node[index].el;
 
-		status = ocfs2_journal_access(handle, inode, eb_bh,
-					      OCFS2_JOURNAL_ACCESS_WRITE);
-		if (status < 0) {
-			mlog_errno(status);
-			goto bail;
-		}
+		mlog(0, "traveling tree (index = %d, extent block: %llu)\n",
+		     index,  (unsigned long long)eb_bh->b_blocknr);
 
 		BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0);
-		BUG_ON(depth != (le16_to_cpu(el->l_tree_depth) + 1));
+		if (index !=
+		    (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) {
+			ocfs2_error(inode->i_sb,
+				    "Inode %lu has invalid ext. block %llu\n",
+				    inode->i_ino,
+				    (unsigned long long)eb_bh->b_blocknr);
+			goto bail;
+		}
 
 		i = le16_to_cpu(el->l_next_free_rec) - 1;
 
@@ -1680,7 +3149,6 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
 		BUG_ON(le32_to_cpu(el->l_recs[i].e_clusters) < clusters_to_del);
 		le32_add_cpu(&el->l_recs[i].e_clusters, -clusters_to_del);
 
-		next_eb = le64_to_cpu(el->l_recs[i].e_blkno);
 		/* bottom-most block requires us to delete data.*/
 		if (!el->l_tree_depth)
 			delete_blk = le64_to_cpu(el->l_recs[i].e_blkno)
@@ -1692,6 +3160,12 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
 			BUG_ON(!el->l_next_free_rec);
 			le16_add_cpu(&el->l_next_free_rec, -1);
 		}
+		if (i == 1 && ocfs2_is_empty_extent(&el->l_recs[0])) {
+			el->l_recs[0].e_cpos = 0;
+			el->l_recs[0].e_blkno = 0;
+			el->l_next_free_rec = 0;
+		}
+
 		mlog(0, "extent block %llu, after: record %d: "
 		     "(%u, %u, %llu), next = %u\n",
 		     (unsigned long long)le64_to_cpu(eb->h_blkno), i,
@@ -1714,6 +3188,22 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
 			BUG_ON(el->l_recs[0].e_clusters);
 			BUG_ON(el->l_recs[0].e_cpos);
 			BUG_ON(el->l_recs[0].e_blkno);
+
+			/*
+			 * We need to remove this extent block from
+			 * the list above it.
+			 *
+			 * Since we've passed it already in this loop,
+			 * no need to worry about journaling.
+			 */
+			el = path->p_node[index - 1].el;
+			i = le16_to_cpu(el->l_next_free_rec) - 1;
+			BUG_ON(i < 0);
+			el->l_recs[i].e_cpos = 0;
+			el->l_recs[i].e_clusters = 0;
+			el->l_recs[i].e_blkno = 0;
+			le16_add_cpu(&el->l_next_free_rec, -1);
+
 			if (eb->h_suballoc_slot == 0) {
 				/*
 				 * This code only understands how to
@@ -1736,9 +3226,7 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
 				}
 			}
 		}
-		brelse(eb_bh);
-		eb_bh = NULL;
-		depth--;
+		index++;
 	}
 
 	BUG_ON(!delete_blk);
@@ -1750,10 +3238,7 @@ static int ocfs2_do_truncate(struct ocfs2_super *osb,
 	}
 	status = 0;
 bail:
-	if (!status)
-		ocfs2_extent_map_trunc(inode, le32_to_cpu(fe->i_clusters));
-	else
-		ocfs2_extent_map_drop(inode, 0);
+
 	mlog_exit(status);
 	return status;
 }
@@ -1770,82 +3255,70 @@ int ocfs2_commit_truncate(struct ocfs2_super *osb,
 			  struct ocfs2_truncate_context *tc)
 {
 	int status, i, credits, tl_sem = 0;
-	u32 clusters_to_del, target_i_clusters;
-	u64 last_eb = 0;
-	struct ocfs2_dinode *fe;
-	struct ocfs2_extent_block *eb;
+	u32 clusters_to_del, new_highest_cpos, range;
 	struct ocfs2_extent_list *el;
-	struct buffer_head *last_eb_bh;
 	handle_t *handle = NULL;
 	struct inode *tl_inode = osb->osb_tl_inode;
+	struct ocfs2_path *path = NULL;
 
 	mlog_entry_void();
 
 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
 
-	target_i_clusters = ocfs2_clusters_for_bytes(osb->sb,
+	new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb,
 						     i_size_read(inode));
 
-	last_eb_bh = tc->tc_last_eb_bh;
-	tc->tc_last_eb_bh = NULL;
-
-	fe = (struct ocfs2_dinode *) fe_bh->b_data;
-
-	if (fe->id2.i_list.l_tree_depth) {
-		eb = (struct ocfs2_extent_block *) last_eb_bh->b_data;
-		el = &eb->h_list;
-	} else
-		el = &fe->id2.i_list;
-	last_eb = le64_to_cpu(fe->i_last_eb_blk);
+	path = ocfs2_new_inode_path(fe_bh);
+	if (!path) {
+		status = -ENOMEM;
+		mlog_errno(status);
+		goto bail;
+	}
 start:
-	mlog(0, "ocfs2_commit_truncate: fe->i_clusters = %u, "
-	     "last_eb = %llu, fe->i_last_eb_blk = %llu, "
-	     "fe->id2.i_list.l_tree_depth = %u last_eb_bh = %p\n",
-	     le32_to_cpu(fe->i_clusters), (unsigned long long)last_eb,
-	     (unsigned long long)le64_to_cpu(fe->i_last_eb_blk),
-	     le16_to_cpu(fe->id2.i_list.l_tree_depth), last_eb_bh);
-
-	if (last_eb != le64_to_cpu(fe->i_last_eb_blk)) {
-		mlog(0, "last_eb changed!\n");
-		BUG_ON(!fe->id2.i_list.l_tree_depth);
-		last_eb = le64_to_cpu(fe->i_last_eb_blk);
-		/* i_last_eb_blk may have changed, read it if
-		 * necessary. We don't have to worry about the
-		 * truncate to zero case here (where there becomes no
-		 * last_eb) because we never loop back after our work
-		 * is done. */
-		if (last_eb_bh) {
-			brelse(last_eb_bh);
-			last_eb_bh = NULL;
-		}
-
-		status = ocfs2_read_block(osb, last_eb,
-					  &last_eb_bh, OCFS2_BH_CACHED,
-					  inode);
-		if (status < 0) {
-			mlog_errno(status);
-			goto bail;
-		}
-		eb = (struct ocfs2_extent_block *) last_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;
-		}
-		el = &(eb->h_list);
+	/*
+	 * Truncate always works against the rightmost tree branch.
+	 */
+	status = ocfs2_find_path(inode, path, UINT_MAX);
+	if (status) {
+		mlog_errno(status);
+		goto bail;
 	}
 
-	/* by now, el will point to the extent list on the bottom most
-	 * portion of this tree. */
+	mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n",
+	     OCFS2_I(inode)->ip_clusters, path->p_tree_depth);
+
+	/*
+	 * By now, el will point to the extent list on the bottom most
+	 * portion of this tree. Only the tail record is considered in
+	 * each pass.
+	 *
+	 * We handle the following cases, in order:
+	 * - empty extent: delete the remaining branch
+	 * - remove the entire record
+	 * - remove a partial record
+	 * - no record needs to be removed (truncate has completed)
+	 */
+	el = path_leaf_el(path);
 	i = le16_to_cpu(el->l_next_free_rec) - 1;
-	if (le32_to_cpu(el->l_recs[i].e_cpos) >= target_i_clusters)
+	range = le32_to_cpu(el->l_recs[i].e_cpos) +
+		le32_to_cpu(el->l_recs[i].e_clusters);
+	if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) {
+		clusters_to_del = 0;
+	} else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) {
 		clusters_to_del = le32_to_cpu(el->l_recs[i].e_clusters);
-	else
+	} else if (range > new_highest_cpos) {
 		clusters_to_del = (le32_to_cpu(el->l_recs[i].e_clusters) +
 				   le32_to_cpu(el->l_recs[i].e_cpos)) -
-				  target_i_clusters;
+				  new_highest_cpos;
+	} else {
+		status = 0;
+		goto bail;
+	}
 
-	mlog(0, "clusters_to_del = %u in this pass\n", clusters_to_del);
+	mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n",
+	     clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr);
+
+	BUG_ON(clusters_to_del == 0);
 
 	mutex_lock(&tl_inode->i_mutex);
 	tl_sem = 1;
@@ -1861,7 +3334,8 @@ start:
 	}
 
 	credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del,
-						fe, el);
+						(struct ocfs2_dinode *)fe_bh->b_data,
+						el);
 	handle = ocfs2_start_trans(osb, credits);
 	if (IS_ERR(handle)) {
 		status = PTR_ERR(handle);
@@ -1870,13 +3344,8 @@ start:
 		goto bail;
 	}
 
-	inode->i_ctime = inode->i_mtime = CURRENT_TIME;
-	status = ocfs2_mark_inode_dirty(handle, inode, fe_bh);
-	if (status < 0)
-		mlog_errno(status);
-
-	status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh,
-				   last_eb_bh, handle, tc);
+	status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle,
+				   tc, path);
 	if (status < 0) {
 		mlog_errno(status);
 		goto bail;
@@ -1888,8 +3357,12 @@ start:
 	ocfs2_commit_trans(osb, handle);
 	handle = NULL;
 
-	BUG_ON(le32_to_cpu(fe->i_clusters) < target_i_clusters);
-	if (le32_to_cpu(fe->i_clusters) > target_i_clusters)
+	ocfs2_reinit_path(path, 1);
+
+	/*
+	 * Only loop if we still have allocation.
+	 */
+	if (OCFS2_I(inode)->ip_clusters)
 		goto start;
 bail:
 	up_write(&OCFS2_I(inode)->ip_alloc_sem);
@@ -1902,8 +3375,7 @@ bail:
 	if (handle)
 		ocfs2_commit_trans(osb, handle);
 
-	if (last_eb_bh)
-		brelse(last_eb_bh);
+	ocfs2_free_path(path);
 
 	/* This will drop the ext_alloc cluster lock for us */
 	ocfs2_free_truncate_context(tc);
@@ -1912,7 +3384,6 @@ bail:
 	return status;
 }
 
-
 /*
  * Expects the inode to already be locked. This will figure out which
  * inodes need to be locked and will put them on the returned truncate
@@ -1923,7 +3394,7 @@ int ocfs2_prepare_truncate(struct ocfs2_super *osb,
 			   struct buffer_head *fe_bh,
 			   struct ocfs2_truncate_context **tc)
 {
-	int status, metadata_delete;
+	int status, metadata_delete, i;
 	unsigned int new_i_clusters;
 	struct ocfs2_dinode *fe;
 	struct ocfs2_extent_block *eb;
@@ -1944,7 +3415,8 @@ int ocfs2_prepare_truncate(struct ocfs2_super *osb,
 	     "%llu\n", fe->i_clusters, new_i_clusters,
 	     (unsigned long long)fe->i_size);
 
-	if (le32_to_cpu(fe->i_clusters) <= new_i_clusters) {
+	if (!ocfs2_sparse_alloc(osb) &&
+	    le32_to_cpu(fe->i_clusters) <= new_i_clusters) {
 		ocfs2_error(inode->i_sb, "Dinode %llu has cluster count "
 			    "%u and size %llu whereas struct inode has "
 			    "cluster count %u and size %llu which caused an "
@@ -1986,7 +3458,15 @@ int ocfs2_prepare_truncate(struct ocfs2_super *osb,
 			goto bail;
 		}
 		el = &(eb->h_list);
-		if (le32_to_cpu(el->l_recs[0].e_cpos) >= new_i_clusters)
+
+		i = 0;
+		if (ocfs2_is_empty_extent(&el->l_recs[0]))
+			i = 1;
+		/*
+		 * XXX: Should we check that next_free_rec contains
+		 * the extent?
+		 */
+		if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_i_clusters)
 			metadata_delete = 1;
 	}
 

fs/ocfs2/alloc.h

@@ -31,7 +31,8 @@ int ocfs2_insert_extent(struct ocfs2_super *osb,
 			handle_t *handle,
 			struct inode *inode,
 			struct buffer_head *fe_bh,
-			u64 blkno,
+			u32 cpos,
+			u64 start_blk,
 			u32 new_clusters,
 			struct ocfs2_alloc_context *meta_ac);
 int ocfs2_num_free_extents(struct ocfs2_super *osb,

fs/ocfs2/dir.c

@@ -365,8 +365,10 @@ int ocfs2_do_extend_dir(struct super_block *sb,
 	spin_unlock(&OCFS2_I(dir)->ip_lock);
 
 	if (extend) {
-		status = ocfs2_do_extend_allocation(OCFS2_SB(sb), dir, 1,
-						    parent_fe_bh, handle,
+		u32 offset = OCFS2_I(dir)->ip_clusters;
+
+		status = ocfs2_do_extend_allocation(OCFS2_SB(sb), dir, &offset,
+						    1, parent_fe_bh, handle,
 						    data_ac, meta_ac, NULL);
 		BUG_ON(status == -EAGAIN);
 		if (status < 0) {

fs/ocfs2/file.c

@@ -397,6 +397,7 @@ bail:
  */
 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
 			       struct inode *inode,
+			       u32 *logical_offset,
 			       u32 clusters_to_add,
 			       struct buffer_head *fe_bh,
 			       handle_t *handle,
@@ -460,18 +461,14 @@ int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
 	block = ocfs2_clusters_to_blocks(osb->sb, bit_off);
 	mlog(0, "Allocating %u clusters at block %u for inode %llu\n",
 	     num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno);
-	status = ocfs2_insert_extent(osb, handle, inode, fe_bh, block,
-				     num_bits, meta_ac);
+	status = ocfs2_insert_extent(osb, handle, inode, fe_bh,
+				     *logical_offset, block, num_bits,
+				     meta_ac);
 	if (status < 0) {
 		mlog_errno(status);
 		goto leave;
 	}
 
-	le32_add_cpu(&fe->i_clusters, num_bits);
-	spin_lock(&OCFS2_I(inode)->ip_lock);
-	OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters);
-	spin_unlock(&OCFS2_I(inode)->ip_lock);
-
 	status = ocfs2_journal_dirty(handle, fe_bh);
 	if (status < 0) {
 		mlog_errno(status);
@@ -479,6 +476,7 @@ int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
 	}
 
 	clusters_to_add -= num_bits;
+	*logical_offset += num_bits;
 
 	if (clusters_to_add) {
 		mlog(0, "need to alloc once more, clusters = %u, wanted = "
@@ -501,7 +499,7 @@ static int ocfs2_extend_allocation(struct inode *inode,
 	int restart_func = 0;
 	int drop_alloc_sem = 0;
 	int credits, num_free_extents;
-	u32 prev_clusters;
+	u32 prev_clusters, logical_start;
 	struct buffer_head *bh = NULL;
 	struct ocfs2_dinode *fe = NULL;
 	handle_t *handle = NULL;
@@ -512,6 +510,12 @@ static int ocfs2_extend_allocation(struct inode *inode,
 
 	mlog_entry("(clusters_to_add = %u)\n", clusters_to_add);
 
+	/*
+	 * This function only exists for file systems which don't
+	 * support holes.
+	 */
+	BUG_ON(ocfs2_sparse_alloc(osb));
+
 	status = ocfs2_read_block(osb, OCFS2_I(inode)->ip_blkno, &bh,
 				  OCFS2_BH_CACHED, inode);
 	if (status < 0) {
@@ -526,6 +530,8 @@ static int ocfs2_extend_allocation(struct inode *inode,
 		goto leave;
 	}
 
+	logical_start = OCFS2_I(inode)->ip_clusters;
+
 restart_all:
 	BUG_ON(le32_to_cpu(fe->i_clusters) != OCFS2_I(inode)->ip_clusters);
 
@@ -590,6 +596,7 @@ restarted_transaction:
 
 	status = ocfs2_do_extend_allocation(osb,
 					    inode,
+					    &logical_start,
 					    clusters_to_add,
 					    bh,
 					    handle,

fs/ocfs2/file.h

@@ -39,6 +39,7 @@ enum ocfs2_alloc_restarted {
 };
 int ocfs2_do_extend_allocation(struct ocfs2_super *osb,
 			       struct inode *inode,
+			       u32 *cluster_start,
 			       u32 clusters_to_add,
 			       struct buffer_head *fe_bh,
 			       handle_t *handle,

fs/ocfs2/namei.c

@@ -1671,8 +1671,11 @@ static int ocfs2_symlink(struct inode *dir,
 	inode->i_rdev = 0;
 	newsize = l - 1;
 	if (l > ocfs2_fast_symlink_chars(sb)) {
+		u32 offset = 0;
+
 		inode->i_op = &ocfs2_symlink_inode_operations;
-		status = ocfs2_do_extend_allocation(osb, inode, 1, new_fe_bh,
+		status = ocfs2_do_extend_allocation(osb, inode, &offset, 1,
+						    new_fe_bh,
 						    handle, data_ac, NULL,
 						    NULL);
 		if (status < 0) {

fs/ocfs2/ocfs2.h

@@ -303,6 +303,13 @@ static inline int ocfs2_should_order_data(struct inode *inode)
 	return 1;
 }
 
+static inline int ocfs2_sparse_alloc(struct ocfs2_super *osb)
+{
+	if (osb->s_feature_incompat & OCFS2_FEATURE_INCOMPAT_SPARSE_ALLOC)
+		return 1;
+	return 0;
+}
+
 /* set / clear functions because cluster events can make these happen
  * in parallel so we want the transitions to be atomic. this also
  * means that any future flags osb_flags must be protected by spinlock

fs/ocfs2/ocfs2_fs.h

@@ -86,7 +86,8 @@
 	OCFS2_SB(sb)->s_feature_incompat &= ~(mask)
 
 #define OCFS2_FEATURE_COMPAT_SUPP	OCFS2_FEATURE_COMPAT_BACKUP_SB
-#define OCFS2_FEATURE_INCOMPAT_SUPP	OCFS2_FEATURE_INCOMPAT_LOCAL_MOUNT
+#define OCFS2_FEATURE_INCOMPAT_SUPP	(OCFS2_FEATURE_INCOMPAT_LOCAL_MOUNT \
+					 | OCFS2_FEATURE_INCOMPAT_SPARSE_ALLOC)
 #define OCFS2_FEATURE_RO_COMPAT_SUPP	0
 
 /*
@@ -311,7 +312,10 @@ struct ocfs2_extent_list {
 /*00*/	__le16 l_tree_depth;		/* Extent tree depth from this
 					   point.  0 means data extents
 					   hang directly off this
-					   header (a leaf) */
+					   header (a leaf)
+					   NOTE: The high 8 bits cannot be
+					   used - tree_depth is never that big.
+					*/
 	__le16 l_count;			/* Number of extent records */
 	__le16 l_next_free_rec;		/* Next unused extent slot */
 	__le16 l_reserved1;