ha_innopart.cc

/*****************************************************************************

Copyright (c) 2014, 2017, Oracle and/or its affiliates. All rights reserved.
Copyright (c) 2016, 2017, MariaDB Corporation.

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; version 2 of the License.

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.,
51 Franklin Street, Suite 500, Boston, MA 02110-1335 USA

*****************************************************************************/

/** @file ha_innopart.cc
Code for native partitioning in InnoDB.

Created Nov 22, 2013 Mattias Jonsson */

#include "univ.i"

/* Include necessary SQL headers */
#include <debug_sync.h>
#include <log.h>
#include <strfunc.h>
#include <sql_acl.h>
#include <sql_class.h>
#include <sql_show.h>
#include <sql_table.h>
#include <my_check_opt.h>

/* Include necessary InnoDB headers */
#include "btr0sea.h"
#include "dict0dict.h"
#include "dict0stats.h"
#include "lock0lock.h"
#include "row0import.h"
#include "row0merge.h"
#include "row0mysql.h"
#include "row0quiesce.h"
#include "row0sel.h"
#include "row0ins.h"
#include "row0upd.h"
#include "fsp0sysspace.h"
#include "ut0ut.h"

#include "ha_innodb.h"
#include "ha_innopart.h"
#include "partition_info.h"
#include "key.h"

#define INSIDE_HA_INNOPART_CC

/* To be backwards compatible we also fold partition separator on windows. */
#ifdef _WIN32
static const char* part_sep = "#p#";
static const char* sub_sep = "#sp#";
#else
static const char* part_sep = "#P#";
static const char* sub_sep = "#SP#";
#endif /* _WIN32 */

/* Partition separator for *nix platforms */
const char* part_sep_nix = "#P#";
const char* sub_sep_nix = "#SP#";

extern char*	innobase_file_format_max;

Ha_innopart_share::Ha_innopart_share(
	TABLE_SHARE*	table_share)
	:
	Partition_share(),
	m_table_parts(),
	m_index_mapping(),
	m_tot_parts(),
	m_index_count(),
	m_ref_count(),
	m_table_share(table_share)
{}

Ha_innopart_share::~Ha_innopart_share()
{
	ut_ad(m_ref_count == 0);
	if (m_table_parts != NULL) {
		ut_free(m_table_parts);
		m_table_parts = NULL;
	}
	if (m_index_mapping != NULL) {
		ut_free(m_index_mapping);
		m_index_mapping = NULL;
	}
}

/** Fold to lower case if windows or lower_case_table_names == 1.
@param[in,out]	s	String to fold.*/
void
Ha_innopart_share::partition_name_casedn_str(
	char*	s)
{
#ifdef _WIN32
	innobase_casedn_str(s);
#endif
}

/** Translate and append partition name.
@param[out]	to	String to write in filesystem charset
@param[in]	from	Name in system charset
@param[in]	sep	Separator
@param[in]	len	Max length of to buffer
@return	length of written string. */
size_t
Ha_innopart_share::append_sep_and_name(
	char*		to,
	const char*	from,
	const char*	sep,
	size_t		len)
{
	size_t	ret;
	size_t	sep_len = strlen(sep);

	ut_ad(len > sep_len + strlen(from));
	ut_ad(to != NULL);
	ut_ad(from != NULL);
	ut_ad(from[0] != '\0');
	memcpy(to, sep, sep_len);

	ret = tablename_to_filename(from, to + sep_len,
		len - sep_len);

	/* Don't convert to lower case for nix style name. */
	if (strcmp(sep, part_sep_nix) != 0
	    && strcmp(sep, sub_sep_nix) != 0) {

		partition_name_casedn_str(to);
	}

	return(ret + sep_len);
}

/** Copy a cached MySQL row.
If requested, also avoids overwriting non-read columns.
@param[out]	buf		Row in MySQL format.
@param[in]	cached_row	Which row to copy. */
inline
void
ha_innopart::copy_cached_row(
	uchar*		buf,
	const uchar*	cached_row)
{
	if (m_prebuilt->keep_other_fields_on_keyread) {
		row_sel_copy_cached_fields_for_mysql(buf, cached_row,
			m_prebuilt);
	} else {
		memcpy(buf, cached_row, m_rec_length);
	}
}

/** Open one partition.
@param[in]	part_id		Partition id to open.
@param[in]	partition_name	Name of internal innodb table to open.
@return	false on success else true. */
bool
Ha_innopart_share::open_one_table_part(
	uint		part_id,
	const char*	partition_name)
{
	char	norm_name[FN_REFLEN];

	normalize_table_name(norm_name, partition_name);
	m_table_parts[part_id] =
		ha_innobase::open_dict_table(partition_name, norm_name,
					     TRUE, DICT_ERR_IGNORE_NONE);

	if (m_table_parts[part_id] == NULL) {
		return(true);
	}

	dict_table_t *ib_table = m_table_parts[part_id];
	if ((!DICT_TF2_FLAG_IS_SET(ib_table, DICT_TF2_FTS_HAS_DOC_ID)
	     && m_table_share->fields
		 != (dict_table_get_n_user_cols(ib_table)
		     + dict_table_get_n_v_cols(ib_table)))
	    || (DICT_TF2_FLAG_IS_SET(ib_table, DICT_TF2_FTS_HAS_DOC_ID)
		&& (m_table_share->fields
		    != dict_table_get_n_user_cols(ib_table)
		       + dict_table_get_n_v_cols(ib_table) - 1))) {
		ib::warn() << "Partition `" << get_partition_name(part_id)
			<< "` contains " << dict_table_get_n_user_cols(ib_table)
			<< " user defined columns in InnoDB, but "
			<< m_table_share->fields
			<< " columns in MySQL. Please check"
			" INFORMATION_SCHEMA.INNODB_SYS_COLUMNS and " REFMAN
			"innodb-troubleshooting.html for how to resolve the"
			" issue.";

		/* Mark this partition as corrupted, so the drop table
		or force recovery can still use it, but not others.
		TODO: persist table->corrupted so it will be retained on
		restart and out-of-bounds operations will see it. */

		ib_table->corrupted = true;
		dict_table_close(ib_table, FALSE, FALSE);
	}

	/* TODO: To save memory, compare with first partition and reuse
	the column names etc. in the internal InnoDB meta-data cache. */

	return(false);
}

/** Set up the virtual column template for partition table, and points
all m_table_parts[]->vc_templ to it.
@param[in]	table		MySQL TABLE object
@param[in]	ib_table	InnoDB dict_table_t
@param[in]	table_name	Table name (db/table_name) */
void
Ha_innopart_share::set_v_templ(
	TABLE*		table,
	dict_table_t*	ib_table,
	const char*	name)
{
	ut_ad(mutex_own(&dict_sys->mutex));

	if (ib_table->n_v_cols > 0) {
		for (ulint i = 0; i < m_tot_parts; i++) {
			if (m_table_parts[i]->vc_templ == NULL) {
				m_table_parts[i]->vc_templ
					= UT_NEW_NOKEY(dict_vcol_templ_t());
				m_table_parts[i]->vc_templ->vtempl = NULL;
			} else if (m_table_parts[i]->get_ref_count() == 1) {
				/* Clean and refresh the template */
				dict_free_vc_templ(m_table_parts[i]->vc_templ);
				m_table_parts[i]->vc_templ->vtempl = NULL;
			}

			if (m_table_parts[i]->vc_templ->vtempl == NULL) {
				innobase_build_v_templ(
					table, ib_table,
					m_table_parts[i]->vc_templ,
					NULL, true, name);
			}
		}
	}
}

/** Initialize the share with table and indexes per partition.
@param[in]	part_info	Partition info (partition names to use).
@param[in]	table_name	Table name (db/table_name).
@return	false on success else true. */
bool
Ha_innopart_share::open_table_parts(
	partition_info*	part_info,
	const char*	table_name)
{
	size_t	table_name_len;
	size_t	len;
	uint	ib_num_index;
	uint	mysql_num_index;
	char	partition_name[FN_REFLEN];
	bool	index_loaded = true;

#ifndef DBUG_OFF
	if (m_table_share->tmp_table == NO_TMP_TABLE) {
		mysql_mutex_assert_owner(&m_table_share->LOCK_ha_data);
	}
#endif /* DBUG_OFF */
	m_ref_count++;
	if (m_table_parts != NULL) {
		ut_ad(m_ref_count > 1);
		ut_ad(m_tot_parts > 0);

		/* Increment dict_table_t reference count for all partitions */
		mutex_enter(&dict_sys->mutex);
		for (uint i = 0; i < m_tot_parts; i++) {
			dict_table_t*	table = m_table_parts[i];
			table->acquire();
			ut_ad(table->get_ref_count() >= m_ref_count);
		}
		mutex_exit(&dict_sys->mutex);

		return(false);
	}
	ut_ad(m_ref_count == 1);
	m_tot_parts = part_info->get_tot_partitions();
	size_t	table_parts_size = sizeof(dict_table_t*) * m_tot_parts;
	m_table_parts = static_cast<dict_table_t**>(
		ut_zalloc(table_parts_size, mem_key_partitioning));
	if (m_table_parts == NULL) {
		m_ref_count--;
		return(true);
	}

	/* Set up the array over all table partitions. */
	table_name_len = strlen(table_name);
	memcpy(partition_name, table_name, table_name_len);
	List_iterator<partition_element>
				part_it(part_info->partitions);
	partition_element*	part_elem;
	uint			i = 0;

	while ((part_elem = part_it++)) {
		len = append_sep_and_name(
				partition_name + table_name_len,
				part_elem->partition_name,
				part_sep_nix,
				FN_REFLEN - table_name_len);
		if (part_info->is_sub_partitioned()) {
			List_iterator<partition_element>
				sub_it(part_elem->subpartitions);
			partition_element*	sub_elem;
			while ((sub_elem = sub_it++)) {
				append_sep_and_name(
					partition_name
					+ table_name_len + len,
					sub_elem->partition_name,
					sub_sep_nix,
					FN_REFLEN - table_name_len - len);
				if (open_one_table_part(i, partition_name)) {
					goto err;
				}
				i++;
			}
		} else {
			if (open_one_table_part(i, partition_name)) {
				goto err;
			}
			i++;
		}
	}
	ut_ad(i == m_tot_parts);

	/* Create the mapping of mysql index number to innodb indexes. */

	ib_num_index = (uint) UT_LIST_GET_LEN(m_table_parts[0]->indexes);
	mysql_num_index = part_info->table->s->keys;

	/* If there exists inconsistency between MySQL and InnoDB dictionary
	(metadata) information, the number of index defined in MySQL
	could exceed that in InnoDB, do not build index translation
	table in such case. */

	if (ib_num_index < mysql_num_index) {
		ut_ad(0);
		goto err;
	}

	if (mysql_num_index != 0) {
		size_t	alloc_size = mysql_num_index * m_tot_parts
			* sizeof(*m_index_mapping);
		m_index_mapping = static_cast<dict_index_t**>(
			ut_zalloc(alloc_size, mem_key_partitioning));
		if (m_index_mapping == NULL) {

			/* Report an error if index_mapping continues to be
			NULL and mysql_num_index is a non-zero value. */

			ib::error() << "Failed to allocate memory for"
				" index translation table. Number of"
				" Index:" << mysql_num_index;
			goto err;
		}
	}

	/* For each index in the mysql key_info array, fetch its
	corresponding InnoDB index pointer into index_mapping
	array. */

	for (ulint idx = 0; idx < mysql_num_index; idx++) {
		for (ulint part = 0; part < m_tot_parts; part++) {
			ulint	count = part * mysql_num_index + idx;

			/* Fetch index pointers into index_mapping according
			to mysql index sequence. */

			m_index_mapping[count] = dict_table_get_index_on_name(
				m_table_parts[part],
				part_info->table->key_info[idx].name);

			if (m_index_mapping[count] == NULL) {
				ib::error() << "Cannot find index `"
					<< part_info->table->key_info[idx].name
					<< "` in InnoDB index dictionary"
					" partition `"
					<< get_partition_name(part) << "`.";
				index_loaded = false;
				break;
			}

			/* Double check fetched index has the same
			column info as those in mysql key_info. */

			if (!innobase_match_index_columns(
					&part_info->table->key_info[idx],
					m_index_mapping[count])) {
				ib::error() << "Found index `"
					<< part_info->table->key_info[idx].name
					<< "` whose column info does not match"
					" that of MySQL.";
				index_loaded = false;
				break;
			}
		}
	}
	if (!index_loaded && m_index_mapping != NULL) {
		ut_free(m_index_mapping);
		m_index_mapping = NULL;
	}

	/* Successfully built the translation table. */
	m_index_count = mysql_num_index;

	return(false);
err:
	close_table_parts();

	return(true);
}

/** Close all partitions. */
void
Ha_innopart_share::close_table_parts()
{
#ifndef DBUG_OFF
	if (m_table_share->tmp_table == NO_TMP_TABLE) {
		mysql_mutex_assert_owner(&m_table_share->LOCK_ha_data);
	}
#endif /* DBUG_OFF */
	m_ref_count--;
	if (m_ref_count != 0) {

		/* Decrement dict_table_t reference count for all partitions */
		mutex_enter(&dict_sys->mutex);
		for (uint i = 0; i < m_tot_parts; i++) {
			dict_table_t*	table = m_table_parts[i];
			table->release();
			ut_ad(table->get_ref_count() >= m_ref_count);
		}
		mutex_exit(&dict_sys->mutex);

		return;
	}

	/* Last instance closed, close all table partitions and
	free the memory. */

	mutex_enter(&dict_sys->mutex);
	if (m_table_parts != NULL) {
		for (uint i = 0; i < m_tot_parts; i++) {
			if (m_table_parts[i] != NULL) {
				dict_table_close(m_table_parts[i], TRUE, TRUE);
			}
		}
		ut_free(m_table_parts);
		m_table_parts = NULL;
	}
	mutex_exit(&dict_sys->mutex);
	if (m_index_mapping != NULL) {
		ut_free(m_index_mapping);
		m_index_mapping = NULL;
	}

	m_tot_parts = 0;
	m_index_count = 0;
}

/** Get index.
Find the index of the specified partition and key number.
@param[in]	part_id	Partition number.
@param[in]	keynr	Key number.
@return	Index pointer or NULL. */
inline
dict_index_t*
Ha_innopart_share::get_index(
	uint	part_id,
	uint	keynr)
{
	ut_a(part_id < m_tot_parts);
	ut_ad(keynr < m_index_count || keynr == MAX_KEY);
	if (m_index_mapping == NULL
	    || keynr >= m_index_count) {

		if (keynr == MAX_KEY) {
			return(dict_table_get_first_index(
				get_table_part(part_id)));
		}
		return(NULL);
	}
	return(m_index_mapping[m_index_count * part_id + keynr]);
}

/** Get MySQL key number corresponding to InnoDB index.
Calculates the key number used inside MySQL for an Innobase index. We will
first check the "index translation table" for a match of the index to get
the index number. If there does not exist an "index translation table",
or not able to find the index in the translation table, then we will fall back
to the traditional way of looping through dict_index_t list to find a
match. In this case, we have to take into account if we generated a
default clustered index for the table
@param[in]	part_id	Partition the index belongs to.
@param[in]	index	Index to return MySQL key number for.
@return	the key number used inside MySQL or UINT_MAX if key is not found. */
inline
uint
Ha_innopart_share::get_mysql_key(
	uint			part_id,
	const dict_index_t*	index)
{
	ut_ad(index != NULL);
	ut_ad(m_index_mapping != NULL);
	ut_ad(m_tot_parts);

	if (index != NULL && m_index_mapping != NULL) {
		uint	start;
		uint	end;

		if (part_id < m_tot_parts) {
			start = part_id * m_index_count;
			end = start + m_index_count;
		} else {
			start = 0;
			end = m_tot_parts * m_index_count;
		}
		for (uint i = start; i < end; i++) {
			if (m_index_mapping[i] == index) {
				return(i % m_index_count);
			}
		}

		/* Print an error message if we cannot find the index
		in the "index translation table". */

		if (index->is_committed()) {
			ib::error() << "Cannot find index "
				<< index->name
				<< " in InnoDB index translation table.";
		}
	}

	return(UINT_MAX);
}

/** Helper function for set bit in bitmap.
@param[in,out]	buf	Bitmap buffer to update bit in.
@param[in]	bit_pos	Bit number (index starts at 0). */
static
inline
void
set_bit(
	byte*	buf,
	size_t	pos)
{
	buf[pos/8] |= (0x1 << (pos & 0x7));
}

/** Helper function for clear bit in bitmap.
@param[in,out]	buf	Bitmap buffer to update bit in.
@param[in]	bit_pos	Bit number (index starts at 0). */
static
inline
void
clear_bit(
	byte*	buf,
	size_t	pos)
{
	buf[pos/8] &= ~(0x1 << (pos & 0x7));
}

/** Helper function for get bit in bitmap.
@param[in,out]	buf	Bitmap buffer.
@param[in]	bit_pos	Bit number (index starts at 0).
@return	byte set to 0x0 or 0x1.
@retval	0x0 bit not set.
@retval	0x1 bet set. */
static
inline
byte
get_bit(
	byte*	buf,
	size_t	pos)
{
	return((buf[pos/8] >> (pos & 0x7)) & 0x1);
}

/** Helper class for encapsulating new/altered partitions during
ADD/REORG/... PARTITION. */
class Altered_partitions
{
private:
	/** New partitions during ADD/REORG/... PARTITION. */
	dict_table_t**	m_new_table_parts;

	/** Insert nodes per partition. */
	ins_node_t**	m_ins_nodes;

	/** sql_stat_start per partition. */
	byte*		m_sql_stat_start;

	/** Trx id per partition. */
	trx_id_t*	m_trx_ids;

	/** Number of new partitions. */
	size_t		m_num_new_parts;

	/** Only need to create the partitions (no open/lock). */
	bool		m_only_create;

public:
	Altered_partitions(
		uint n_partitions,
		bool only_create);

	~Altered_partitions();

	bool
	initialize();

	bool
	only_create() const
	{
		return(m_only_create);
	}

	/** Set currently used partition.
	@param[in]	new_part_id	Partition id to set.
	@param[in]	part	InnoDB table to use. */
	inline
	void
	set_part(
		ulint		new_part_id,
		dict_table_t*	part)
	{
		ut_ad(m_new_table_parts[new_part_id] == NULL);
		m_new_table_parts[new_part_id] = part;
		set_bit(m_sql_stat_start, new_part_id);
	}

	/** Get lower level InnoDB table for partition.
	@param[in]	part_id	Partition id.
	@return Lower level InnoDB table for the partition id. */
	inline
	dict_table_t*
	part(
		uint	part_id) const
	{
		ut_ad(part_id < m_num_new_parts);
		return(m_new_table_parts[part_id]);
	}

	/** Set up prebuilt for using a specified partition.
	@param[in]	prebuilt	Prebuilt to update.
	@param[in]	new_part_id	Partition to use. */
	inline
	void
	get_prebuilt(
		row_prebuilt_t*	prebuilt,
		uint		new_part_id) const
	{
		ut_ad(m_new_table_parts[new_part_id]);
		prebuilt->table = m_new_table_parts[new_part_id];
		prebuilt->ins_node = m_ins_nodes[new_part_id];
		prebuilt->trx_id = m_trx_ids[new_part_id];
		prebuilt->sql_stat_start = get_bit(m_sql_stat_start,
						new_part_id);
	}

	/** Update cached values for a partition from prebuilt.
	@param[in]	prebuilt	Prebuilt to copy from.
	@param[in]	new_part_id	Partition id to copy. */
	inline
	void
	set_from_prebuilt(
		row_prebuilt_t*	prebuilt,
		uint		new_part_id)
	{
		ut_ad(m_new_table_parts[new_part_id] == prebuilt->table);
		m_ins_nodes[new_part_id] = prebuilt->ins_node;
		m_trx_ids[new_part_id] = prebuilt->trx_id;
		if (prebuilt->sql_stat_start == 0) {
			clear_bit(m_sql_stat_start, new_part_id);
		}
	}
};

Altered_partitions::Altered_partitions(
		uint n_partitions,
		bool only_create)
		:
		m_new_table_parts(),
		m_ins_nodes(),
		m_sql_stat_start(),
		m_trx_ids(),
		m_num_new_parts(n_partitions),
		m_only_create(only_create)
	{}

Altered_partitions::~Altered_partitions()
{
	if (m_new_table_parts != NULL) {
		for (ulint i = 0; i < m_num_new_parts; i++) {
			if (m_new_table_parts[i] != NULL) {
				dict_table_close(m_new_table_parts[i],
					false, true);
			}
		}
		ut_free(m_new_table_parts);
		m_new_table_parts = NULL;
	}
	if (m_ins_nodes != NULL) {
		for (ulint i = 0; i < m_num_new_parts; i++) {
			if (m_ins_nodes[i] != NULL) {
				ins_node_t*	ins = m_ins_nodes[i];
				ut_ad(ins->select == NULL);
				que_graph_free_recursive(ins->select);
				ins->select = NULL;
				if (ins->entry_sys_heap != NULL) {
					mem_heap_free(ins->entry_sys_heap);
					ins->entry_sys_heap = NULL;
				}
			}
		}
		ut_free(m_ins_nodes);
		m_ins_nodes = NULL;
	}
	if (m_sql_stat_start != NULL) {
		ut_free(m_sql_stat_start);
		m_sql_stat_start = NULL;
	}
	if (m_trx_ids != NULL) {
		ut_free(m_trx_ids);
		m_trx_ids = NULL;
	}
}

/** Initialize the object.
@return false on success else true. */
bool
Altered_partitions::initialize()
{
	size_t	alloc_size = sizeof(*m_new_table_parts) * m_num_new_parts;
	m_new_table_parts = static_cast<dict_table_t**>(
		ut_zalloc(alloc_size, mem_key_partitioning));
	if (m_new_table_parts == NULL) {
		return(true);
	}

	alloc_size = sizeof(*m_ins_nodes) * m_num_new_parts;
	m_ins_nodes = static_cast<ins_node_t**>(
		ut_zalloc(alloc_size, mem_key_partitioning));
	if (m_ins_nodes == NULL) {
		ut_free(m_new_table_parts);
		m_new_table_parts = NULL;
		return(true);
	}

	alloc_size = sizeof(*m_sql_stat_start)
		* UT_BITS_IN_BYTES(m_num_new_parts);
	m_sql_stat_start = static_cast<byte*>(
		ut_zalloc(alloc_size, mem_key_partitioning));
	if (m_sql_stat_start == NULL) {
		ut_free(m_new_table_parts);
		m_new_table_parts = NULL;
		ut_free(m_ins_nodes);
		m_ins_nodes = NULL;
		return(true);
	}

	alloc_size = sizeof(*m_trx_ids) * m_num_new_parts;
	m_trx_ids = static_cast<trx_id_t*>(
		ut_zalloc(alloc_size, mem_key_partitioning));
	if (m_trx_ids == NULL) {
		ut_free(m_new_table_parts);
		m_new_table_parts = NULL;
		ut_free(m_ins_nodes);
		m_ins_nodes = NULL;
		ut_free(m_sql_stat_start);
		m_sql_stat_start = NULL;
		return(true);
	}

	return(false);
}

/** Construct ha_innopart handler.
@param[in]	hton		Handlerton.
@param[in]	table_arg	MySQL Table.
@return	a new ha_innopart handler. */
ha_innopart::ha_innopart(
	handlerton*	hton,
	TABLE_SHARE*	table_arg)
	:
	ha_innobase(hton, table_arg),
	Partition_helper(this),
	m_ins_node_parts(),
	m_upd_node_parts(),
	m_blob_heap_parts(),
	m_trx_id_parts(),
	m_row_read_type_parts(),
	m_sql_stat_start_parts(),
	m_pcur(),
	m_clust_pcur(),
	m_new_partitions()
{
	m_int_table_flags &=	~(HA_INNOPART_DISABLED_TABLE_FLAGS);

	/* INNOBASE_SHARE is not used in ha_innopart.
	This also flags for ha_innobase that it is a partitioned table.
	And make it impossible to use legacy share functionality. */

	m_share = NULL;
}

/** Destruct ha_innopart handler. */
ha_innopart::~ha_innopart()
{}

/** Returned supported alter table flags.
@param[in]	flags	Flags to support.
@return	Supported flags. */
uint
ha_innopart::alter_table_flags(
	uint	flags)
{
	return(HA_PARTITION_FUNCTION_SUPPORTED | HA_FAST_CHANGE_PARTITION);
}

/** Set the autoinc column max value.
This should only be called once from ha_innobase::open().
Therefore there's no need for a covering lock.
@param[in]	no_lock	Ignored!
@return	0 for success or error code. */
inline
int
ha_innopart::initialize_auto_increment(
	bool	/* no_lock */)
{
	int		error = 0;
	ulonglong	auto_inc = 0;
	const Field*	field = table->found_next_number_field;

#ifndef DBUG_OFF
	if (table_share->tmp_table == NO_TMP_TABLE)
	{
		mysql_mutex_assert_owner(m_part_share->auto_inc_mutex);
	}
#endif

	/* Since a table can already be "open" in InnoDB's internal
	data dictionary, we only init the autoinc counter once, the
	first time the table is loaded. We can safely reuse the
	autoinc value from a previous MySQL open. */

	if (m_part_share->auto_inc_initialized) {
		/* Already initialized, nothing to do. */
		return(0);
	}

	if (field == NULL) {
		ib::info() << "Unable to determine the AUTOINC column name";
	}

	if (srv_force_recovery >= SRV_FORCE_NO_IBUF_MERGE) {
		/* If the recovery level is set so high that writes
		are disabled we force the AUTOINC counter to 0
		value effectively disabling writes to the table.
		Secondly, we avoid reading the table in case the read
		results in failure due to a corrupted table/index.

		We will not return an error to the client, so that the
		tables can be dumped with minimal hassle. If an error
		were returned in this case, the first attempt to read
		the table would fail and subsequent SELECTs would succeed. */

	} else if (field == NULL) {
		/* This is a far more serious error, best to avoid
		opening the table and return failure. */

		my_error(ER_AUTOINC_READ_FAILED, MYF(0));
		error = HA_ERR_AUTOINC_READ_FAILED;
	} else {
		ib_uint64_t	col_max_value = field->get_max_int_value();

		update_thd(ha_thd());

		for (uint part = 0; part < m_tot_parts; part++) {
			dict_table_t*	ib_table
				= m_part_share->get_table_part(part);
			dict_table_autoinc_lock(ib_table);
			ut_ad(ib_table->persistent_autoinc);
			ib_uint64_t	read_auto_inc
				= dict_table_autoinc_read(ib_table);
			if (read_auto_inc == 0) {
				read_auto_inc = btr_read_autoinc(
					dict_table_get_first_index(ib_table));

				/* At the this stage we do not know the
				increment nor the offset,
				so use a default increment of 1. */

				read_auto_inc = innobase_next_autoinc(
					read_auto_inc, 1, 1, 0, col_max_value);
				dict_table_autoinc_initialize(ib_table,
					read_auto_inc);
			}
			set_if_bigger(auto_inc, read_auto_inc);
			dict_table_autoinc_unlock(ib_table);
		}
	}

done:
	m_part_share->next_auto_inc_val = auto_inc;
	m_part_share->auto_inc_initialized = true;
	return(error);
}

/** Opens a partitioned InnoDB table.
Initializes needed data and opens the table which already exists
in an InnoDB database.
@param[in]	name		Table name (db/tablename)
@param[in]	mode		Not used
@param[in]	test_if_locked	Not used
@return	0 or error number. */
int
ha_innopart::open(
	const char*	name,
	int		/*mode*/,
	uint		/*test_if_locked*/)
{
	dict_table_t*	ib_table;
	char		norm_name[FN_REFLEN];
	THD*		thd;

	DBUG_ENTER("ha_innopart::open");

	ut_ad(table);
	if (m_part_info == NULL) {
		/* Must be during ::clone()! */
		ut_ad(table->part_info != NULL);
		m_part_info = table->part_info;
	}
	thd = ha_thd();

	/* Under some cases MySQL seems to call this function while
	holding search latch(es). This breaks the latching order as
	we acquire dict_sys->mutex below and leads to a deadlock. */

	if (thd != NULL) {
		innobase_release_temporary_latches(ht, thd);
	}

	normalize_table_name(norm_name, name);

	m_user_thd = NULL;

	/* Get the Ha_innopart_share from the TABLE_SHARE. */
	lock_shared_ha_data();
	m_part_share = static_cast<Ha_innopart_share*>(get_ha_share_ptr());
	if (m_part_share == NULL) {
		m_part_share = new (std::nothrow)
				Ha_innopart_share(table_share);
		if (m_part_share == NULL) {
share_error:
			unlock_shared_ha_data();
			DBUG_RETURN(HA_ERR_INTERNAL_ERROR);
		}
		set_ha_share_ptr(static_cast<Handler_share*>(m_part_share));
	}
	if (m_part_share->open_table_parts(m_part_info, name)
	    || m_part_share->populate_partition_name_hash(m_part_info)) {
		goto share_error;
	}
	if (m_part_share->auto_inc_mutex == NULL
	    && table->found_next_number_field != NULL) {
		if (m_part_share->init_auto_inc_mutex(table_share)) {
			goto share_error;
		}
	}
	unlock_shared_ha_data();

	/* Will be allocated if it is needed in ::update_row(). */
	m_upd_buf = NULL;
	m_upd_buf_size = 0;

	/* Get pointer to a table object in InnoDB dictionary cache. */
	ib_table = m_part_share->get_table_part(0);

	m_pcur_parts = NULL;
	m_clust_pcur_parts = NULL;
	m_pcur_map = NULL;

	/* TODO: Handle mismatching #P# vs #p# in upgrading to new DD instead!
	See bug#58406, The problem exists when moving partitioned tables
	between Windows and Unix-like platforms. InnoDB always folds the name
	on windows, partitioning never folds partition (and #P# separator).
	I.e. non of it follows lower_case_table_names correctly :( */

	if (open_partitioning(m_part_share))
	{
		close();
		DBUG_RETURN(HA_ERR_INITIALIZATION);
	}

	/* Currently we track statistics for all partitions, but for
	the secondary indexes we only use the biggest partition. */

	for (uint part_id = 0; part_id < m_tot_parts; part_id++) {
		innobase_copy_frm_flags_from_table_share(
			m_part_share->get_table_part(part_id),
			table->s);
		dict_stats_init(m_part_share->get_table_part(part_id));
	}

	MONITOR_INC(MONITOR_TABLE_OPEN);

	bool	no_tablespace;

	/* TODO: Should we do this check for every partition during ::open()? */
	/* TODO: refactor this in ha_innobase so it can increase code reuse. */
	if (dict_table_is_discarded(ib_table)) {

		ib_senderrf(thd,
			IB_LOG_LEVEL_WARN, ER_TABLESPACE_DISCARDED,
			table->s->table_name.str);

		/* Allow an open because a proper DISCARD should have set
		all the flags and index root page numbers to FIL_NULL that
		should prevent any DML from running but it should allow DDL
		operations. */

		no_tablespace = false;

	} else if (ib_table->ibd_file_missing) {

		ib_senderrf(
			thd, IB_LOG_LEVEL_WARN,
			ER_TABLESPACE_MISSING, norm_name);

		/* This means we have no idea what happened to the tablespace
		file, best to play it safe. */

		no_tablespace = true;
	} else {
		no_tablespace = false;
	}

	if (!thd_tablespace_op(thd) && no_tablespace) {
                set_my_errno(ENOENT);

		lock_shared_ha_data();
		m_part_share->close_table_parts();
		unlock_shared_ha_data();
		m_part_share = NULL;

		DBUG_RETURN(HA_ERR_NO_SUCH_TABLE);
	}

	m_prebuilt = row_create_prebuilt(ib_table, table->s->reclength);

	m_prebuilt->default_rec = table->s->default_values;
	ut_ad(m_prebuilt->default_rec);

	DBUG_ASSERT(table != NULL);
	m_prebuilt->m_mysql_table = table;

	if (ib_table->n_v_cols > 0) {
		mutex_enter(&dict_sys->mutex);
		m_part_share->set_v_templ(table, ib_table, name);
		mutex_exit(&dict_sys->mutex);
	}

	/* Looks like MySQL-3.23 sometimes has primary key number != 0. */
	m_primary_key = table->s->primary_key;
	key_used_on_scan = m_primary_key;

	/* Allocate a buffer for a 'row reference'. A row reference is
	a string of bytes of length ref_length which uniquely specifies
	a row in our table. Note that MySQL may also compare two row
	references for equality by doing a simple memcmp on the strings
	of length ref_length! */

	if (!row_table_got_default_clust_index(ib_table)) {

		m_prebuilt->clust_index_was_generated = FALSE;

		if (UNIV_UNLIKELY(m_primary_key >= MAX_KEY)) {
			table_name_t table_name;
			table_name.m_name = const_cast<char*>(name);
			ib::error() << "Table " << table_name
				<< " has a primary key in InnoDB data"
				" dictionary, but not in MySQL!";

			/* This mismatch could cause further problems
			if not attended, bring this to the user's attention
			by printing a warning in addition to log a message
			in the errorlog. */

			push_warning_printf(thd, Sql_condition::SL_WARNING,
					    ER_NO_SUCH_INDEX,
					    "Table %s has a"
					    " primary key in InnoDB data"
					    " dictionary, but not in"
					    " MySQL!", name);

			/* If m_primary_key >= MAX_KEY, its (m_primary_key)
			value could be out of bound if continue to index
			into key_info[] array. Find InnoDB primary index,
			and assign its key_length to ref_length.
			In addition, since MySQL indexes are sorted starting
			with primary index, unique index etc., initialize
			ref_length to the first index key length in
			case we fail to find InnoDB cluster index.

			Please note, this will not resolve the primary
			index mismatch problem, other side effects are
			possible if users continue to use the table.
			However, we allow this table to be opened so
			that user can adopt necessary measures for the
			mismatch while still being accessible to the table
			date. */

			if (table->key_info == NULL) {
				ut_ad(table->s->keys == 0);
				ref_length = 0;
			} else {
				ref_length = table->key_info[0].key_length;
			}

			/* Find corresponding cluster index
			key length in MySQL's key_info[] array. */

			for (uint i = 0; i < table->s->keys; i++) {
				dict_index_t*	index;
				index = innopart_get_index(0, i);
				if (dict_index_is_clust(index)) {
					ref_length =
						 table->key_info[i].key_length;
				}
			}
			ut_a(ref_length);
			ref_length += PARTITION_BYTES_IN_POS;
		} else {
			/* MySQL allocates the buffer for ref.
			key_info->key_length includes space for all key
			columns + one byte for each column that may be
			NULL. ref_length must be as exact as possible to
			save space, because all row reference buffers are
			allocated based on ref_length. */

			ref_length = table->key_info[m_primary_key].key_length;
			ref_length += PARTITION_BYTES_IN_POS;
		}
	} else {
		if (m_primary_key != MAX_KEY) {
			table_name_t table_name;
			table_name.m_name = const_cast<char*>(name);
			ib::error() << "Table " << table_name
				<< " has no primary key in InnoDB data"
				" dictionary, but has one in MySQL! If you"
				" created the table with a MySQL version <"
				" 3.23.54 and did not define a primary key,"
				" but defined a unique key with all non-NULL"
				" columns, then MySQL internally treats that"
				" key as the primary key. You can fix this"
				" error by dump + DROP + CREATE + reimport"
				" of the table.";

			/* This mismatch could cause further problems
			if not attended, bring this to the user attention
			by printing a warning in addition to log a message
			in the errorlog. */

			push_warning_printf(thd, Sql_condition::SL_WARNING,
					    ER_NO_SUCH_INDEX,
					    "InnoDB: Table %s has no"
					    " primary key in InnoDB data"
					    " dictionary, but has one in"
					    " MySQL!", name);
		}

		m_prebuilt->clust_index_was_generated = TRUE;

		ref_length = DATA_ROW_ID_LEN;
		ref_length += PARTITION_BYTES_IN_POS;

		/* If we automatically created the clustered index, then
		MySQL does not know about it, and MySQL must NOT be aware
		of the index used on scan, to make it avoid checking if we
		update the column of the index. That is why we assert below
		that key_used_on_scan is the undefined value MAX_KEY.
		The column is the row id in the automatical generation case,
		and it will never be updated anyway. */

		if (key_used_on_scan != MAX_KEY) {
			table_name_t table_name;
			table_name.m_name = const_cast<char*>(name);
			ib::warn() << "Table " << table_name
				<< " key_used_on_scan is "
				<< key_used_on_scan << " even though there is"
				" no primary key inside InnoDB.";
		}
	}

	/* Index block size in InnoDB: used by MySQL in query optimization. */
	stats.block_size = UNIV_PAGE_SIZE;

	if (m_prebuilt->table != NULL) {
		/* We update the highest file format in the system table
		space, if this table has higher file format setting. */

		trx_sys_file_format_max_upgrade(
			(const char**) &innobase_file_format_max,
			dict_table_get_format(m_prebuilt->table));
	}

	/* Only if the table has an AUTOINC column. */
	if (m_prebuilt->table != NULL
	    && !m_prebuilt->table->ibd_file_missing
	    && table->found_next_number_field != NULL) {
		int	error;

		/* Since a table can already be "open" in InnoDB's internal
		data dictionary, we only init the autoinc counter once, the
		first time the table is loaded,
		see ha_innopart::initialize_auto_increment.
		We can safely reuse the autoinc value from a previous MySQL
		open. */

		lock_auto_increment();
		error = initialize_auto_increment(false);
		unlock_auto_increment();
		if (error != 0) {
			close();
			DBUG_RETURN(error);
		}
	}

#ifdef HA_INNOPART_SUPPORTS_FULLTEXT
	/* Set plugin parser for fulltext index. */
	for (uint i = 0; i < table->s->keys; i++) {
		if (table->key_info[i].flags & HA_USES_PARSER) {
			dict_index_t*	index = innobase_get_index(i);
			plugin_ref	parser = table->key_info[i].parser;

			ut_ad(index->type & DICT_FTS);
			index->parser =
				static_cast<st_mysql_ftparser *>(
					plugin_decl(parser)->info);

			DBUG_EXECUTE_IF("fts_instrument_use_default_parser",
				index->parser = &fts_default_parser;);
		}
	}
#endif /* HA_INNOPART_SUPPORTS_FULLTEXT */

	size_t	alloc_size = sizeof(*m_ins_node_parts) * m_tot_parts;
	m_ins_node_parts = static_cast<ins_node_t**>(
		ut_zalloc(alloc_size, mem_key_partitioning));

	alloc_size = sizeof(*m_upd_node_parts) * m_tot_parts;
	m_upd_node_parts = static_cast<upd_node_t**>(
		ut_zalloc(alloc_size, mem_key_partitioning));

	alloc_blob_heap_array();

	alloc_size = sizeof(*m_trx_id_parts) * m_tot_parts;
	m_trx_id_parts = static_cast<trx_id_t*>(
		ut_zalloc(alloc_size, mem_key_partitioning));

	alloc_size = sizeof(*m_row_read_type_parts) * m_tot_parts;
	m_row_read_type_parts = static_cast<ulint*>(
		ut_zalloc(alloc_size, mem_key_partitioning));

	alloc_size = UT_BITS_IN_BYTES(m_tot_parts);
	m_sql_stat_start_parts = static_cast<uchar*>(
		ut_zalloc(alloc_size, mem_key_partitioning));
	if (m_ins_node_parts == NULL
	    || m_upd_node_parts == NULL
	    || m_blob_heap_parts == NULL
	    || m_trx_id_parts == NULL
	    || m_row_read_type_parts == NULL
	    || m_sql_stat_start_parts == NULL) {
		close();  // Frees all the above.
		DBUG_RETURN(HA_ERR_OUT_OF_MEM);
	}
	info(HA_STATUS_NO_LOCK | HA_STATUS_VARIABLE | HA_STATUS_CONST);

	DBUG_RETURN(0);
}

/** Get a cloned ha_innopart handler.
@param[in]	name		Table name.
@param[in]	mem_root	MySQL mem_root to use.
@return	new ha_innopart handler. */
handler*
ha_innopart::clone(
	const char*	name,
	MEM_ROOT*	mem_root)
{
	ha_innopart*	new_handler;

	DBUG_ENTER("ha_innopart::clone");

	new_handler = dynamic_cast<ha_innopart*>(handler::clone(name,
							mem_root));
	if (new_handler != NULL) {
		ut_ad(new_handler->m_prebuilt != NULL);

		new_handler->m_prebuilt->select_lock_type =
			m_prebuilt->select_lock_type;
	}

	DBUG_RETURN(new_handler);
}

/** Clear used ins_nodes and upd_nodes. */
void ha_innopart::clear_ins_upd_nodes()
{
	/* Free memory from insert nodes. */
	if (m_ins_node_parts != NULL) {
		for (uint i = 0; i < m_tot_parts; i++) {
			if (m_ins_node_parts[i] != NULL) {
				ins_node_t*	ins = m_ins_node_parts[i];
				if (ins->select != NULL) {
					que_graph_free_recursive(ins->select);
					ins->select = NULL;
				}

				if (ins->entry_sys_heap != NULL) {
					mem_heap_free(ins->entry_sys_heap);
					ins->entry_sys_heap = NULL;
				}
				m_ins_node_parts[i] = NULL;
			}
		}
	}

	/* Free memory from update nodes. */
	if (m_upd_node_parts != NULL) {
		for (uint i = 0; i < m_tot_parts; i++) {
			if (m_upd_node_parts[i] != NULL) {
				upd_node_t*	upd = m_upd_node_parts[i];
				if (upd->cascade_top) {
					mem_heap_free(upd->cascade_heap);
					upd->cascade_top = false;
					upd->cascade_heap = NULL;
				}
				if (upd->in_mysql_interface) {
					btr_pcur_free_for_mysql(upd->pcur);
					upd->in_mysql_interface = FALSE;
				}

				if (upd->select != NULL) {
					que_graph_free_recursive(upd->select);
					upd->select = NULL;
				}
				if (upd->heap != NULL) {
					mem_heap_free(upd->heap);
					upd->heap = NULL;
				}
				m_upd_node_parts[i] = NULL;
			}
		}
	}
}

/** Closes a handle to an InnoDB table.
@return	0 */
int
ha_innopart::close()
{
	THD*	thd;

	DBUG_ENTER("ha_innopart::close");

	thd = ha_thd();
	if (thd != NULL) {
		innobase_release_temporary_latches(ht, thd);
	}

	ut_ad(m_pcur_parts == NULL);
	ut_ad(m_clust_pcur_parts == NULL);
	close_partitioning();

	ut_ad(m_part_share != NULL);
	if (m_part_share != NULL) {
		lock_shared_ha_data();
		m_part_share->close_table_parts();
		unlock_shared_ha_data();
		m_part_share = NULL;
	}
	clear_ins_upd_nodes();
	free_blob_heap_array();

	/* Prevent double close of m_prebuilt->table. The real one was done
	done in m_part_share->close_table_parts(). */
	m_prebuilt->table = NULL;
	row_prebuilt_free(m_prebuilt, FALSE);

	if (m_upd_buf != NULL) {
		ut_ad(m_upd_buf_size != 0);
		/* Allocated with my_malloc! */
		my_free(m_upd_buf);
		m_upd_buf = NULL;
		m_upd_buf_size = 0;
	}

	if (m_ins_node_parts != NULL) {
		ut_free(m_ins_node_parts);
		m_ins_node_parts = NULL;
	}
	if (m_upd_node_parts != NULL) {
		ut_free(m_upd_node_parts);
		m_upd_node_parts = NULL;
	}
	if (m_trx_id_parts != NULL) {
		ut_free(m_trx_id_parts);
		m_trx_id_parts = NULL;
	}
	if (m_row_read_type_parts != NULL) {
		ut_free(m_row_read_type_parts);
		m_row_read_type_parts = NULL;
	}
	if (m_sql_stat_start_parts != NULL) {
		ut_free(m_sql_stat_start_parts);
		m_sql_stat_start_parts = NULL;
	}

	MONITOR_INC(MONITOR_TABLE_CLOSE);

	/* Tell InnoDB server that there might be work for
	utility threads: */

	srv_active_wake_master_thread();

	DBUG_RETURN(0);
}

/** Change active partition.
Copies needed info into m_prebuilt from the partition specific memory.
@param[in]	part_id	Partition to set as active. */
void
ha_innopart::set_partition(
	uint	part_id)
{
	DBUG_ENTER("ha_innopart::set_partition");

	DBUG_PRINT("ha_innopart", ("partition id: %u", part_id));

	if (part_id >= m_tot_parts) {
		ut_ad(0);
		DBUG_VOID_RETURN;
	}
	if (m_pcur_parts != NULL) {
		m_prebuilt->pcur = &m_pcur_parts[m_pcur_map[part_id]];
	}
	if (m_clust_pcur_parts != NULL) {
		m_prebuilt->clust_pcur =
			&m_clust_pcur_parts[m_pcur_map[part_id]];
	}
	m_prebuilt->ins_node = m_ins_node_parts[part_id];
	m_prebuilt->upd_node = m_upd_node_parts[part_id];

	/* For unordered scan and table scan, use blob_heap from first
	partition as we need exactly one blob. */
	m_prebuilt->blob_heap = m_blob_heap_parts[m_ordered ? part_id : 0];

#ifdef UNIV_DEBUG
	if (m_prebuilt->blob_heap != NULL) {
		DBUG_PRINT("ha_innopart", ("validating blob_heap: %p",
					   m_prebuilt->blob_heap));
		mem_heap_validate(m_prebuilt->blob_heap);
	}
#endif

	m_prebuilt->trx_id = m_trx_id_parts[part_id];
	m_prebuilt->row_read_type = m_row_read_type_parts[part_id];
	m_prebuilt->sql_stat_start = get_bit(m_sql_stat_start_parts, part_id);
	m_prebuilt->table = m_part_share->get_table_part(part_id);
	m_prebuilt->index = innopart_get_index(part_id, active_index);

	DBUG_VOID_RETURN;
}

/** Update active partition.
Copies needed info from m_prebuilt into the partition specific memory.
@param[in]	part_id	Partition to set as active. */
void
ha_innopart::update_partition(
	uint	part_id)
{
	DBUG_ENTER("ha_innopart::update_partition");
	DBUG_PRINT("ha_innopart", ("partition id: %u", part_id));

	if (part_id >= m_tot_parts) {
		ut_ad(0);
		DBUG_VOID_RETURN;
	}
	m_ins_node_parts[part_id] = m_prebuilt->ins_node;
	m_upd_node_parts[part_id] = m_prebuilt->upd_node;

#ifdef UNIV_DEBUG
	if (m_prebuilt->blob_heap != NULL) {
		DBUG_PRINT("ha_innopart", ("validating blob_heap: %p",
					   m_prebuilt->blob_heap));
		mem_heap_validate(m_prebuilt->blob_heap);
	}
#endif

	/* For unordered scan and table scan, use blob_heap from first
	partition as we need exactly one blob anytime. */
	m_blob_heap_parts[m_ordered ? part_id : 0] = m_prebuilt->blob_heap;

	m_trx_id_parts[part_id] = m_prebuilt->trx_id;
	m_row_read_type_parts[part_id] = m_prebuilt->row_read_type;
	if (m_prebuilt->sql_stat_start == 0) {
		clear_bit(m_sql_stat_start_parts, part_id);
	}
	m_last_part = part_id;
	DBUG_VOID_RETURN;
}

/** Was the last returned row semi consistent read.
In an UPDATE or DELETE, if the row under the cursor was locked by
another transaction, and the engine used an optimistic read of the last
committed row value under the cursor, then the engine returns 1 from
this function. MySQL must NOT try to update this optimistic value. If
the optimistic value does not match the WHERE condition, MySQL can
decide to skip over this row. This can be used to avoid unnecessary
lock waits.

If this method returns true, it will also signal the storage
engine that the next read will be a locking re-read of the row.
@see handler.h and row0mysql.h
@return	true if last read was semi consistent else false. */
bool
ha_innopart::was_semi_consistent_read()
{
	return(m_row_read_type_parts[m_last_part]
		== ROW_READ_DID_SEMI_CONSISTENT);
}

/** Try semi consistent read.
Tell the engine whether it should avoid unnecessary lock waits.
If yes, in an UPDATE or DELETE, if the row under the cursor was locked
by another transaction, the engine may try an optimistic read of
the last committed row value under the cursor.
@see handler.h and row0mysql.h
@param[in]	yes	Should semi-consistent read be used. */
void
ha_innopart::try_semi_consistent_read(
	bool	yes)
{
	ha_innobase::try_semi_consistent_read(yes);
	for (uint i = m_part_info->get_first_used_partition();
	     i < m_tot_parts;
	     i = m_part_info->get_next_used_partition(i)) {

		m_row_read_type_parts[i] = m_prebuilt->row_read_type;
	}
}

/** Removes a lock on a row.
Removes a new lock set on a row, if it was not read optimistically.
This can be called after a row has been read in the processing of
an UPDATE or a DELETE query. @see ha_innobase::unlock_row(). */
void
ha_innopart::unlock_row()
{
	ut_ad(m_last_part < m_tot_parts);
	set_partition(m_last_part);
	ha_innobase::unlock_row();
	update_partition(m_last_part);
}

/** Write a row in partition.
Stores a row in an InnoDB database, to the table specified in this
handle.
@param[in]	part_id	Partition to write to.
@param[in]	record	A row in MySQL format.
@return	0 or error code. */
int
ha_innopart::write_row_in_part(
	uint	part_id,
	uchar*	record)
{
	int	error;
	Field*	saved_next_number_field = table->next_number_field;
	DBUG_ENTER("ha_innopart::write_row_in_part");
	set_partition(part_id);

	/* Prevent update_auto_increment to be called
	again in ha_innobase::write_row(). */

	table->next_number_field = NULL;

	/* TODO: try to avoid creating a new dtuple
	(in row_get_prebuilt_insert_row()) for each partition).
	Might be needed due to ins_node implementation. */

	error = ha_innobase::write_row(record);
	update_partition(part_id);
	table->next_number_field = saved_next_number_field;
	DBUG_RETURN(error);
}

/** Update a row in partition.
Updates a row given as a parameter to a new value.
@param[in]	part_id	Partition to update row in.
@param[in]	old_row	Old row in MySQL format.
@param[in]	new_row	New row in MySQL format.
@return	0 or error number. */
int
ha_innopart::update_row_in_part(
	uint		part_id,
	const uchar*	old_row,
	uchar*		new_row)
{
	int	     error;
	DBUG_ENTER("ha_innopart::update_row_in_part");

	set_partition(part_id);
	error = ha_innobase::update_row(old_row, new_row);
	update_partition(part_id);
	DBUG_RETURN(error);
}

/** Deletes a row in partition.
@param[in]	part_id	Partition to delete from.
@param[in]	record	Row to delete in MySQL format.
@return	0 or error number. */
int
ha_innopart::delete_row_in_part(
	uint		part_id,
	const uchar*	record)
{
	int	error;
	DBUG_ENTER("ha_innopart::delete_row_in_part");
	m_err_rec = NULL;

	m_last_part = part_id;
	set_partition(part_id);
	error = ha_innobase::delete_row(record);
	update_partition(part_id);
	DBUG_RETURN(error);
}

/** Initializes a handle to use an index.
@param[in]	keynr	Key (index) number.
@param[in]	sorted	True if result MUST be sorted according to index.
@return	0 or error number. */
int
ha_innopart::index_init(
	uint	keynr,
	bool	sorted)
{
	int	error;
	uint	part_id = m_part_info->get_first_used_partition();
	DBUG_ENTER("ha_innopart::index_init");

	active_index = keynr;
	if (part_id == MY_BIT_NONE) {
		DBUG_RETURN(0);
	}

	error = ph_index_init_setup(keynr, sorted);
	if (error != 0) {
		DBUG_RETURN(error);
	}

	if (sorted) {
		error = init_record_priority_queue();
		if (error != 0) {
			/* Needs cleanup in case it returns error. */
			destroy_record_priority_queue();
			DBUG_RETURN(error);
		}
		/* Disable prefetch.
		The prefetch buffer is not partitioning aware, so it may return
		rows from a different partition if either the prefetch buffer is
		full, or it is non-empty and the partition is exhausted. */
		m_prebuilt->m_no_prefetch = true;
	}

	/* For scan across partitions, the keys needs to be materialized */
	m_prebuilt->m_read_virtual_key = true;

	error = change_active_index(part_id, keynr);
	if (error != 0) {
		destroy_record_priority_queue();
		DBUG_RETURN(error);
	}

	DBUG_EXECUTE_IF("partition_fail_index_init", {
		destroy_record_priority_queue();
		DBUG_RETURN(HA_ERR_NO_PARTITION_FOUND);
	});

	DBUG_RETURN(0);
}

/** End index cursor.
@return	0 or error code. */
int
ha_innopart::index_end()
{
	uint	part_id = m_part_info->get_first_used_partition();
	DBUG_ENTER("ha_innopart::index_end");

	if (part_id == MY_BIT_NONE) {
		/* Never initialized any index. */
		active_index = MAX_KEY;
		DBUG_RETURN(0);
	}
	if (m_ordered) {
		destroy_record_priority_queue();
		m_prebuilt->m_no_prefetch = false;
	}
	m_prebuilt->m_read_virtual_key = false;

	DBUG_RETURN(ha_innobase::index_end());
}

/* Partitioning support functions. */

/** Setup the ordered record buffer and the priority queue.
@param[in]	used_parts	Number of used partitions in query.
@return	false for success else true. */
int
ha_innopart::init_record_priority_queue_for_parts(
	uint	used_parts)
{
	size_t	alloc_size;
	void*	buf;

	DBUG_ENTER("ha_innopart::init_record_priority_queue_for_parts");
	ut_ad(used_parts >= 1);
	/* TODO: Don't use this if only one partition is used! */
	//ut_ad(used_parts > 1);

	/* We could reuse current m_prebuilt->pcur/clust_pcur for the first
	used partition, but it would complicate and affect performance,
	so we trade some extra memory instead. */

	m_pcur = m_prebuilt->pcur;
	m_clust_pcur = m_prebuilt->clust_pcur;

	/* If we searching for secondary key or doing a write/update
	we will need two pcur, one for the active (secondary) index and
	one for the clustered index. */

	bool	need_clust_index =
			m_curr_key_info[1] != NULL
			|| get_lock_type() != F_RDLCK;

	/* pcur and clust_pcur per partition.
	By using zalloc, we do not need to initialize the pcur's! */

	alloc_size = used_parts * sizeof(btr_pcur_t);
	if (need_clust_index) {
		alloc_size *= 2;
	}
	buf = ut_zalloc(alloc_size, mem_key_partitioning);
	if (buf == NULL) {
		DBUG_RETURN(true);
	}
	m_pcur_parts = static_cast<btr_pcur_t*>(buf);
	if (need_clust_index) {
		m_clust_pcur_parts = &m_pcur_parts[used_parts];
	}
	/* mapping from part_id to pcur. */
	alloc_size = m_tot_parts * sizeof(*m_pcur_map);
	buf = ut_zalloc(alloc_size, mem_key_partitioning);
	if (buf == NULL) {
		DBUG_RETURN(true);
	}
	m_pcur_map = static_cast<uint16_t*>(buf);
	{
		uint16_t pcur_count = 0;
		for (uint i = m_part_info->get_first_used_partition();
		     i < m_tot_parts;
		     i = m_part_info->get_next_used_partition(i)) {
			m_pcur_map[i] = pcur_count++;
		}
	}

	DBUG_RETURN(false);
}

/** Destroy the ordered record buffer and the priority queue. */
inline
void
ha_innopart::destroy_record_priority_queue_for_parts()
{
	DBUG_ENTER("ha_innopart::destroy_record_priority_queue");
	if (m_pcur_parts != NULL) {
		uint	used_parts;
		used_parts = bitmap_bits_set(&m_part_info->read_partitions);
		for (uint i = 0; i < used_parts; i++) {
			btr_pcur_free(&m_pcur_parts[i]);
			if (m_clust_pcur_parts != NULL) {
				btr_pcur_free(&m_clust_pcur_parts[i]);
			}
		}
		ut_free(m_pcur_parts);
		m_clust_pcur_parts = NULL;
		m_pcur_parts = NULL;
		/* Reset the original m_prebuilt->pcur. */
		m_prebuilt->pcur = m_pcur;
		m_prebuilt->clust_pcur = m_clust_pcur;
	}
	if (m_pcur_map != NULL) {
		ut_free(m_pcur_map);
		m_pcur_map = NULL;
	}
	DBUG_VOID_RETURN;
}

/** Print error information.
@param[in]	error	Error code (MySQL).
@param[in]	errflag	Flags. */
void
ha_innopart::print_error(
	int	error,
	myf	errflag)
{
	DBUG_ENTER("ha_innopart::print_error");
	if (print_partition_error(error, errflag)) {
		ha_innobase::print_error(error, errflag);
	}

	DBUG_VOID_RETURN;
}

/** Can error be ignored.
@param[in]	error	Error code to check.
@return	true if ignorable else false. */
bool
ha_innopart::is_ignorable_error(
	int	error)
{
	if (ha_innobase::is_ignorable_error(error)
	    || error == HA_ERR_NO_PARTITION_FOUND
	    || error == HA_ERR_NOT_IN_LOCK_PARTITIONS) {

		return(true);
	}
	return(false);
}

/** Get the index for the current partition
@param[in]	keynr	MySQL index number.
@return	InnoDB index or NULL. */
inline
dict_index_t*
ha_innopart::innobase_get_index(
	uint	keynr)
{
	uint	part_id = m_last_part;
	if (part_id >= m_tot_parts) {
		ut_ad(0);
		part_id = 0;
	}
	return(innopart_get_index(part_id, keynr));
}

/** Get the index for a handle.
Does not change active index.
@param[in]	keynr	Use this index; MAX_KEY means always clustered index,
even if it was internally generated by InnoDB.
@param[in]	part_id	From this partition.
@return	NULL or index instance. */
inline
dict_index_t*
ha_innopart::innopart_get_index(
	uint	part_id,
	uint	keynr)
{
	KEY*		key = NULL;
	dict_index_t*	index = NULL;

	DBUG_ENTER("innopart_get_index");

	if (keynr != MAX_KEY && table->s->keys > 0) {
		key = table->key_info + keynr;

		index = m_part_share->get_index(part_id, keynr);

		if (index != NULL) {
			ut_a(ut_strcmp(index->name, key->name) == 0);
		} else {
			/* Can't find index with keynr in the translation
			table. Only print message if the index translation
			table exists. */

			ib::warn() << "InnoDB could not find index "
				<< (key ? key->name : "NULL")
				<< " key no " << keynr << " for table "
				<< m_prebuilt->table->name
				<< " through its index translation table";

			index = dict_table_get_index_on_name(m_prebuilt->table,
							     key->name);
		}
	} else {
		/* Get the generated index. */
		ut_ad(keynr == MAX_KEY);
		index = dict_table_get_first_index(
				m_part_share->get_table_part(part_id));
	}

	if (index == NULL) {
		ib::error() << "InnoDB could not find key n:o "
			<< keynr << " with name " << (key ? key->name : "NULL")
			<< " from dict cache for table "
			<< m_prebuilt->table->name << " partition n:o "
			<< part_id;
	}

	DBUG_RETURN(index);
}

/** Changes the active index of a handle.
@param[in]	part_id	Use this partition.
@param[in]	keynr	Use this index; MAX_KEY means always clustered index,
even if it was internally generated by InnoDB.
@return	0 or error number. */
int
ha_innopart::change_active_index(
	uint	part_id,
	uint	keynr)
{
	DBUG_ENTER("ha_innopart::change_active_index");

	ut_ad(m_user_thd == ha_thd());
	ut_a(m_prebuilt->trx == thd_to_trx(m_user_thd));

	active_index = keynr;
	set_partition(part_id);

	if (UNIV_UNLIKELY(m_prebuilt->index == NULL)) {
		ib::warn() << "change_active_index(" << part_id
			<< "," << keynr << ") failed";
		m_prebuilt->index_usable = FALSE;
		DBUG_RETURN(1);
	}

	m_prebuilt->index_usable = row_merge_is_index_usable(m_prebuilt->trx,
							   m_prebuilt->index);

	if (UNIV_UNLIKELY(!m_prebuilt->index_usable)) {
		if (dict_index_is_corrupted(m_prebuilt->index)) {
			char table_name[MAX_FULL_NAME_LEN + 1];

			innobase_format_name(
				table_name, sizeof table_name,
				m_prebuilt->index->table->name.m_name);

			push_warning_printf(
				m_user_thd, Sql_condition::SL_WARNING,
				HA_ERR_INDEX_CORRUPT,
				"InnoDB: Index %s for table %s is"
				" marked as corrupted"
				" (partition %u)",
				m_prebuilt->index->name(), table_name, part_id);
			DBUG_RETURN(HA_ERR_INDEX_CORRUPT);
		} else {
			push_warning_printf(
				m_user_thd, Sql_condition::SL_WARNING,
				HA_ERR_TABLE_DEF_CHANGED,
				"InnoDB: insufficient history for index %u",
				keynr);
		}

		/* The caller seems to ignore this. Thus, we must check
		this again in row_search_for_mysql(). */

		DBUG_RETURN(HA_ERR_TABLE_DEF_CHANGED);
	}

	ut_a(m_prebuilt->search_tuple != NULL);

	/* If too expensive, cache the keynr and only update search_tuple when
	keynr changes. Remember that the clustered index is also used for
	MAX_KEY. */
	dtuple_set_n_fields(m_prebuilt->search_tuple,
		m_prebuilt->index->n_fields);

	dict_index_copy_types(m_prebuilt->search_tuple, m_prebuilt->index,
			m_prebuilt->index->n_fields);

	/* MySQL changes the active index for a handle also during some
	queries, for example SELECT MAX(a), SUM(a) first retrieves the
	MAX() and then calculates the sum. Previously we played safe
	and used the flag ROW_MYSQL_WHOLE_ROW below, but that caused
	unnecessary copying. Starting from MySQL-4.1 we use a more
	efficient flag here. */

	/* TODO: Is this really needed?
	Will it not be built in index_read? */

	build_template(false);

	DBUG_RETURN(0);
}

/** Return first record in index from a partition.
@param[in]	part	Partition to read from.
@param[out]	record	First record in index in the partition.
@return	error number or 0. */
int
ha_innopart::index_first_in_part(
	uint	part,
	uchar*	record)
{
	int	error;
	DBUG_ENTER("ha_innopart::index_first_in_part");

	set_partition(part);
	error = ha_innobase::index_first(record);
	update_partition(part);

	DBUG_RETURN(error);
}

/** Return next record in index from a partition.
@param[in]	part	Partition to read from.
@param[out]	record	Last record in index in the partition.
@return	error number or 0. */
int
ha_innopart::index_next_in_part(
	uint	part,
	uchar*	record)
{
	DBUG_ENTER("ha_innopart::index_next_in_part");

	int	error;

	set_partition(part);
	error = ha_innobase::index_next(record);
	update_partition(part);

	ut_ad(m_ordered_scan_ongoing
	      || m_ordered_rec_buffer == NULL
	      || m_prebuilt->used_in_HANDLER
	      || m_part_spec.start_part >= m_part_spec.end_part);

	DBUG_RETURN(error);
}

/** Return next same record in index from a partition.
This routine is used to read the next record, but only if the key is
the same as supplied in the call.
@param[in]	part	Partition to read from.
@param[out]	record	Last record in index in the partition.
@param[in]	key	Key to match.
@param[in]	length	Length of key.
@return	error number or 0. */
int
ha_innopart::index_next_same_in_part(
	uint		part,
	uchar*		record,
	const uchar*	key,
	uint		length)
{
	int	error;

	set_partition(part);
	error = ha_innobase::index_next_same(record, key, length);
	update_partition(part);
	return(error);
}

/** Return last record in index from a partition.
@param[in]	part	Partition to read from.
@param[out]	record	Last record in index in the partition.
@return	error number or 0. */
int
ha_innopart::index_last_in_part(
	uint	part,
	uchar*	record)
{
	int	error;

	set_partition(part);
	error = ha_innobase::index_last(record);
	update_partition(part);
	return(error);
}

/** Return previous record in index from a partition.
@param[in]	part	Partition to read from.
@param[out]	record	Last record in index in the partition.
@return	error number or 0. */
int
ha_innopart::index_prev_in_part(
	uint	part,
	uchar*	record)
{
	int	error;

	set_partition(part);
	error = ha_innobase::index_prev(record);
	update_partition(part);

	ut_ad(m_ordered_scan_ongoing
	      || m_ordered_rec_buffer == NULL
	      || m_prebuilt->used_in_HANDLER
	      || m_part_spec.start_part >= m_part_spec.end_part);

	return(error);
}

/** Start index scan and return first record from a partition.
This routine starts an index scan using a start key. The calling
function will check the end key on its own.
@param[in]	part		Partition to read from.
@param[out]	record		First matching record in index in the partition.
@param[in]	key		Key to match.
@param[in]	keypart_map	Which part of the key to use.
@param[in]	find_flag	Key condition/direction to use.
@return	error number or 0. */
int
ha_innopart::index_read_map_in_part(
	uint			part,
	uchar*			record,
	const uchar*		key,
	key_part_map		keypart_map,
	enum ha_rkey_function	find_flag)
{
	int	error;

	set_partition(part);
	error = ha_innobase::index_read_map(
			record,
			key,
			keypart_map,
			find_flag);
	update_partition(part);
	return(error);
}

/** Start index scan and return first record from a partition.
This routine starts an index scan using a start key. The calling
function will check the end key on its own.
@param[in]	part		Partition to read from.
@param[out]	record		First matching record in index in the partition.
@param[in]	index		Index to read from.
@param[in]	key		Key to match.
@param[in]	keypart_map	Which part of the key to use.
@param[in]	find_flag	Key condition/direction to use.
@return	error number or 0. */
int
ha_innopart::index_read_idx_map_in_part(
	uint			part,
	uchar*			record,
	uint			index,
	const uchar*		key,
	key_part_map		keypart_map,
	enum ha_rkey_function	find_flag)
{
	int	error;

	set_partition(part);
	error = ha_innobase::index_read_idx_map(
			record,
			index,
			key,
			keypart_map,
			find_flag);
	update_partition(part);
	return(error);
}

/** Return last matching record in index from a partition.
@param[in]	part		Partition to read from.
@param[out]	record		Last matching record in index in the partition.
@param[in]	key		Key to match.
@param[in]	keypart_map	Which part of the key to use.
@return	error number or 0. */
int
ha_innopart::index_read_last_map_in_part(
	uint		part,
	uchar*		record,
	const uchar*	key,
	key_part_map	keypart_map)
{
	int	error;
	set_partition(part);
	error = ha_innobase::index_read_last_map(record, key, keypart_map);
	update_partition(part);
	return(error);
}

/** Start index scan and return first record from a partition.
This routine starts an index scan using a start and end key.
@param[in]	part		Partition to read from.
@param[in,out]	record		First matching record in index in the partition,
if NULL use table->record[0] as return buffer.
@param[in]	start_key	Start key to match.
@param[in]	end_key		End key to match.
@param[in]	eq_range	Is equal range, start_key == end_key.
@param[in]	sorted		Return rows in sorted order.
@return	error number or 0. */
int
ha_innopart::read_range_first_in_part(
	uint			part,
	uchar*			record,
	const key_range*	start_key,
	const key_range*	end_key,
	bool			eq_range,
	bool			sorted)
{
	int	error;
	uchar*	read_record = record;
	set_partition(part);
	if (read_record == NULL) {
		read_record = table->record[0];
	}
	if (m_start_key.key != NULL) {
		error = ha_innobase::index_read(
				read_record,
				m_start_key.key,
				m_start_key.length,
				m_start_key.flag);
	} else {
		error = ha_innobase::index_first(read_record);
	}
	if (error == HA_ERR_KEY_NOT_FOUND) {
		error = HA_ERR_END_OF_FILE;
	} else if (error == 0 && !in_range_check_pushed_down) {
		/* compare_key uses table->record[0], so we
		need to copy the data if not already there. */

		if (record != NULL) {
			copy_cached_row(table->record[0], read_record);
		}
		if (compare_key(end_range) > 0) {
			/* must use ha_innobase:: due to set/update_partition
			could overwrite states if ha_innopart::unlock_row()
			was used. */
			ha_innobase::unlock_row();
			error = HA_ERR_END_OF_FILE;
		}
	}
	update_partition(part);
	return(error);
}

/** Return next record in index range scan from a partition.
@param[in]	part	Partition to read from.
@param[in,out]	record	First matching record in index in the partition,
if NULL use table->record[0] as return buffer.
@return	error number or 0. */
int
ha_innopart::read_range_next_in_part(
	uint	part,
	uchar*	record)
{
	int	error;
	uchar*	read_record = record;

	set_partition(part);
	if (read_record == NULL) {
		read_record = table->record[0];
	}

	/* TODO: Implement ha_innobase::read_range*?
	So it will return HA_ERR_END_OF_FILE or
	HA_ERR_KEY_NOT_FOUND when passing end_range. */

	error = ha_innobase::index_next(read_record);
	if (error == 0 && !in_range_check_pushed_down) {
		/* compare_key uses table->record[0], so we
		need to copy the data if not already there. */

		if (record != NULL) {
			copy_cached_row(table->record[0], read_record);
		}
		if (compare_key(end_range) > 0) {
			/* must use ha_innobase:: due to set/update_partition
			could overwrite states if ha_innopart::unlock_row()
			was used. */
			ha_innobase::unlock_row();
			error = HA_ERR_END_OF_FILE;
		}
	}
	update_partition(part);

	return(error);
}

/** Initialize a table scan in a specific partition.
@param[in]	part_id	Partition to initialize.
@param[in]	scan	True if table/index scan false otherwise (for rnd_pos)
@return	0 or error number. */
int
ha_innopart::rnd_init_in_part(
	uint	part_id,
	bool	scan)
{
	int	err;

	if (m_prebuilt->clust_index_was_generated) {
		err = change_active_index(part_id, MAX_KEY);
	} else {
		err = change_active_index(part_id, m_primary_key);
	}

	m_start_of_scan = 1;

	/* Don't use semi-consistent read in random row reads (by position).
	This means we must disable semi_consistent_read if scan is false. */

	if (!scan) {
		try_semi_consistent_read(false);
	}

	return(err);
}

/** Ends a table scan.
@param[in]	part_id	Partition to end table scan in.
@param[in]	scan	True for scan else random access.
@return	0 or error number. */
int
ha_innopart::rnd_end_in_part(
	uint	part_id,
	bool	scan)
{
	return(index_end());
}

/** Read next row in partition.
Reads the next row in a table scan (also used to read the FIRST row
in a table scan).
@param[in]	part_id	Partition to end table scan in.
@param[out]	buf	Returns the row in this buffer, in MySQL format.
@return	0, HA_ERR_END_OF_FILE or error number. */
int
ha_innopart::rnd_next_in_part(
	uint	part_id,
	uchar*	buf)
{
	int	error;

	DBUG_ENTER("ha_innopart::rnd_next_in_part");

	set_partition(part_id);
	if (m_start_of_scan) {
		error = ha_innobase::index_first(buf);

		if (error == HA_ERR_KEY_NOT_FOUND) {
			error = HA_ERR_END_OF_FILE;
		}
		m_start_of_scan = 0;
	} else {
		ha_statistic_increment(&SSV::ha_read_rnd_next_count);
		error = ha_innobase::general_fetch(buf, ROW_SEL_NEXT, 0);
	}

	update_partition(part_id);
	DBUG_RETURN(error);
}

/** Get a row from a position.
Fetches a row from the table based on a row reference.
@param[out]	buf	Returns the row in this buffer, in MySQL format.
@param[in]	pos	Position, given as primary key value or DB_ROW_ID
(if no primary key) of the row in MySQL format.  The length of data in pos has
to be ref_length.
@return	0, HA_ERR_KEY_NOT_FOUND or error code. */
int
ha_innopart::rnd_pos(
	uchar*	buf,
	uchar*	pos)
{
	int	error;
	uint	part_id;
	DBUG_ENTER("ha_innopart::rnd_pos");
	ut_ad(PARTITION_BYTES_IN_POS == 2);
	DBUG_DUMP("pos", pos, ref_length);

	ha_statistic_increment(&SSV::ha_read_rnd_count);

	ut_a(m_prebuilt->trx == thd_to_trx(ha_thd()));

	/* Restore used partition. */
	part_id = uint2korr(pos);

	set_partition(part_id);

	/* Note that we assume the length of the row reference is fixed
	for the table, and it is == ref_length. */

	error = ha_innobase::index_read(buf, pos + PARTITION_BYTES_IN_POS,
				ref_length - PARTITION_BYTES_IN_POS,
				HA_READ_KEY_EXACT);
	DBUG_PRINT("info", ("part %u index_read returned %d", part_id, error));
	DBUG_DUMP("buf", buf, table_share->reclength);

	update_partition(part_id);

	DBUG_RETURN(error);
}

/** Return position for cursor in last used partition.
Stores a reference to the current row to 'ref' field of the handle. Note
that in the case where we have generated the clustered index for the
table, the function parameter is illogical: we MUST ASSUME that 'record'
is the current 'position' of the handle, because if row ref is actually
the row id internally generated in InnoDB, then 'record' does not contain
it. We just guess that the row id must be for the record where the handle
was positioned the last time.
@param[out]	ref_arg	Pointer to buffer where to write the position.
@param[in]	record	Record to position for. */
void
ha_innopart::position_in_last_part(
	uchar*		ref_arg,
	const uchar*	record)
{
	if (m_prebuilt->clust_index_was_generated) {
		/* No primary key was defined for the table and we
		generated the clustered index from row id: the
		row reference will be the row id, not any key value
		that MySQL knows of. */

		memcpy(ref_arg, m_prebuilt->row_id, DATA_ROW_ID_LEN);
	} else {

		/* Copy primary key as the row reference */
		KEY*	key_info = table->key_info + m_primary_key;
		key_copy(ref_arg, (uchar*)record, key_info,
			 key_info->key_length);
	}
}

/** Fill in data_dir_path and tablespace name from internal data
dictionary.
@param	part_elem	Partition element to fill.
@param	ib_table	InnoDB table to copy from. */
void
ha_innopart::update_part_elem(
	partition_element*	part_elem,
	dict_table_t*		ib_table)
{
	dict_get_and_save_data_dir_path(ib_table, false);
	if (ib_table->data_dir_path != NULL) {
		if (part_elem->data_file_name == NULL
		    || strcmp(ib_table->data_dir_path,
			part_elem->data_file_name) != 0) {

			/* Play safe and allocate memory from TABLE and copy
			instead of expose the internal data dictionary. */
			part_elem->data_file_name =
				strdup_root(&table->mem_root,
					ib_table->data_dir_path);
		}
	} else {
		part_elem->data_file_name = NULL;
	}

	part_elem->index_file_name = NULL;
}

/** Update create_info.
Used in SHOW CREATE TABLE et al.
@param[in,out]	create_info	Create info to update. */
void
ha_innopart::update_create_info(
	HA_CREATE_INFO*	create_info)
{
	uint		num_subparts	= m_part_info->num_subparts;
	uint		num_parts;
	uint		part;
	dict_table_t*	table;
	List_iterator<partition_element>
				part_it(m_part_info->partitions);
	partition_element*	part_elem;
	partition_element*	sub_elem;
	DBUG_ENTER("ha_innopart::update_create_info");
	if ((create_info->used_fields & HA_CREATE_USED_AUTO) == 0) {
		info(HA_STATUS_AUTO);
		create_info->auto_increment_value = stats.auto_increment_value;
	}

	num_parts = (num_subparts != 0) ? m_tot_parts / num_subparts : m_tot_parts;

	/* DATA/INDEX DIRECTORY are never applied to the whole partitioned
	table, only to its parts. */

	create_info->data_file_name = NULL;
	create_info->index_file_name = NULL;

	/* Since update_create_info() can be called from
	mysql_prepare_alter_table() when not all partitions are set up,
	we look for that condition first.
	If all partitions are not available then simply return,
	since it does not need any updated partitioning info. */

	if (!m_part_info->temp_partitions.is_empty()) {
		DBUG_VOID_RETURN;
	}
	part = 0;
	while ((part_elem = part_it++)) {
		if (part >= num_parts) {
			DBUG_VOID_RETURN;
		}
		if (m_part_info->is_sub_partitioned()) {
			List_iterator<partition_element>
				subpart_it(part_elem->subpartitions);
			uint	subpart = 0;
			while ((sub_elem = subpart_it++)) {
				if (subpart >= num_subparts) {
					DBUG_VOID_RETURN;
				}
				subpart++;
			}
			if (subpart != num_subparts) {
				DBUG_VOID_RETURN;
			}
		}
		part++;
	}
	if (part != num_parts) {
		DBUG_VOID_RETURN;
	}

	/* part_elem->data_file_name should be correct from
	the .frm, but may have been changed, so update from SYS_DATAFILES.
	index_file_name is ignored, so remove it. */

	part = 0;
	part_it.rewind();
	while ((part_elem = part_it++)) {
		if (m_part_info->is_sub_partitioned()) {
			List_iterator<partition_element>
				subpart_it(part_elem->subpartitions);
			while ((sub_elem = subpart_it++)) {
				table = m_part_share->get_table_part(part++);
				update_part_elem(sub_elem, table);
			}
		} else {
			table = m_part_share->get_table_part(part++);
			update_part_elem(part_elem, table);
		}
	}
	DBUG_VOID_RETURN;
}

/** Set create_info->data_file_name.
@param[in]	part_elem	Partition to copy from.
@param[in,out]	info		Create info to set. */
static
void
set_create_info_dir(
	partition_element*	part_elem,
	HA_CREATE_INFO*		info)
{
	if (part_elem->data_file_name != NULL
	    && part_elem->data_file_name[0] != '\0') {
		info->data_file_name = part_elem->data_file_name;
	}
	if (part_elem->index_file_name != NULL
	    && part_elem->index_file_name[0] != '\0') {
		info->index_file_name = part_elem->index_file_name;
	}
}

/** Set flags and append '/' to remote path if necessary. */
void
create_table_info_t::set_remote_path_flags()
{
	if (m_remote_path[0] != '\0') {
		ut_ad(DICT_TF_HAS_DATA_DIR(m_flags) != 0);

		/* os_file_make_remote_pathname will truncate
		everything after the last '/', so append '/'
		if it is not the last character. */

		size_t len = strlen(m_remote_path);
		if (m_remote_path[len - 1] != OS_PATH_SEPARATOR) {
			m_remote_path[len] = OS_PATH_SEPARATOR;
			m_remote_path[len + 1] = '\0';
		}
	} else {
		ut_ad(DICT_TF_HAS_DATA_DIR(m_flags) == 0);
	}
}

/** Creates a new table to an InnoDB database.
@param[in]	name		Table name (in filesystem charset).
@param[in]	form		MySQL Table containing information of
partitions, columns and indexes etc.
@param[in]	create_info	Additional create information, like
create statement string.
@return	0 or error number. */
int
ha_innopart::create(
	const char*	name,
	TABLE*		form,
	HA_CREATE_INFO*	create_info)
{
	int		error;
	/** {database}/{tablename} */
	char		table_name[FN_REFLEN];
	/** absolute path of table */
	char		remote_path[FN_REFLEN];
	char		partition_name[FN_REFLEN];
	char*		table_name_end;
	size_t		table_name_len;
	char*		partition_name_start;
	char		table_data_file_name[FN_REFLEN];
	const char*	index_file_name;
	size_t		len;

	create_table_info_t	info(ha_thd(),
				     form,
				     create_info,
				     table_name,
				     remote_path);

	DBUG_ENTER("ha_innopart::create");
	ut_ad(create_info != NULL);
	ut_ad(m_part_info == form->part_info);
	ut_ad(table_share != NULL);

	/* Not allowed to create temporary partitioned tables. */
	if (create_info != NULL
	    && (create_info->options & HA_LEX_CREATE_TMP_TABLE) != 0) {
		my_error(ER_PARTITION_NO_TEMPORARY, MYF(0));
		ut_ad(0); // Can we support partitioned temporary tables?
		DBUG_RETURN(HA_ERR_INTERNAL_ERROR);
	}

	error = info.initialize();
	if (error != 0) {
		DBUG_RETURN(error);
	}

	/* Setup and check table level options. */
	error = info.prepare_create_table(name);
	if (error != 0) {
		DBUG_RETURN(error);
	}
	strcpy(partition_name, table_name);
	partition_name_start = partition_name + strlen(partition_name);
	table_name_len = strlen(table_name);
	table_name_end = table_name + table_name_len;
	if (create_info->data_file_name != NULL) {
		/* Strip the tablename from the path. */
		strncpy(table_data_file_name, create_info->data_file_name,
			FN_REFLEN-1);
		table_data_file_name[FN_REFLEN - 1] = '\0';
		char* ptr = strrchr(table_data_file_name, OS_PATH_SEPARATOR);
		ut_ad(ptr != NULL);
		if (ptr != NULL) {
			ptr++;
			*ptr = '\0';
			create_info->data_file_name = table_data_file_name;
		}
	} else {
		table_data_file_name[0] = '\0';
	}
	index_file_name = create_info->index_file_name;

	info.allocate_trx();

	/* Latch the InnoDB data dictionary exclusively so that no deadlocks
	or lock waits can happen in it during a table create operation.
	Drop table etc. do this latching in row0mysql.cc. */

	row_mysql_lock_data_dictionary(info.trx());

	/* TODO: use the new DD tables instead to decrease duplicate info. */
	List_iterator_fast <partition_element>
		part_it(form->part_info->partitions);
	partition_element* part_elem;
	while ((part_elem = part_it++)) {
		/* Append the partition name to the table name. */
		len = Ha_innopart_share::append_sep_and_name(
				partition_name_start,
				part_elem->partition_name,
				part_sep,
				FN_REFLEN - table_name_len);
		if ((table_name_len + len) >= FN_REFLEN) {
			ut_ad(0);
			goto cleanup;
		}

		/* Override table level DATA/INDEX DIRECTORY. */
		set_create_info_dir(part_elem, create_info);

		if (!form->part_info->is_sub_partitioned()) {
			error = info.prepare_create_table(partition_name);
			if (error != 0) {
				goto cleanup;
			}
			info.set_remote_path_flags();
			error = info.create_table();
			if (error != 0) {
				goto cleanup;
			}
		} else {
			size_t	part_name_len = strlen(partition_name_start)
						+ table_name_len;
			char*	part_name_end = partition_name + part_name_len;
			List_iterator_fast <partition_element>
				sub_it(part_elem->subpartitions);
			partition_element* sub_elem;

			while ((sub_elem = sub_it++)) {
				ut_ad(sub_elem->partition_name != NULL);

				/* 'table' will be
				<name>#P#<part_name>#SP#<subpart_name>.
				Append the sub-partition name to
				the partition name. */

				len = Ha_innopart_share::append_sep_and_name(
					part_name_end,
					sub_elem->partition_name,
					sub_sep,
					FN_REFLEN - part_name_len);
				if ((len + part_name_len) >= FN_REFLEN) {
					ut_ad(0);
					goto cleanup;
				}
				/* Override part level DATA/INDEX DIRECTORY. */
				set_create_info_dir(sub_elem, create_info);

				Ha_innopart_share::partition_name_casedn_str(
					part_name_end + 4);
				error = info.prepare_create_table(partition_name);
				if (error != 0) {
					goto cleanup;
				}
				info.set_remote_path_flags();
				error = info.create_table();
				if (error != 0) {
					goto cleanup;
				}

				/* Reset partition level
				DATA/INDEX DIRECTORY. */

				create_info->data_file_name =
					table_data_file_name;
				create_info->index_file_name =
					index_file_name;
				set_create_info_dir(part_elem, create_info);
			}
		}
		/* Reset table level DATA/INDEX DIRECTORY. */
		create_info->data_file_name = table_data_file_name;
		create_info->index_file_name = index_file_name;
	}

	innobase_commit_low(info.trx());

	row_mysql_unlock_data_dictionary(info.trx());

	/* Flush the log to reduce probability that the .frm files and
	the InnoDB data dictionary get out-of-sync if the user runs
	with innodb_flush_log_at_trx_commit = 0. */

	log_buffer_flush_to_disk();

	part_it.rewind();
	/* No need to use these now, only table_name will be used. */
	create_info->data_file_name = NULL;
	create_info->index_file_name = NULL;
	while ((part_elem = part_it++)) {
		Ha_innopart_share::append_sep_and_name(
			table_name_end,
			part_elem->partition_name,
			part_sep,
			FN_REFLEN - table_name_len);
		if (!form->part_info->is_sub_partitioned()) {
			error = info.create_table_update_dict();
			if (error != 0) {
				ut_ad(0);
				goto end;
			}
		} else {
			size_t	part_name_len = strlen(table_name_end);
			char*	part_name_end = table_name_end + part_name_len;
			List_iterator_fast <partition_element>
				sub_it(part_elem->subpartitions);
			partition_element* sub_elem;
			while ((sub_elem = sub_it++)) {
				Ha_innopart_share::append_sep_and_name(
					part_name_end,
					sub_elem->partition_name,
					sub_sep,
					FN_REFLEN - table_name_len
					- part_name_len);
				error = info.create_table_update_dict();
				if (error != 0) {
					ut_ad(0);
					goto end;
				}
			}
		}
	}

end:
	/* Tell the InnoDB server that there might be work for
	utility threads: */

	srv_active_wake_master_thread();

	trx_free_for_mysql(info.trx());

	DBUG_RETURN(error);

cleanup:
	trx_rollback_for_mysql(info.trx());

	row_mysql_unlock_data_dictionary(info.trx());

	trx_free_for_mysql(info.trx());

	DBUG_RETURN(error);
}

/** Discards or imports an InnoDB tablespace.
@param[in]	discard	True if discard, else import.
@return	0 or error number. */
int
ha_innopart::discard_or_import_tablespace(
	my_bool	discard)
{
	int	error = 0;
	uint	i;
	DBUG_ENTER("ha_innopart::discard_or_import_tablespace");

	for (i= m_part_info->get_first_used_partition();
	     i < m_tot_parts;
	     i= m_part_info->get_next_used_partition(i)) {

		m_prebuilt->table = m_part_share->get_table_part(i);
		error= ha_innobase::discard_or_import_tablespace(discard);
		if (error != 0) {
			break;
		}
	}
	m_prebuilt->table = m_part_share->get_table_part(0);

	/* IMPORT/DISCARD also means resetting auto_increment. Make sure
	that auto_increment initialization is done after all partitions
	are imported. */
	if (table->found_next_number_field != NULL) {
		lock_auto_increment();
		m_part_share->next_auto_inc_val = 0;
		m_part_share->auto_inc_initialized = false;
		unlock_auto_increment();
	}

	DBUG_RETURN(error);
}

/** Compare key and rowid.
Helper function for sorting records in the priority queue.
a/b points to table->record[0] rows which must have the
key fields set. The bytes before a and b store the rowid.
This is used for comparing/sorting rows first according to
KEY and if same KEY, by rowid (ref).
@param[in]	key_info	Null terminated array of index information.
@param[in]	a		Pointer to record+ref in first record.
@param[in]	b		Pointer to record+ref in second record.
@return	Return value is SIGN(first_rec - second_rec)
@retval	0	Keys are equal.
@retval	-1	second_rec is greater than first_rec.
@retval	+1	first_rec is greater than second_rec. */
int
ha_innopart::key_and_rowid_cmp(
	KEY**	key_info,
	uchar	*a,
	uchar	*b)
{
	int	cmp = key_rec_cmp(key_info, a, b);
	if (cmp != 0) {
		return(cmp);
	}

	/* We must compare by rowid, which is added before the record,
	in the priority queue. */

	return(memcmp(a - DATA_ROW_ID_LEN, b - DATA_ROW_ID_LEN,
		DATA_ROW_ID_LEN));
}

/** Extra hints from MySQL.
@param[in]	operation	Operation hint.
@return	0 or error number. */
int
ha_innopart::extra(
	enum ha_extra_function	operation)
{
	if (operation == HA_EXTRA_SECONDARY_SORT_ROWID) {
		/* index_init(sorted=true) must have been called! */
		ut_ad(m_ordered);
		ut_ad(m_ordered_rec_buffer != NULL);
		/* No index_read call must have been done! */
		ut_ad(m_queue->empty());

		/* If not PK is set as secondary sort, do secondary sort by
		rowid/ref. */

		ut_ad(m_curr_key_info[1] != NULL
		      || m_prebuilt->clust_index_was_generated != 0
		      || m_curr_key_info[0]
			 == table->key_info + table->s->primary_key);

		if (m_curr_key_info[1] == NULL
		    && m_prebuilt->clust_index_was_generated) {
			m_ref_usage = Partition_helper::REF_USED_FOR_SORT;
			m_queue->m_fun = key_and_rowid_cmp;
		}
		return(0);
	}
	return(ha_innobase::extra(operation));
}

/** Delete all rows in a partition.
@return	0 or error number. */
int
ha_innopart::truncate_partition_low()
{
	return(truncate());
}

/** Deletes all rows of a partitioned InnoDB table.
@return	0 or error number. */
int
ha_innopart::truncate()
{
	dberr_t		err = DB_SUCCESS;
	int		error;

	DBUG_ENTER("ha_innopart::truncate");

	if (high_level_read_only) {
		DBUG_RETURN(HA_ERR_TABLE_READONLY);
	}

	/* TRUNCATE also means resetting auto_increment. Hence, reset
	it so that it will be initialized again at the next use. */

	if (table->found_next_number_field != NULL) {
		lock_auto_increment();
		m_part_share->next_auto_inc_val= 0;
		m_part_share->auto_inc_initialized= false;
		unlock_auto_increment();
	}

	/* Get the transaction associated with the current thd, or create one
	if not yet created, and update m_prebuilt->trx. */

	update_thd(ha_thd());

	if (!trx_is_started(m_prebuilt->trx)) {
		++m_prebuilt->trx->will_lock;
	}
	/* Truncate the table in InnoDB. */

	for (uint i = m_part_info->get_first_used_partition();
	     i < m_tot_parts;
	     i = m_part_info->get_next_used_partition(i)) {

		set_partition(i);
		err = row_truncate_table_for_mysql(m_prebuilt->table,
				m_prebuilt->trx);
		update_partition(i);
		if (err != DB_SUCCESS) {
			break;
		}
	}

	switch (err) {

	case DB_TABLESPACE_DELETED:
	case DB_TABLESPACE_NOT_FOUND:
		ib_senderrf(
			m_prebuilt->trx->mysql_thd, IB_LOG_LEVEL_ERROR,
			(err == DB_TABLESPACE_DELETED ?
			ER_TABLESPACE_DISCARDED : ER_TABLESPACE_MISSING),
			table->s->table_name.str);
		table->status = STATUS_NOT_FOUND;
		error = HA_ERR_NO_SUCH_TABLE;
		break;

	default:
		error = convert_error_code_to_mysql(
			err, m_prebuilt->table->flags,
			m_prebuilt->trx->mysql_thd);
		table->status = STATUS_NOT_FOUND;
		break;
	}
	DBUG_RETURN(error);
}

/** Estimates the number of index records in a range.
@param[in]	keynr	Index number.
@param[in]	min_key	Start key value (or NULL).
@param[in]	max_key	End key value (or NULL).
@return	estimated number of rows. */
ha_rows
ha_innopart::records_in_range(
	uint		keynr,
	key_range*	min_key,
	key_range*	max_key)
{
	KEY*		key;
	dict_index_t*	index;
	dtuple_t*	range_start;
	dtuple_t*	range_end;
	int64_t		n_rows = 0;
	page_cur_mode_t	mode1;
	page_cur_mode_t	mode2;
	mem_heap_t*	heap;
	uint		part_id;

	DBUG_ENTER("ha_innopart::records_in_range");
	DBUG_PRINT("info", ("keynr %u min %p max %p", keynr, min_key, max_key));

	ut_a(m_prebuilt->trx == thd_to_trx(ha_thd()));

	m_prebuilt->trx->op_info = (char*)"estimating records in index range";

	/* In case MySQL calls this in the middle of a SELECT query, release
	possible adaptive hash latch to avoid deadlocks of threads. */

	trx_assert_no_search_latch(m_prebuilt->trx);

	active_index = keynr;

	key = table->key_info + active_index;

	part_id = m_part_info->get_first_used_partition();
	if (part_id == MY_BIT_NONE) {
		DBUG_RETURN(0);
	}
	/* This also sets m_prebuilt->index! */
	set_partition(part_id);
	index = m_prebuilt->index;

	/* Only validate the first partition, to avoid too much overhead. */

	/* There exists possibility of not being able to find requested
	index due to inconsistency between MySQL and InoDB dictionary info.
	Necessary message should have been printed in innopart_get_index(). */
	if (index == NULL
	    || dict_table_is_discarded(m_prebuilt->table)
	    || dict_index_is_corrupted(index)
	    || !row_merge_is_index_usable(m_prebuilt->trx, index)) {

		n_rows = HA_POS_ERROR;
		goto func_exit;
	}

	heap = mem_heap_create(2 * (key->actual_key_parts * sizeof(dfield_t)
				    + sizeof(dtuple_t)));

	range_start = dtuple_create(heap, key->actual_key_parts);
	dict_index_copy_types(range_start, index, key->actual_key_parts);

	range_end = dtuple_create(heap, key->actual_key_parts);
	dict_index_copy_types(range_end, index, key->actual_key_parts);

	row_sel_convert_mysql_key_to_innobase(
		range_start,
		m_prebuilt->srch_key_val1,
		m_prebuilt->srch_key_val_len,
		index,
		(byte*) (min_key ? min_key->key : (const uchar*) 0),
		(ulint) (min_key ? min_key->length : 0),
		m_prebuilt->trx);

	ut_ad(min_key != NULL
	      ? range_start->n_fields > 0
	      : range_start->n_fields == 0);

	row_sel_convert_mysql_key_to_innobase(
		range_end,
		m_prebuilt->srch_key_val2,
		m_prebuilt->srch_key_val_len,
		index,
		(byte*) (max_key != NULL ? max_key->key : (const uchar*) 0),
		(ulint) (max_key != NULL ? max_key->length : 0),
		m_prebuilt->trx);

	ut_ad(max_key != NULL
	      ? range_end->n_fields > 0
	      : range_end->n_fields == 0);

	mode1 = convert_search_mode_to_innobase(min_key ? min_key->flag :
						HA_READ_KEY_EXACT);
	mode2 = convert_search_mode_to_innobase(max_key ? max_key->flag :
						HA_READ_KEY_EXACT);

	if (mode1 != PAGE_CUR_UNSUPP && mode2 != PAGE_CUR_UNSUPP) {

		n_rows = btr_estimate_n_rows_in_range(index, range_start,
						      mode1, range_end,
						      mode2);
		DBUG_PRINT("info", ("part_id %u rows %ld", part_id,
					(long int) n_rows));
		for (part_id = m_part_info->get_next_used_partition(part_id);
		     part_id < m_tot_parts;
		     part_id = m_part_info->get_next_used_partition(part_id)) {

			index = m_part_share->get_index(part_id, keynr);
			int64_t n = btr_estimate_n_rows_in_range(index,
							       range_start,
							       mode1,
							       range_end,
							       mode2);
			n_rows += n;
			DBUG_PRINT("info", ("part_id %u rows %ld (%ld)",
						part_id,
						(long int) n,
						(long int) n_rows));
		}
	} else {

		n_rows = HA_POS_ERROR;
	}

	mem_heap_free(heap);

func_exit:

	m_prebuilt->trx->op_info = (char*)"";

	/* The MySQL optimizer seems to believe an estimate of 0 rows is
	always accurate and may return the result 'Empty set' based on that.
	The accuracy is not guaranteed, and even if it were, for a locking
	read we should anyway perform the search to set the next-key lock.
	Add 1 to the value to make sure MySQL does not make the assumption! */

	if (n_rows == 0) {
		n_rows = 1;
	}

	DBUG_RETURN((ha_rows) n_rows);
}

/** Gives an UPPER BOUND to the number of rows in a table.
This is used in filesort.cc.
@return	upper bound of rows. */
ha_rows
ha_innopart::estimate_rows_upper_bound()
{
	const dict_index_t*	index;
	ulonglong		estimate = 0;
	ulonglong		local_data_file_length;
	ulint			stat_n_leaf_pages;

	DBUG_ENTER("ha_innopart::estimate_rows_upper_bound");

	/* We do not know if MySQL can call this function before calling
	external_lock(). To be safe, update the thd of the current table
	handle. */

	update_thd(ha_thd());

	m_prebuilt->trx->op_info = "calculating upper bound for table rows";

	/* In case MySQL calls this in the middle of a SELECT query, release
	possible adaptive hash latch to avoid deadlocks of threads. */

	trx_assert_no_search_latch(m_prebuilt->trx);

	for (uint i = m_part_info->get_first_used_partition();
	     i < m_tot_parts;
	     i = m_part_info->get_next_used_partition(i)) {

		m_prebuilt->table = m_part_share->get_table_part(i);
		index = dict_table_get_first_index(m_prebuilt->table);

		stat_n_leaf_pages = index->stat_n_leaf_pages;

		ut_a(stat_n_leaf_pages > 0);

		local_data_file_length =
			((ulonglong) stat_n_leaf_pages) * UNIV_PAGE_SIZE;

		/* Calculate a minimum length for a clustered index record
		and from that an upper bound for the number of rows.
		Since we only calculate new statistics in row0mysql.cc when a
		table has grown by a threshold factor,
		we must add a safety factor 2 in front of the formula below. */

		estimate += 2 * local_data_file_length
			/ dict_index_calc_min_rec_len(index);
	}

	m_prebuilt->trx->op_info = "";

	DBUG_RETURN((ha_rows) estimate);
}

/** Time estimate for full table scan.
How many seeks it will take to read through the table. This is to be
comparable to the number returned by records_in_range so that we can
decide if we should scan the table or use keys.
@return	estimated time measured in disk seeks. */
double
ha_innopart::scan_time()
{
	double	scan_time = 0.0;
	DBUG_ENTER("ha_innopart::scan_time");

	for (uint i = m_part_info->get_first_used_partition();
	     i < m_tot_parts;
	     i = m_part_info->get_next_used_partition(i)) {
		m_prebuilt->table = m_part_share->get_table_part(i);
		scan_time += ha_innobase::scan_time();
	}
	DBUG_RETURN(scan_time);
}

/** Updates the statistics for one partition (table).
@param[in]	table		Table to update the statistics for.
@param[in]	is_analyze	True if called from ::analyze().
@return	error code. */
static
int
update_table_stats(
	dict_table_t*	table,
	bool		is_analyze)
{
	dict_stats_upd_option_t	opt;
	dberr_t			ret;

	if (dict_stats_is_persistent_enabled(table)) {
		if (is_analyze) {
			opt = DICT_STATS_RECALC_PERSISTENT;
		} else {
			/* This is e.g. 'SHOW INDEXES',
			fetch the persistent stats from disk. */
			opt = DICT_STATS_FETCH_ONLY_IF_NOT_IN_MEMORY;
		}
	} else {
		opt = DICT_STATS_RECALC_TRANSIENT;
	}

	ut_ad(!mutex_own(&dict_sys->mutex));
	ret = dict_stats_update(table, opt);

	if (ret != DB_SUCCESS) {
		return(HA_ERR_GENERIC);
	}
	return(0);
}

/** Updates and return statistics.
Returns statistics information of the table to the MySQL interpreter,
in various fields of the handle object.
@param[in]	flag		Flags for what to update and return.
@param[in]	is_analyze	True if called from ::analyze().
@return	HA_ERR_* error code or 0. */
int
ha_innopart::info_low(
	uint	flag,
	bool	is_analyze)
{
	dict_table_t*	ib_table;
	ib_uint64_t	max_rows = 0;
	uint		biggest_partition = 0;
	int		error = 0;

	DBUG_ENTER("ha_innopart::info_low");

	/* If we are forcing recovery at a high level, we will suppress
	statistics calculation on tables, because that may crash the
	server if an index is badly corrupted. */

	/* We do not know if MySQL can call this function before calling
	external_lock(). To be safe, update the thd of the current table
	handle. */

	update_thd(ha_thd());

	/* In case MySQL calls this in the middle of a SELECT query, release
	possible adaptive hash latch to avoid deadlocks of threads. */

	m_prebuilt->trx->op_info = (char*)"returning various info to MySQL";

	trx_assert_no_search_latch(m_prebuilt->trx);

	ut_ad(m_part_share->get_table_part(0)->n_ref_count > 0);

	if ((flag & HA_STATUS_TIME) != 0) {
		stats.update_time = 0;

		if (is_analyze) {
			/* Only analyze the given partitions. */
			int	error = set_altered_partitions();
			if (error != 0) {
				/* Already checked in mysql_admin_table! */
				ut_ad(0);
				DBUG_RETURN(error);
			}
		}
		if (is_analyze || innobase_stats_on_metadata) {
			m_prebuilt->trx->op_info = "updating table statistics";
		}

		/* TODO: Only analyze the PK for all partitions,
		then the secondary indexes only for the largest partition! */
		for (uint i = m_part_info->get_first_used_partition();
		     i < m_tot_parts;
		     i = m_part_info->get_next_used_partition(i)) {

			ib_table = m_part_share->get_table_part(i);
			if (is_analyze || innobase_stats_on_metadata) {
				error = update_table_stats(ib_table, is_analyze);
				if (error != 0) {
					m_prebuilt->trx->op_info = "";
					DBUG_RETURN(error);
				}
			}
			set_if_bigger(stats.update_time,
				(ulong) ib_table->update_time);
		}

		if (is_analyze || innobase_stats_on_metadata) {
			m_prebuilt->trx->op_info =
				"returning various info to MySQL";
		}
	}

	if ((flag & HA_STATUS_VARIABLE) != 0) {

		/* TODO: If this is called after pruning, then we could
		also update the statistics according to the non-pruned
		partitions, by allocating new rec_per_key on the TABLE,
		instead of using the info from the TABLE_SHARE. */
		ulint		stat_clustered_index_size = 0;
		ulint		stat_sum_of_other_index_sizes = 0;
		ib_uint64_t	n_rows = 0;
		ulint		avail_space = 0;
		bool		checked_sys_tablespace = false;

		if ((flag & HA_STATUS_VARIABLE_EXTRA) != 0) {
			stats.delete_length = 0;
		}

		for (uint i = m_part_info->get_first_used_partition();
		     i < m_tot_parts;
		     i = m_part_info->get_next_used_partition(i)) {

			ib_table = m_part_share->get_table_part(i);
			if ((flag & HA_STATUS_NO_LOCK) == 0) {
				dict_table_stats_lock(ib_table, RW_S_LATCH);
			}

			ut_a(ib_table->stat_initialized);

			n_rows += ib_table->stat_n_rows;
			if (ib_table->stat_n_rows > max_rows) {
				max_rows = ib_table->stat_n_rows;
				biggest_partition = i;
			}

			stat_clustered_index_size +=
				ib_table->stat_clustered_index_size;

			stat_sum_of_other_index_sizes +=
				ib_table->stat_sum_of_other_index_sizes;

			if ((flag & HA_STATUS_NO_LOCK) == 0) {
				dict_table_stats_unlock(ib_table, RW_S_LATCH);
			}

			if ((flag & HA_STATUS_VARIABLE_EXTRA) != 0
			    && (flag & HA_STATUS_NO_LOCK) == 0
			    && srv_force_recovery < SRV_FORCE_NO_IBUF_MERGE
			    && avail_space != ULINT_UNDEFINED) {

				/* Only count system tablespace once! */
				if (is_system_tablespace(ib_table->space)) {
					if (checked_sys_tablespace) {
						continue;
					}
					checked_sys_tablespace = true;
				}

				uintmax_t	space =
					fsp_get_available_space_in_free_extents(
						ib_table->space);
				if (space == UINTMAX_MAX) {
					THD*	thd = ha_thd();
					const char* table_name
						= ib_table->name.m_name;

					push_warning_printf(
						thd,
						Sql_condition::SL_WARNING,
						ER_CANT_GET_STAT,
						"InnoDB: Trying to get the"
						" free space for partition %s"
						" but its tablespace has been"
						" discarded or the .ibd file"
						" is missing. Setting the free"
						" space of the partition to"
						" zero.",
						ut_get_name(
							m_prebuilt->trx,
							table_name).c_str());
				} else {
					avail_space +=
						static_cast<ulint>(space);
				}
			}
		}

		/*
		The MySQL optimizer seems to assume in a left join that n_rows
		is an accurate estimate if it is zero. Of course, it is not,
		since we do not have any locks on the rows yet at this phase.
		Since SHOW TABLE STATUS seems to call this function with the
		HA_STATUS_TIME flag set, while the left join optimizer does not
		set that flag, we add one to a zero value if the flag is not
		set. That way SHOW TABLE STATUS will show the best estimate,
		while the optimizer never sees the table empty. */

		if (n_rows == 0 && (flag & HA_STATUS_TIME) == 0) {
			n_rows++;
		}

		/* Fix bug#40386: Not flushing query cache after truncate.
		n_rows can not be 0 unless the table is empty, set to 1
		instead. The original problem of bug#29507 is actually
		fixed in the server code. */
		if (thd_sql_command(m_user_thd) == SQLCOM_TRUNCATE) {

			n_rows = 1;

			/* We need to reset the m_prebuilt value too, otherwise
			checks for values greater than the last value written
			to the table will fail and the autoinc counter will
			not be updated. This will force write_row() into
			attempting an update of the table's AUTOINC counter. */

			m_prebuilt->autoinc_last_value = 0;
		}

		/* Take page_size from first partition. */
		ib_table = m_part_share->get_table_part(0);
		const page_size_t&	page_size =
			dict_table_page_size(ib_table);

		stats.records = (ha_rows) n_rows;
		stats.deleted = 0;
		stats.data_file_length =
			((ulonglong) stat_clustered_index_size)
			* page_size.physical();
		stats.index_file_length =
			((ulonglong) stat_sum_of_other_index_sizes)
			* page_size.physical();

		/* See ha_innobase::info_low() for comments! */
		if ((flag & HA_STATUS_NO_LOCK) == 0
		    && (flag & HA_STATUS_VARIABLE_EXTRA) != 0
		    && srv_force_recovery < SRV_FORCE_NO_IBUF_MERGE) {
			stats.delete_length = avail_space * 1024;
		}

		stats.check_time = 0;
		stats.mrr_length_per_rec = ref_length + sizeof(void*)
						- PARTITION_BYTES_IN_POS;

		if (stats.records == 0) {
			stats.mean_rec_length = 0;
		} else {
			stats.mean_rec_length = (ulong)
				(stats.data_file_length / stats.records);
		}
	}

	if ((flag & HA_STATUS_CONST) != 0) {
		/* Find max rows and biggest partition. */
		for (uint i = 0; i < m_tot_parts; i++) {
			/* Skip partitions from above. */
			if ((flag & HA_STATUS_VARIABLE) == 0
			    || !bitmap_is_set(&(m_part_info->read_partitions),
					i)) {

				ib_table = m_part_share->get_table_part(i);
				if (ib_table->stat_n_rows > max_rows) {
					max_rows = ib_table->stat_n_rows;
					biggest_partition = i;
				}
			}
		}
		ib_table = m_part_share->get_table_part(biggest_partition);
		/* Verify the number of index in InnoDB and MySQL
		matches up. If m_prebuilt->clust_index_was_generated
		holds, InnoDB defines GEN_CLUST_INDEX internally. */
		ulint	num_innodb_index = UT_LIST_GET_LEN(ib_table->indexes)
			- m_prebuilt->clust_index_was_generated;
		if (table->s->keys < num_innodb_index) {
			/* If there are too many indexes defined
			inside InnoDB, ignore those that are being
			created, because MySQL will only consider
			the fully built indexes here. */

			for (const dict_index_t* index =
					UT_LIST_GET_FIRST(ib_table->indexes);
			     index != NULL;
			     index = UT_LIST_GET_NEXT(indexes, index)) {

				/* First, online index creation is
				completed inside InnoDB, and then
				MySQL attempts to upgrade the
				meta-data lock so that it can rebuild
				the .frm file. If we get here in that
				time frame, dict_index_is_online_ddl()
				would not hold and the index would
				still not be included in TABLE_SHARE. */
				if (!index->is_committed()) {
					num_innodb_index--;
				}
			}

			if (table->s->keys < num_innodb_index
			    && (innobase_fts_check_doc_id_index(ib_table,
							NULL, NULL)
				 == FTS_EXIST_DOC_ID_INDEX)) {
				num_innodb_index--;
			}
		}

		if (table->s->keys != num_innodb_index) {
			ib::error() << "Table "
				<< ib_table->name << " contains "
				<< num_innodb_index
				<< " indexes inside InnoDB, which"
				" is different from the number of"
				" indexes " << table->s->keys
				<< " defined in the MySQL";
		}

		if ((flag & HA_STATUS_NO_LOCK) == 0) {
			dict_table_stats_lock(ib_table, RW_S_LATCH);
		}

		ut_a(ib_table->stat_initialized);

		for (ulong i = 0; i < table->s->keys; i++) {
			ulong	j;
			/* We could get index quickly through internal
			index mapping with the index translation table.
			The identity of index (match up index name with
			that of table->key_info[i]) is already verified in
			innopart_get_index(). */
			dict_index_t*	index = innopart_get_index(
							biggest_partition, i);

			if (index == NULL) {
				ib::error() << "Table "
					<< ib_table->name << " contains fewer"
					" indexes inside InnoDB than"
					" are defined in the MySQL"
					" .frm file. Have you mixed up"
					" .frm files from different"
					" installations? "
					<< TROUBLESHOOTING_MSG;
				break;
			}

			KEY*	key = &table->key_info[i];
			for (j = 0;
			     j < key->actual_key_parts;
			     j++) {

				if ((key->flags & HA_FULLTEXT) != 0) {
					/* The whole concept has no validity
					for FTS indexes. */
					key->rec_per_key[j] = 1;
					continue;
				}

				if ((j + 1) > index->n_uniq) {
					ib::error() << "Index " << index->name
						<< " of " << ib_table->name
						<< " has " << index->n_uniq
						<< " columns unique inside"
						" InnoDB, but MySQL is"
						" asking statistics for "
						<< j + 1 << " columns. Have"
						" you mixed up .frm files"
						" from different"
						" installations? "
						<< TROUBLESHOOTING_MSG;
					break;
				}

				/* innodb_rec_per_key() will use
				index->stat_n_diff_key_vals[] and the value we
				pass index->table->stat_n_rows. Both are
				calculated by ANALYZE and by the background
				stats gathering thread (which kicks in when too
				much of the table has been changed). In
				addition table->stat_n_rows is adjusted with
				each DML (e.g. ++ on row insert). Those
				adjustments are not MVCC'ed and not even
				reversed on rollback. So,
				index->stat_n_diff_key_vals[] and
				index->table->stat_n_rows could have been
				calculated at different time. This is
				acceptable. */
				const rec_per_key_t	rec_per_key =
					innodb_rec_per_key(
						index, j,
						max_rows);

				key->set_records_per_key(j, rec_per_key);

				/* The code below is legacy and should be
				removed together with this comment once we
				are sure the new floating point rec_per_key,
				set via set_records_per_key(), works fine. */

				ulong	rec_per_key_int = static_cast<ulong>(
					innodb_rec_per_key(index, j,
							   max_rows));

				/* Since MySQL seems to favor table scans
				too much over index searches, we pretend
				index selectivity is 2 times better than
				our estimate: */

				rec_per_key_int = rec_per_key_int / 2;

				if (rec_per_key_int == 0) {
					rec_per_key_int = 1;
				}

				key->rec_per_key[j] = rec_per_key_int;
			}
		}

		if ((flag & HA_STATUS_NO_LOCK) == 0) {
			dict_table_stats_unlock(ib_table, RW_S_LATCH);
		}

		char		path[FN_REFLEN];
		os_file_stat_t	stat_info;
		/* Use the first partition for create time until new DD. */
		ib_table = m_part_share->get_table_part(0);
		my_snprintf(path, sizeof(path), "%s/%s%s",
			    mysql_data_home,
			    table->s->normalized_path.str,
			    reg_ext);

		unpack_filename(path,path);

		if (os_file_get_status(path, &stat_info, false, true) == DB_SUCCESS) {
			stats.create_time = (ulong) stat_info.ctime;
		}
	}

	if (srv_force_recovery >= SRV_FORCE_NO_IBUF_MERGE) {

		goto func_exit;
	}

	if ((flag & HA_STATUS_ERRKEY) != 0) {
		const dict_index_t*	err_index;

		ut_a(m_prebuilt->trx);
		ut_a(m_prebuilt->trx->magic_n == TRX_MAGIC_N);

		err_index = trx_get_error_info(m_prebuilt->trx);

		if (err_index != NULL) {
			errkey = m_part_share->get_mysql_key(m_last_part,
							err_index);
		} else {
			errkey = (unsigned int) (
				(m_prebuilt->trx->error_key_num
				 == ULINT_UNDEFINED)
					? UINT_MAX
					: m_prebuilt->trx->error_key_num);
		}
	}

	if ((flag & HA_STATUS_AUTO) != 0) {
		/* auto_inc is only supported in first key for InnoDB! */
		ut_ad(table_share->next_number_keypart == 0);
		DBUG_PRINT("info", ("HA_STATUS_AUTO"));
		if (table->found_next_number_field == NULL) {
			stats.auto_increment_value = 0;
		} else {
			/* Lock to avoid two concurrent initializations. */
			lock_auto_increment();
			if (m_part_share->auto_inc_initialized) {
				stats.auto_increment_value =
					m_part_share->next_auto_inc_val;
			} else {
				/* The auto-inc mutex in the table_share is
				locked, so we do not need to have the handlers
				locked. */

				error = initialize_auto_increment(
					(flag & HA_STATUS_NO_LOCK) != 0);
				stats.auto_increment_value =
						m_part_share->next_auto_inc_val;
			}
			unlock_auto_increment();
		}
	}

func_exit:
	m_prebuilt->trx->op_info = (char*)"";

	DBUG_RETURN(error);
}

/** Optimize table.
This is mapped to "ALTER TABLE tablename ENGINE=InnoDB", which rebuilds
the table in MySQL.
@param[in]	thd		Connection thread handle.
@param[in]	check_opt	Currently ignored.
@return	0 for success else error code. */
int
ha_innopart::optimize(
	THD*		thd,
	HA_CHECK_OPT*	check_opt)
{
	return(HA_ADMIN_TRY_ALTER);
}

/** Checks a partitioned table.
Tries to check that an InnoDB table is not corrupted. If corruption is
noticed, prints to stderr information about it. In case of corruption
may also assert a failure and crash the server. Also checks for records
in wrong partition.
@param[in]	thd		MySQL THD object/thread handle.
@param[in]	check_opt	Check options.
@return	HA_ADMIN_CORRUPT or HA_ADMIN_OK. */
int
ha_innopart::check(
	THD*		thd,
	HA_CHECK_OPT*	check_opt)
{
	uint	error = HA_ADMIN_OK;
	uint	i;

	DBUG_ENTER("ha_innopart::check");
	/* TODO: Enhance this to:
	- Every partition has the same structure.
	- The names are correct (partition names checked in ::open()?)
	Currently it only does normal InnoDB check of each partition. */

	if (set_altered_partitions()) {
		ut_ad(0);   // Already checked by set_part_state()!
		DBUG_RETURN(HA_ADMIN_INVALID);
	}
	for (i = m_part_info->get_first_used_partition();
	     i < m_tot_parts;
	     i = m_part_info->get_next_used_partition(i)) {

		m_prebuilt->table = m_part_share->get_table_part(i);
		error = ha_innobase::check(thd, check_opt);
		if (error != 0) {
			break;
		}
		if ((check_opt->flags & (T_MEDIUM | T_EXTEND)) != 0) {
			error = Partition_helper::check_misplaced_rows(i, false);
			if (error != 0) {
				break;
			}
		}
	}
	if (error != 0) {
		print_admin_msg(
			thd,
			256,
			"error",
			table_share->db.str,
			table->alias,
			"check",
			m_is_sub_partitioned ?
			  "Subpartition %s returned error"
			  : "Partition %s returned error",
			m_part_share->get_partition_name(i));
	}

	DBUG_RETURN(error);
}

/** Repair a partitioned table.
Only repairs records in wrong partitions (moves them to the correct
partition or deletes them if not in any partition).
@param[in]	thd		MySQL THD object/thread handle.
@param[in]	repair_opt	Repair options.
@return	0 or error code. */
int
ha_innopart::repair(
	THD*		thd,
	HA_CHECK_OPT*	repair_opt)
{
	uint	error = HA_ADMIN_OK;

	DBUG_ENTER("ha_innopart::repair");

	/* TODO: enable this warning to be clear about what is repaired.
	Currently disabled to generate smaller test diffs. */
#ifdef ADD_WARNING_FOR_REPAIR_ONLY_PARTITION
	push_warning_printf(thd, Sql_condition::SL_WARNING,
			    ER_ILLEGAL_HA,
			    "Only moving rows from wrong partition to correct"
			    " partition is supported,"
			    " repairing InnoDB indexes is not yet supported!");
#endif

	/* Only repair partitions for MEDIUM or EXTENDED options. */
	if ((repair_opt->flags & (T_MEDIUM | T_EXTEND)) == 0) {
		DBUG_RETURN(HA_ADMIN_OK);
	}
	if (set_altered_partitions()) {
		ut_ad(0);   // Already checked by set_part_state()!
		DBUG_RETURN(HA_ADMIN_INVALID);
	}
	for (uint i = m_part_info->get_first_used_partition();
	     i < m_tot_parts;
	     i = m_part_info->get_next_used_partition(i)) {

		/* TODO: Implement and use ha_innobase::repair()! */
		error = Partition_helper::check_misplaced_rows(i, true);
		if (error != 0) {
			print_admin_msg(
				thd,
				256,
				"error",
				table_share->db.str,
				table->alias,
				"repair",
				m_is_sub_partitioned ?
				  "Subpartition %s returned error"
				  : "Partition %s returned error",
				m_part_share->get_partition_name(i));
			break;
		}
	}

	DBUG_RETURN(error);
}

/** Check if possible to switch engine (no foreign keys).
Checks if ALTER TABLE may change the storage engine of the table.
Changing storage engines is not allowed for tables for which there
are foreign key constraints (parent or child tables).
@return	true if can switch engines. */
bool
ha_innopart::can_switch_engines()
{
	bool	can_switch;

	DBUG_ENTER("ha_innopart::can_switch_engines");
	can_switch = ha_innobase::can_switch_engines();
	ut_ad(can_switch);

	DBUG_RETURN(can_switch);
}

/** Checks if a table is referenced by a foreign key.
The MySQL manual states that a REPLACE is either equivalent to an INSERT,
or DELETE(s) + INSERT. Only a delete is then allowed internally to resolve
a duplicate key conflict in REPLACE, not an update.
@return	> 0 if referenced by a FOREIGN KEY. */
uint
ha_innopart::referenced_by_foreign_key()
{
	if (dict_table_is_referenced_by_foreign_key(m_prebuilt->table)) {

#ifndef HA_INNOPART_SUPPORTS_FOREIGN_KEYS
		ut_ad(0);
#endif /* HA_INNOPART_SUPPORTS_FOREIGN_KEYS */
		return(1);
	}

	return(0);
}

/** Start statement.
MySQL calls this function at the start of each SQL statement inside LOCK
TABLES. Inside LOCK TABLES the ::external_lock method does not work to
mark SQL statement borders. Note also a special case: if a temporary table
is created inside LOCK TABLES, MySQL has not called external_lock() at all
on that table.
MySQL-5.0 also calls this before each statement in an execution of a stored
procedure. To make the execution more deterministic for binlogging, MySQL-5.0
locks all tables involved in a stored procedure with full explicit table
locks (thd_in_lock_tables(thd) holds in store_lock()) before executing the
procedure.
@param[in]	thd		Handle to the user thread.
@param[in]	lock_type	Lock type.
@return	0 or error code. */
int
ha_innopart::start_stmt(
	THD*		thd,
	thr_lock_type	lock_type)
{
	int	error = 0;

	if (m_part_info->get_first_used_partition() == MY_BIT_NONE) {
		/* All partitions pruned away, do nothing! */
		return(error);
	}

	error = ha_innobase::start_stmt(thd, lock_type);
	if (m_prebuilt->sql_stat_start) {
		memset(m_sql_stat_start_parts, 0xff,
		       UT_BITS_IN_BYTES(m_tot_parts));
	} else {
		memset(m_sql_stat_start_parts, 0,
		       UT_BITS_IN_BYTES(m_tot_parts));
	}
	return(error);
}

/** Function to store lock for all partitions in native partitioned table. Also
look at ha_innobase::store_lock for more details.
@param[in]	thd		user thread handle
@param[in]	to		pointer to the current element in an array of
pointers to lock structs
@param[in]	lock_type	lock type to store in 'lock'; this may also be
TL_IGNORE
@retval	to	pointer to the current element in the 'to' array */
THR_LOCK_DATA**
ha_innopart::store_lock(
	THD*			thd,
	THR_LOCK_DATA**		to,
	thr_lock_type		lock_type)
{
	trx_t*  trx = m_prebuilt->trx;
	const uint sql_command = thd_sql_command(thd);

	ha_innobase::store_lock(thd, to, lock_type);

	if (sql_command == SQLCOM_FLUSH
	    && lock_type == TL_READ_NO_INSERT) {
		for (uint i = 1; i < m_tot_parts; i++) {
			dict_table_t* table = m_part_share->get_table_part(i);

			dberr_t err = row_quiesce_set_state(
				table, QUIESCE_START, trx);
			ut_a(err == DB_SUCCESS || err == DB_UNSUPPORTED);
		}
	}

	return to;
}

/** Lock/prepare to lock table.
As MySQL will execute an external lock for every new table it uses when it
starts to process an SQL statement (an exception is when MySQL calls
start_stmt for the handle) we can use this function to store the pointer to
the THD in the handle. We will also use this function to communicate
to InnoDB that a new SQL statement has started and that we must store a
savepoint to our transaction handle, so that we are able to roll back
the SQL statement in case of an error.
@param[in]	thd		Handle to the user thread.
@param[in]	lock_type	Lock type.
@return	0 or error number. */
int
ha_innopart::external_lock(
	THD*	thd,
	int	lock_type)
{
	int	error = 0;

	if (m_part_info->get_first_used_partition() == MY_BIT_NONE
		&& !(m_mysql_has_locked
		     && lock_type == F_UNLCK)) {

		/* All partitions pruned away, do nothing! */
		ut_ad(!m_mysql_has_locked);
		return(error);
	}
	ut_ad(m_mysql_has_locked || lock_type != F_UNLCK);

	m_prebuilt->table = m_part_share->get_table_part(0);
	error = ha_innobase::external_lock(thd, lock_type);

        for (uint i = 0; i < m_tot_parts; i++) {
		dict_table_t* table = m_part_share->get_table_part(i);

		switch (table->quiesce) {
		case QUIESCE_START:
			/* Check for FLUSH TABLE t WITH READ LOCK */
			if (!srv_read_only_mode
			    && thd_sql_command(thd) == SQLCOM_FLUSH
			    && lock_type == F_RDLCK) {

				ut_ad(table->quiesce == QUIESCE_START);

				row_quiesce_table_start(table,
							m_prebuilt->trx);

				/* Use the transaction instance to track
				UNLOCK TABLES. It can be done via START
				TRANSACTION; too implicitly. */

				++m_prebuilt->trx->flush_tables;
			}
			break;

		case QUIESCE_COMPLETE:
			/* Check for UNLOCK TABLES; implicit or explicit
			or trx interruption. */
			if (m_prebuilt->trx->flush_tables > 0
			    && (lock_type == F_UNLCK
				|| trx_is_interrupted(m_prebuilt->trx))) {

				ut_ad(table->quiesce == QUIESCE_COMPLETE);
				row_quiesce_table_complete(table,
							   m_prebuilt->trx);

				ut_a(m_prebuilt->trx->flush_tables > 0);
				--m_prebuilt->trx->flush_tables;
			}
			break;

		case QUIESCE_NONE:
			break;

		default:
			ut_ad(0);
		}
	}

	ut_ad(!m_auto_increment_lock);
	ut_ad(!m_auto_increment_safe_stmt_log_lock);

	if (m_prebuilt->sql_stat_start) {
		memset(m_sql_stat_start_parts, 0xff,
		       UT_BITS_IN_BYTES(m_tot_parts));
	} else {
		memset(m_sql_stat_start_parts, 0,
		       UT_BITS_IN_BYTES(m_tot_parts));
	}
	return(error);
}

/** Get the current auto_increment value.
@param[in]	offset			Table auto-inc offset.
@param[in]	increment		Table auto-inc increment.
@param[in]	nb_desired_values	Number of required values.
@param[out]	first_value		The auto increment value.
@param[out]	nb_reserved_values	Number of reserved values.
@return	Auto increment value, or ~0 on failure. */
void
ha_innopart::get_auto_increment(
	ulonglong	offset,
	ulonglong	increment,
	ulonglong	nb_desired_values,
	ulonglong*	first_value,
	ulonglong*	nb_reserved_values)
{
	DBUG_ENTER("ha_innopart::get_auto_increment");
	if (table_share->next_number_keypart != 0) {
		/* Only first key part allowed as autoinc for InnoDB tables! */
		ut_ad(0);
		*first_value = ULLONG_MAX;
		DBUG_VOID_RETURN;
	}
	get_auto_increment_first_field(
		increment,
		nb_desired_values,
		first_value,
		nb_reserved_values);
	DBUG_VOID_RETURN;
}

/** Compares two 'refs'.
A 'ref' is the (internal) primary key value of the row.
If there is no explicitly declared non-null unique key or a primary key, then
InnoDB internally uses the row id as the primary key.
It will use the partition id as secondary compare.
@param[in]	ref1	An (internal) primary key value in the MySQL key value
format.
@param[in]	ref2	Reference to compare with (same type as ref1).
@return	< 0 if ref1 < ref2, 0 if equal, else > 0. */
int
ha_innopart::cmp_ref(
	const uchar*	ref1,
	const uchar*	ref2)
{
	int	cmp;

	cmp = ha_innobase::cmp_ref(ref1 + PARTITION_BYTES_IN_POS,
				   ref2 + PARTITION_BYTES_IN_POS);

	if (cmp != 0) {
		return(cmp);
	}

	cmp = static_cast<int>(uint2korr(ref1))
		- static_cast<int>(uint2korr(ref2));

	return(cmp);
}

/** Prepare for creating new partitions during ALTER TABLE ... PARTITION.
@param[in]	num_partitions	Number of new partitions to be created.
@param[in]	only_create	True if only creating the partition
(no open/lock is needed).
@return	0 for success else error code. */
int
ha_innopart::prepare_for_new_partitions(
	uint	num_partitions,
	bool	only_create)
{
	m_new_partitions = UT_NEW(Altered_partitions(num_partitions,
						     only_create),
				  mem_key_partitioning);
	if (m_new_partitions == NULL) {
		return(HA_ERR_OUT_OF_MEM);
	}
	if (m_new_partitions->initialize()) {
		UT_DELETE(m_new_partitions);
		m_new_partitions = NULL;
		return(HA_ERR_OUT_OF_MEM);
	}
	return(0);
}

/** Create a new partition to be filled during ALTER TABLE ... PARTITION.
@param[in]	table		Table to create the partition in.
@param[in]	create_info	Table/partition specific create info.
@param[in]	part_name	Partition name.
@param[in]	new_part_id	Partition id in new table.
@param[in]	part_elem	Partition element.
@return	0 for success else error code. */
int
ha_innopart::create_new_partition(
	TABLE*			table,
	HA_CREATE_INFO*		create_info,
	const char*		part_name,
	uint			new_part_id,
	partition_element*	part_elem)
{
	int		error;
	char		norm_name[FN_REFLEN];
	const char*	data_file_name_backup = create_info->data_file_name;
	DBUG_ENTER("ha_innopart::create_new_partition");
	/* Delete by ddl_log on failure. */
	normalize_table_name(norm_name, part_name);
	set_create_info_dir(part_elem, create_info);

	error = ha_innobase::create(norm_name, table, create_info);
	create_info->data_file_name = data_file_name_backup;
	if (error == HA_ERR_FOUND_DUPP_KEY) {
		DBUG_RETURN(HA_ERR_TABLE_EXIST);
	}
	if (error != 0) {
		DBUG_RETURN(error);
	}
	if (!m_new_partitions->only_create())
	{
		dict_table_t* part;
		part = dict_table_open_on_name(norm_name,
					       false,
					       true,
					       DICT_ERR_IGNORE_NONE);
		if (part == NULL) {
			DBUG_RETURN(HA_ERR_INTERNAL_ERROR);
		}
		m_new_partitions->set_part(new_part_id, part);
	}
	DBUG_RETURN(0);
}

/** Close and finalize new partitions. */
void
ha_innopart::close_new_partitions()
{
	if (m_new_partitions != NULL) {
		UT_DELETE(m_new_partitions);
		m_new_partitions = NULL;
	}
}

/** write row to new partition.
@param[in]	new_part	New partition to write to.
@return	0 for success else error code. */
int
ha_innopart::write_row_in_new_part(
	uint	new_part)
{
	int	result;
	DBUG_ENTER("ha_innopart::write_row_in_new_part");

	m_last_part = new_part;
	if (m_new_partitions->part(new_part) == NULL) {
		/* Altered partition contains misplaced row. */
		m_err_rec = table->record[0];
		DBUG_RETURN(HA_ERR_ROW_IN_WRONG_PARTITION);
	}
	m_new_partitions->get_prebuilt(m_prebuilt, new_part);
	result = ha_innobase::write_row(table->record[0]);
	m_new_partitions->set_from_prebuilt(m_prebuilt, new_part);
	DBUG_RETURN(result);
}

/** Allocate the array to hold blob heaps for all partitions */
mem_heap_t**
ha_innopart::alloc_blob_heap_array()
{
	DBUG_ENTER("ha_innopart::alloc_blob_heap_array");

	const ulint	len = sizeof(mem_heap_t*) * m_tot_parts;
	m_blob_heap_parts = static_cast<mem_heap_t**>(
		ut_zalloc(len, mem_key_partitioning));
	if (m_blob_heap_parts == NULL) {
		DBUG_RETURN(NULL);
	}

	DBUG_RETURN(m_blob_heap_parts);
}

/** Free the array that holds blob heaps for all partitions */
void
ha_innopart::free_blob_heap_array()
{
	DBUG_ENTER("ha_innopart::free_blob_heap_array");

	if (m_blob_heap_parts != NULL) {
		clear_blob_heaps();
		ut_free(m_blob_heap_parts);
		m_blob_heap_parts = NULL;
	}

	DBUG_VOID_RETURN;
}

void
ha_innopart::clear_blob_heaps()
{
	DBUG_ENTER("ha_innopart::clear_blob_heaps");

	if (m_blob_heap_parts == NULL) {
		DBUG_VOID_RETURN;
	}

	for (uint i = 0; i < m_tot_parts; i++) {
		if (m_blob_heap_parts[i] != NULL) {
			DBUG_PRINT("ha_innopart", ("freeing blob_heap: %p",
						   m_blob_heap_parts[i]));
			mem_heap_free(m_blob_heap_parts[i]);
			m_blob_heap_parts[i] = NULL;
		}
	}

	/* Reset blob_heap in m_prebuilt after freeing all heaps. It is set in
	ha_innopart::set_partition to the blob heap of current partition. */
	m_prebuilt->blob_heap = NULL;

	DBUG_VOID_RETURN;
}

/** Reset state of file to after 'open'. This function is called
after every statement for all tables used by that statement. */
int
ha_innopart::reset()
{
	DBUG_ENTER("ha_innopart::reset");

	clear_blob_heaps();

	DBUG_RETURN(ha_innobase::reset());
}

/****************************************************************************
 * DS-MRR implementation
 ***************************************************************************/

/* TODO: move the default implementations into the base handler class! */
/* TODO: See if it could be optimized for partitioned tables? */
/* Use default ha_innobase implementation for now... */