packages/mysql5
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
mysql5/mysql-5.7.27/sql/ndb_conflict.cc

2960 lines
93 KiB
C++
Raw Blame History

/*
Copyright (c) 2012, 2015, Oracle and/or its affiliates. All rights reserved.
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; 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 St, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <my_global.h> /* For config defines */
#include "ha_ndbcluster_glue.h"
#include "ndb_conflict.h"
#include "ndb_binlog_extra_row_info.h"
#include "ndb_table_guard.h"
extern st_ndb_slave_state g_ndb_slave_state;
#ifdef HAVE_NDB_BINLOG
#include "ndb_mi.h"
extern ulong opt_ndb_slave_conflict_role;
extern ulong opt_ndb_extra_logging;
#define NDBTAB NdbDictionary::Table
#define NDBCOL NdbDictionary::Column
#define NDB_EXCEPTIONS_TABLE_SUFFIX "$EX"
#define NDB_EXCEPTIONS_TABLE_SUFFIX_LOWER "$ex"
#define NDB_EXCEPTIONS_TABLE_COLUMN_PREFIX "NDB$"
#define NDB_EXCEPTIONS_TABLE_OP_TYPE "NDB$OP_TYPE"
#define NDB_EXCEPTIONS_TABLE_CONFLICT_CAUSE "NDB$CFT_CAUSE"
#define NDB_EXCEPTIONS_TABLE_ORIG_TRANSID "NDB$ORIG_TRANSID"
#define NDB_EXCEPTIONS_TABLE_COLUMN_OLD_SUFFIX "$OLD"
#define NDB_EXCEPTIONS_TABLE_COLUMN_NEW_SUFFIX "$NEW"
/*
Return true if a column has a specific prefix.
*/
bool
ExceptionsTableWriter::has_prefix_ci(const char *col_name,
const char *prefix,
CHARSET_INFO *cs)
{
uint col_len= strlen(col_name);
uint prefix_len= strlen(prefix);
if (col_len < prefix_len)
return false;
char col_name_prefix[FN_HEADLEN];
strncpy(col_name_prefix, col_name, prefix_len);
col_name_prefix[prefix_len]= '\0';
return (my_strcasecmp(cs,
col_name_prefix,
prefix) == 0);
}
/*
Return true if a column has a specific suffix
and sets the column_real_name to the column name
without the suffix.
*/
bool
ExceptionsTableWriter::has_suffix_ci(const char *col_name,
const char *suffix,
CHARSET_INFO *cs,
char *col_name_real)
{
uint col_len= strlen(col_name);
uint suffix_len= strlen(suffix);
const char *col_name_endp= col_name + col_len;
strcpy(col_name_real, col_name);
if (col_len > suffix_len &&
my_strcasecmp(cs,
col_name_endp - suffix_len,
suffix) == 0)
{
col_name_real[col_len - suffix_len]= '\0';
return true;
}
return false;
}
/*
Search for column_name in table and
return true if found. Also return what
position column was found in pos and possible
position in the primary key in key_pos.
*/
bool
ExceptionsTableWriter::find_column_name_ci(CHARSET_INFO *cs,
const char *col_name,
const NdbDictionary::Table* table,
int *pos,
int *key_pos)
{
int ncol= table->getNoOfColumns();
for(int m= 0; m < ncol; m++)
{
const NdbDictionary::Column* col= table->getColumn(m);
const char *tcol_name= col->getName();
if (col->getPrimaryKey())
(*key_pos)++;
if (my_strcasecmp(cs, col_name, tcol_name) == 0)
{
*pos= m;
return true;
}
}
return false;
}
bool
ExceptionsTableWriter::check_mandatory_columns(const NdbDictionary::Table* exceptionsTable)
{
DBUG_ENTER("ExceptionsTableWriter::check_mandatory_columns");
if (/* server id */
exceptionsTable->getColumn(0)->getType() == NDBCOL::Unsigned &&
exceptionsTable->getColumn(0)->getPrimaryKey() &&
/* master_server_id */
exceptionsTable->getColumn(1)->getType() == NDBCOL::Unsigned &&
exceptionsTable->getColumn(1)->getPrimaryKey() &&
/* master_epoch */
exceptionsTable->getColumn(2)->getType() == NDBCOL::Bigunsigned &&
exceptionsTable->getColumn(2)->getPrimaryKey() &&
/* count */
exceptionsTable->getColumn(3)->getType() == NDBCOL::Unsigned &&
exceptionsTable->getColumn(3)->getPrimaryKey()
)
DBUG_RETURN(true);
else
DBUG_RETURN(false);
}
bool
ExceptionsTableWriter::check_pk_columns(const NdbDictionary::Table* mainTable,
const NdbDictionary::Table* exceptionsTable,
int &k)
{
DBUG_ENTER("ExceptionsTableWriter::check_pk_columns");
const int fixed_cols= 4;
int ncol= mainTable->getNoOfColumns();
int nkey= mainTable->getNoOfPrimaryKeys();
/* Check columns that are part of the primary key */
for (int i= k= 0; i < ncol && k < nkey; i++)
{
const NdbDictionary::Column* col= mainTable->getColumn(i);
if (col->getPrimaryKey())
{
const NdbDictionary::Column* ex_col=
exceptionsTable->getColumn(fixed_cols + k);
if(!(ex_col != NULL &&
col->getType() == ex_col->getType() &&
col->getLength() == ex_col->getLength() &&
col->getNullable() == ex_col->getNullable()))
{
/*
Primary key type of the original table doesn't match
the primary key column of the execption table.
Assume that the table format has been extended and
check more below.
*/
DBUG_PRINT("info", ("Primary key column columns don't match, assume extended table"));
m_extended= true;
break;
}
/*
Store mapping of Exception table key# to
orig table attrid
*/
DBUG_PRINT("info", ("%u: Setting m_key_attrids[%i]= %i", __LINE__, k, i));
m_key_attrids[k]= i;
k++;
}
}
DBUG_RETURN(true);
}
bool
ExceptionsTableWriter::check_optional_columns(const NdbDictionary::Table* mainTable,
const NdbDictionary::Table* exceptionsTable,
char* msg_buf,
uint msg_buf_len,
const char** msg,
int &k,
char *error_details,
uint error_details_len)
{
DBUG_ENTER("ExceptionsTableWriter::check_optional_columns");
/*
Check optional columns.
Check if table has been extended by looking for
the NDB$ prefix. By looking at the columns in
reverse order we can determine if table has been
extended and then double check that the original
mandatory columns also have the NDB$ prefix.
If an incomplete primary key has been found or
additional non-primary key attributes from the
original table then table is also assumed to be
extended.
*/
const char* ex_tab_name= exceptionsTable->getName();
const int fixed_cols= 4;
bool ok= true;
int xncol= exceptionsTable->getNoOfColumns();
int i;
for (i= xncol - 1; i >= 0; i--)
{
const NdbDictionary::Column* col= exceptionsTable->getColumn(i);
const char* col_name= col->getName();
/*
We really need the CHARSET_INFO from when the table was
created but NdbDictionary::Table doesn't save this. This
means we cannot handle tables and execption tables defined
with a charset different than the system charset.
*/
CHARSET_INFO *cs= system_charset_info;
bool has_prefix= false;
if (has_prefix_ci(col_name, NDB_EXCEPTIONS_TABLE_COLUMN_PREFIX, cs))
{
has_prefix= true;
m_extended= true;
DBUG_PRINT("info",
("Exceptions table %s is extended with column %s",
ex_tab_name, col_name));
}
/* Check that mandatory columns have NDB$ prefix */
if (i < 4)
{
if (m_extended && !has_prefix)
{
my_snprintf(msg_buf, msg_buf_len,
"Exceptions table %s is extended, but mandatory column %s doesn't have the \'%s\' prefix",
ex_tab_name,
col_name,
NDB_EXCEPTIONS_TABLE_COLUMN_PREFIX);
*msg= msg_buf;
DBUG_RETURN(false);
}
}
k= i - fixed_cols;
/* Check for extended columns */
if (my_strcasecmp(cs,
col_name,
NDB_EXCEPTIONS_TABLE_OP_TYPE) == 0)
{
/* Check if ENUM or INT UNSIGNED */
if (exceptionsTable->getColumn(i)->getType() != NDBCOL::Char &&
exceptionsTable->getColumn(i)->getType() != NDBCOL::Unsigned)
{
my_snprintf(error_details, error_details_len,
"Table %s has incorrect type %u for NDB$OP_TYPE",
exceptionsTable->getName(),
exceptionsTable->getColumn(i)->getType());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
m_extended= true;
m_op_type_pos= i;
continue;
}
if (my_strcasecmp(cs,
col_name,
NDB_EXCEPTIONS_TABLE_CONFLICT_CAUSE) == 0)
{
/* Check if ENUM or INT UNSIGNED */
if (exceptionsTable->getColumn(i)->getType() != NDBCOL::Char &&
exceptionsTable->getColumn(i)->getType() != NDBCOL::Unsigned)
{
my_snprintf(error_details, error_details_len,
"Table %s has incorrect type %u for NDB$CFT_CAUSE",
exceptionsTable->getName(),
exceptionsTable->getColumn(i)->getType());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
m_extended= true;
m_conflict_cause_pos= i;
continue;
}
if (my_strcasecmp(cs,
col_name,
NDB_EXCEPTIONS_TABLE_ORIG_TRANSID) == 0)
{
if (exceptionsTable->getColumn(i)->getType() != NDBCOL::Bigunsigned)
{
my_snprintf(error_details, error_details_len,
"Table %s has incorrect type %u for NDB$ORIG_TRANSID",
exceptionsTable->getName(),
exceptionsTable->getColumn(i)->getType());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
m_extended= true;
m_orig_transid_pos= i;
continue;
}
/*
Check for any optional columns from the original table in the extended
table. Compare column types of columns with names matching a column in
the original table. If a non-primary key column is found we assume that
the table is extended.
*/
if (i >= fixed_cols)
{
int match= -1;
int match_k= -1;
COLUMN_VERSION column_version= DEFAULT;
char col_name_real[FN_HEADLEN];
/* Check for old or new column reference */
if (has_suffix_ci(col_name,
NDB_EXCEPTIONS_TABLE_COLUMN_OLD_SUFFIX,
cs,
col_name_real))
{
DBUG_PRINT("info", ("Found reference to old column %s", col_name));
column_version= OLD;
}
else if (has_suffix_ci(col_name,
NDB_EXCEPTIONS_TABLE_COLUMN_NEW_SUFFIX,
cs,
col_name_real))
{
DBUG_PRINT("info", ("Found reference to new column %s", col_name));
column_version= NEW;
}
DBUG_PRINT("info", ("Checking for original column %s", col_name_real));
/*
We really need the CHARSET_INFO from when the table was
created but NdbDictionary::Table doesn't save this. This
means we cannot handle tables end execption tables defined
with a charset different than the system charset.
*/
CHARSET_INFO *mcs= system_charset_info;
if (! find_column_name_ci(mcs, col_name_real, mainTable, &match, &match_k))
{
if (! strcmp(col_name, col_name_real))
{
/*
Column did have $OLD or $NEW suffix, but it didn't
match. Check if that is the real name of the column.
*/
match_k= -1;
if (find_column_name_ci(mcs, col_name, mainTable, &match, &match_k))
{
DBUG_PRINT("info", ("Column %s in main table %s has an unfortunate name",
col_name, mainTable->getName()));
}
}
}
/*
Check that old or new references are nullable
or have a default value.
*/
if (column_version != DEFAULT &&
match_k != -1)
{
if ((! col->getNullable()) &&
col->getDefaultValue() == NULL)
{
my_snprintf(error_details, error_details_len,
"Old or new column reference %s in table %s is not nullable and doesn't have a default value",
col->getName(), exceptionsTable->getName());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
}
if (match == -1)
{
/*
Column do not have the same name, could be allowed
if column is nullable or has a default value,
continue checking, but give a warning to user
*/
if ((! col->getNullable()) &&
col->getDefaultValue() == NULL)
{
my_snprintf(error_details, error_details_len,
"Extra column %s in table %s is not nullable and doesn't have a default value",
col->getName(), exceptionsTable->getName());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
my_snprintf(error_details, error_details_len,
"Column %s in extension table %s not found in %s",
col->getName(), exceptionsTable->getName(),
mainTable->getName());
DBUG_PRINT("info", ("%s", error_details));
my_snprintf(msg_buf, msg_buf_len,
"NDB Slave: exceptions table %s has suspicious "
"definition ((column %d): %s",
ex_tab_name, fixed_cols + k, error_details);
continue;
}
/* We have a matching name */
const NdbDictionary::Column* mcol= mainTable->getColumn(match);
if (col->getType() == mcol->getType())
{
DBUG_PRINT("info", ("Comparing column %s in exceptions table with column %s in main table", col->getName(), mcol->getName()));
/* We have matching type */
if (!mcol->getPrimaryKey())
{
/*
Matching non-key column found.
Check that column is nullable
or has a default value.
*/
if (col->getNullable() ||
col->getDefaultValue() != NULL)
{
DBUG_PRINT("info", ("Mapping column %s %s(%i) to %s(%i)",
col->getName(),
mainTable->getName(), match,
exceptionsTable->getName(), i));
/* Save position */
m_data_pos[i]= match;
m_column_version[i]= column_version;
}
else
{
my_snprintf(error_details, error_details_len,
"Data column %s in table %s is not nullable and doesn't have a default value",
col->getName(), exceptionsTable->getName());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
}
else
{
/* Column is part of the primary key */
if (column_version != DEFAULT)
{
my_snprintf(error_details, error_details_len,
"Old or new values of primary key columns cannot be referenced since primary keys cannot be updated, column %s in table %s",
col->getName(), exceptionsTable->getName());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
if (col->getNullable() == mcol->getNullable())
{
/*
Columns are both nullable or not nullable.
Save position.
*/
if (m_key_data_pos[match_k] != -1)
{
my_snprintf(error_details, error_details_len,
"Multiple references to the same key column %s in table %s",
col->getName(), exceptionsTable->getName());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
DBUG_PRINT("info", ("Setting m_key_data_pos[%i]= %i", match_k, i));
m_key_data_pos[match_k]= i;
if (i == fixed_cols + match_k)
{
/* Found key column in correct position */
if (!m_extended)
continue;
}
/*
Store mapping of Exception table key# to
orig table attrid
*/
DBUG_PRINT("info", ("%u: Setting m_key_attrids[%i]= %i", __LINE__, match_k, match));
m_key_attrids[match_k]= match;
m_extended= true;
}
else if (column_version == DEFAULT)
{
/*
Columns have same name and same type
Column with this name is part of primary key,
but both columns are not declared not null
*/
my_snprintf(error_details, error_details_len,
"Pk column %s not declared not null in both tables",
col->getName());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
}
}
else
{
/*
Columns have same name, but not the same type
*/
my_snprintf(error_details, error_details_len,
"Column %s has matching name to column %s for table %s, but wrong type, %u versus %u",
col->getName(), mcol->getName(),
mainTable->getName(),
col->getType(), mcol->getType());
DBUG_PRINT("info", ("%s", error_details));
ok= false;
break;
}
}
}
DBUG_RETURN(ok);
}
int
ExceptionsTableWriter::init(const NdbDictionary::Table* mainTable,
const NdbDictionary::Table* exceptionsTable,
char* msg_buf,
uint msg_buf_len,
const char** msg)
{
DBUG_ENTER("ExceptionsTableWriter::init");
const char* ex_tab_name= exceptionsTable->getName();
const int fixed_cols= 4;
*msg= NULL;
*msg_buf= '\0';
DBUG_PRINT("info", ("Checking definition of exceptions table %s",
ex_tab_name));
/*
Check that the table have the corrct number of columns
and the mandatory columns.
*/
bool ok=
exceptionsTable->getNoOfColumns() >= fixed_cols &&
exceptionsTable->getNoOfPrimaryKeys() == 4 &&
check_mandatory_columns(exceptionsTable);
if (ok)
{
char error_details[ FN_REFLEN ];
uint error_details_len= sizeof(error_details);
error_details[0]= '\0';
int ncol= mainTable->getNoOfColumns();
int nkey= mainTable->getNoOfPrimaryKeys();
int xncol= exceptionsTable->getNoOfColumns();
int i, k;
/* Initialize position arrays */
for(k=0; k < nkey; k++)
m_key_data_pos[k]= -1;
for(i=0; i < xncol; i++)
m_data_pos[i]= -1;
/* Initialize nullability information */
for(i=0; i < ncol; i++)
{
const NdbDictionary::Column* col= mainTable->getColumn(i);
m_col_nullable[i]= col->getNullable();
}
/*
Check that the primary key columns in the main table
are referenced correctly.
Then check if the table is extended with optional
columns.
*/
ok=
check_pk_columns(mainTable, exceptionsTable, k) &&
check_optional_columns(mainTable,
exceptionsTable,
msg_buf,
msg_buf_len,
msg,
k,
error_details,
error_details_len);
if (ok)
{
m_ex_tab= exceptionsTable;
m_pk_cols= nkey;
m_cols= ncol;
m_xcols= xncol;
if (m_extended && strlen(msg_buf) > 0)
*msg= msg_buf;
DBUG_RETURN(0);
}
else
my_snprintf(msg_buf, msg_buf_len,
"NDB Slave: exceptions table %s has wrong "
"definition (column %d): %s",
ex_tab_name, fixed_cols + k, error_details);
}
else
my_snprintf(msg_buf, msg_buf_len,
"NDB Slave: exceptions table %s has wrong "
"definition (initial %d columns)",
ex_tab_name, fixed_cols);
*msg= msg_buf;
DBUG_RETURN(-1);
}
void
ExceptionsTableWriter::mem_free(Ndb* ndb)
{
if (m_ex_tab)
{
NdbDictionary::Dictionary* dict = ndb->getDictionary();
dict->removeTableGlobal(*m_ex_tab, 0);
m_ex_tab= 0;
}
}
int
ExceptionsTableWriter::writeRow(NdbTransaction* trans,
const NdbRecord* keyRecord,
const NdbRecord* dataRecord,
uint32 server_id,
uint32 master_server_id,
uint64 master_epoch,
const uchar* oldRowPtr,
const uchar* newRowPtr,
enum_conflicting_op_type op_type,
enum_conflict_cause conflict_cause,
uint64 orig_transid,
const MY_BITMAP *write_set,
NdbError& err)
{
DBUG_ENTER("ExceptionsTableWriter::writeRow");
DBUG_PRINT("info", ("op_type(pos):%u(%u), conflict_cause(pos):%u(%u), orig_transid:%llu(%u)",
op_type, m_op_type_pos,
conflict_cause, m_conflict_cause_pos,
orig_transid, m_orig_transid_pos));
DBUG_ASSERT(write_set != NULL);
assert(err.code == 0);
const uchar* rowPtr= (op_type == DELETE_ROW)? oldRowPtr : newRowPtr;
do
{
/* Have exceptions table, add row to it */
const NDBTAB *ex_tab= m_ex_tab;
/* get insert op */
NdbOperation *ex_op= trans->getNdbOperation(ex_tab);
if (ex_op == NULL)
{
err= trans->getNdbError();
break;
}
if (ex_op->insertTuple() == -1)
{
err= ex_op->getNdbError();
break;
}
{
uint32 count= (uint32)++m_count;
/* Set mandatory columns */
if (ex_op->setValue((Uint32)0, (const char *)&(server_id)) ||
ex_op->setValue((Uint32)1, (const char *)&(master_server_id)) ||
ex_op->setValue((Uint32)2, (const char *)&(master_epoch)) ||
ex_op->setValue((Uint32)3, (const char *)&(count)))
{
err= ex_op->getNdbError();
break;
}
/* Set optional columns */
if (m_extended)
{
if (m_op_type_pos)
{
if (m_ex_tab->getColumn(m_op_type_pos)->getType()
== NDBCOL::Char)
{
/* Defined as ENUM */
char op_type_val= (char)op_type;
if (ex_op->setValue((Uint32)m_op_type_pos,
(const char *)&(op_type_val)))
{
err= ex_op->getNdbError();
break;
}
}
else
{
uint32 op_type_val= op_type;
if (ex_op->setValue((Uint32)m_op_type_pos,
(const char *)&(op_type_val)))
{
err= ex_op->getNdbError();
break;
}
}
}
if (m_conflict_cause_pos)
{
if (m_ex_tab->getColumn(m_conflict_cause_pos)->getType()
== NDBCOL::Char)
{
/* Defined as ENUM */
char conflict_cause_val= (char)conflict_cause;
if (ex_op->setValue((Uint32)m_conflict_cause_pos,
(const char *)&(conflict_cause_val)))
{
err= ex_op->getNdbError();
break;
}
}
else
{
uint32 conflict_cause_val= conflict_cause;
if (ex_op->setValue((Uint32)m_conflict_cause_pos,
(const char *)&(conflict_cause_val)))
{
err= ex_op->getNdbError();
break;
}
}
}
if (m_orig_transid_pos != 0)
{
const NdbDictionary::Column* col= m_ex_tab->getColumn(m_orig_transid_pos);
if (orig_transid == Ndb_binlog_extra_row_info::InvalidTransactionId
&&
col->getNullable())
{
if (ex_op->setValue((Uint32) m_orig_transid_pos, (char*)NULL))
{
err= ex_op->getNdbError();
break;
}
}
else
{
DBUG_PRINT("info", ("Setting orig_transid (%u) for table %s", m_orig_transid_pos, ex_tab->getName()));
uint64 orig_transid_val= orig_transid;
if (ex_op->setValue((Uint32)m_orig_transid_pos,
(const char *)&(orig_transid_val)))
{
err= ex_op->getNdbError();
break;
}
}
}
}
}
/* copy primary keys */
{
int nkey= m_pk_cols;
int k;
for (k= 0; k < nkey; k++)
{
DBUG_ASSERT(rowPtr != NULL);
if (m_key_data_pos[k] != -1)
{
const uchar* data=
(const uchar*) NdbDictionary::getValuePtr(keyRecord,
(const char*) rowPtr,
m_key_attrids[k]);
if (ex_op->setValue((Uint32) m_key_data_pos[k], (const char*)data) == -1)
{
err= ex_op->getNdbError();
break;
}
}
}
}
/* Copy additional data */
if (m_extended)
{
int xncol= m_xcols;
int i;
for (i= 0; i < xncol; i++)
{
const NdbDictionary::Column* col= m_ex_tab->getColumn(i);
const uchar* default_value= (const uchar*) col->getDefaultValue();
DBUG_PRINT("info", ("Checking column %s(%i)%s", col->getName(), i,
(default_value)?", has default value":""));
DBUG_ASSERT(rowPtr != NULL);
if (m_data_pos[i] != -1)
{
const uchar* row_vPtr= NULL;
switch (m_column_version[i]) {
case DEFAULT:
row_vPtr= rowPtr;
break;
case OLD:
if (op_type != WRITE_ROW)
row_vPtr= oldRowPtr;
break;
case NEW:
if (op_type != DELETE_ROW)
row_vPtr= newRowPtr;
}
if (row_vPtr == NULL ||
(m_col_nullable[m_data_pos[i]] &&
NdbDictionary::isNull(dataRecord,
(const char*) row_vPtr,
m_data_pos[i])))
{
DBUG_PRINT("info", ("Column %s is set to NULL because it is NULL", col->getName()));
if (ex_op->setValue((Uint32) i, (char*)NULL))
{
err= ex_op->getNdbError();
break;
}
}
else if (write_set != NULL && bitmap_is_set(write_set, m_data_pos[i]))
{
DBUG_PRINT("info", ("Column %s is set", col->getName()));
const uchar* data=
(const uchar*) NdbDictionary::getValuePtr(dataRecord,
(const char*) row_vPtr,
m_data_pos[i]);
if (ex_op->setValue((Uint32) i, (const char*)data) == -1)
{
err= ex_op->getNdbError();
break;
}
}
else if (default_value != NULL)
{
DBUG_PRINT("info", ("Column %s is not set to NULL because it has a default value", col->getName()));
/*
* Column has a default value
* Since no value was set in write_set
* we let the default value be set from
* Ndb instead.
*/
}
else
{
DBUG_PRINT("info", ("Column %s is set to NULL because it not in write_set", col->getName()));
if (ex_op->setValue((Uint32) i, (char*)NULL))
{
err= ex_op->getNdbError();
break;
}
}
}
}
}
} while (0);
if (err.code != 0)
{
if (err.classification == NdbError::SchemaError)
{
/*
* Something up with Exceptions table schema, forget it.
* No further exceptions will be recorded.
* Caller will log this and slave will stop.
*/
NdbDictionary::Dictionary* dict= trans->getNdb()->getDictionary();
dict->removeTableGlobal(*m_ex_tab, false);
m_ex_tab= NULL;
DBUG_RETURN(0);
}
DBUG_RETURN(-1);
}
DBUG_RETURN(0);
}
/* HAVE_NDB_BINLOG */
#endif
/**
st_ndb_slave_state constructor
Initialise Ndb Slave state object
*/
st_ndb_slave_state::st_ndb_slave_state()
: current_delete_delete_count(0),
current_reflect_op_prepare_count(0),
current_reflect_op_discard_count(0),
current_refresh_op_count(0),
current_master_server_epoch(0),
current_master_server_epoch_committed(false),
current_max_rep_epoch(0),
conflict_flags(0),
retry_trans_count(0),
current_trans_row_conflict_count(0),
current_trans_row_reject_count(0),
current_trans_in_conflict_count(0),
last_conflicted_epoch(0),
last_stable_epoch(0),
total_delete_delete_count(0),
total_reflect_op_prepare_count(0),
total_reflect_op_discard_count(0),
total_refresh_op_count(0),
max_rep_epoch(0),
sql_run_id(~Uint32(0)),
trans_row_conflict_count(0),
trans_row_reject_count(0),
trans_detect_iter_count(0),
trans_in_conflict_count(0),
trans_conflict_commit_count(0),
trans_conflict_apply_state(SAS_NORMAL),
trans_dependency_tracker(NULL)
{
memset(current_violation_count, 0, sizeof(current_violation_count));
memset(total_violation_count, 0, sizeof(total_violation_count));
/* Init conflict handling state memroot */
const size_t CONFLICT_MEMROOT_BLOCK_SIZE = 32768;
init_alloc_root(PSI_INSTRUMENT_ME,
&conflict_mem_root, CONFLICT_MEMROOT_BLOCK_SIZE, 0);
}
st_ndb_slave_state::~st_ndb_slave_state()
{
free_root(&conflict_mem_root, 0);
}
/**
resetPerAttemptCounters
Reset the per-epoch-transaction-application-attempt counters
*/
void
st_ndb_slave_state::resetPerAttemptCounters()
{
memset(current_violation_count, 0, sizeof(current_violation_count));
current_delete_delete_count = 0;
current_reflect_op_prepare_count = 0;
current_reflect_op_discard_count = 0;
current_refresh_op_count = 0;
current_trans_row_conflict_count = 0;
current_trans_row_reject_count = 0;
current_trans_in_conflict_count = 0;
conflict_flags = 0;
current_max_rep_epoch = 0;
}
/**
atTransactionAbort()
Called by Slave SQL thread during transaction abort.
*/
void
st_ndb_slave_state::atTransactionAbort()
{
#ifdef HAVE_NDB_BINLOG
/* Reset any gathered transaction dependency information */
atEndTransConflictHandling();
trans_conflict_apply_state = SAS_NORMAL;
#endif
/* Reset current-transaction counters + state */
resetPerAttemptCounters();
}
/**
atTransactionCommit()
Called by Slave SQL thread after transaction commit
*/
void
st_ndb_slave_state::atTransactionCommit(Uint64 epoch)
{
assert( ((trans_dependency_tracker == NULL) &&
(trans_conflict_apply_state == SAS_NORMAL)) ||
((trans_dependency_tracker != NULL) &&
(trans_conflict_apply_state == SAS_TRACK_TRANS_DEPENDENCIES)) );
assert( trans_conflict_apply_state != SAS_APPLY_TRANS_DEPENDENCIES );
/* Merge committed transaction counters into total state
* Then reset current transaction counters
*/
Uint32 total_conflicts = 0;
for (int i=0; i < CFT_NUMBER_OF_CFTS; i++)
{
total_conflicts+= current_violation_count[i];
total_violation_count[i]+= current_violation_count[i];
}
total_delete_delete_count+= current_delete_delete_count;
total_reflect_op_prepare_count+= current_reflect_op_prepare_count;
total_reflect_op_discard_count+= current_reflect_op_discard_count;
total_refresh_op_count+= current_refresh_op_count;
trans_row_conflict_count+= current_trans_row_conflict_count;
trans_row_reject_count+= current_trans_row_reject_count;
trans_in_conflict_count+= current_trans_in_conflict_count;
if (current_trans_in_conflict_count)
trans_conflict_commit_count++;
if (current_max_rep_epoch > max_rep_epoch)
{
DBUG_PRINT("info", ("Max replicated epoch increases from %llu to %llu",
max_rep_epoch,
current_max_rep_epoch));
max_rep_epoch = current_max_rep_epoch;
}
{
bool hadConflict = false;
if (total_conflicts > 0)
{
/**
* Conflict detected locally
*/
DBUG_PRINT("info", ("Last conflicted epoch increases from %llu to %llu",
last_conflicted_epoch,
epoch));
hadConflict = true;
}
else
{
/**
* Update last_conflicted_epoch if we applied reflected or refresh ops
* (Implies Secondary role in asymmetric algorithms)
*/
assert(current_reflect_op_prepare_count >= current_reflect_op_discard_count);
Uint32 current_reflect_op_apply_count = current_reflect_op_prepare_count -
current_reflect_op_discard_count;
if (current_reflect_op_apply_count > 0 ||
current_refresh_op_count > 0)
{
DBUG_PRINT("info", ("Reflected (%u) or Refresh (%u) operations applied this "
"epoch, increasing last conflicted epoch from %llu to %llu.",
current_reflect_op_apply_count,
current_refresh_op_count,
last_conflicted_epoch,
epoch));
hadConflict = true;
}
}
/* Update status vars */
if (hadConflict)
{
last_conflicted_epoch = epoch;
}
else
{
if (max_rep_epoch >= last_conflicted_epoch)
{
/**
* This epoch which has looped the circle was stable -
* no new conflicts have been found / corrected since
* it was logged
*/
last_stable_epoch = max_rep_epoch;
/**
* Note that max_rep_epoch >= last_conflicted_epoch
* implies that there are no currently known-about
* conflicts.
* On the primary this is a definitive fact as it
* finds out about all conflicts immediately.
* On the secondary it does not mean that there
* are not committed conflicts, just that they
* have not started being corrected yet.
*/
}
}
}
resetPerAttemptCounters();
/* Clear per-epoch-transaction retry_trans_count */
retry_trans_count = 0;
current_master_server_epoch_committed = true;
DBUG_EXECUTE_IF("ndb_slave_fail_marking_epoch_committed",
{
fprintf(stderr,
"Slave clearing epoch committed flag "
"for epoch %llu/%llu (%llu)\n",
current_master_server_epoch >> 32,
current_master_server_epoch & 0xffffffff,
current_master_server_epoch);
current_master_server_epoch_committed = false;
});
}
/**
verifyNextEpoch
Check that a new incoming epoch from the relay log
is expected given the current slave state, previous
epoch etc.
This is checking Generic replication errors, with
a user warning thrown in too.
*/
bool
st_ndb_slave_state::verifyNextEpoch(Uint64 next_epoch,
Uint32 master_server_id) const
{
DBUG_ENTER("verifyNextEpoch");
#ifdef HAVE_NDB_BINLOG
/**
WRITE_ROW to ndb_apply_status injected by MySQLD
immediately upstream of us.
Now we do some validation of the incoming epoch transaction's
epoch - to make sure that we are getting a sensible sequence
of epochs.
*/
bool first_epoch_since_slave_start = (ndb_mi_get_slave_run_id() != sql_run_id);
DBUG_PRINT("info", ("ndb_apply_status write from upstream master."
"ServerId %u, Epoch %llu/%llu (%llu) "
"Current master server epoch %llu/%llu (%llu)"
"Current master server epoch committed? %u",
master_server_id,
next_epoch >> 32,
next_epoch & 0xffffffff,
next_epoch,
current_master_server_epoch >> 32,
current_master_server_epoch & 0xffffffff,
current_master_server_epoch,
current_master_server_epoch_committed));
DBUG_PRINT("info", ("mi_slave_run_id=%u, ndb_slave_state_run_id=%u",
ndb_mi_get_slave_run_id(), sql_run_id));
DBUG_PRINT("info", ("First epoch since slave start : %u",
first_epoch_since_slave_start));
/* Analysis of nextEpoch generally depends on whether it's the first or not */
if (first_epoch_since_slave_start)
{
/**
First epoch since slave start - might've had a CHANGE MASTER command,
since we were last running, so we are not too strict about epoch
changes, but we will warn.
*/
if (next_epoch < current_master_server_epoch)
{
sql_print_warning("NDB Slave : At SQL thread start "
"applying epoch %llu/%llu "
"(%llu) from Master ServerId %u which is lower than previously "
"applied epoch %llu/%llu (%llu). "
"Group Master Log : %s Group Master Log Pos : %llu. "
"Check slave positioning.",
next_epoch >> 32,
next_epoch & 0xffffffff,
next_epoch,
master_server_id,
current_master_server_epoch >> 32,
current_master_server_epoch & 0xffffffff,
current_master_server_epoch,
ndb_mi_get_group_master_log_name(),
ndb_mi_get_group_master_log_pos());
/* Slave not stopped */
}
else if (next_epoch == current_master_server_epoch)
{
/**
Could warn that started on already applied epoch,
but this is often harmless.
*/
}
else
{
/* next_epoch > current_master_server_epoch - fine. */
}
}
else
{
/**
! first_epoch_since_slave_start
Slave has already applied some epoch in this run, so we expect
either :
a) previous epoch committed ok and next epoch is higher
or
b) previous epoch not committed and next epoch is the same
(Retry case)
*/
if (next_epoch < current_master_server_epoch)
{
/* Should never happen */
sql_print_error("NDB Slave : SQL thread stopped as "
"applying epoch %llu/%llu "
"(%llu) from Master ServerId %u which is lower than previously "
"applied epoch %llu/%llu (%llu). "
"Group Master Log : %s Group Master Log Pos : %llu",
next_epoch >> 32,
next_epoch & 0xffffffff,
next_epoch,
master_server_id,
current_master_server_epoch >> 32,
current_master_server_epoch & 0xffffffff,
current_master_server_epoch,
ndb_mi_get_group_master_log_name(),
ndb_mi_get_group_master_log_pos());
/* Stop the slave */
DBUG_RETURN(false);
}
else if (next_epoch == current_master_server_epoch)
{
/**
This is ok if we are retrying - e.g. the
last epoch was not committed
*/
if (current_master_server_epoch_committed)
{
/* This epoch is committed already, why are we replaying it? */
sql_print_error("NDB Slave : SQL thread stopped as attempted "
"to reapply already committed epoch %llu/%llu (%llu) "
"from server id %u. "
"Group Master Log : %s Group Master Log Pos : %llu.",
current_master_server_epoch >> 32,
current_master_server_epoch & 0xffffffff,
current_master_server_epoch,
master_server_id,
ndb_mi_get_group_master_log_name(),
ndb_mi_get_group_master_log_pos());
/* Stop the slave */
DBUG_RETURN(false);
}
else
{
/* Probably a retry, no problem. */
}
}
else
{
/**
next_epoch > current_master_server_epoch
This is the normal case, *unless* the previous epoch
did not commit - in which case it may be a bug in
transaction retry.
*/
if (!current_master_server_epoch_committed)
{
/**
We've moved onto a new epoch without committing
the last - probably a bug in transaction retry
*/
sql_print_error("NDB Slave : SQL thread stopped as attempting to "
"apply new epoch %llu/%llu (%llu) while lower "
"received epoch %llu/%llu (%llu) has not been "
"committed. Master server id : %u. "
"Group Master Log : %s Group Master Log Pos : %llu.",
next_epoch >> 32,
next_epoch & 0xffffffff,
next_epoch,
current_master_server_epoch >> 32,
current_master_server_epoch & 0xffffffff,
current_master_server_epoch,
master_server_id,
ndb_mi_get_group_master_log_name(),
ndb_mi_get_group_master_log_pos());
/* Stop the slave */
DBUG_RETURN(false);
}
else
{
/* Normal case of next epoch after committing last */
}
}
}
#endif
/* Epoch looks ok */
DBUG_RETURN(true);
}
/**
atApplyStatusWrite
Called by Slave SQL thread when applying an event to the
ndb_apply_status table
*/
int
st_ndb_slave_state::atApplyStatusWrite(Uint32 master_server_id,
Uint32 row_server_id,
Uint64 row_epoch,
bool is_row_server_id_local)
{
DBUG_ENTER("atApplyStatusWrite");
if (row_server_id == master_server_id)
{
/* This is an apply status write from the immediate master */
if (!verifyNextEpoch(row_epoch,
master_server_id))
{
/* Problem with the next epoch, stop the slave SQL thread */
DBUG_RETURN(HA_ERR_ROWS_EVENT_APPLY);
}
/* Epoch ok, record that we're working on it now... */
current_master_server_epoch = row_epoch;
current_master_server_epoch_committed = false;
assert(! is_row_server_id_local);
}
else if (is_row_server_id_local)
{
DBUG_PRINT("info", ("Recording application of local server %u epoch %llu "
" which is %s.",
row_server_id, row_epoch,
(row_epoch > current_max_rep_epoch)?
" new highest." : " older than previously applied"));
if (row_epoch > current_max_rep_epoch)
{
/*
Store new highest epoch in thdvar. If we commit successfully
then this can become the new global max
*/
current_max_rep_epoch = row_epoch;
}
}
DBUG_RETURN(0);
}
/**
atResetSlave()
Called when RESET SLAVE command issued - in context of command client.
*/
void
st_ndb_slave_state::atResetSlave()
{
/* Reset the Maximum replicated epoch vars
* on slave reset
* No need to touch the sql_run_id as that
* will increment if the slave is started
* again.
*/
resetPerAttemptCounters();
retry_trans_count = 0;
max_rep_epoch = 0;
last_conflicted_epoch = 0;
last_stable_epoch = 0;
/* Reset current master server epoch
* This avoids warnings when replaying a lower
* epoch number after a RESET SLAVE - in this
* case we assume the user knows best.
*/
current_master_server_epoch = 0;
current_master_server_epoch_committed = false;
}
/**
atStartSlave()
Called by Slave SQL thread when first applying a row to Ndb after
a START SLAVE command.
*/
void
st_ndb_slave_state::atStartSlave()
{
#ifdef HAVE_NDB_BINLOG
if (trans_conflict_apply_state != SAS_NORMAL)
{
/*
Remove conflict handling state on a SQL thread
restart
*/
atEndTransConflictHandling();
trans_conflict_apply_state = SAS_NORMAL;
}
#endif
}
bool
st_ndb_slave_state::checkSlaveConflictRoleChange(enum_slave_conflict_role old_role,
enum_slave_conflict_role new_role,
const char** failure_cause)
{
if (old_role == new_role)
return true;
/**
* Initial role is SCR_NONE
* Allowed transitions :
* SCR_NONE -> SCR_PASS
* SCR_NONE -> SCR_PRIMARY
* SCR_NONE -> SCR_SECONDARY
* SCR_PRIMARY -> SCR_NONE
* SCR_PRIMARY -> SCR_SECONDARY
* SCR_SECONDARY -> SCR_NONE
* SCR_SECONDARY -> SCR_PRIMARY
* SCR_PASS -> SCR_NONE
*
* Disallowed transitions
* SCR_PASS -> SCR_PRIMARY
* SCR_PASS -> SCR_SECONDARY
* SCR_PRIMARY -> SCR_PASS
* SCR_SECONDARY -> SCR_PASS
*/
bool bad_transition = false;
*failure_cause = "Internal error";
switch (old_role)
{
case SCR_NONE:
break;
case SCR_PRIMARY:
case SCR_SECONDARY:
bad_transition = (new_role == SCR_PASS);
break;
case SCR_PASS:
bad_transition = ((new_role == SCR_PRIMARY) ||
(new_role == SCR_SECONDARY));
break;
default:
assert(false);
return false;
}
if (bad_transition)
{
*failure_cause = "Invalid role change.";
return false;
}
#ifdef HAVE_NDB_BINLOG
/* Check that Slave SQL thread is not running */
if (ndb_mi_get_slave_sql_running())
{
*failure_cause = "Cannot change role while Slave SQL "
"thread is running. Use STOP SLAVE first.";
return false;
}
#endif
return true;
}
#ifdef HAVE_NDB_BINLOG
/**
atEndTransConflictHandling
Called when transactional conflict handling has completed.
*/
void
st_ndb_slave_state::atEndTransConflictHandling()
{
DBUG_ENTER("atEndTransConflictHandling");
/* Release any conflict handling state */
if (trans_dependency_tracker)
{
current_trans_in_conflict_count =
trans_dependency_tracker->get_conflict_count();
trans_dependency_tracker = NULL;
free_root(&conflict_mem_root, MY_MARK_BLOCKS_FREE);
}
DBUG_VOID_RETURN;
}
/**
atBeginTransConflictHandling()
Called by Slave SQL thread when it determines that Transactional
Conflict handling is required
*/
void
st_ndb_slave_state::atBeginTransConflictHandling()
{
DBUG_ENTER("atBeginTransConflictHandling");
/*
Allocate and initialise Transactional Conflict
Resolution Handling Structures
*/
assert(trans_dependency_tracker == NULL);
trans_dependency_tracker = DependencyTracker::newDependencyTracker(&conflict_mem_root);
DBUG_VOID_RETURN;
}
/**
atPrepareConflictDetection
Called by Slave SQL thread prior to defining an operation on
a table with conflict detection defined.
*/
int
st_ndb_slave_state::atPrepareConflictDetection(const NdbDictionary::Table* table,
const NdbRecord* key_rec,
const uchar* row_data,
Uint64 transaction_id,
bool& handle_conflict_now)
{
DBUG_ENTER("atPrepareConflictDetection");
/*
Slave is preparing to apply an operation with conflict detection.
If we're performing Transactional Conflict Resolution, take
extra steps
*/
switch( trans_conflict_apply_state )
{
case SAS_NORMAL:
DBUG_PRINT("info", ("SAS_NORMAL : No special handling"));
/* No special handling */
break;
case SAS_TRACK_TRANS_DEPENDENCIES:
{
DBUG_PRINT("info", ("SAS_TRACK_TRANS_DEPENDENCIES : Tracking operation"));
/*
Track this operation and its transaction id, to determine
inter-transaction dependencies by {table, primary key}
*/
assert( trans_dependency_tracker );
int res = trans_dependency_tracker
->track_operation(table,
key_rec,
row_data,
transaction_id);
if (res != 0)
{
sql_print_error("%s", trans_dependency_tracker->get_error_text());
DBUG_RETURN(res);
}
/* Proceed as normal */
break;
}
case SAS_APPLY_TRANS_DEPENDENCIES:
{
DBUG_PRINT("info", ("SAS_APPLY_TRANS_DEPENDENCIES : Deciding whether to apply"));
/*
Check if this operation's transaction id is marked in-conflict.
If it is, we tell the caller to perform conflict resolution now instead
of attempting to apply the operation.
*/
assert( trans_dependency_tracker );
if (trans_dependency_tracker->in_conflict(transaction_id))
{
DBUG_PRINT("info", ("Event for transaction %llu is conflicting. Handling.",
transaction_id));
current_trans_row_reject_count++;
handle_conflict_now = true;
DBUG_RETURN(0);
}
/*
This transaction is not marked in-conflict, so continue with normal
processing.
Note that normal processing may subsequently detect a conflict which
didn't exist at the time of the previous TRACK_DEPENDENCIES pass.
In this case, we will rollback and repeat the TRACK_DEPENDENCIES
stage.
*/
DBUG_PRINT("info", ("Event for transaction %llu is OK, applying",
transaction_id));
break;
}
}
DBUG_RETURN(0);
}
/**
atTransConflictDetected
Called by the Slave SQL thread when a conflict is detected on
an executed operation.
*/
int
st_ndb_slave_state::atTransConflictDetected(Uint64 transaction_id)
{
DBUG_ENTER("atTransConflictDetected");
/*
The Slave has detected a conflict on an operation applied
to a table with Transactional Conflict Resolution defined.
Handle according to current state.
*/
conflict_flags |= SCS_TRANS_CONFLICT_DETECTED_THIS_PASS;
current_trans_row_conflict_count++;
switch (trans_conflict_apply_state)
{
case SAS_NORMAL:
{
DBUG_PRINT("info", ("SAS_NORMAL : Conflict on op on table with trans detection."
"Requires multi-pass resolution. Will transition to "
"SAS_TRACK_TRANS_DEPENDENCIES at Commit."));
/*
Conflict on table with transactional conflict resolution
defined.
This is the trigger that we will do transactional conflict
resolution.
Record that we need to do multiple passes to correctly
perform resolution.
TODO : Early exit from applying epoch?
*/
break;
}
case SAS_TRACK_TRANS_DEPENDENCIES:
{
DBUG_PRINT("info", ("SAS_TRACK_TRANS_DEPENDENCIES : Operation in transaction %llu "
"had conflict",
transaction_id));
/*
Conflict on table with transactional conflict resolution
defined.
We will mark the operation's transaction_id as in-conflict,
so that any other operations on the transaction are also
considered in-conflict, and any dependent transactions are also
considered in-conflict.
*/
assert(trans_dependency_tracker != NULL);
int res = trans_dependency_tracker
->mark_conflict(transaction_id);
if (res != 0)
{
sql_print_error("%s", trans_dependency_tracker->get_error_text());
DBUG_RETURN(res);
}
break;
}
case SAS_APPLY_TRANS_DEPENDENCIES:
{
/*
This must be a new conflict, not noticed on the previous
pass.
*/
DBUG_PRINT("info", ("SAS_APPLY_TRANS_DEPENDENCIES : Conflict detected. "
"Must be further conflict. Will return to "
"SAS_TRACK_TRANS_DEPENDENCIES state at commit."));
// TODO : Early exit from applying epoch
break;
}
default:
break;
}
DBUG_RETURN(0);
}
/**
atConflictPreCommit
Called by the Slave SQL thread prior to committing a Slave transaction.
This method can request that the Slave transaction is retried.
State transitions :
START SLAVE /
RESET SLAVE /
STARTUP
|
|
v
****************
* SAS_NORMAL *
****************
^ |
No transactional | | Conflict on transactional table
conflicts | | (Rollback)
(Commit) | |
| v
**********************************
* SAS_TRACK_TRANS_DEPENDENCIES *
**********************************
^ I ^
More I I Dependencies |
conflicts I I determined | No new conflicts
found I I (Rollback) | (Commit)
(Rollback) I I |
I v |
**********************************
* SAS_APPLY_TRANS_DEPENDENCIES *
**********************************
Operation
The initial state is SAS_NORMAL.
On detecting a conflict on a transactional conflict detetecing table,
SAS_TRACK_TRANS_DEPENDENCIES is entered, and the epoch transaction is
rolled back and reapplied.
In SAS_TRACK_TRANS_DEPENDENCIES state, transaction dependencies and
conflicts are tracked as the epoch transaction is applied.
Then the Slave transitions to SAS_APPLY_TRANS_DEPENDENCIES state, and
the epoch transaction is rolled back and reapplied.
In the SAS_APPLY_TRANS_DEPENDENCIES state, operations for transactions
marked as in-conflict are not applied.
If this results in no new conflicts, the epoch transaction is committed,
and the SAS_TRACK_TRANS_DEPENDENCIES state is re-entered for processing
the next replicated epch transaction.
If it results in new conflicts, the epoch transactions is rolled back, and
the SAS_TRACK_TRANS_DEPENDENCIES state is re-entered again, to determine
the new set of dependencies.
If no conflicts are found in the SAS_TRACK_TRANS_DEPENDENCIES state, then
the epoch transaction is committed, and the Slave transitions to SAS_NORMAL
state.
Properties
1) Normally, there is no transaction dependency tracking overhead paid by
the slave.
2) On first detecting a transactional conflict, the epoch transaction must be
applied at least three times, with two rollbacks.
3) Transactional conflicts detected in subsequent epochs require the epoch
transaction to be applied two times, with one rollback.
4) A loop between states SAS_TRACK_TRANS_DEPENDENCIES and SAS_APPLY_TRANS_
DEPENDENCIES occurs when further transactional conflicts are discovered
in SAS_APPLY_TRANS_DEPENDENCIES state. This implies that the conflicts
discovered in the SAS_TRACK_TRANS_DEPENDENCIES state must not be complete,
so we revisit that state to get a more complete picture.
5) The number of iterations of this loop is fixed to a hard coded limit, after
which the Slave will stop with an error. This should be an unlikely
occurrence, as it requires not just n conflicts, but at least 1 new conflict
appearing between the transactions in the epoch transaction and the
database between the two states, n times in a row.
6) Where conflicts are occasional, as expected, the post-commit transition to
SAS_TRACK_TRANS_DEPENDENCIES rather than SAS_NORMAL results in one epoch
transaction having its transaction dependencies needlessly tracked.
*/
int
st_ndb_slave_state::atConflictPreCommit(bool& retry_slave_trans)
{
DBUG_ENTER("atConflictPreCommit");
/*
Prior to committing a Slave transaction, we check whether
Transactional conflicts have been detected which require
us to retry the slave transaction
*/
retry_slave_trans = false;
switch(trans_conflict_apply_state)
{
case SAS_NORMAL:
{
DBUG_PRINT("info", ("SAS_NORMAL"));
/*
Normal case. Only if we defined conflict detection on a table
with transactional conflict detection, and saw conflicts (on any table)
do we go to another state
*/
if (conflict_flags & SCS_TRANS_CONFLICT_DETECTED_THIS_PASS)
{
DBUG_PRINT("info", ("Conflict(s) detected this pass, transitioning to "
"SAS_TRACK_TRANS_DEPENDENCIES."));
assert(conflict_flags & SCS_OPS_DEFINED);
/* Transactional conflict resolution required, switch state */
atBeginTransConflictHandling();
resetPerAttemptCounters();
trans_conflict_apply_state = SAS_TRACK_TRANS_DEPENDENCIES;
retry_slave_trans = true;
}
break;
}
case SAS_TRACK_TRANS_DEPENDENCIES:
{
DBUG_PRINT("info", ("SAS_TRACK_TRANS_DEPENDENCIES"));
if (conflict_flags & SCS_TRANS_CONFLICT_DETECTED_THIS_PASS)
{
/*
Conflict on table with transactional detection
this pass, we have collected the details and
dependencies, now transition to
SAS_APPLY_TRANS_DEPENDENCIES and
reapply the epoch transaction without the
conflicting transactions.
*/
assert(conflict_flags & SCS_OPS_DEFINED);
DBUG_PRINT("info", ("Transactional conflicts, transitioning to "
"SAS_APPLY_TRANS_DEPENDENCIES"));
trans_conflict_apply_state = SAS_APPLY_TRANS_DEPENDENCIES;
trans_detect_iter_count++;
retry_slave_trans = true;
break;
}
else
{
/*
No transactional conflicts detected this pass, lets
return to SAS_NORMAL state after commit for more efficient
application of epoch transactions
*/
DBUG_PRINT("info", ("No transactional conflicts, transitioning to "
"SAS_NORMAL"));
atEndTransConflictHandling();
trans_conflict_apply_state = SAS_NORMAL;
break;
}
}
case SAS_APPLY_TRANS_DEPENDENCIES:
{
DBUG_PRINT("info", ("SAS_APPLY_TRANS_DEPENDENCIES"));
assert(conflict_flags & SCS_OPS_DEFINED);
/*
We've applied the Slave epoch transaction subject to the
conflict detection. If any further transactional
conflicts have been observed, then we must repeat the
process.
*/
atEndTransConflictHandling();
atBeginTransConflictHandling();
trans_conflict_apply_state = SAS_TRACK_TRANS_DEPENDENCIES;
if (unlikely(conflict_flags & SCS_TRANS_CONFLICT_DETECTED_THIS_PASS))
{
DBUG_PRINT("info", ("Further conflict(s) detected, repeating the "
"TRACK_TRANS_DEPENDENCIES pass"));
/*
Further conflict observed when applying, need
to re-determine dependencies
*/
resetPerAttemptCounters();
retry_slave_trans = true;
break;
}
DBUG_PRINT("info", ("No further conflicts detected, committing and "
"returning to SAS_TRACK_TRANS_DEPENDENCIES state"));
/*
With dependencies taken into account, no further
conflicts detected, can now proceed to commit
*/
break;
}
}
/*
Clear conflict flags, to ensure that we detect any new conflicts
*/
conflict_flags = 0;
if (retry_slave_trans)
{
DBUG_PRINT("info", ("Requesting transaction restart"));
DBUG_RETURN(1);
}
DBUG_PRINT("info", ("Allowing commit to proceed"));
DBUG_RETURN(0);
}
/**
* Conflict function interpreted programs
*/
/**
CFT_NDB_OLD
To perform conflict detection, an interpreted program is used to read
the timestamp stored locally and compare to what was on the master.
If timestamp is not equal, an error for this operation (9998) will be raised,
and new row will not be applied. The error codes for the operations will
be checked on return. For this to work is is vital that the operation
is run with ignore error option.
As an independent feature, phase 2 also saves the
conflicts into the table's exceptions table.
*/
static int
row_conflict_fn_old(NDB_CONFLICT_FN_SHARE* cfn_share,
enum_conflicting_op_type op_type,
const NdbRecord* data_record,
const uchar* old_data,
const uchar* new_data,
const MY_BITMAP* bi_cols,
const MY_BITMAP* ai_cols,
NdbInterpretedCode* code)
{
DBUG_ENTER("row_conflict_fn_old");
uint32 resolve_column= cfn_share->m_resolve_column;
uint32 resolve_size= cfn_share->m_resolve_size;
const uchar* field_ptr = (const uchar*)
NdbDictionary::getValuePtr(data_record,
(const char*) old_data,
cfn_share->m_resolve_column);
assert((resolve_size == 4) || (resolve_size == 8));
if (unlikely(!bitmap_is_set(bi_cols, resolve_column)))
{
sql_print_information("NDB Slave: missing data for %s "
"timestamp column %u.",
cfn_share->m_conflict_fn->name,
resolve_column);
DBUG_RETURN(1);
}
const uint label_0= 0;
const Uint32 RegOldValue= 1, RegCurrentValue= 2;
int r;
DBUG_PRINT("info",
("Adding interpreted filter, existing value must eq event old value"));
/*
* read old value from record
*/
union {
uint32 old_value_32;
uint64 old_value_64;
};
{
if (resolve_size == 4)
{
memcpy(&old_value_32, field_ptr, resolve_size);
DBUG_PRINT("info", (" old_value_32: %u", old_value_32));
}
else
{
memcpy(&old_value_64, field_ptr, resolve_size);
DBUG_PRINT("info", (" old_value_64: %llu",
(unsigned long long) old_value_64));
}
}
/*
* Load registers RegOldValue and RegCurrentValue
*/
if (resolve_size == 4)
r= code->load_const_u32(RegOldValue, old_value_32);
else
r= code->load_const_u64(RegOldValue, old_value_64);
DBUG_ASSERT(r == 0);
r= code->read_attr(RegCurrentValue, resolve_column);
DBUG_ASSERT(r == 0);
/*
* if RegOldValue == RegCurrentValue goto label_0
* else raise error for this row
*/
r= code->branch_eq(RegOldValue, RegCurrentValue, label_0);
DBUG_ASSERT(r == 0);
r= code->interpret_exit_nok(error_conflict_fn_violation);
DBUG_ASSERT(r == 0);
r= code->def_label(label_0);
DBUG_ASSERT(r == 0);
r= code->interpret_exit_ok();
DBUG_ASSERT(r == 0);
r= code->finalise();
DBUG_ASSERT(r == 0);
DBUG_RETURN(r);
}
static int
row_conflict_fn_max_update_only(NDB_CONFLICT_FN_SHARE* cfn_share,
enum_conflicting_op_type op_type,
const NdbRecord* data_record,
const uchar* old_data,
const uchar* new_data,
const MY_BITMAP* bi_cols,
const MY_BITMAP* ai_cols,
NdbInterpretedCode* code)
{
DBUG_ENTER("row_conflict_fn_max_update_only");
uint32 resolve_column= cfn_share->m_resolve_column;
uint32 resolve_size= cfn_share->m_resolve_size;
const uchar* field_ptr = (const uchar*)
NdbDictionary::getValuePtr(data_record,
(const char*) new_data,
cfn_share->m_resolve_column);
assert((resolve_size == 4) || (resolve_size == 8));
if (unlikely(!bitmap_is_set(ai_cols, resolve_column)))
{
sql_print_information("NDB Slave: missing data for %s "
"timestamp column %u.",
cfn_share->m_conflict_fn->name,
resolve_column);
DBUG_RETURN(1);
}
const uint label_0= 0;
const Uint32 RegNewValue= 1, RegCurrentValue= 2;
int r;
DBUG_PRINT("info",
("Adding interpreted filter, existing value must be lt event new"));
/*
* read new value from record
*/
union {
uint32 new_value_32;
uint64 new_value_64;
};
{
if (resolve_size == 4)
{
memcpy(&new_value_32, field_ptr, resolve_size);
DBUG_PRINT("info", (" new_value_32: %u", new_value_32));
}
else
{
memcpy(&new_value_64, field_ptr, resolve_size);
DBUG_PRINT("info", (" new_value_64: %llu",
(unsigned long long) new_value_64));
}
}
/*
* Load registers RegNewValue and RegCurrentValue
*/
if (resolve_size == 4)
r= code->load_const_u32(RegNewValue, new_value_32);
else
r= code->load_const_u64(RegNewValue, new_value_64);
DBUG_ASSERT(r == 0);
r= code->read_attr(RegCurrentValue, resolve_column);
DBUG_ASSERT(r == 0);
/*
* if RegNewValue > RegCurrentValue goto label_0
* else raise error for this row
*/
r= code->branch_gt(RegNewValue, RegCurrentValue, label_0);
DBUG_ASSERT(r == 0);
r= code->interpret_exit_nok(error_conflict_fn_violation);
DBUG_ASSERT(r == 0);
r= code->def_label(label_0);
DBUG_ASSERT(r == 0);
r= code->interpret_exit_ok();
DBUG_ASSERT(r == 0);
r= code->finalise();
DBUG_ASSERT(r == 0);
DBUG_RETURN(r);
}
/**
CFT_NDB_MAX
To perform conflict resolution, an interpreted program is used to read
the timestamp stored locally and compare to what is going to be applied.
If timestamp is lower, an error for this operation (9999) will be raised,
and new row will not be applied. The error codes for the operations will
be checked on return. For this to work is is vital that the operation
is run with ignore error option.
Note that for delete, this algorithm reverts to the OLD algorithm.
*/
static int
row_conflict_fn_max(NDB_CONFLICT_FN_SHARE* cfn_share,
enum_conflicting_op_type op_type,
const NdbRecord* data_record,
const uchar* old_data,
const uchar* new_data,
const MY_BITMAP* bi_cols,
const MY_BITMAP* ai_cols,
NdbInterpretedCode* code)
{
switch(op_type)
{
case WRITE_ROW:
abort();
return 1;
case UPDATE_ROW:
return row_conflict_fn_max_update_only(cfn_share,
op_type,
data_record,
old_data,
new_data,
bi_cols,
ai_cols,
code);
case DELETE_ROW:
/* Can't use max of new image, as there's no new image
* for DELETE
* Use OLD instead
*/
return row_conflict_fn_old(cfn_share,
op_type,
data_record,
old_data,
new_data,
bi_cols,
ai_cols,
code);
default:
abort();
return 1;
}
}
/**
CFT_NDB_MAX_DEL_WIN
To perform conflict resolution, an interpreted program is used to read
the timestamp stored locally and compare to what is going to be applied.
If timestamp is lower, an error for this operation (9999) will be raised,
and new row will not be applied. The error codes for the operations will
be checked on return. For this to work is is vital that the operation
is run with ignore error option.
In this variant, replicated DELETEs alway succeed - no filter is added
to them.
*/
static int
row_conflict_fn_max_del_win(NDB_CONFLICT_FN_SHARE* cfn_share,
enum_conflicting_op_type op_type,
const NdbRecord* data_record,
const uchar* old_data,
const uchar* new_data,
const MY_BITMAP* bi_cols,
const MY_BITMAP* ai_cols,
NdbInterpretedCode* code)
{
switch(op_type)
{
case WRITE_ROW:
abort();
return 1;
case UPDATE_ROW:
return row_conflict_fn_max_update_only(cfn_share,
op_type,
data_record,
old_data,
new_data,
bi_cols,
ai_cols,
code);
case DELETE_ROW:
/* This variant always lets a received DELETE_ROW
* succeed.
*/
return 0;
default:
abort();
return 1;
}
}
/**
CFT_NDB_EPOCH
*/
static int
row_conflict_fn_epoch(NDB_CONFLICT_FN_SHARE* cfn_share,
enum_conflicting_op_type op_type,
const NdbRecord* data_record,
const uchar* old_data,
const uchar* new_data,
const MY_BITMAP* bi_cols,
const MY_BITMAP* ai_cols,
NdbInterpretedCode* code)
{
DBUG_ENTER("row_conflict_fn_epoch");
switch(op_type)
{
case WRITE_ROW:
abort();
DBUG_RETURN(1);
case UPDATE_ROW:
case DELETE_ROW:
case READ_ROW: /* Read tracking */
{
const uint label_0= 0;
const Uint32
RegAuthor= 1, RegZero= 2,
RegMaxRepEpoch= 1, RegRowEpoch= 2;
int r;
r= code->load_const_u32(RegZero, 0);
assert(r == 0);
r= code->read_attr(RegAuthor, NdbDictionary::Column::ROW_AUTHOR);
assert(r == 0);
/* If last author was not local, assume no conflict */
r= code->branch_ne(RegZero, RegAuthor, label_0);
assert(r == 0);
/*
* Load registers RegMaxRepEpoch and RegRowEpoch
*/
r= code->load_const_u64(RegMaxRepEpoch, g_ndb_slave_state.max_rep_epoch);
assert(r == 0);
r= code->read_attr(RegRowEpoch, NdbDictionary::Column::ROW_GCI64);
assert(r == 0);
/*
* if RegRowEpoch <= RegMaxRepEpoch goto label_0
* else raise error for this row
*/
r= code->branch_le(RegRowEpoch, RegMaxRepEpoch, label_0);
assert(r == 0);
r= code->interpret_exit_nok(error_conflict_fn_violation);
assert(r == 0);
r= code->def_label(label_0);
assert(r == 0);
r= code->interpret_exit_ok();
assert(r == 0);
r= code->finalise();
assert(r == 0);
DBUG_RETURN(r);
}
default:
abort();
DBUG_RETURN(1);
}
}
/**
* CFT_NDB_EPOCH2
*/
static int
row_conflict_fn_epoch2_primary(NDB_CONFLICT_FN_SHARE* cfn_share,
enum_conflicting_op_type op_type,
const NdbRecord* data_record,
const uchar* old_data,
const uchar* new_data,
const MY_BITMAP* bi_cols,
const MY_BITMAP* ai_cols,
NdbInterpretedCode* code)
{
DBUG_ENTER("row_conflict_fn_epoch2_primary");
/* We use the normal NDB$EPOCH detection function */
DBUG_RETURN(row_conflict_fn_epoch(cfn_share,
op_type,
data_record,
old_data,
new_data,
bi_cols,
ai_cols,
code));
}
static int
row_conflict_fn_epoch2_secondary(NDB_CONFLICT_FN_SHARE* cfn_share,
enum_conflicting_op_type op_type,
const NdbRecord* data_record,
const uchar* old_data,
const uchar* new_data,
const MY_BITMAP* bi_cols,
const MY_BITMAP* ai_cols,
NdbInterpretedCode* code)
{
DBUG_ENTER("row_conflict_fn_epoch2_secondary");
/* Only called for reflected update and delete operations
* on the secondary.
* These are returning operations which should only be
* applied if the row in the database was last written
* remotely (by the Primary)
*/
switch(op_type)
{
case WRITE_ROW:
abort();
DBUG_RETURN(1);
case UPDATE_ROW:
case DELETE_ROW:
{
const uint label_0= 0;
const Uint32
RegAuthor= 1, RegZero= 2;
int r;
r= code->load_const_u32(RegZero, 0);
assert(r == 0);
r= code->read_attr(RegAuthor, NdbDictionary::Column::ROW_AUTHOR);
assert(r == 0);
r= code->branch_eq(RegZero, RegAuthor, label_0);
assert(r == 0);
/* Last author was not local, no conflict, apply */
r= code->interpret_exit_ok();
assert(r == 0);
r= code->def_label(label_0);
assert(r == 0);
/* Last author was secondary-local, conflict, do not apply */
r= code->interpret_exit_nok(error_conflict_fn_violation);
assert(r == 0);
r= code->finalise();
assert(r == 0);
DBUG_RETURN(r);
}
default:
abort();
DBUG_RETURN(1);
}
}
static int
row_conflict_fn_epoch2(NDB_CONFLICT_FN_SHARE* cfn_share,
enum_conflicting_op_type op_type,
const NdbRecord* data_record,
const uchar* old_data,
const uchar* new_data,
const MY_BITMAP* bi_cols,
const MY_BITMAP* ai_cols,
NdbInterpretedCode* code)
{
DBUG_ENTER("row_conflict_fn_epoch2");
/**
* NdbEpoch2 behaviour depends on the Slave conflict role variable
*
*/
switch(opt_ndb_slave_conflict_role)
{
case SCR_NONE:
/* This is a problem */
DBUG_RETURN(1);
case SCR_PRIMARY:
DBUG_RETURN(row_conflict_fn_epoch2_primary(cfn_share,
op_type,
data_record,
old_data,
new_data,
bi_cols,
ai_cols,
code));
case SCR_SECONDARY:
DBUG_RETURN(row_conflict_fn_epoch2_secondary(cfn_share,
op_type,
data_record,
old_data,
new_data,
bi_cols,
ai_cols,
code));
case SCR_PASS:
/* Do nothing */
DBUG_RETURN(0);
default:
break;
}
abort();
DBUG_RETURN(1);
}
/**
* Conflict function setup infrastructure
*/
static const st_conflict_fn_arg_def resolve_col_args[]=
{
/* Arg type Optional */
{ CFAT_COLUMN_NAME, false },
{ CFAT_END, false }
};
static const st_conflict_fn_arg_def epoch_fn_args[]=
{
/* Arg type Optional */
{ CFAT_EXTRA_GCI_BITS, true },
{ CFAT_END, false }
};
static const st_conflict_fn_def conflict_fns[]=
{
{ "NDB$MAX_DELETE_WIN", CFT_NDB_MAX_DEL_WIN,
&resolve_col_args[0], row_conflict_fn_max_del_win, 0 },
{ "NDB$MAX", CFT_NDB_MAX,
&resolve_col_args[0], row_conflict_fn_max, 0 },
{ "NDB$OLD", CFT_NDB_OLD,
&resolve_col_args[0], row_conflict_fn_old, 0 },
{ "NDB$EPOCH2_TRANS", CFT_NDB_EPOCH2_TRANS,
&epoch_fn_args[0], row_conflict_fn_epoch2,
CF_REFLECT_SEC_OPS | CF_USE_ROLE_VAR | CF_TRANSACTIONAL},
{ "NDB$EPOCH2", CFT_NDB_EPOCH2,
&epoch_fn_args[0], row_conflict_fn_epoch2,
CF_REFLECT_SEC_OPS | CF_USE_ROLE_VAR
},
{ "NDB$EPOCH_TRANS", CFT_NDB_EPOCH_TRANS,
&epoch_fn_args[0], row_conflict_fn_epoch, CF_TRANSACTIONAL},
{ "NDB$EPOCH", CFT_NDB_EPOCH,
&epoch_fn_args[0], row_conflict_fn_epoch, 0 }
};
static unsigned n_conflict_fns=
sizeof(conflict_fns) / sizeof(struct st_conflict_fn_def);
int
parse_conflict_fn_spec(const char* conflict_fn_spec,
const st_conflict_fn_def** conflict_fn,
st_conflict_fn_arg* args,
Uint32* max_args,
char *msg, uint msg_len)
{
DBUG_ENTER("parse_conflict_fn_spec");
Uint32 no_args = 0;
const char *ptr= conflict_fn_spec;
const char *error_str= "unknown conflict resolution function";
/* remove whitespace */
while (*ptr == ' ' && *ptr != '\0') ptr++;
DBUG_PRINT("info", ("parsing %s", conflict_fn_spec));
for (unsigned i= 0; i < n_conflict_fns; i++)
{
const st_conflict_fn_def &fn= conflict_fns[i];
uint len= (uint)strlen(fn.name);
if (strncmp(ptr, fn.name, len))
continue;
DBUG_PRINT("info", ("found function %s", fn.name));
/* skip function name */
ptr+= len;
/* remove whitespace */
while (*ptr == ' ' && *ptr != '\0') ptr++;
/* next '(' */
if (*ptr != '(')
{
error_str= "missing '('";
DBUG_PRINT("info", ("parse error %s", error_str));
break;
}
ptr++;
/* find all arguments */
for (;;)
{
if (no_args >= *max_args)
{
error_str= "too many arguments";
DBUG_PRINT("info", ("parse error %s", error_str));
break;
}
/* expected type */
enum enum_conflict_fn_arg_type type=
conflict_fns[i].arg_defs[no_args].arg_type;
/* remove whitespace */
while (*ptr == ' ' && *ptr != '\0') ptr++;
if (type == CFAT_END)
{
args[no_args].type= type;
error_str= NULL;
break;
}
/* arg */
/* Todo : Should support comma as an arg separator? */
const char *start_arg= ptr;
while (*ptr != ')' && *ptr != ' ' && *ptr != '\0') ptr++;
const char *end_arg= ptr;
bool optional_arg = conflict_fns[i].arg_defs[no_args].optional;
/* any arg given? */
if (start_arg == end_arg)
{
if (!optional_arg)
{
error_str= "missing function argument";
DBUG_PRINT("info", ("parse error %s", error_str));
break;
}
else
{
/* Arg was optional, and not present
* Must be at end of args, finish parsing
*/
args[no_args].type= CFAT_END;
error_str= NULL;
break;
}
}
uint len= (uint)(end_arg - start_arg);
args[no_args].type= type;
DBUG_PRINT("info", ("found argument %s %u", start_arg, len));
bool arg_processing_error = false;
switch (type)
{
case CFAT_COLUMN_NAME:
{
/* Copy column name out into argument's buffer */
char* dest= &args[no_args].resolveColNameBuff[0];
memcpy(dest, start_arg, (len < (uint) NAME_CHAR_LEN ?
len :
NAME_CHAR_LEN));
dest[len]= '\0';
break;
}
case CFAT_EXTRA_GCI_BITS:
{
/* Map string to number and check it's in range etc */
char* end_of_arg = (char*) end_arg;
Uint32 bits = strtoul(start_arg, &end_of_arg, 0);
DBUG_PRINT("info", ("Using %u as the number of extra bits", bits));
if (bits > 31)
{
arg_processing_error= true;
error_str= "Too many extra Gci bits";
DBUG_PRINT("info", ("%s", error_str));
break;
}
/* Num bits seems ok */
args[no_args].extraGciBits = bits;
break;
}
case CFAT_END:
abort();
}
if (arg_processing_error)
break;
no_args++;
}
if (error_str)
break;
/* remove whitespace */
while (*ptr == ' ' && *ptr != '\0') ptr++;
/* next ')' */
if (*ptr != ')')
{
error_str= "missing ')'";
break;
}
ptr++;
/* remove whitespace */
while (*ptr == ' ' && *ptr != '\0') ptr++;
/* garbage in the end? */
if (*ptr != '\0')
{
error_str= "garbage in the end";
break;
}
/* Update ptrs to conflict fn + # of args */
*conflict_fn = &conflict_fns[i];
*max_args = no_args;
DBUG_RETURN(0);
}
/* parse error */
my_snprintf(msg, msg_len, "%s, %s at '%s'",
conflict_fn_spec, error_str, ptr);
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
static uint
slave_check_resolve_col_type(const NDBTAB *ndbtab,
uint field_index)
{
DBUG_ENTER("slave_check_resolve_col_type");
const NDBCOL *c= ndbtab->getColumn(field_index);
uint sz= 0;
switch (c->getType())
{
case NDBCOL::Unsigned:
sz= sizeof(Uint32);
DBUG_PRINT("info", ("resolve column Uint32 %u",
field_index));
break;
case NDBCOL::Bigunsigned:
sz= sizeof(Uint64);
DBUG_PRINT("info", ("resolve column Uint64 %u",
field_index));
break;
default:
DBUG_PRINT("info", ("resolve column %u has wrong type",
field_index));
break;
}
DBUG_RETURN(sz);
}
static int
slave_set_resolve_fn(Ndb* ndb,
NDB_CONFLICT_FN_SHARE** ppcfn_share,
const char* dbName,
const char* tabName,
const NDBTAB *ndbtab, uint field_index,
uint resolve_col_sz,
const st_conflict_fn_def* conflict_fn,
uint8 flags)
{
DBUG_ENTER("slave_set_resolve_fn");
NdbDictionary::Dictionary *dict= ndb->getDictionary();
NDB_CONFLICT_FN_SHARE *cfn_share= *ppcfn_share;
const char *ex_suffix= (char *)NDB_EXCEPTIONS_TABLE_SUFFIX;
if (cfn_share == NULL)
{
*ppcfn_share= cfn_share=
(NDB_CONFLICT_FN_SHARE*) my_malloc(PSI_INSTRUMENT_ME,
sizeof(NDB_CONFLICT_FN_SHARE),
MYF(MY_WME | ME_FATALERROR));
slave_reset_conflict_fn(cfn_share);
}
cfn_share->m_conflict_fn= conflict_fn;
/* Calculate resolve col stuff (if relevant) */
cfn_share->m_resolve_size= resolve_col_sz;
cfn_share->m_resolve_column= field_index;
cfn_share->m_flags = flags;
/* Init Exceptions Table Writer */
new (&cfn_share->m_ex_tab_writer) ExceptionsTableWriter();
/* Check for '$EX' or '$ex' suffix in table name */
for (int tries= 2;
tries-- > 0;
ex_suffix=
(tries == 1)
? (const char *)NDB_EXCEPTIONS_TABLE_SUFFIX_LOWER
: NullS)
{
/* get exceptions table */
char ex_tab_name[FN_REFLEN];
strxnmov(ex_tab_name, sizeof(ex_tab_name), tabName,
ex_suffix, NullS);
ndb->setDatabaseName(dbName);
Ndb_table_guard ndbtab_g(dict, ex_tab_name);
const NDBTAB *ex_tab= ndbtab_g.get_table();
if (ex_tab)
{
char msgBuf[ FN_REFLEN ];
const char* msg = NULL;
if (cfn_share->m_ex_tab_writer.init(ndbtab,
ex_tab,
msgBuf,
sizeof(msgBuf),
&msg) == 0)
{
/* Ok */
/* Hold our table reference outside the table_guard scope */
ndbtab_g.release();
/* Table looked suspicious, warn user */
if (msg)
sql_print_warning("%s", msg);
if (opt_ndb_extra_logging)
{
sql_print_information("NDB Slave: Table %s.%s logging exceptions to %s.%s",
dbName,
tabName,
dbName,
ex_tab_name);
}
}
else
{
sql_print_warning("%s", msg);
}
break;
} /* if (ex_tab) */
}
DBUG_RETURN(0);
}
bool
is_exceptions_table(const char *table_name)
{
size_t len = strlen(table_name);
size_t suffixlen = strlen(NDB_EXCEPTIONS_TABLE_SUFFIX);
if(len > suffixlen &&
(strcmp(table_name + len - suffixlen,
lower_case_table_names ? NDB_EXCEPTIONS_TABLE_SUFFIX_LOWER :
NDB_EXCEPTIONS_TABLE_SUFFIX) == 0))
{
return true;
}
return false;
}
int
setup_conflict_fn(Ndb* ndb,
NDB_CONFLICT_FN_SHARE** ppcfn_share,
const char* dbName,
const char* tabName,
bool tableUsesBlobs,
bool tableBinlogUseUpdate,
const NDBTAB *ndbtab,
char *msg, uint msg_len,
const st_conflict_fn_def* conflict_fn,
const st_conflict_fn_arg* args,
const Uint32 num_args)
{
DBUG_ENTER("setup_conflict_fn");
if(is_exceptions_table(tabName))
{
my_snprintf(msg, msg_len,
"Ndb Slave: Table %s.%s is exceptions table: not using conflict function %s",
dbName,
tabName,
conflict_fn->name);
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(0);
}
/* setup the function */
switch (conflict_fn->type)
{
case CFT_NDB_MAX:
case CFT_NDB_OLD:
case CFT_NDB_MAX_DEL_WIN:
{
if (num_args != 1)
{
my_snprintf(msg, msg_len,
"Incorrect arguments to conflict function");
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
/* Now try to find the column in the table */
int colNum = -1;
const char* resolveColName = args[0].resolveColNameBuff;
int resolveColNameLen = (int)strlen(resolveColName);
for (int j=0; j< ndbtab->getNoOfColumns(); j++)
{
const char* colName = ndbtab->getColumn(j)->getName();
if (strncmp(colName,
resolveColName,
resolveColNameLen) == 0 &&
colName[resolveColNameLen] == '\0')
{
colNum = j;
break;
}
}
if (colNum == -1)
{
my_snprintf(msg, msg_len,
"Could not find resolve column %s.",
resolveColName);
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
const uint resolve_col_sz= slave_check_resolve_col_type(ndbtab, colNum);
if (resolve_col_sz == 0)
{
/* wrong data type */
slave_reset_conflict_fn(*ppcfn_share);
my_snprintf(msg, msg_len,
"Column '%s' has wrong datatype",
resolveColName);
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
if (slave_set_resolve_fn(ndb,
ppcfn_share,
dbName,
tabName,
ndbtab,
colNum, resolve_col_sz,
conflict_fn, CFF_NONE))
{
my_snprintf(msg, msg_len,
"Unable to setup conflict resolution using column '%s'",
resolveColName);
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
/* Success, update message */
my_snprintf(msg, msg_len,
"NDB Slave: Table %s.%s using conflict_fn %s on attribute %s.",
dbName,
tabName,
conflict_fn->name,
resolveColName);
break;
}
case CFT_NDB_EPOCH2:
case CFT_NDB_EPOCH2_TRANS:
{
/* Check how updates will be logged... */
bool log_update_as_write = (!tableBinlogUseUpdate);
if (log_update_as_write)
{
my_snprintf(msg, msg_len,
"Table %s.%s configured to log updates as writes. "
"Not suitable for %s.",
dbName,
tabName,
conflict_fn->name);
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
/* Fall through for the rest of the EPOCH* processing... */
}
case CFT_NDB_EPOCH:
case CFT_NDB_EPOCH_TRANS:
{
if (num_args > 1)
{
my_snprintf(msg, msg_len,
"Too many arguments to conflict function");
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
/* Check that table doesn't have Blobs as we don't support that */
if (tableUsesBlobs)
{
my_snprintf(msg, msg_len, "Table has Blob column(s), not suitable for %s.",
conflict_fn->name);
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
/* Check that table has required extra meta-columns */
/* Todo : Could warn if extra gcibits is insufficient to
* represent SavePeriod/EpochPeriod
*/
if (ndbtab->getExtraRowGciBits() == 0)
sql_print_information("NDB Slave: Table %s.%s : %s, low epoch resolution",
dbName,
tabName,
conflict_fn->name);
if (ndbtab->getExtraRowAuthorBits() == 0)
{
my_snprintf(msg, msg_len, "No extra row author bits in table.");
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
if (slave_set_resolve_fn(ndb,
ppcfn_share,
dbName,
tabName,
ndbtab,
0, // field_no
0, // resolve_col_sz
conflict_fn, CFF_REFRESH_ROWS))
{
my_snprintf(msg, msg_len,
"unable to setup conflict resolution");
DBUG_PRINT("info", ("%s", msg));
DBUG_RETURN(-1);
}
/* Success, update message */
my_snprintf(msg, msg_len,
"NDB Slave: Table %s.%s using conflict_fn %s.",
dbName,
tabName,
conflict_fn->name);
break;
}
case CFT_NUMBER_OF_CFTS:
case CFT_NDB_UNDEF:
abort();
}
DBUG_RETURN(0);
}
void
teardown_conflict_fn(Ndb* ndb, NDB_CONFLICT_FN_SHARE* cfn_share)
{
if (cfn_share &&
cfn_share->m_ex_tab_writer.hasTable() &&
ndb)
{
cfn_share->m_ex_tab_writer.mem_free(ndb);
}
// Release the NDB_CONFLICT_FN_SHARE which was allocated
// in setup_conflict_fn()
my_free(cfn_share);
}
void slave_reset_conflict_fn(NDB_CONFLICT_FN_SHARE *cfn_share)
{
if (cfn_share)
{
memset(cfn_share, 0, sizeof(*cfn_share));
}
}
#endif
/**
* Variables related to conflict handling
* All prefixed 'ndb_conflict'
*/
SHOW_VAR ndb_status_conflict_variables[]= {
{"fn_max", (char*) &g_ndb_slave_state.total_violation_count[CFT_NDB_MAX], SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"fn_old", (char*) &g_ndb_slave_state.total_violation_count[CFT_NDB_OLD], SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"fn_max_del_win", (char*) &g_ndb_slave_state.total_violation_count[CFT_NDB_MAX_DEL_WIN], SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"fn_epoch", (char*) &g_ndb_slave_state.total_violation_count[CFT_NDB_EPOCH], SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"fn_epoch_trans", (char*) &g_ndb_slave_state.total_violation_count[CFT_NDB_EPOCH_TRANS], SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"fn_epoch2", (char*) &g_ndb_slave_state.total_violation_count[CFT_NDB_EPOCH2], SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"fn_epoch2_trans", (char*) &g_ndb_slave_state.total_violation_count[CFT_NDB_EPOCH2_TRANS], SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"trans_row_conflict_count", (char*) &g_ndb_slave_state.trans_row_conflict_count, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"trans_row_reject_count", (char*) &g_ndb_slave_state.trans_row_reject_count, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"trans_reject_count", (char*) &g_ndb_slave_state.trans_in_conflict_count, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"trans_detect_iter_count", (char*) &g_ndb_slave_state.trans_detect_iter_count, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"trans_conflict_commit_count",
(char*) &g_ndb_slave_state.trans_conflict_commit_count, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"epoch_delete_delete_count", (char*) &g_ndb_slave_state.total_delete_delete_count, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"reflected_op_prepare_count", (char*) &g_ndb_slave_state.total_reflect_op_prepare_count, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"reflected_op_discard_count", (char*) &g_ndb_slave_state.total_reflect_op_discard_count, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"refresh_op_count", (char*) &g_ndb_slave_state.total_refresh_op_count, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"last_conflict_epoch", (char*) &g_ndb_slave_state.last_conflicted_epoch, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{"last_stable_epoch", (char*) &g_ndb_slave_state.last_stable_epoch, SHOW_LONGLONG, SHOW_SCOPE_GLOBAL},
{NullS, NullS, SHOW_LONG, SHOW_SCOPE_GLOBAL}
};
int
show_ndb_conflict_status_vars(THD *thd, struct st_mysql_show_var *var, char *buff)
{
/* Just a function to allow moving array into this file */
var->type = SHOW_ARRAY;
var->value = (char*) &ndb_status_conflict_variables;
return 0;
}
Reference in New Issue View Git Blame Copy Permalink