/*
* Establish persistent connection to backend
*/
static void establish_persistent_connection(void)
{
int i;
BackendInfo *bkinfo;
POOL_CONNECTION_POOL_SLOT *s;
for (i=0;i<NUM_BACKENDS;i++)
{
if (!VALID_BACKEND(i))
continue;
if (slots[i] == NULL)
{
bkinfo = pool_get_node_info(i);
s = make_persistent_db_connection(bkinfo->backend_hostname,
bkinfo->backend_port,
"postgres",
pool_config->sr_check_user,
pool_config->sr_check_password, true);
if (s)
slots[i] = s;
else
slots[i] = NULL;
}
}
}
static void
inform_node_info(PCP_CONNECTION *frontend,char *buf)
{
int node_id;
int wsize;
char port_str[6];
char status[2];
char weight_str[20];
char code[] = "CommandComplete";
BackendInfo *bi = NULL;
node_id = atoi(buf);
bi = pool_get_node_info(node_id);
if (bi == NULL)
ereport(ERROR,
(errmsg("informing node info failed"),
errdetail("invalid node ID")));
ereport(DEBUG2,
(errmsg("PCP: informing node info"),
errdetail("retrieved node information from shared memory")));
snprintf(port_str, sizeof(port_str), "%d", bi->backend_port);
snprintf(status, sizeof(status), "%d", bi->backend_status);
snprintf(weight_str, sizeof(weight_str), "%f", bi->backend_weight);
pcp_write(frontend, "i", 1);
wsize = htonl(sizeof(code) +
strlen(bi->backend_hostname)+1 +
strlen(port_str)+1 +
strlen(status)+1 +
strlen(weight_str)+1 +
sizeof(int));
pcp_write(frontend, &wsize, sizeof(int));
pcp_write(frontend, code, sizeof(code));
pcp_write(frontend, bi->backend_hostname, strlen(bi->backend_hostname)+1);
pcp_write(frontend, port_str, strlen(port_str)+1);
pcp_write(frontend, status, strlen(status)+1);
pcp_write(frontend, weight_str, strlen(weight_str)+1);
do_pcp_flush(frontend);
}
POOL_REPORT_NODES* get_nodes(int *nrows)
{
int i;
POOL_REPORT_NODES* nodes = malloc(NUM_BACKENDS * sizeof(POOL_REPORT_NODES));
BackendInfo *bi = NULL;
for (i = 0; i < NUM_BACKENDS; i++)
{
bi = pool_get_node_info(i);
snprintf(nodes[i].node_id, POOLCONFIG_MAXSTATLEN, "%d", i);
strncpy(nodes[i].hostname, bi->backend_hostname, strlen(bi->backend_hostname)+1);
snprintf(nodes[i].port, POOLCONFIG_MAXIDENTLEN, "%d", bi->backend_port);
snprintf(nodes[i].status, POOLCONFIG_MAXSTATLEN, "%d", bi->backend_status);
snprintf(nodes[i].lb_weight, POOLCONFIG_MAXWEIGHTLEN, "%f", bi->backend_weight/RAND_MAX);
}
*nrows = i;
return nodes;
}
/*
* Decide where to send queries(thus expecting response)
*/
void pool_where_to_send(POOL_QUERY_CONTEXT *query_context, char *query, Node *node)
{
POOL_SESSION_CONTEXT *session_context;
POOL_CONNECTION_POOL *backend;
int i;
if (!query_context)
{
pool_error("pool_where_to_send: no query context");
return;
}
session_context = pool_get_session_context();
backend = session_context->backend;
/*
* Zap out DB node map
*/
pool_clear_node_to_be_sent(query_context);
/*
* If there is "NO LOAD BALANCE" comment, we send only to master node.
*/
if (!strncasecmp(query, NO_LOAD_BALANCE, NO_LOAD_BALANCE_COMMENT_SZ))
{
pool_set_node_to_be_sent(query_context,
MASTER_SLAVE ? PRIMARY_NODE_ID : REAL_MASTER_NODE_ID);
for (i=0;i<NUM_BACKENDS;i++)
{
if (query_context->where_to_send[i])
{
query_context->virtual_master_node_id = i;
break;
}
}
return;
}
/*
* In raw mode, we send only to master node. Simple enough.
*/
if (RAW_MODE)
{
pool_set_node_to_be_sent(query_context, REAL_MASTER_NODE_ID);
}
else if (MASTER_SLAVE && query_context->is_multi_statement)
{
/*
* If we are in master/slave mode and we have multi stametemt
* query, we should send it to primary server only. Otherwise
* it is possible to send a write query to standby servers
* because we only use the first element of the multi
* statement query and don't care about the rest. Typical
* situation where we are bugged by this is, "BEGIN;DELETE
* FROM table;END". Note that from pgpool-II 3.1.0
* transactional statements such as "BEGIN" is unconditionaly
* sent to all nodes(see send_to_where() for more details).
* Someday we might be able to understand all part of multi
* statement queries, but until that day we need this band
* aid.
*/
if (query_context->is_multi_statement)
{
pool_set_node_to_be_sent(query_context, PRIMARY_NODE_ID);
}
}
else if (MASTER_SLAVE)
{
POOL_DEST dest;
POOL_MEMORY_POOL *old_context;
old_context = pool_memory_context_switch_to(query_context->memory_context);
dest = send_to_where(node, query);
pool_memory_context_switch_to(old_context);
pool_debug("send_to_where: %d query: %s", dest, query);
/* Should be sent to primary only? */
if (dest == POOL_PRIMARY)
{
pool_set_node_to_be_sent(query_context, PRIMARY_NODE_ID);
}
/* Should be sent to both primary and standby? */
else if (dest == POOL_BOTH)
{
pool_setall_node_to_be_sent(query_context);
}
/*
* Ok, we might be able to load balance the SELECT query.
*/
else
{
if (pool_config->load_balance_mode &&
is_select_query(node, query) &&
MAJOR(backend) == PROTO_MAJOR_V3)
{
/*
* If (we are outside of an explicit transaction) OR
* (the transaction has not issued a write query yet, AND
* transaction isolation level is not SERIALIZABLE)
* we might be able to load balance.
*/
if (TSTATE(backend, PRIMARY_NODE_ID) == 'I' ||
(!pool_is_writing_transaction() &&
!pool_is_failed_transaction() &&
pool_get_transaction_isolation() != POOL_SERIALIZABLE))
{
BackendInfo *bkinfo = pool_get_node_info(session_context->load_balance_node_id);
/*
* Load balance if possible
*/
/*
* If replication delay is too much, we prefer to send to the primary.
*/
if (!strcmp(pool_config->master_slave_sub_mode, MODE_STREAMREP) &&
pool_config->delay_threshold &&
bkinfo->standby_delay > pool_config->delay_threshold)
{
pool_set_node_to_be_sent(query_context, PRIMARY_NODE_ID);
}
/*
* If a writing function call is used,
* we prefer to send to the primary.
*/
else if (pool_has_function_call(node))
{
pool_set_node_to_be_sent(query_context, PRIMARY_NODE_ID);
}
/*
* If system catalog is used in the SELECT, we
* prefer to send to the primary. Example: SELECT
* * FROM pg_class WHERE relname = 't1'; Because
* 't1' is a constant, it's hard to recognize as
* table name. Most use case such query is
* against system catalog, and the table name can
* be a temporary table, it's best to query
* against primary system catalog.
* Please note that this test must be done *before*
* test using pool_has_temp_table.
*/
else if (pool_has_system_catalog(node))
{
pool_set_node_to_be_sent(query_context, PRIMARY_NODE_ID);
}
/*
* If temporary table is used in the SELECT,
* we prefer to send to the primary.
*/
else if (pool_config->check_temp_table && pool_has_temp_table(node))
{
pool_set_node_to_be_sent(query_context, PRIMARY_NODE_ID);
}
/*
* If unlogged table is used in the SELECT,
* we prefer to send to the primary.
*/
else if (pool_has_unlogged_table(node))
{
pool_set_node_to_be_sent(query_context, PRIMARY_NODE_ID);
}
else
{
pool_set_node_to_be_sent(query_context,
session_context->load_balance_node_id);
}
}
else
{
/* Send to the primary only */
pool_set_node_to_be_sent(query_context, PRIMARY_NODE_ID);
}
}
else
{
/* Send to the primary only */
pool_set_node_to_be_sent(query_context, PRIMARY_NODE_ID);
}
}
}
else if (REPLICATION || PARALLEL_MODE)
{
if (is_select_query(node, query))
{
/*
* If a writing function call is used or replicate_select is true,
* we prefer to send to all nodes.
*/
if ((pool_has_function_call(node) || pool_config->replicate_select))
{
pool_setall_node_to_be_sent(query_context);
}
else if (pool_config->load_balance_mode &&
MAJOR(backend) == PROTO_MAJOR_V3 &&
TSTATE(backend, MASTER_NODE_ID) == 'I')
{
/* load balance */
pool_set_node_to_be_sent(query_context,
session_context->load_balance_node_id);
}
else
{
/* only send to master node */
pool_set_node_to_be_sent(query_context, REAL_MASTER_NODE_ID);
}
}
else if (IsA(node, DeclareCursorStmt) || IsA(node, ClosePortalStmt) || IsA(node, FetchStmt))
{
if (query_context->loadbalance_cursor)
{
if (pool_config->load_balance_mode &&
MAJOR(backend) == PROTO_MAJOR_V3 &&
TSTATE(backend, MASTER_NODE_ID) == 'I')
{
/* load balance */
pool_set_node_to_be_sent(query_context,
session_context->load_balance_node_id);
}
else
{
/* only send to master node */
pool_set_node_to_be_sent(query_context, REAL_MASTER_NODE_ID);
}
}
else
{
/* send to all nodes */
pool_setall_node_to_be_sent(query_context);
}
}
else
{
/* send to all nodes */
pool_setall_node_to_be_sent(query_context);
}
}
else
{
pool_error("pool_where_to_send: unknown mode");
return;
}
/*
* EXECUTE?
*/
if (IsA(node, ExecuteStmt))
{
POOL_SENT_MESSAGE *msg;
msg = pool_get_sent_message('Q', ((ExecuteStmt *)node)->name);
if (!msg)
msg = pool_get_sent_message('P', ((ExecuteStmt *)node)->name);
if (msg)
pool_copy_prep_where(msg->query_context->where_to_send,
query_context->where_to_send);
}
/*
* DEALLOCATE?
*/
else if (IsA(node, DeallocateStmt))
{
where_to_send_deallocate(query_context, node);
}
for (i=0;i<NUM_BACKENDS;i++)
{
if (query_context->where_to_send[i])
{
query_context->virtual_master_node_id = i;
break;
}
}
return;
}
/*
* trigger_failover_command: execute specified command at failover.
* command_line is null-terminated string.
*/
static int trigger_failover_command(int node, const char *command_line)
{
int r = 0;
String *exec_cmd;
char port_buf[6];
char buf[2];
BackendInfo *info;
if (command_line == NULL || (strlen(command_line) == 0))
return 0;
/* check nodeID */
if (node < 0 || node > NUM_BACKENDS)
return -1;
info = pool_get_node_info(node);
if (!info)
return -1;
buf[1] = '\0';
pool_memory = pool_memory_create(PREPARE_BLOCK_SIZE);
if (!pool_memory)
{
pool_error("trigger_failover_command: pool_memory_create() failed");
return -1;
}
exec_cmd = init_string("");
while (*command_line)
{
if (*command_line == '%')
{
if (*(command_line + 1))
{
char val = *(command_line + 1);
switch (val)
{
case 'p': /* port */
snprintf(port_buf, sizeof(port_buf), "%d", info->backend_port);
string_append_char(exec_cmd, port_buf);
break;
case 'D': /* database directory */
string_append_char(exec_cmd, info->backend_data_directory);
break;
case 'd': /* node id */
snprintf(port_buf, sizeof(port_buf), "%d", node);
string_append_char(exec_cmd, port_buf);
break;
case 'h': /* host name */
string_append_char(exec_cmd, info->backend_hostname);
break;
case 'm': /* new master node id */
snprintf(port_buf, sizeof(port_buf), "%d", get_next_master_node());
string_append_char(exec_cmd, port_buf);
break;
case 'M': /* old master node id */
snprintf(port_buf, sizeof(port_buf), "%d", MASTER_NODE_ID);
string_append_char(exec_cmd, port_buf);
break;
case '%': /* escape */
string_append_char(exec_cmd, "%");
break;
default: /* ignore */
break;
}
command_line++;
}
} else {
buf[0] = *command_line;
string_append_char(exec_cmd, buf);
}
command_line++;
}
if (strlen(exec_cmd->data) != 0)
{
pool_log("execute command: %s", exec_cmd->data);
r = system(exec_cmd->data);
}
pool_memory_delete(pool_memory, 0);
pool_memory = NULL;
return r;
}
/*
* Check replicaton time lag
*/
static void check_replication_time_lag(void)
{
int i;
int active_nodes = 0;
POOL_STATUS sts;
POOL_SELECT_RESULT *res;
unsigned long long int lsn[MAX_NUM_BACKENDS];
char *query;
BackendInfo *bkinfo;
unsigned long long int lag;
if (NUM_BACKENDS <= 1)
{
/* If there's only one node, there's no point to do checking */
return;
}
/* Count healthy nodes */
for (i=0;i<NUM_BACKENDS;i++)
{
if (VALID_BACKEND(i))
active_nodes++;
}
if (active_nodes <= 1)
{
/* If there's only one or less active node, there's no point
* to do checking */
return;
}
for (i=0;i<NUM_BACKENDS;i++)
{
if (!VALID_BACKEND(i))
continue;
if (!slots[i])
{
pool_debug("check_replication_time_lag: DB node is valid but no persistent connection");
pool_error("check_replication_time_lag: could not connect to DB node %d, check sr_check_user and sr_check_password", i);
return;
}
if (PRIMARY_NODE_ID == i)
{
query = "SELECT pg_current_xlog_location()";
}
else
{
query = "SELECT pg_last_xlog_replay_location()";
}
sts = do_query(slots[i]->con, query, &res, PROTO_MAJOR_V3);
if (sts != POOL_CONTINUE)
{
pool_error("check_replication_time_lag: %s failed", query);
return;
}
if (!res)
{
pool_error("check_replication_time_lag: %s result is null", query);
return;
}
if (res->numrows <= 0)
{
pool_error("check_replication_time_lag: %s returns no rows", query);
free_select_result(res);
return;
}
if (res->data[0] == NULL)
{
pool_error("check_replication_time_lag: %s returns no data", query);
free_select_result(res);
return;
}
if (res->nullflags[0] == -1)
{
pool_log("check_replication_time_lag: %s returns NULL", query);
free_select_result(res);
lsn[i] = 0;
}
else
{
lsn[i] = text_to_lsn(res->data[0]);
free_select_result(res);
}
}
for (i=0;i<NUM_BACKENDS;i++)
{
if (!VALID_BACKEND(i))
continue;
/* Set standby delay value */
bkinfo = pool_get_node_info(i);
lag = (lsn[PRIMARY_NODE_ID] > lsn[i]) ? lsn[PRIMARY_NODE_ID] - lsn[i] : 0;
if (PRIMARY_NODE_ID == i)
{
bkinfo->standby_delay = 0;
}
else
{
bkinfo->standby_delay = lag;
/* Log delay if necessary */
if ((!strcmp(pool_config->log_standby_delay, "always") && lag > 0) ||
(pool_config->delay_threshold &&
!strcmp(pool_config->log_standby_delay, "if_over_threshold") &&
lag > pool_config->delay_threshold))
{
pool_log("Replication of node:%d is behind %llu bytes from the primary server (node:%d)",
i, lsn[PRIMARY_NODE_ID] - lsn[i], PRIMARY_NODE_ID);
}
}
}
}