[thread_pool] Call thd_wait_* callbacks for admitting new query in th…

…read pool plugin

Summary:
1. Change the code to admit new query through `thd_wait_begin/end` callbacks using new `THD_WAIT_ADMIT` wait type.
2. Move yield counting to the callbacks to unify the code between the server and the plugin. `yield_cond` exposes the yield condition to check once the yield counter says that yield is needed.
3. Expose all wait events through `admission_control_wait_events` var. This is especially needed for `USER_LOCK` which still controls innodb thread concurrency yield.
4. Copy and fix up admission control tests to the thread_pool suite. Only the var and table names are changed to refer to thread_pool plugin. The results have been diffed with original tests to make sure they are identical (except for var/table names).

Reviewed By: lth

Differential Revision: D27593084

---------------------------------------------------------------------------------------------

use lambda instead of std::bind (facebook#1243)

Summary:
std::bind(yield_condition, table) will generate a functor which size is larger than std::function's local buf, thus std::function needs to new/delete memory to store the functor.

This PR use the lambda which just capture one pointer(table), which size can fit into std::function's local buf thus new/delete is not needed.

Pull Request resolved: facebook#1243

Reviewed By: lth

Differential Revision: D40858532

Pulled By: hermanlee

---------------------------------------------------------------------------------------------

Add yields to hash join iterator

Summary: Queries performing joins are lacking yields and monopolizing thread pool scheduler.

Differential Revision: D39948376

------------------------------------------------------------------------------

Add yield to union iterator

Summary:
This was detected by the stall detection (8.0.23 based callstack):
```
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431] Scheduler 12 stalled by worker 0x7f6367d5d640
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431] Stack trace for 1 thread(s) [1422783 mysqld]:
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 00000000001204d0 ppoll
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 000000000289d67f vio_io_wait(Vio*, enum_vio_io_event, timeout_t)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 0000000002952296 vio_socket_io_wait(Vio*, enum_vio_io_event)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 0000000002385fb2 net_write_packet(NET*, unsigned char const*, unsigned long)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 00000000025b2c44 SELECT_LEX_UNIT::ExecuteIteratorQuery(THD*)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 000000000258d2db Sql_cmd_dml::execute(THD*)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 000000000254f179 mysql_execute_command(THD*, bool, unsigned long long*)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 0000000002545936 dispatch_sql_command(THD*, Parser_state*, unsigned long long*)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 000000000269d931 dispatch_command(THD*, COM_DATA const*, enum_server_command)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 000000000269c3e0 do_command(THD*)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 0000000005c4ffa7 mysql::thread_pool::TpConnHandler::processEvent(void*)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 0000000005c5395f mysql::thread_pool::TpTask::execute()
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 0000000005c5c299 mysql::thread_pool::TpWorkerPool::processWorker(void*)
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 0000000004867855 pfs_spawn_thread(void*) [clone .llvm.11916604799980067416]
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 000000000009ac0e start_thread
W0205 14:38:40.101111 2319640 TpScheduler.cpp:431]     @ 000000000012d1db __clone3
```

Cleaned up existing yield code to use a common default yield condition.

Reviewed By: preritj24

Differential Revision: D43071441

------------------------------------------------------------------------------

Add yields to Item_func::fix_fields

Summary:
When resolving an expression with many arguments (e.g. a long `IN (...)` list) we may not yield the scheduler in a timely fashion.

This should resolve stalls like
```
W0127 03:04:04.517819 2608142 TpScheduler::checkStall: Conn state: acExitLevel=0 isGoverned=1 readmissionMode=0 admitCount=1 waitEvents=0xa7 acFilter=0x0
W0127 03:04:04.517819 2608142 TpScheduler::checkStall: Collect time 748751 nsec, symbolize time 6554425 nsec, digest dc6129f5efe09474cfacb020cdc3521b
W0127 03:04:04.517819 2608142 TpScheduler::checkStall: Stack trace for 1 thread(s) [2608144 tp_conn-0]:
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 0000000002c134bd Item_func_in::resolve_type(THD*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 0000000002839275 Item_func::fix_fields(THD*, Item**)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 0000000002eca7d5 Item_func_in::fix_fields(THD*, Item**)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 000000000295fd80 Query_block::setup_conds(THD*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 0000000002ca4b77 Sql_cmd_delete::prepare_inner(THD*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 000000000282d295 Sql_cmd_dml::prepare(THD*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 0000000002827053 Sql_cmd_dml::execute(THD*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 000000000282177d mysql_execute_command(THD*, bool, unsigned long long*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 0000000002807d26 dispatch_sql_command(THD*, Parser_state*, unsigned long long*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 00000000027f8238 dispatch_command(THD*, COM_DATA const*, enum_server_command)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 00000000027f6d20 do_command(THD*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 00000000062dedd7 mysql::thread_pool::TpConnHandler::processEvent(void*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 00000000062f01f6 mysql::thread_pool::TpWorkerPool::processWorker(void*)
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 000000000475447a pfs_spawn_thread(void*) [clone .__uniq.73903348312165179054628174044725776814] [clone .llvm.15823314186845019609]
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 000000000009abae start_thread
W0127 03:04:04.517819 2608142 TpScheduler::checkStall:     @ 000000000012d17b __clone3
```

Differential Revision: D53241252

fbshipit-source-id: 7fee664

include/mysql/service_thd_wait.h

@@ -56,7 +56,7 @@
 class THD;
 #define MYSQL_THD THD *
 
-/*
+/**
   One should only report wait events that could potentially block for a
   long time. A mutex wait is too short of an event to report. The reason
   is that an event which is reported leads to a new thread starts
@@ -69,9 +69,13 @@ class THD;
   holds true for global read locks, table locks and other meta data locks.
   Another event of interest is going to sleep for an extended time.
 
-  Note that user-level locks no longer use THD_WAIT_USER_LOCK wait type.
+  @note User-level locks no longer use THD_WAIT_USER_LOCK wait type.
   Since their implementation relies on metadata locks manager it uses
   THD_WAIT_META_DATA_LOCK instead.
+
+  @note THD_WAIT_ADMIT is a fake wait type communicating that query has been
+  parsed and is ready for execution. The query attributes and sql command
+  are available at this point.
 */
 typedef enum _thd_wait_type_e {
   THD_WAIT_SLEEP = 1,
@@ -80,14 +84,15 @@ typedef enum _thd_wait_type_e {
   THD_WAIT_GLOBAL_LOCK = 4,
   THD_WAIT_META_DATA_LOCK = 5,
   THD_WAIT_TABLE_LOCK = 6,
-  THD_WAIT_USER_LOCK = 7,
+  THD_WAIT_INNODB_CONC = 7,
   THD_WAIT_BINLOG = 8,
   THD_WAIT_GROUP_COMMIT = 9,
   THD_WAIT_SYNC = 10,
   THD_WAIT_FOR_HLC = 11,
   THD_WAIT_NET_IO = 12,
   THD_WAIT_YIELD = 13,
-  THD_WAIT_LAST = 14
+  THD_WAIT_ADMIT = 14,
+  THD_WAIT_LAST = 15
 } thd_wait_type;
 
 extern "C" struct thd_wait_service_st {

mysql-test/suite/sys_vars/r/admission_control_wait_events_basic.result

@@ -18,13 +18,17 @@ SET GLOBAL admission_control_wait_events='YIELD';
 SELECT @@global.admission_control_wait_events;
 @@global.admission_control_wait_events
 YIELD
-SET GLOBAL admission_control_wait_events='SLEEP,NET_IO,META_DATA_LOCK,ROW_LOCK,YIELD';
+SET GLOBAL admission_control_wait_events='SLEEP,ROW_LOCK,META_DATA_LOCK,INNODB_CONC,NET_IO,YIELD';
 SELECT @@global.admission_control_wait_events;
 @@global.admission_control_wait_events
-SLEEP,ROW_LOCK,META_DATA_LOCK,NET_IO,YIELD
+SLEEP,ROW_LOCK,META_DATA_LOCK,INNODB_CONC,NET_IO,YIELD
+SET GLOBAL admission_control_wait_events='SLEEP,SLEEP';
+SELECT @@global.admission_control_wait_events;
+@@global.admission_control_wait_events
+SLEEP
 SET GLOBAL admission_control_wait_events='YIELD,NONEXISTING_BIT';
 ERROR 42000: Variable 'admission_control_wait_events' can't be set to the value of 'NONEXISTING_BIT'
 SELECT @@global.admission_control_wait_events;
 @@global.admission_control_wait_events
-SLEEP,ROW_LOCK,META_DATA_LOCK,NET_IO,YIELD
+SLEEP
 set global admission_control_wait_events = @saved_admission_control_wait_events;

mysql-test/suite/sys_vars/t/admission_control_wait_events_basic.test

@@ -12,7 +12,11 @@ SELECT @@global.admission_control_wait_events;
 SET GLOBAL admission_control_wait_events='YIELD';
 SELECT @@global.admission_control_wait_events;
 
-SET GLOBAL admission_control_wait_events='SLEEP,NET_IO,META_DATA_LOCK,ROW_LOCK,YIELD';
+SET GLOBAL admission_control_wait_events='SLEEP,ROW_LOCK,META_DATA_LOCK,INNODB_CONC,NET_IO,YIELD';
+SELECT @@global.admission_control_wait_events;
+
+# check that repeating the same value is OK
+SET GLOBAL admission_control_wait_events='SLEEP,SLEEP';
 SELECT @@global.admission_control_wait_events;
 
 # checking that setting variable to a non existing value raises error

mysql-test/suite/thread_pool/r/admission_control.result

@@ -0,0 +1,160 @@
+create database test_db;
+create database test_db2;
+create user 'test_user'@'localhost';
+grant all on test_db.* to 'test_user'@'localhost';
+grant all on test_db2.* to 'test_user'@'localhost';
+grant all on test.* to 'test_user'@'localhost';
+use test_db2;
+use test_db;
+set @start_max_running_queries= @@global.thread_pool_max_running_queries;
+set @start_max_waiting_queries= @@global.thread_pool_max_waiting_queries;
+set @@global.thread_pool_max_running_queries=10;
+set @@global.thread_pool_max_waiting_queries=5;
+create table t1(a int) engine=InnoDB;
+lock table t1 write;
+Threads waiting for admission will have appropriate state set in processlist.
+Super user is exempted from admission control checks.
+select * from t1;
+a
+set @@global.thread_pool_admission_control_filter = 'USE';
+select @@global.thread_pool_admission_control_filter;
+@@global.thread_pool_admission_control_filter
+USE
+Maximum waiting queries reached. So this would hit an error.
+use test_db;
+select * from t1||||
+ERROR HY000: Maximum waiting queries 5 reached for database `test_db`
+Maximum waiting queries reached. So this would hit an error.
+use test_db2;
+create table t1_test(aaa int);
+insert into t1_test values (1);
+select aaa from t1_test;
+drop table t1_test;
+use test_db;
+select * from t1||||
+aaa
+1
+ERROR HY000: Maximum waiting queries 5 reached for database `test_db`
+use test_db;
+select * from t1;
+ERROR HY000: Maximum waiting queries 5 reached for database `test_db`
+set @@global.thread_pool_admission_control_filter = '';
+select @@global.thread_pool_admission_control_filter;
+@@global.thread_pool_admission_control_filter
+
+Check status variables
+aborted_queries = 3
+running_queries = 10
+waiting_queries = 5
+select * from information_schema.tp_admission_control_entities where schema_name like 'test_db%' order by schema_name;
+SCHEMA_NAME	WAITING_QUERIES	RUNNING_QUERIES	ABORTED_QUERIES	TIMEOUT_QUERIES	CONNECTIONS	REJECTED_CONNECTIONS
+test_db	5	10	3	0	15	0
+test_db2	0	0	0	0	0	0
+Filled up queues on one db doesn't affect queries on other db.
+use test_db2;
+select * from information_schema.tp_admission_control_entities where schema_name like 'test_db%' order by schema_name;
+SCHEMA_NAME	WAITING_QUERIES	RUNNING_QUERIES	ABORTED_QUERIES	TIMEOUT_QUERIES	CONNECTIONS	REJECTED_CONNECTIONS
+test_db	5	10	3	0	15	0
+test_db2	0	1	0	0	1	0
+set @@global.thread_pool_max_waiting_queries=6;
+Kill a thread that is waiting for admission.
+select count(*) from t1;
+kill ID;
+use test_db;
+unlock tables;
+Verify the waiting queries received wakeup signal.
+select count(*) from t1;
+count(*)
+15
+set @save_admission_control_by_trx = @@global.thread_pool_admission_control_by_trx;
+select @save_admission_control_by_trx;
+@save_admission_control_by_trx
+0
+set @@global.thread_pool_max_running_queries=5;
+set @@global.thread_pool_max_waiting_queries=10;
+# By default, open transaction has no effect on running queries
+select count(*) from t1;
+count(*)
+15
+# Test: open transactions will take slots in running queries,
+#       and will not be blocked
+set @@global.thread_pool_admission_control_filter = 'BEGIN,COMMIT,ROLLBACK';
+select @@global.thread_pool_admission_control_filter;
+@@global.thread_pool_admission_control_filter
+BEGIN,COMMIT,ROLLBACK
+set @@global.thread_pool_admission_control_by_trx = true;
+SELECT @@global.thread_pool_admission_control_by_trx;
+@@global.thread_pool_admission_control_by_trx
+1
+Open transaction is able to continue running queries
+connection con_max_wait;
+New queries will be rejected (waiting queue is full)
+select * from t1;
+ERROR HY000: Maximum waiting queries 10 reached for database `test_db`
+New transactions will be rejected (waiting queue is full)
+begin;
+select * from t1;
+ERROR HY000: Maximum waiting queries 10 reached for database `test_db`
+aborted_queries will increase by 2
+Committing a transaction will free up the running query slots
+The waiting queries will be unblocked
+Check status variables
+include/assert.inc [DB Admission control waiting queries should be zero]
+include/assert.inc [DB Admission control running queries should be zero]
+include/assert.inc [DB Admission control aborted queries should be five]
+select * from information_schema.tp_admission_control_entities where schema_name like 'test_db%' order by schema_name;
+SCHEMA_NAME	WAITING_QUERIES	RUNNING_QUERIES	ABORTED_QUERIES	TIMEOUT_QUERIES	CONNECTIONS	REJECTED_CONNECTIONS
+test_db	0	0	5	0	13	0
+test_db2	0	0	0	0	0	0
+set @@global.thread_pool_admission_control_by_trx = @save_admission_control_by_trx;
+select @@global.thread_pool_admission_control_by_trx;
+@@global.thread_pool_admission_control_by_trx
+0
+set @@global.thread_pool_admission_control_filter = '';
+select @@global.thread_pool_admission_control_filter;
+@@global.thread_pool_admission_control_filter
+
+# End of open transaction test
+# Test admission_control_queue_timeout
+use test_db;
+set @@global.thread_pool_max_running_queries=1;
+set @@global.thread_pool_max_waiting_queries=5;
+set @save_admission_control_filter = @@global.thread_pool_admission_control_filter;
+set @save_admission_control_queue_timeout = @@global.thread_pool_admission_control_queue_timeout;
+set @@global.thread_pool_admission_control_queue_timeout = 100;
+set @@global.thread_pool_admission_control_filter = 'BEGIN,COMMIT,ROLLBACK';
+create table t2(a int primary key) engine=InnoDB;
+begin;
+insert into t2 values (1);
+begin;
+insert into t2 values (1);
+insert into t2 values (2);
+ERROR HY000: Got timeout while waiting on admission control queue for database `test_db`
+rollback;
+rollback;
+drop table t2;
+set @@global.thread_pool_admission_control_filter = @save_admission_control_filter;
+set @@global.thread_pool_admission_control_queue_timeout = @save_admission_control_queue_timeout;
+timeout_queries should be 1
+timeout_queries = 1
+waiting_queries should be 0
+waiting_queries = 0
+select * from information_schema.tp_admission_control_entities where schema_name like 'test_db%' order by schema_name;
+SCHEMA_NAME	WAITING_QUERIES	RUNNING_QUERIES	ABORTED_QUERIES	TIMEOUT_QUERIES	CONNECTIONS	REJECTED_CONNECTIONS
+test_db	0	0	5	1	15	0
+test_db2	0	0	0	0	0	0
+reset global.thread_pool_max_running_queries and global.thread_pool_max_waiting_queries
+set @@global.thread_pool_max_running_queries=10;
+set @@global.thread_pool_max_waiting_queries=5;
+Run parallel load and drop the database.
+set @@global.thread_pool_max_waiting_queries=0;
+Cleanup.
+Verify there are no waiting threads.
+select count(*) from information_schema.processlist where state='waiting for admission';
+count(*)
+0
+select * from information_schema.tp_admission_control_entities where schema_name like 'test_db%' order by schema_name;
+SCHEMA_NAME	WAITING_QUERIES	RUNNING_QUERIES	ABORTED_QUERIES	TIMEOUT_QUERIES	CONNECTIONS	REJECTED_CONNECTIONS
+set @@global.thread_pool_max_running_queries=@start_max_running_queries;
+set @@global.thread_pool_max_waiting_queries=@start_max_waiting_queries;
+drop user test_user@localhost;

mysql-test/suite/thread_pool/r/admission_control_hang.result

@@ -0,0 +1,93 @@
+create database test_db;
+create user test_user@localhost;
+grant all on test_db.* to test_user@localhost;
+set @start_max_running_queries = @@global.max_running_queries;
+set @@global.max_running_queries = 4;
+set @start_innodb_lock_wait_timeout = @@global.innodb_lock_wait_timeout;
+set @@global.innodb_lock_wait_timeout = 10000;
+set @start_admission_control_filter = @@global.admission_control_filter;
+set @@global.admission_control_filter = 'COMMIT';
+create table t1 (a int) engine=innodb;
+insert into t1 values(1);
+begin;
+update t1 set a=2 where a=1;
+update t1 set a=2 where a=1;
+update t1 set a=2 where a=1;
+update t1 set a=2 where a=1;
+update t1 set a=2 where a=1;
+set @@global.admission_control_filter = 'USE';
+select @@global.admission_control_filter;
+@@global.admission_control_filter
+USE
+use test;
+use test_db;
+set @@global.admission_control_filter = 'ALTER,BEGIN,COMMIT,CREATE,DELETE,DROP,INSERT,LOAD,SELECT,SET,REPLACE,TRUNCATE,UPDATE,SHOW,ROLLBACK';
+select @@global.admission_control_filter;
+@@global.admission_control_filter
+ALTER,BEGIN,COMMIT,CREATE,DELETE,DROP,INSERT,LOAD,SELECT,SET,REPLACE,ROLLBACK,TRUNCATE,UPDATE,SHOW
+create table t2(a int) engine=innodb;
+begin;
+insert into t2 values(1);
+update t2 set a=2 where a=1;
+commit;
+SHOW TABLES LIKE 't2';
+Tables_in_test_db (t2)
+t2
+begin;
+alter table t2 rename t3;
+select * from t3;
+a
+2
+delete from t3;
+set @val = 1;
+truncate table t3;
+rollback;
+drop table t3;
+set @save_admission_control_by_trx = @@global.admission_control_by_trx;
+select @save_admission_control_by_trx;
+@save_admission_control_by_trx
+0
+# Turn on admission_control_by_trx
+set @@global.admission_control_by_trx = true;
+SELECT @@global.admission_control_by_trx;
+@@global.admission_control_by_trx
+1
+create table t2(a int) engine=innodb;
+begin;
+insert into t2 values(1);
+update t2 set a=2 where a=1;
+commit;
+SHOW TABLES LIKE 't2';
+Tables_in_test_db (t2)
+t2
+begin;
+alter table t2 rename t3;
+select * from t3;
+a
+2
+delete from t3;
+set @val = 1;
+truncate table t3;
+rollback;
+drop table t3;
+set @@global.admission_control_filter = default;
+select @@global.admission_control_filter;
+@@global.admission_control_filter
+
+select count(*) from t1;
+count(*)
+1
+set @@global.admission_control_by_trx = @save_admission_control_by_trx;
+select @@global.admission_control_by_trx;
+@@global.admission_control_by_trx
+0
+set @@global.admission_control_filter = 'COMMIT';
+select @@global.admission_control_filter;
+@@global.admission_control_filter
+COMMIT
+commit;
+set @@global.max_running_queries = @start_max_running_queries;
+set @@global.innodb_lock_wait_timeout = @start_innodb_lock_wait_timeout;
+set @@global.admission_control_filter = @start_admission_control_filter;
+drop database test_db;
+drop user test_user@localhost;

mysql-test/suite/thread_pool/r/admission_control_multi_query.result

@@ -0,0 +1,8 @@
+create database test_db;
+create user test_user@localhost identified with 'mysql_native_password' BY '';
+grant all on test_db.* to test_user@localhost;
+grant all on test.* to test_user@localhost;
+use test_db;
+create table t1 (a int primary key, b int) engine=InnoDB;
+drop database test_db;
+drop user test_user@localhost;

mysql-test/suite/thread_pool/r/admission_control_multi_query_wait_events.result

@@ -0,0 +1,8 @@
+create database test_db;
+create user test_user@localhost identified with 'mysql_native_password' BY '';
+grant all on test_db.* to test_user@localhost;
+grant all on test.* to test_user@localhost;
+use test_db;
+create table t1 (a int primary key, b int) engine=InnoDB;
+drop database test_db;
+drop user test_user@localhost;

mysql-test/suite/thread_pool/r/admission_control_multi_query_weighted.result

@@ -0,0 +1,8 @@
+create database test_db;
+create user test_user@localhost identified with 'mysql_native_password' BY '';
+grant all on test_db.* to test_user@localhost;
+grant all on test.* to test_user@localhost;
+use test_db;
+create table t1 (a int primary key, b int) engine=InnoDB;
+drop database test_db;
+drop user test_user@localhost;