MDEV-24789: Reduce lock_sys.wait_mutex contention

A performance regression was introduced by
commit e71e6133 (MDEV-24671)
and mostly addressed by
commit 455514c8.

The regression is likely caused by increased contention
lock_sys.latch (former lock_sys.mutex), possibly indirectly
caused by contention on lock_sys.wait_mutex. This change aims to
reduce both, but further improvements will be needed.

lock_wait(): Minimize the lock_sys.wait_mutex hold time.

lock_sys_t::deadlock_check(): Add a parameter for indicating
whether lock_sys.latch is exclusively locked.

trx_t::was_chosen_as_deadlock_victim: Always use atomics.

lock_wait_wsrep(): Assume that no mutex is being held.

Deadlock::report(): Always kill the victim transaction.

lock_sys_t::timeout: New counter to back MONITOR_TIMEOUT.

mysql-test/suite/innodb/r/innodb_skip_innodb_is_tables.result

@@ -40,7 +40,7 @@ metadata_table_handles_opened	metadata	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NU
 metadata_table_handles_closed	metadata	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of table handles closed
 metadata_table_reference_count	metadata	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Table reference counter
 lock_deadlocks	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	value	Number of deadlocks
-lock_timeouts	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of lock timeouts
+lock_timeouts	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	value	Number of lock timeouts
 lock_rec_lock_waits	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of times enqueued into record lock wait queue
 lock_table_lock_waits	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of times enqueued into table lock wait queue
 lock_rec_lock_requests	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of record locks requested

storage/innobase/handler/ha_innodb.cc

@@ -4494,8 +4494,8 @@ static void innobase_kill_query(handlerton*, THD *thd, enum thd_kill_levels)
           lock_cancel_waiting_and_release(lock);
           trx->mutex_unlock();
         }
+        lock_sys.deadlock_check(true);
       }
-      lock_sys.deadlock_check();
       mysql_mutex_unlock(&lock_sys.wait_mutex);
     }
   }

storage/innobase/include/lock0lock.h

@@ -710,6 +710,8 @@ class lock_sys_t
 public:
   /** number of deadlocks detected; protected by wait_mutex */
   ulint deadlocks;
+  /** number of lock wait timeouts; protected by wait_mutex */
+  ulint timeouts;
   /**
     Constructor.
 
@@ -814,8 +816,9 @@ class lock_sys_t
   void close();
 
 
-  /** Check for deadlocks */
-  static void deadlock_check();
+  /** Check for deadlocks
+  @param locked lock_sys.is_writer() */
+  static void deadlock_check(bool locked);
 
 
   /** Note that a record lock wait started */

storage/innobase/include/trx0trx.h

@@ -429,19 +429,12 @@ struct trx_lock_t
   trx_t *wait_trx;
   /** condition variable for !wait_lock; used with lock_sys.wait_mutex */
   pthread_cond_t cond;
-  /** lock wait start time, protected only by lock_sys.wait_mutex */
-  my_hrtime_t suspend_time;
+  /** lock wait start time */
+  Atomic_relaxed<my_hrtime_t> suspend_time;
 
-#ifdef WITH_WSREP
-  /** 2=high priority wsrep thread has marked this trx to abort;
+  /** 2=high priority WSREP thread has marked this trx to abort;
   1=another transaction chose this as a victim in deadlock resolution. */
   Atomic_relaxed<byte> was_chosen_as_deadlock_victim;
-#else
-  /** When the transaction decides to wait for a lock, it clears this;
-  set if another transaction chooses this transaction as a victim in deadlock
-  resolution. Protected by lock_sys.latch and lock_sys.wait_mutex. */
-  bool was_chosen_as_deadlock_victim;
-#endif
 
   /** Next available rec_pool[] entry */
   byte rec_cached;

storage/innobase/srv/srv0mon.cc

@@ -96,7 +96,8 @@ static monitor_info_t	innodb_counter_info[] =
 	 MONITOR_DEFAULT_START, MONITOR_DEADLOCK},
 
 	{"lock_timeouts", "lock", "Number of lock timeouts",
-	 MONITOR_DEFAULT_ON,
+	 static_cast<monitor_type_t>(
+	 MONITOR_EXISTING | MONITOR_DEFAULT_ON | MONITOR_DISPLAY_CURRENT),
 	 MONITOR_DEFAULT_START, MONITOR_TIMEOUT},
 
 	{"lock_rec_lock_waits", "lock",
@@ -1883,6 +1884,9 @@ srv_mon_process_existing_counter(
 	case MONITOR_DEADLOCK:
 		value = lock_sys.deadlocks;
 		break;
+	case MONITOR_TIMEOUT:
+		value = lock_sys.timeouts;
+		break;
 
 	default:
 		ut_error;

storage/innobase/trx/trx0i_s.cc

@@ -444,9 +444,8 @@ fill_trx_row(
 
 	ut_ad(!wait_lock == !requested_lock_row);
 
-	row->trx_wait_started = wait_lock
-		? hrtime_to_time(trx->lock.suspend_time)
-		: 0;
+	const my_hrtime_t suspend_time= trx->lock.suspend_time;
+	row->trx_wait_started = wait_lock ? hrtime_to_time(suspend_time) : 0;
 
 	row->trx_weight = static_cast<uintmax_t>(TRX_WEIGHT(trx));
 

storage/rocksdb/mysql-test/rocksdb/r/innodb_i_s_tables_disabled.result

@@ -22,7 +22,7 @@ metadata_table_handles_opened	metadata	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NU
 metadata_table_handles_closed	metadata	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of table handles closed
 metadata_table_reference_count	metadata	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Table reference counter
 lock_deadlocks	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	value	Number of deadlocks
-lock_timeouts	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of lock timeouts
+lock_timeouts	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	value	Number of lock timeouts
 lock_rec_lock_waits	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of times enqueued into record lock wait queue
 lock_table_lock_waits	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of times enqueued into table lock wait queue
 lock_rec_lock_requests	lock	0	NULL	NULL	NULL	0	NULL	NULL	NULL	NULL	NULL	NULL	NULL	0	counter	Number of record locks requested