MDEV-14482 - Cache line contention on ut_rnd_ulint_counter()

InnoDB RNG maintains global state, causing otherwise unnecessary bus
traffic. Even worse this is cross-mutex traffic. That is different
mutexes suffer from contention.

Fixed delay of 4 was verified to give best throughput by OLTP update
index and read-write benchmarks on Intel Broadwell (2/20/40) and
ARM (1/46/46).

10 files changed, 23 insertions, 86 deletions

diff --git a/mysql-test/suite/sys_vars/r/innodb_spin_wait_delay_basic.result b/mysql-test/suite/sys_vars/r/innodb_spin_wait_delay_basic.result
index 88516a854fe..b76ba4933f2 100644
--- a/mysql-test/suite/sys_vars/r/innodb_spin_wait_delay_basic.result
+++ b/mysql-test/suite/sys_vars/r/innodb_spin_wait_delay_basic.result
@@ -1,28 +1,28 @@
 SET @start_global_value = @@global.innodb_spin_wait_delay;
 SELECT @start_global_value;
 @start_global_value
-6
+4
 Valid values are zero or above
 select @@global.innodb_spin_wait_delay >=0;
 @@global.innodb_spin_wait_delay >=0
 1
 select @@global.innodb_spin_wait_delay;
 @@global.innodb_spin_wait_delay
-6
+4
 select @@session.innodb_spin_wait_delay;
 ERROR HY000: Variable 'innodb_spin_wait_delay' is a GLOBAL variable
 show global variables like 'innodb_spin_wait_delay';
 Variable_name	Value
-innodb_spin_wait_delay	6
+innodb_spin_wait_delay	4
 show session variables like 'innodb_spin_wait_delay';
 Variable_name	Value
-innodb_spin_wait_delay	6
+innodb_spin_wait_delay	4
 select * from information_schema.global_variables where variable_name='innodb_spin_wait_delay';
 VARIABLE_NAME	VARIABLE_VALUE
-INNODB_SPIN_WAIT_DELAY	6
+INNODB_SPIN_WAIT_DELAY	4
 select * from information_schema.session_variables where variable_name='innodb_spin_wait_delay';
 VARIABLE_NAME	VARIABLE_VALUE
-INNODB_SPIN_WAIT_DELAY	6
+INNODB_SPIN_WAIT_DELAY	4
 set global innodb_spin_wait_delay=10;
 select @@global.innodb_spin_wait_delay;
 @@global.innodb_spin_wait_delay
@@ -38,7 +38,7 @@ ERROR HY000: Variable 'innodb_spin_wait_delay' is a GLOBAL variable and should b
 set global innodb_spin_wait_delay=DEFAULT;
 select @@global.innodb_spin_wait_delay;
 @@global.innodb_spin_wait_delay
-6
+4
 set global innodb_spin_wait_delay=0;
 select @@global.innodb_spin_wait_delay;
 @@global.innodb_spin_wait_delay
@@ -111,4 +111,4 @@ INNODB_SPIN_WAIT_DELAY	0
 SET @@global.innodb_spin_wait_delay = @start_global_value;
 SELECT @@global.innodb_spin_wait_delay;
 @@global.innodb_spin_wait_delay
-6
+4
diff --git a/mysql-test/suite/sys_vars/r/sysvars_innodb.result b/mysql-test/suite/sys_vars/r/sysvars_innodb.result
index 720d6a1458a..ad972212d7b 100644
--- a/mysql-test/suite/sys_vars/r/sysvars_innodb.result
+++ b/mysql-test/suite/sys_vars/r/sysvars_innodb.result
@@ -1996,12 +1996,12 @@ READ_ONLY	YES
 COMMAND_LINE_ARGUMENT	REQUIRED
 VARIABLE_NAME	INNODB_SPIN_WAIT_DELAY
 SESSION_VALUE	NULL
-GLOBAL_VALUE	6
+GLOBAL_VALUE	4
 GLOBAL_VALUE_ORIGIN	COMPILE-TIME
-DEFAULT_VALUE	6
+DEFAULT_VALUE	4
 VARIABLE_SCOPE	GLOBAL
 VARIABLE_TYPE	INT UNSIGNED
-VARIABLE_COMMENT	Maximum delay between polling for a spin lock (6 by default)
+VARIABLE_COMMENT	Maximum delay between polling for a spin lock (4 by default)
 NUMERIC_MIN_VALUE	0
 NUMERIC_MAX_VALUE	6000
 NUMERIC_BLOCK_SIZE	0
diff --git a/storage/innobase/dict/dict0stats.cc b/storage/innobase/dict/dict0stats.cc
index 82bdf1f92ce..a45c752ffc0 100644
--- a/storage/innobase/dict/dict0stats.cc
+++ b/storage/innobase/dict/dict0stats.cc
@@ -1758,11 +1758,8 @@ dict_stats_analyze_index_for_n_prefix(
 		ut_a(left <= right);
 		ut_a(right <= last_idx_on_level);
 
-		/* we do not pass (left, right) because we do not want to ask
-		ut_rnd_interval() to work with too big numbers since
-		ib_uint64_t could be bigger than ulint */
-		const ulint	rnd = ut_rnd_interval(
-			0, static_cast<ulint>(right - left));
+		const ulint	rnd = right == left ? 0 :
+			ut_rnd_gen_ulint() % (right - left);
 
 		const ib_uint64_t	dive_below_idx
 			= boundaries->at(static_cast<unsigned>(left + rnd));
diff --git a/storage/innobase/fil/fil0fil.cc b/storage/innobase/fil/fil0fil.cc
index fa65bda4490..b93489a10e6 100644
--- a/storage/innobase/fil/fil0fil.cc
+++ b/storage/innobase/fil/fil0fil.cc
@@ -5978,17 +5978,6 @@ fil_tablespace_iterate(
 		innodb_data_file_key, filepath,
 		OS_FILE_OPEN, OS_FILE_READ_WRITE, srv_read_only_mode, &success);
 
-	DBUG_EXECUTE_IF("fil_tablespace_iterate_failure",
-	{
-		static bool once;
-
-		if (!once || ut_rnd_interval(0, 10) == 5) {
-			once = true;
-			success = false;
-			os_file_close(file);
-		}
-	});
-
 	if (!success) {
 		/* The following call prints an error message */
 		os_file_get_last_error(true);
diff --git a/storage/innobase/handler/ha_innodb.cc b/storage/innobase/handler/ha_innodb.cc
index b1f858a5927..41ce69d9ca6 100644
--- a/storage/innobase/handler/ha_innodb.cc
+++ b/storage/innobase/handler/ha_innodb.cc
@@ -20559,8 +20559,8 @@ static MYSQL_SYSVAR_ULONG(sync_spin_loops, srv_n_spin_wait_rounds,
 
 static MYSQL_SYSVAR_UINT(spin_wait_delay, srv_spin_wait_delay,
   PLUGIN_VAR_OPCMDARG,
-  "Maximum delay between polling for a spin lock (6 by default)",
-  NULL, NULL, 6, 0, 6000, 0);
+  "Maximum delay between polling for a spin lock (4 by default)",
+  NULL, NULL, 4, 0, 6000, 0);
 
 static MYSQL_SYSVAR_ULONG(thread_concurrency, srv_thread_concurrency,
   PLUGIN_VAR_RQCMDARG,
diff --git a/storage/innobase/include/ib0mutex.h b/storage/innobase/include/ib0mutex.h
index 76f02cc1521..20703a4a933 100644
--- a/storage/innobase/include/ib0mutex.h
+++ b/storage/innobase/include/ib0mutex.h
@@ -225,7 +225,7 @@ struct TTASFutexMutex {
 				return;
 			}
 
-			ut_delay(ut_rnd_interval(0, max_delay));
+			ut_delay(max_delay);
 		}
 
 		for (n_waits= 0;; n_waits++) {
@@ -362,7 +362,7 @@ struct TTASMutex {
 		uint32_t n_spins = 0;
 
 		while (!try_lock()) {
-			ut_delay(ut_rnd_interval(0, max_delay));
+			ut_delay(max_delay);
 			if (++n_spins == max_spins) {
 				os_thread_yield();
 				max_spins+= step;
@@ -516,7 +516,7 @@ struct TTASEventMutex {
 					sync_array_wait_event(sync_arr, cell);
 				}
 			} else {
-				ut_delay(ut_rnd_interval(0, max_delay));
+				ut_delay(max_delay);
 			}
 		}
 
diff --git a/storage/innobase/include/ut0rnd.h b/storage/innobase/include/ut0rnd.h
index 49ae3c81356..5baf8684d23 100644
--- a/storage/innobase/include/ut0rnd.h
+++ b/storage/innobase/include/ut0rnd.h
@@ -61,16 +61,6 @@ UNIV_INLINE
 ulint
 ut_rnd_gen_ulint(void);
 /*==================*/
-/********************************************************//**
-Generates a random integer from a given interval.
-@return the 'random' number */
-UNIV_INLINE
-ulint
-ut_rnd_interval(
-/*============*/
-	ulint	low,	/*!< in: low limit; can generate also this value */
-	ulint	high);	/*!< in: high limit; can generate also this value */
-
 /*******************************************************//**
 The following function generates a hash value for a ulint integer
 to a hash table of size table_size, which should be a prime or some
diff --git a/storage/innobase/include/ut0rnd.ic b/storage/innobase/include/ut0rnd.ic
index 16dccb545d8..1e4915dd0f9 100644
--- a/storage/innobase/include/ut0rnd.ic
+++ b/storage/innobase/include/ut0rnd.ic
@@ -97,30 +97,6 @@ ut_rnd_gen_ulint(void)
 	return(rnd);
 }
 
-/********************************************************//**
-Generates a random integer from a given interval.
-@return the 'random' number */
-UNIV_INLINE
-ulint
-ut_rnd_interval(
-/*============*/
-	ulint	low,	/*!< in: low limit; can generate also this value */
-	ulint	high)	/*!< in: high limit; can generate also this value */
-{
-	ulint	rnd;
-
-	ut_ad(high >= low);
-
-	if (low == high) {
-
-		return(low);
-	}
-
-	rnd = ut_rnd_gen_ulint();
-
-	return(low + (rnd % (high - low)));
-}
-
 /*******************************************************//**
 The following function generates a hash value for a ulint integer
 to a hash table of size table_size, which should be a prime
diff --git a/storage/innobase/lock/lock0lock.cc b/storage/innobase/lock/lock0lock.cc
index 8049e6d6d01..1440ad9eae8 100644
--- a/storage/innobase/lock/lock0lock.cc
+++ b/storage/innobase/lock/lock0lock.cc
@@ -7285,7 +7285,7 @@ lock_trx_release_locks(
 
 			/** Doing an implicit to explicit conversion
 			should not be expensive. */
-			ut_delay(ut_rnd_interval(0, srv_spin_wait_delay));
+			ut_delay(srv_spin_wait_delay);
 		}
 
 		lock_mutex_enter();
diff --git a/storage/innobase/sync/sync0rw.cc b/storage/innobase/sync/sync0rw.cc
index 13c81d7333d..9996bd7262c 100644
--- a/storage/innobase/sync/sync0rw.cc
+++ b/storage/innobase/sync/sync0rw.cc
@@ -317,10 +317,7 @@ lock_loop:
 	while (i < srv_n_spin_wait_rounds &&
 	       my_atomic_load32_explicit(&lock->lock_word,
 					 MY_MEMORY_ORDER_RELAXED) <= 0) {
-		if (srv_spin_wait_delay) {
-			ut_delay(ut_rnd_interval(0, srv_spin_wait_delay));
-		}
-
+		ut_delay(srv_spin_wait_delay);
 		i++;
 	}
 
@@ -441,9 +438,7 @@ rw_lock_x_lock_wait_func(
 
 	HMT_low();
 	while (my_atomic_load32_explicit(&lock->lock_word, MY_MEMORY_ORDER_RELAXED) < threshold) {
-		if (srv_spin_wait_delay) {
-			ut_delay(ut_rnd_interval(0, srv_spin_wait_delay));
-		}
+		ut_delay(srv_spin_wait_delay);
 
 		if (i < srv_n_spin_wait_rounds) {
 			i++;
@@ -714,12 +709,7 @@ lock_loop:
 		HMT_low();
 		while (i < srv_n_spin_wait_rounds
 		       && my_atomic_load32_explicit(&lock->lock_word, MY_MEMORY_ORDER_RELAXED) <= X_LOCK_HALF_DECR) {
-
-			if (srv_spin_wait_delay) {
-				ut_delay(ut_rnd_interval(
-						0, srv_spin_wait_delay));
-			}
-
+			ut_delay(srv_spin_wait_delay);
 			i++;
 		}
 
@@ -820,12 +810,7 @@ lock_loop:
 		/* Spin waiting for the lock_word to become free */
 		while (i < srv_n_spin_wait_rounds
 		       && my_atomic_load32_explicit(&lock->lock_word, MY_MEMORY_ORDER_RELAXED) <= X_LOCK_HALF_DECR) {
-
-			if (srv_spin_wait_delay) {
-				ut_delay(ut_rnd_interval(
-						0, srv_spin_wait_delay));
-			}
-
+			ut_delay(srv_spin_wait_delay);
 			i++;
 		}