mcexec: memory policy control by environmental variable

Refs: #1470
Change-Id: I3d556cae90d31d81572b1c4e5c680e826577d428

executer/include/defs.h

@@ -4671,7 +4671,7 @@ void cmd_ipcs(void);         /* ipcs.c */
 /*
  *  main.c
  */
-void main_loop(void);
+//void main_loop(void);
 void exec_command(void);
 struct command_table_entry *get_command_table_entry(char *);
 void program_usage(int);

executer/include/uprotocol.h

@@ -113,6 +113,18 @@ typedef unsigned long __cpu_set_unit;
 #define MPOL_NO_BSS               0x04
 #define MPOL_SHM_PREMAP           0x08
 
+/* should be the same as process.h */
+#define PLD_PROCESS_NUMA_MASK_BITS 256
+
+enum {
+	PLD_MPOL_DEFAULT,
+	PLD_MPOL_PREFERRED,
+	PLD_MPOL_BIND,
+	PLD_MPOL_INTERLEAVE,
+	PLD_MPOL_LOCAL,
+	PLD_MPOL_MAX,	/* always last member of enum */
+};
+
 #define PLD_MAGIC 0xcafecafe44332211UL
 
 struct program_load_desc {
@@ -147,6 +159,10 @@ struct program_load_desc {
 	unsigned long heap_extension;
 	long stack_premap;
 	unsigned long mpol_bind_mask;
+	int mpol_mode;
+	unsigned long mpol_nodemask[PLD_PROCESS_NUMA_MASK_BITS /
+				    (sizeof(unsigned long) * 8)];
+
 	int thp_disable;
 	int uti_thread_rank; /* N-th clone() spawns a thread on Linux CPU */
 	int uti_use_last_cpu; /* Work-around not to share CPU with OpenMP thread */

executer/user/mcexec.c

@@ -84,7 +84,11 @@
 #include "../include/pmi.h"
 #include "../include/qlmpi.h"
 #include <sys/xattr.h>
+#include "../include/defs.h"
 #include "../../lib/include/list.h"
+#include "../../lib/include/bitops-set_bit.h"
+#include "../../lib/include/bitops-clear_bit.h"
+#include "../../lib/include/bitops-test_bit.h"
 
 //#define DEBUG
 #define ADD_ENVS_OPTION
@@ -1056,6 +1060,64 @@ static inline cpu_set_t *numa_node_set(int n)
 	return (cpu_set_t *)(numa_nodes + n * cpu_set_size);
 }
 
+static inline void _numa_local(__cpu_set_unit *localset,
+			       unsigned long *nodemask, int nonlocal)
+{
+	int i;
+
+	memset(nodemask, 0, PLD_PROCESS_NUMA_MASK_BITS / 8);
+
+	for (i = 0; i < nnodes; i++) {
+		cpu_set_t *nodeset = numa_node_set(i);
+		int j;
+
+		if (nonlocal) {
+			set_bit(i, nodemask);
+		}
+
+		for (j = 0; j < ncpu; j++) {
+			if (test_bit(j, localset)) {
+				__dprintf("%d belongs to local set\n", j);
+			}
+
+			if (CPU_ISSET_S(j, cpu_set_size, nodeset)) {
+				__dprintf("%d belongs to node %d\n", j, i);
+			}
+
+			if (test_bit(j, localset) &&
+			    CPU_ISSET_S(j, cpu_set_size, nodeset)) {
+				if (nonlocal) {
+					clear_bit(i, nodemask);
+				} else {
+					set_bit(i, nodemask);
+				}
+			}
+		}
+	}
+}
+
+static inline void numa_local(__cpu_set_unit *localset, unsigned long *nodemask)
+{
+	_numa_local(localset, nodemask, 0);
+}
+
+static inline void numa_nonlocal(__cpu_set_unit *localset,
+				 unsigned long *nodemask)
+{
+	_numa_local(localset, nodemask, 1);
+}
+
+static inline void numa_all(unsigned long *nodemask)
+{
+	int i;
+
+	memset(nodemask, 0, PLD_PROCESS_NUMA_MASK_BITS / 8);
+
+	for (i = 0; i < nnodes; i++) {
+		set_bit(i, nodemask);
+	}
+}
+
 pid_t master_tid;
 
 pthread_mutex_t lock = PTHREAD_MUTEX_INITIALIZER;
@@ -2685,6 +2747,7 @@ int main(int argc, char **argv)
 	desc->heap_extension = heap_extension;
 
 	desc->mpol_bind_mask = 0;
+	desc->mpol_mode = PLD_MPOL_MAX; /* not specified */
 	if (mpol_bind_nodes) {
 		struct bitmask *bind_mask;
 		bind_mask = numa_parse_nodestring_all(mpol_bind_nodes);
@@ -2698,6 +2761,54 @@ int main(int argc, char **argv)
 			}
 		}
 	}
+	/* Fujitsu TCS specific: mempolicy */
+	else if (getenv("OMPI_MCA_plm_ple_memory_allocation_policy")) {
+		char *mpol =
+			getenv("OMPI_MCA_plm_ple_memory_allocation_policy");
+
+		__dprintf("OMPI_MCA_plm_ple_memory_allocation_policy: %s\n",
+			  mpol);
+
+		if (!strncmp(mpol, "localalloc", 10)) {
+			/* MPOL_DEFAULT has the same effect as MPOL_LOCAL */
+			desc->mpol_mode = MPOL_DEFAULT;
+		}
+		else if (!strncmp(mpol, "interleave_local", 16)) {
+			desc->mpol_mode = MPOL_INTERLEAVE;
+			numa_local(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "interleave_nonlocal", 19)) {
+			desc->mpol_mode = MPOL_INTERLEAVE;
+			numa_nonlocal(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "interleave_all", 14)) {
+			desc->mpol_mode = MPOL_INTERLEAVE;
+			numa_all(desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "bind_local", 10)) {
+			desc->mpol_mode = MPOL_BIND;
+			numa_local(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "bind_nonlocal", 13)) {
+			desc->mpol_mode = MPOL_BIND;
+			numa_nonlocal(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "bind_all", 8)) {
+			desc->mpol_mode = MPOL_BIND;
+			numa_all(desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "prefer_local", 12)) {
+			desc->mpol_mode = MPOL_PREFERRED;
+			numa_local(desc->cpu_set, desc->mpol_nodemask);
+		}
+		else if (!strncmp(mpol, "prefer_nonlocal", 15)) {
+			desc->mpol_mode = MPOL_PREFERRED;
+			numa_nonlocal(desc->cpu_set, desc->mpol_nodemask);
+		}
+
+		__dprintf("mpol_mode: %d, mpol_nodemask: %ld\n",
+			  desc->mpol_mode, desc->mpol_nodemask[0]);
+	}
 
 	desc->uti_thread_rank = uti_thread_rank;
 	desc->uti_use_last_cpu = uti_use_last_cpu;

kernel/host.c

@@ -542,6 +542,26 @@ static int process_msg_prepare_process(unsigned long rphys)
 		}
 		vm->numa_mem_policy = MPOL_BIND;
 	}
+	else if (pn->mpol_mode != MPOL_MAX) {
+		int bit;
+
+		vm->numa_mem_policy = pn->mpol_mode;
+
+		memset(&vm->numa_mask, 0, sizeof(vm->numa_mask));
+
+		for_each_set_bit(bit, pn->mpol_nodemask,
+				PLD_PROCESS_NUMA_MASK_BITS) {
+			if (bit >= ihk_mc_get_nr_numa_nodes()) {
+				kprintf("%s: error: NUMA id %d is larger than mask size!\n",
+						__func__, bit);
+				return -EINVAL;
+			}
+			set_bit(bit, &vm->numa_mask[0]);
+		}
+
+		dkprintf("%s: numa_mem_policy: %d, numa_mask: %ld\n",
+			 __func__, vm->numa_mem_policy, vm->numa_mask[0]);
+	}
 
 	proc->uti_thread_rank = pn->uti_thread_rank;
 	proc->uti_use_last_cpu = pn->uti_use_last_cpu;

kernel/include/process.h

@@ -559,6 +559,8 @@ struct process {
 	size_t mpol_threshold;
 	unsigned long heap_extension;
 	unsigned long mpol_bind_mask;
+	int mpol_mode;
+
 	int uti_thread_rank; /* Spawn on Linux CPU when clone_count reaches this */
 	int uti_use_last_cpu; /* Work-around not to share CPU with OpenMP thread */
 	int clone_count;

kernel/include/syscall.h

@@ -183,6 +183,18 @@ typedef unsigned long __cpu_set_unit;
 #define MPOL_NO_BSS               0x04
 #define MPOL_SHM_PREMAP           0x08
 
+/* should be the same as process.h */
+#define PLD_PROCESS_NUMA_MASK_BITS 256
+
+enum {
+	PLD_MPOL_DEFAULT,
+	PLD_MPOL_PREFERRED,
+	PLD_MPOL_BIND,
+	PLD_MPOL_INTERLEAVE,
+	PLD_MPOL_LOCAL,
+	PLD_MPOL_MAX,	/* always last member of enum */
+};
+
 #define PLD_MAGIC 0xcafecafe44332211UL
 
 struct program_load_desc {
@@ -217,6 +229,10 @@ struct program_load_desc {
 	unsigned long heap_extension;
 	long stack_premap;
 	unsigned long mpol_bind_mask;
+	int mpol_mode;
+	unsigned long mpol_nodemask[PLD_PROCESS_NUMA_MASK_BITS /
+				    (sizeof(unsigned long) * 8)];
+
 	int thp_disable;
 	int uti_thread_rank; /* N-th clone() spawns a thread on Linux CPU */
 	int uti_use_last_cpu; /* Work-around not to share CPU with OpenMP thread */

lib/include/bitops-test_bit.h

@@ -0,0 +1,11 @@
+#ifndef INCLUDE_BITOPS_TEST_BIT_H
+#define INCLUDE_BITOPS_TEST_BIT_H
+
+static inline int test_bit(int nr, const void *addr)
+{
+	const uint32_t *p = (const uint32_t *)addr;
+
+	return ((1UL << (nr & 31)) & (p[nr >> 5])) != 0;
+}
+
+#endif

lib/include/bitops.h

@@ -27,11 +27,7 @@ unsigned long find_first_bit(const unsigned long *addr,
 unsigned long find_first_zero_bit(const unsigned long *addr, 
 				  unsigned long size);
 
-static inline int test_bit(int nr, const void *addr)
+#include <bitops-test_bit.h>
-{
-	const uint32_t *p = (const uint32_t *)addr;
-	return ((1UL << (nr & 31)) & (p[nr >> 5])) != 0;
-}
 
 extern unsigned int __sw_hweight32(unsigned int w);
 extern unsigned int __sw_hweight16(unsigned int w);

test/issues/1470/CMakeLists.txt

@@ -0,0 +1,123 @@
+execute_process(COMMAND bash -c "gawk '/CPU implementer/ { print \$4; exit; }' /proc/cpuinfo"
+  OUTPUT_VARIABLE CPU_IMPLEMENTER OUTPUT_STRIP_TRAILING_WHITESPACE)
+execute_process(COMMAND bash -c "gawk '/CPU architecture/ { print \$3; exit; }' /proc/cpuinfo"
+  OUTPUT_VARIABLE CPU_ARCH OUTPUT_STRIP_TRAILING_WHITESPACE)
+execute_process(COMMAND bash -c "gawk '/CPU variant/ { print \$4; exit; }' /proc/cpuinfo"
+  OUTPUT_VARIABLE CPU_VARIANT OUTPUT_STRIP_TRAILING_WHITESPACE)
+execute_process(COMMAND bash -c "gawk '/CPU part/ { print \$4; exit; }' /proc/cpuinfo"
+  OUTPUT_VARIABLE CPU_PART OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+if(CPU_IMPLEMENTER STREQUAL "0x46" AND CPU_ARCH STREQUAL "8" AND
+    CPU_VARIANT STREQUAL "0x1" AND CPU_PART STREQUAL "0x001")
+  message("A64FX detected")
+  set(CPU_MODEL "a64fx")
+  add_definitions(-D__a64fx__)
+endif()
+
+# find first NUMA available to user (0 or 4 now)
+execute_process(COMMAND bash -c "awk -v keyword=nr_free_pages -f ${CMAKE_CURRENT_SOURCE_DIR}/src/zoneinfo.awk /proc/zoneinfo | awk -v page_size=$(getconf PAGE_SIZE) -f ${CMAKE_CURRENT_SOURCE_DIR}/src/zoneinfo_filter.awk | head -n1" OUTPUT_VARIABLE FIRST_USER_NUMA OUTPUT_STRIP_TRAILING_WHITESPACE)
+message("FIRST_USER_NUMA: ${FIRST_USER_NUMA}")
+add_definitions(-DFIRST_USER_NUMA=${FIRST_USER_NUMA})
+
+if (FIRST_USER_NUMA STREQUAL "4")
+  execute_process(COMMAND sudo bash -c "echo 0-7 > /sys/fs/cgroup/cpuset/system.slice/cpuset.mems")
+endif()
+
+cmake_policy(SET CMP0005 NEW)
+
+# Options: -DWITH_MCK=<McKernel install directory>
+add_definitions(-DWITH_MCK=${WITH_MCK})
+
+# Options: -DWITH_MCK_SRC=<McKernel source directory>
+add_definitions(-DWITH_MCK_SRC=${WITH_MCK_SRC})
+
+# for autotest
+if(NOT DEFINED CMAKE_INSTALL_PREFIX_SCRIPTS)
+       set(CMAKE_INSTALL_PREFIX_SCRIPTS ${CMAKE_INSTALL_PREFIX}/scripts)
+endif()
+
+cmake_minimum_required(VERSION 3.0)
+
+project(issue1470 C)
+
+# CPPFLAGS
+
+set(UNAME_R ${CMAKE_SYSTEM_VERSION} CACHE STRING "Kernel version to build against")
+set(KERNEL_DIR "/lib/modules/${UNAME_R}/build" CACHE STRING "kernel build directory")
+execute_process(COMMAND awk -F= "$1 == \"CONFIG_ARM64_64K_PAGES\" { print $2; exit; }" "${KERNEL_DIR}/.config"
+	OUTPUT_VARIABLE CONFIG_ARM64_64K_PAGES OUTPUT_STRIP_TRAILING_WHITESPACE)
+if(CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+	if(CONFIG_ARM64_64K_PAGES STREQUAL "y")
+		set(PAGE_SIZE "65536")
+	else()
+		set(PAGE_SIZE "4096")
+	endif()
+else()
+		set(PAGE_SIZE "4096")
+endif()
+message("PAGE_SIZE: ${PAGE_SIZE}")
+
+if (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64")
+  add_definitions(-DBUILD_TARGET=smp-x86)
+  add_definitions(-DKMOD_POSTFIX=smp_x86)
+elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+  add_definitions(-DBUILD_TARGET=smp-arm64)
+  add_definitions(-DKMOD_POSTFIX=smp_arm64)
+endif()
+
+add_definitions(-DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX})
+
+# CFLAGS
+set(CFLAGS_WARNING "-Wall" "-Wextra" "-Wno-unused-parameter" "-Wno-sign-compare" "-Wno-unused-function" ${EXTRA_WARNINGS} CACHE STRING "Warning flags")
+add_compile_options(
+  -O2
+  -g
+  ${CFLAGS_WARNING}
+  )
+
+
+# -L, this must be done before adding dependants
+link_directories("${WITH_MCK}/lib64")
+
+# -Wl,--rpath=, this must be done before adding dependants
+set(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
+set(CMAKE_INSTALL_RPATH "${WITH_MCK}/lib64")
+
+# test driver scripts
+foreach(target IN ITEMS
+    OMPI_MCA_plm_ple_memory_allocation_policy01
+    OMPI_MCA_plm_ple_memory_allocation_policy02
+    OMPI_MCA_plm_ple_memory_allocation_policy03
+    OMPI_MCA_plm_ple_memory_allocation_policy04
+    OMPI_MCA_plm_ple_memory_allocation_policy05
+    OMPI_MCA_plm_ple_memory_allocation_policy06
+    OMPI_MCA_plm_ple_memory_allocation_policy07
+    )
+
+  # String replacement
+  configure_file(src/${target}.sh.in ihklib-${target} @ONLY)
+
+  # Install scripts
+  install(PROGRAMS ${CMAKE_BINARY_DIR}/ihklib-${target} DESTINATION ${CMAKE_INSTALL_PREFIX_SCRIPTS})
+endforeach()
+
+# programs running on McKernel
+foreach(target IN ITEMS
+    check_mempolicy
+    )
+
+  # Add C target
+  add_executable(${target} src/${target}.c)
+
+  # -I
+  target_include_directories(${target}
+    PRIVATE "${PROJECT_SOURCE_DIR}/include"
+    )
+
+  # -l
+  target_link_libraries(${target} PRIVATE numa)
+
+  # Install
+  install(TARGETS ${target} DESTINATION bin)
+
+endforeach()

test/issues/1470/README

@@ -0,0 +1,32 @@
+============
+What to test
+============
+Check if the mode and nodemask obtained by get_mempolicy() are set to the expected values, with different reserved cpus, cpumask and policy requests.
+The following settings are used.
+
+	Memory reserved: "1G@4,1G@5,1G@6,1G@7"
+
+	CPUs reserved: "12-59", "24-59"
+	FLIB_AFFINITY_ON_PROCESS:
+		"12-23", "24-35", "36-47", "48-59",
+		"12-35", "24-47", "36-59",
+		"12-47", "24-59",
+		"12-59"
+	OMPI_MCA_plm_ple_memory_allocation_policy:
+		{interleave,bind,prefer}_{local,nonlocal},
+		{interleave,bind}_all,
+		localalloc
+
+============
+How to build
+============
+cd <mckernel>/test/issues
+mkdir build
+cd build
+cmake ../1470/ -DCMAKE_INSTALL_PREFIX=<mckernel>/test/issues/install -DWITH_MCK=<mckernel-install> -DWITH_MCK_SRC=<mckernel>
+
+===========
+How to test
+===========
+for i in {1..7}; do <mckernel>/test/issues/install/scripts/ihklib-OMPI_MCA_plm_ple_memory_allocation_policy0$i; done
+check if no "[  NG  ]" is shown.

test/issues/1470/include/okng.h

@@ -0,0 +1,33 @@
+#ifndef __OKNG_H_INCLUDED__
+#define __OKNG_H_INCLUDED__
+
+#include <stdio.h>
+
+#define _OKNG(verb, jump, cond, fmt, args...) do {		\
+	if (cond) {						\
+		if (verb)					\
+			printf("[  OK  ] " fmt, ##args);	\
+	} else {						\
+		printf("[  NG  ] " fmt, ##args);		\
+		if (jump) {					\
+			ret = 1;				\
+			goto out;				\
+		}						\
+	}							\
+} while (0)
+
+#define OKNG(args...) _OKNG(1, 1, ##args)
+#define INFO(fmt, args...) printf("[ INFO ] " fmt, ##args)
+#define START(fmt, args...) printf("[ START] " fmt, ##args)
+#define INTERR(cond, fmt, args...) do {	 \
+	if (cond) {			 \
+		char msg[4096];			 \
+		sprintf(msg, fmt, ##args);		 \
+		printf("[INTERR] %s:%d %s", __FILE__, __LINE__, msg);	\
+		ret = 1;					\
+		goto out;					\
+	} \
+} while (0)
+#define ARRAY_SIZE_CHECK(array, size) INTERR(sizeof(array)/sizeof(array[0]) != size, "size of array \"%s\" isn't %d\n", #array, size)
+
+#endif

test/issues/1470/src/OMPI_MCA_plm_ple_memory_allocation_policy01.sh.in

@@ -0,0 +1,53 @@
+#!/usr/bin/bash
+
+# define WORKDIR
+SCRIPT_PATH=$(readlink -m "${BASH_SOURCE[0]}")
+SCRIPT_NAME="${SCRIPT_PATH##*/}"
+TEST_NAME="${SCRIPT_NAME%.sh}"
+AUTOTEST_HOME="${SCRIPT_PATH%/*/*/*}"
+if [ -f ${AUTOTEST_HOME}/bin/config.sh ]; then
+    . ${AUTOTEST_HOME}/bin/config.sh
+else
+    WORKDIR=$(pwd)
+fi
+
+declare -A mode
+mode[interleave_local]="3"
+mode[bind_local]="2"
+mode[prefer_local]="1"
+
+declare -A nodemask
+
+nodemask[$(seq -s, 12 23)]="1"
+nodemask[$(seq -s, 24 35)]="2"
+nodemask[$(seq -s, 36 47)]="4"
+nodemask[$(seq -s, 48 59)]="8"
+nodemask[$(seq -s, 12 35)]="3"
+nodemask[$(seq -s, 24 47)]="6"
+nodemask[$(seq -s, 36 59)]="12"
+nodemask[$(seq -s, 12 47)]="7"
+nodemask[$(seq -s, 24 59)]="14"
+nodemask[$(seq -s, 12 59)]="15"
+
+@WITH_MCK@/sbin/mcstop+release.sh
+@WITH_MCK@/sbin/mcreboot.sh -c 12-59 -m 1G@4,1G@5,1G@6,1G@7
+
+for policy in interleave_local bind_local prefer_local; do
+    for cpuset in \
+	$(seq -s, 12 23) $(seq -s, 24 35) $(seq -s, 36 47) $(seq -s, 48 59) \
+	$(seq -s, 12 35) $(seq -s, 24 47) $(seq -s, 36 59) \
+	$(seq -s, 12 47) $(seq -s, 24 59) \
+	$(seq -s, 12 59); do
+
+	# check if policy is not set when not specified
+	if (( i++ == 0 )); then
+	    FLIB_NUM_PROCESS_ON_NODE=1 FLIB_AFFINITY_ON_PROCESS=$cpuset @WITH_MCK@/bin/mcexec @CMAKE_INSTALL_PREFIX@/bin/check_mempolicy -m 0 || exit $?
+	fi
+
+	FLIB_NUM_PROCESS_ON_NODE=1 FLIB_AFFINITY_ON_PROCESS=$cpuset OMPI_MCA_plm_ple_memory_allocation_policy=$policy @WITH_MCK@/bin/mcexec @CMAKE_INSTALL_PREFIX@/bin/check_mempolicy -m ${mode[$policy]} -n ${nodemask[$cpuset]} || exit $?
+    done
+done
+
+@WITH_MCK@/sbin/mcstop+release.sh
+
+exit 0

test/issues/1470/src/OMPI_MCA_plm_ple_memory_allocation_policy02.sh.in

@@ -0,0 +1,48 @@
+#!/usr/bin/bash
+
+# define WORKDIR
+SCRIPT_PATH=$(readlink -m "${BASH_SOURCE[0]}")
+SCRIPT_NAME="${SCRIPT_PATH##*/}"
+TEST_NAME="${SCRIPT_NAME%.sh}"
+AUTOTEST_HOME="${SCRIPT_PATH%/*/*/*}"
+if [ -f ${AUTOTEST_HOME}/bin/config.sh ]; then
+    . ${AUTOTEST_HOME}/bin/config.sh
+else
+    WORKDIR=$(pwd)
+fi
+
+declare -A mode
+mode[interleave_nonlocal]="3"
+mode[bind_nonlocal]="2"
+mode[prefer_nonlocal]="1"
+
+declare -A nodemask
+
+nodemask[$(seq -s, 12 23)]="14"
+nodemask[$(seq -s, 24 35)]="13"
+nodemask[$(seq -s, 36 47)]="11"
+nodemask[$(seq -s, 48 59)]="7"
+nodemask[$(seq -s, 12 35)]="12"
+nodemask[$(seq -s, 24 47)]="9"
+nodemask[$(seq -s, 36 59)]="3"
+nodemask[$(seq -s, 12 47)]="8"
+nodemask[$(seq -s, 24 59)]="1"
+nodemask[$(seq -s, 12 59)]="0"
+
+@WITH_MCK@/sbin/mcstop+release.sh
+@WITH_MCK@/sbin/mcreboot.sh -c 12-59 -m 1G@4,1G@5,1G@6,1G@7
+
+for policy in interleave_nonlocal bind_nonlocal prefer_nonlocal; do
+    for cpuset in \
+	$(seq -s, 12 23) $(seq -s, 24 35) $(seq -s, 36 47) $(seq -s, 48 59) \
+	$(seq -s, 12 35) $(seq -s, 24 47) $(seq -s, 36 59) \
+	$(seq -s, 12 47) $(seq -s, 24 59) \
+	$(seq -s, 12 59); do
+
+	FLIB_NUM_PROCESS_ON_NODE=1 FLIB_AFFINITY_ON_PROCESS=$cpuset OMPI_MCA_plm_ple_memory_allocation_policy=$policy @WITH_MCK@/bin/mcexec @CMAKE_INSTALL_PREFIX@/bin/check_mempolicy -m ${mode[$policy]} -n ${nodemask[$cpuset]} || exit $?
+    done
+done
+
+@WITH_MCK@/sbin/mcstop+release.sh
+
+exit 0

test/issues/1470/src/OMPI_MCA_plm_ple_memory_allocation_policy03.sh.in

@@ -0,0 +1,34 @@
+#!/usr/bin/bash
+
+# define WORKDIR
+SCRIPT_PATH=$(readlink -m "${BASH_SOURCE[0]}")
+SCRIPT_NAME="${SCRIPT_PATH##*/}"
+TEST_NAME="${SCRIPT_NAME%.sh}"
+AUTOTEST_HOME="${SCRIPT_PATH%/*/*/*}"
+if [ -f ${AUTOTEST_HOME}/bin/config.sh ]; then
+    . ${AUTOTEST_HOME}/bin/config.sh
+else
+    WORKDIR=$(pwd)
+fi
+
+declare -A mode
+mode[interleave_all]="3"
+mode[bind_all]="2"
+
+@WITH_MCK@/sbin/mcstop+release.sh
+@WITH_MCK@/sbin/mcreboot.sh -c 12-59 -m 1G@4,1G@5,1G@6,1G@7
+
+for policy in interleave_all bind_all; do
+    for cpuset in \
+	$(seq -s, 12 23) $(seq -s, 24 35) $(seq -s, 36 47) $(seq -s, 48 59) \
+	$(seq -s, 12 35) $(seq -s, 24 47) $(seq -s, 36 59) \
+	$(seq -s, 12 47) $(seq -s, 24 59) \
+	$(seq -s, 12 59); do
+
+	FLIB_NUM_PROCESS_ON_NODE=1 FLIB_AFFINITY_ON_PROCESS=$cpuset OMPI_MCA_plm_ple_memory_allocation_policy=$policy @WITH_MCK@/bin/mcexec @CMAKE_INSTALL_PREFIX@/bin/check_mempolicy -m ${mode[$policy]} -n 15 || exit $?
+    done
+done
+
+@WITH_MCK@/sbin/mcstop+release.sh
+
+exit 0

test/issues/1470/src/OMPI_MCA_plm_ple_memory_allocation_policy04.sh.in

@@ -0,0 +1,30 @@
+#!/usr/bin/bash
+
+# define WORKDIR
+SCRIPT_PATH=$(readlink -m "${BASH_SOURCE[0]}")
+SCRIPT_NAME="${SCRIPT_PATH##*/}"
+TEST_NAME="${SCRIPT_NAME%.sh}"
+AUTOTEST_HOME="${SCRIPT_PATH%/*/*/*}"
+if [ -f ${AUTOTEST_HOME}/bin/config.sh ]; then
+    . ${AUTOTEST_HOME}/bin/config.sh
+else
+    WORKDIR=$(pwd)
+fi
+
+@WITH_MCK@/sbin/mcstop+release.sh
+@WITH_MCK@/sbin/mcreboot.sh -c 12-59 -m 1G@4,1G@5,1G@6,1G@7
+
+for policy in localalloc; do
+    for cpuset in \
+	$(seq -s, 12 23) $(seq -s, 24 35) $(seq -s, 36 47) $(seq -s, 48 59) \
+	$(seq -s, 12 35) $(seq -s, 24 47) $(seq -s, 36 59) \
+	$(seq -s, 12 47) $(seq -s, 24 59) \
+	$(seq -s, 12 59); do
+
+	FLIB_NUM_PROCESS_ON_NODE=1 FLIB_AFFINITY_ON_PROCESS=$cpuset OMPI_MCA_plm_ple_memory_allocation_policy=$policy @WITH_MCK@/bin/mcexec @CMAKE_INSTALL_PREFIX@/bin/check_mempolicy -m 0 || exit $?
+    done
+done
+
+@WITH_MCK@/sbin/mcstop+release.sh
+
+exit 0

test/issues/1470/src/OMPI_MCA_plm_ple_memory_allocation_policy05.sh.in

@@ -0,0 +1,44 @@
+#!/usr/bin/bash
+
+# define WORKDIR
+SCRIPT_PATH=$(readlink -m "${BASH_SOURCE[0]}")
+SCRIPT_NAME="${SCRIPT_PATH##*/}"
+TEST_NAME="${SCRIPT_NAME%.sh}"
+AUTOTEST_HOME="${SCRIPT_PATH%/*/*/*}"
+if [ -f ${AUTOTEST_HOME}/bin/config.sh ]; then
+    . ${AUTOTEST_HOME}/bin/config.sh
+else
+    WORKDIR=$(pwd)
+fi
+
+declare -A mode
+mode[interleave_local]="3"
+mode[bind_local]="2"
+mode[prefer_local]="1"
+
+declare -A nodemask
+
+nodemask[$(seq -s, 12 23)]="2"
+nodemask[$(seq -s, 24 35)]="4"
+nodemask[$(seq -s, 36 47)]="8"
+nodemask[$(seq -s, 12 35)]="6"
+nodemask[$(seq -s, 24 47)]="12"
+nodemask[$(seq -s, 12 47)]="14"
+
+# reserve the last 36 cpus
+@WITH_MCK@/sbin/mcreboot.sh -c 24-59 -m 1G@4,1G@5,1G@6,1G@7
+
+for policy in interleave_local bind_local prefer_local; do
+    for cpuset in \
+	$(seq -s, 12 23) $(seq -s, 24 35) $(seq -s, 36 47) \
+	$(seq -s, 12 35) $(seq -s, 24 47) \
+	$(seq -s, 12 47); do
+
+	# check nodemask when last 36 cpus are reserved
+	FLIB_NUM_PROCESS_ON_NODE=1 FLIB_AFFINITY_ON_PROCESS=$cpuset OMPI_MCA_plm_ple_memory_allocation_policy=$policy @WITH_MCK@/bin/mcexec @CMAKE_INSTALL_PREFIX@/bin/check_mempolicy -m ${mode[$policy]} -n ${nodemask[$cpuset]} || exit $?
+    done
+done
+
+@WITH_MCK@/sbin/mcstop+release.sh
+
+exit 0

test/issues/1470/src/OMPI_MCA_plm_ple_memory_allocation_policy06.sh.in

@@ -0,0 +1,44 @@
+#!/usr/bin/bash
+
+# define WORKDIR
+SCRIPT_PATH=$(readlink -m "${BASH_SOURCE[0]}")
+SCRIPT_NAME="${SCRIPT_PATH##*/}"
+TEST_NAME="${SCRIPT_NAME%.sh}"
+AUTOTEST_HOME="${SCRIPT_PATH%/*/*/*}"
+if [ -f ${AUTOTEST_HOME}/bin/config.sh ]; then
+    . ${AUTOTEST_HOME}/bin/config.sh
+else
+    WORKDIR=$(pwd)
+fi
+
+declare -A mode
+mode[interleave_nonlocal]="3"
+mode[bind_nonlocal]="2"
+mode[prefer_nonlocal]="1"
+
+declare -A nodemask
+
+nodemask[$(seq -s, 12 23)]="13"
+nodemask[$(seq -s, 24 35)]="11"
+nodemask[$(seq -s, 36 47)]="7"
+nodemask[$(seq -s, 12 35)]="9"
+nodemask[$(seq -s, 24 47)]="3"
+nodemask[$(seq -s, 12 47)]="1"
+
+# reserve the last 36 cpus
+@WITH_MCK@/sbin/mcreboot.sh -c 24-59 -m 1G@4,1G@5,1G@6,1G@7
+
+for policy in interleave_nonlocal bind_nonlocal prefer_nonlocal; do
+    for cpuset in \
+	$(seq -s, 12 23) $(seq -s, 24 35) $(seq -s, 36 47) \
+	$(seq -s, 12 35) $(seq -s, 24 47) \
+	$(seq -s, 12 47); do
+
+	# check nodemask when last 36 cpus are reserved
+	FLIB_NUM_PROCESS_ON_NODE=1 FLIB_AFFINITY_ON_PROCESS=$cpuset OMPI_MCA_plm_ple_memory_allocation_policy=$policy @WITH_MCK@/bin/mcexec @CMAKE_INSTALL_PREFIX@/bin/check_mempolicy -m ${mode[$policy]} -n ${nodemask[$cpuset]} || exit $?
+    done
+done
+
+@WITH_MCK@/sbin/mcstop+release.sh
+
+exit 0

test/issues/1470/src/OMPI_MCA_plm_ple_memory_allocation_policy07.sh.in

@@ -0,0 +1,34 @@
+#!/usr/bin/bash
+
+# define WORKDIR
+SCRIPT_PATH=$(readlink -m "${BASH_SOURCE[0]}")
+SCRIPT_NAME="${SCRIPT_PATH##*/}"
+TEST_NAME="${SCRIPT_NAME%.sh}"
+AUTOTEST_HOME="${SCRIPT_PATH%/*/*/*}"
+if [ -f ${AUTOTEST_HOME}/bin/config.sh ]; then
+    . ${AUTOTEST_HOME}/bin/config.sh
+else
+    WORKDIR=$(pwd)
+fi
+
+declare -A mode
+mode[interleave_all]="3"
+mode[bind_all]="2"
+
+# reserve the last 36 cpus
+@WITH_MCK@/sbin/mcreboot.sh -c 24-59 -m 1G@4,1G@5,1G@6,1G@7
+
+for policy in interleave_all bind_all; do
+    for cpuset in \
+	$(seq -s, 12 23) $(seq -s, 24 35) $(seq -s, 36 47) \
+	$(seq -s, 12 35) $(seq -s, 24 47) \
+	$(seq -s, 12 47); do
+
+	# check nodemask when last 36 cpus are reserved
+	FLIB_NUM_PROCESS_ON_NODE=1 FLIB_AFFINITY_ON_PROCESS=$cpuset OMPI_MCA_plm_ple_memory_allocation_policy=$policy @WITH_MCK@/bin/mcexec @CMAKE_INSTALL_PREFIX@/bin/check_mempolicy -m ${mode[$policy]} -n 15 || exit $?
+    done
+done
+
+@WITH_MCK@/sbin/mcstop+release.sh
+
+exit 0

test/issues/1470/src/check_mempolicy.c

@@ -0,0 +1,54 @@
+#include <stdio.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <errno.h>
+#include <numaif.h>
+#include <okng.h>
+
+#define PLD_PROCESS_NUMA_MASK_BITS 256
+
+int main(int argc, char **argv)
+{
+	long ret;
+	int mode;
+	unsigned long nodemask[PLD_PROCESS_NUMA_MASK_BITS /
+			       (sizeof(unsigned long) * 8)] = { 0 };
+	int mode_expected = -1;
+	unsigned long nodemask_expected[PLD_PROCESS_NUMA_MASK_BITS /
+					(sizeof(unsigned long) * 8)] = { 0 };
+	int opt;
+
+	while ((opt = getopt(argc, argv, "m:n:")) != -1) {
+		switch (opt) {
+		case 'm':
+			mode_expected = atol(optarg);
+			break;
+		case 'n':
+			nodemask_expected[0] = atoi(optarg);
+			break;
+		default: /* '?' */
+			INTERR(1, "unknown option %c\n", optopt);
+		}
+	}
+
+	INTERR(mode_expected == -1, "specify -m <mode>\n");
+
+	ret = get_mempolicy(&mode, nodemask, PLD_PROCESS_NUMA_MASK_BITS,
+			    NULL, 0);
+	INTERR(ret, "get_mempolicy failed with %ld\n", ret);
+
+	OKNG(mode == mode_expected, "mode: actual (%d), expected (%d)\n",
+	     mode, mode_expected);
+
+	/* nodemask is "don't care" when mode is MPOL_DEFAULT */
+	if (mode_expected != 0) {
+		OKNG(nodemask[0] == nodemask_expected[0],
+		     "nodemask: actual (%ld), expected (%ld)\n",
+		     nodemask[0],
+		     nodemask_expected[0]);
+	}
+
+	ret = 0;
+ out:
+	return ret;
+}

test/issues/1470/src/zoneinfo.awk

@@ -0,0 +1,9 @@
+BEGIN { id = -1; }
+/Node .*, zone\s*(Normal|DMA32)/ { id = substr($2, 1, length($2) - 1); }
+{
+    if ($0 ~ keyword && id != -1) {
+	printf("id: %d, nr_free_pages: %ld\n", id, $2);
+	id = -1;
+    }
+}
+

test/issues/1470/src/zoneinfo_filter.awk

@@ -0,0 +1,13 @@
+{
+    id = substr($2, 1, length($2) - 1);
+    size = $4;
+    sizes[id] += size;
+}
+
+END {
+    for (i = 0; i <= id; i++) {
+	if (sizes[i] * page_size > 2 * 1024 * 1024 * 1024) {
+	    print i;
+	}
+    }
+}