spec: prerelease 0.3 for testing ihk_reserve_mem and memory policy

Change-Id: I4fbcfa1f93522fd01af42d1ef13d0be075086773

CMakeLists.txt

@@ -10,7 +10,7 @@ project(mckernel C ASM)
 set(MCKERNEL_VERSION "1.7.1")
 
 # See "Fedora Packaging Guidelines -- Versioning"
-set(MCKERNEL_RELEASE "0.1")
+set(MCKERNEL_RELEASE "0.3")
 
 set(CMAKE_MODULE_PATH ${CMAKE_SOURCE_DIR}/cmake/modules)
 # for rpmbuild
@@ -26,10 +26,8 @@ endif()
 
 if (BUILD_TARGET STREQUAL "smp-x86")
 	set(ARCH "x86_64")
-	option(ENABLE_TOFU "Built-in tofu driver support" OFF)
 elseif (BUILD_TARGET STREQUAL "smp-arm64")
 	set(ARCH "arm64")
-	option(ENABLE_TOFU "Built-in tofu driver support" ON)
 endif()
 
 include(GNUInstallDirs)
@@ -52,6 +50,20 @@ if (ENABLE_WERROR)
 	add_compile_options("-Werror")
 endif(ENABLE_WERROR)
 
+execute_process(COMMAND bash -c "ls -ld /proc/tofu/ 2>/dev/null | wc -l"
+  OUTPUT_VARIABLE PROC_TOFU OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+if(PROC_TOFU STREQUAL "1")
+	option(ENABLE_TOFU "Built-in tofu driver support" ON)
+else()
+	option(ENABLE_TOFU "Built-in tofu driver support" OFF)
+endif()
+
+if(ENABLE_TOFU)
+	add_definitions(-DENABLE_TOFU)
+	set(KBUILD_C_FLAGS "${KBUILD_C_FLAGS} -DENABLE_TOFU")
+endif()
+
 option(ENABLE_LINUX_WORK_IRQ_FOR_IKC "Use Linux work IRQ for IKC IPI" ON)
 if (ENABLE_LINUX_WORK_IRQ_FOR_IKC)
 	set(KBUILD_C_FLAGS "${KBUILD_C_FLAGS} -DIHK_IKC_USE_LINUX_WORK_IRQ")

arch/arm64/kernel/fault.c

@@ -223,8 +223,12 @@ static int do_translation_fault(unsigned long addr,
 					  unsigned int esr,
 					  struct pt_regs *regs)
 {
+#ifdef ENABLE_TOFU
 	// XXX: Handle kernel space page faults for Tofu driver
 	//if (addr < USER_END)
+#else
+	if (addr < USER_END)
+#endif
 		return do_page_fault(addr, esr, regs);
 
 	do_bad_area(addr, esr, regs);

arch/arm64/kernel/include/arch-memory.h

@@ -94,7 +94,11 @@ extern char _end[];
 # define LD_TASK_UNMAPPED_BASE	UL(0x0000080000000000)
 # define TASK_UNMAPPED_BASE	UL(0x0000100000000000)
 # define USER_END		UL(0x0000400000000000)
+#ifdef ENABLE_TOFU
 # define MAP_VMAP_START		UL(0xffff7bdfffff0000)
+#else
+# define MAP_VMAP_START		UL(0xffff780000000000)
+#endif
 # define MAP_VMAP_SIZE		UL(0x0000000100000000)
 # define MAP_FIXED_START	UL(0xffff7ffffbdd0000)
 # define MAP_ST_START		UL(0xffff800000000000)

arch/arm64/kernel/memory.c

@@ -217,11 +217,13 @@ static inline int ptl4_index(unsigned long addr)
 	int idx = (addr >> PTL4_SHIFT) & PTL4_INDEX_MASK;
 	return idx;
 }
+#ifdef ENABLE_TOFU
 static inline int ptl3_index_linux(unsigned long addr)
 {
 	int idx = (addr >> PTL3_SHIFT) & PTL3_INDEX_MASK_LINUX;
 	return idx;
 }
+#endif
 static inline int ptl3_index(unsigned long addr)
 {
 	int idx = (addr >> PTL3_SHIFT) & PTL3_INDEX_MASK;
@@ -281,6 +283,7 @@ static inline pte_t* ptl4_offset(const translation_table_t* ptl4, unsigned long
 	return ptep;
 }
 
+#ifdef ENABLE_TOFU
 static inline pte_t* ptl3_offset_linux(const pte_t* l4p, unsigned long addr)
 {
 	pte_t* ptep = NULL;
@@ -311,6 +314,7 @@ static inline pte_t* ptl3_offset_linux(const pte_t* l4p, unsigned long addr)
 	}
 	return ptep;
 }
+#endif
 
 static inline pte_t* ptl3_offset(const pte_t* l4p, unsigned long addr)
 {
@@ -991,10 +995,12 @@ static void init_normal_area(struct page_table *pt)
 	
 	tt = get_translation_table(pt);
 
+#ifdef ENABLE_TOFU
 	setup(tt,
 			arm64_st_phys_base,
 			arm64_st_phys_base  + (1UL << 40));
 	return;
+#endif
 
 	for (i = 0; i < ihk_mc_get_nr_memory_chunks(); i++) {
 		unsigned long map_start, map_end;
@@ -1323,6 +1329,7 @@ out:
 	return ret;
 }
 
+#ifdef ENABLE_TOFU
 int ihk_mc_linux_pt_virt_to_phys_size(struct page_table *pt,
                            const void *virt,
 						   unsigned long *phys,
@@ -1373,7 +1380,7 @@ out:
 	if(size) *size = lsize;
 	return 0;
 }
-
+#endif
 
 int ihk_mc_pt_virt_to_phys_size(struct page_table *pt,
                            const void *virt,

arch/arm64/kernel/traps.c

@@ -174,13 +174,14 @@ void bad_mode(struct pt_regs *regs, int reason, unsigned int esr)
 
 	arch_show_interrupt_context(regs);
 
-#if 0
+#ifdef ENABLE_TOFU
+	info.si_signo = SIGSTOP;
+	info.si_errno = 0;
+#else
 	info.si_signo = SIGILL;
 	info.si_errno = 0;
 	info.si_code  = ILL_ILLOPC;
 #endif
-	info.si_signo = SIGSTOP;
-	info.si_errno = 0;
 	info._sifields._sigfault.si_addr  = (void*)regs->pc;
 
 	arm64_notify_die("Oops - bad mode", regs, &info, 0);

config.h.in

@@ -16,9 +16,6 @@
 /* whether perf is enabled */
 #cmakedefine ENABLE_PERF 1
 
-/* whether built-in tofu driver is enabled */
-#cmakedefine ENABLE_TOFU 1
-
 /* whether qlmpi is enabled */
 #cmakedefine ENABLE_QLMPI 1
 

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 {
@@ -147,12 +159,18 @@ 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 */
 	int straight_map;
 	size_t straight_map_threshold;
+#ifdef ENABLE_TOFU
 	int enable_tofu;
+#endif
 	int nr_processes;
 	int process_rank;
 	__cpu_set_unit cpu_set[PLD_CPU_SET_SIZE];
@@ -199,7 +217,9 @@ struct syscall_response {
 	unsigned long req_thread_status;
 	long ret;
 	unsigned long fault_address;
+#ifdef ENABLE_TOFU
 	void *pde_data;
+#endif
 };
 
 struct syscall_ret_desc {

executer/kernel/mcctrl/arch/arm64/archdeps.c

@@ -28,6 +28,7 @@ void *vdso_end;
 static struct vm_special_mapping (*vdso_spec)[2];
 #endif
 
+#ifdef ENABLE_TOFU
 /* Tofu CQ and barrier gate release functions */
 struct file_operations *mcctrl_tof_utofu_procfs_ops_cq;
 int (*mcctrl_tof_utofu_release_cq)(struct inode *inode,
@@ -35,6 +36,7 @@ int (*mcctrl_tof_utofu_release_cq)(struct inode *inode,
 struct file_operations *mcctrl_tof_utofu_procfs_ops_bch;
 int (*mcctrl_tof_utofu_release_bch)(struct inode *inode,
 		struct file *filp);
+#endif
 
 int arch_symbols_init(void)
 {
@@ -52,6 +54,7 @@ int arch_symbols_init(void)
 		return -EFAULT;
 #endif
 
+#ifdef ENABLE_TOFU
 	mcctrl_tof_utofu_procfs_ops_cq =
 		(void *)kallsyms_lookup_name("tof_utofu_procfs_ops_cq");
 	if (WARN_ON(!mcctrl_tof_utofu_procfs_ops_cq))
@@ -71,6 +74,7 @@ int arch_symbols_init(void)
 		(void *)kallsyms_lookup_name("tof_utofu_release_bch");
 	if (WARN_ON(!mcctrl_tof_utofu_release_bch))
 		return -EFAULT;
+#endif
 
 	return 0;
 }
@@ -448,6 +452,7 @@ out:
 }
 
 
+#ifdef ENABLE_TOFU
 /*
  * Tofu CQ and BCH release handlers
  */
@@ -549,3 +554,4 @@ int __mcctrl_tof_utofu_release_bch(struct inode *inode, struct file *filp)
 	return __mcctrl_tof_utofu_release_handler(inode, filp,
 			mcctrl_tof_utofu_release_bch);
 }
+#endif

executer/kernel/mcctrl/driver.c

@@ -50,7 +50,9 @@ extern void procfs_exit(int);
 extern void uti_attr_finalize(void);
 extern void binfmt_mcexec_init(void);
 extern void binfmt_mcexec_exit(void);
+#ifdef ENABLE_TOFU
 extern void mcctrl_file_to_pidfd_hash_init(void);
+#endif
 
 extern int mcctrl_os_read_cpu_register(ihk_os_t os, int cpu,
 		struct ihk_os_cpu_register *desc);
@@ -325,7 +327,9 @@ static int __init mcctrl_init(void)
 	}
 
 	binfmt_mcexec_init();
+#ifdef ENABLE_TOFU
 	mcctrl_file_to_pidfd_hash_init();
+#endif
 
 	if ((ret = symbols_init()))
 		goto error;

executer/kernel/mcctrl/mcctrl.h

@@ -560,6 +560,7 @@ struct uti_futex_resp {
 	wait_queue_head_t wq;
 };
 
+#ifdef ENABLE_TOFU
 /*
  * Hash table to keep track of files and related processes
  * and file descriptors.
@@ -585,3 +586,4 @@ struct mcctrl_file_to_pidfd *mcctrl_file_to_pidfd_hash_lookup(
 int mcctrl_file_to_pidfd_hash_remove(struct file *filp,
 	ihk_os_t os, struct task_struct *group_leader, int fd);
 #endif
+#endif

executer/kernel/mcctrl/syscall.c

@@ -1843,6 +1843,7 @@ static long pager_call(ihk_os_t os, struct syscall_request *req)
 	return ret;
 }
 
+#ifdef ENABLE_TOFU
 struct list_head mcctrl_file_to_pidfd_hash[MCCTRL_FILE_2_PIDFD_HASH_SIZE];
 spinlock_t mcctrl_file_to_pidfd_hash_lock;
 
@@ -1971,7 +1972,7 @@ unlock_out:
 	spin_unlock_irqrestore(&mcctrl_file_to_pidfd_hash_lock, irqflags);
 	return ret;
 }
-
+#endif
 
 void __return_syscall(ihk_os_t os, struct ikc_scd_packet *packet,
 		long ret, int stid)
@@ -2458,6 +2459,7 @@ int __do_in_kernel_syscall(ihk_os_t os, struct ikc_scd_packet *packet)
 
 	dprintk("%s: system call: %lx\n", __FUNCTION__, sc->args[0]);
 	switch (sc->number) {
+#ifdef ENABLE_TOFU
 	case __NR_close: {
 		struct fd f;
 		int fd;
@@ -2478,6 +2480,7 @@ int __do_in_kernel_syscall(ihk_os_t os, struct ikc_scd_packet *packet)
 
 		break;
 	}
+#endif
 	case __NR_mmap:
 		ret = pager_call(os, sc);
 		break;

executer/user/mcexec.c

@@ -68,13 +68,13 @@
 #include <sys/user.h>
 #endif /* !__aarch64__ */
 #include <sys/prctl.h>
+#include "../../config.h"
 #include "../include/uprotocol.h"
 #include <ihk/ihk_host_user.h>
 #include "../include/uti.h"
 #include <getopt.h>
 #include "archdep.h"
 #include "arch_args.h"
-#include "../../config.h"
 #include <numa.h>
 #include <numaif.h>
 #include <spawn.h>
@@ -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
@@ -200,7 +204,9 @@ static char *mpol_bind_nodes = NULL;
 static int uti_thread_rank = 0;
 static int uti_use_last_cpu = 0;
 static int enable_uti = 0;
+#ifdef ENABLE_TOFU
 static int enable_tofu = 0;
+#endif
 
 /* Partitioned execution (e.g., for MPI) */
 static int nr_processes = 0;
@@ -1056,6 +1062,64 @@ 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;
@@ -1725,12 +1789,14 @@ static struct option mcexec_options[] = {
 		.flag =		&enable_uti,
 		.val =		1,
 	},
+#ifdef ENABLE_TOFU
 	{
 		.name =		"enable-tofu",
 		.has_arg =	no_argument,
 		.flag =		&enable_tofu,
 		.val =		1,
 	},
+#endif
 	{
 		.name =		"debug-mcexec",
 		.has_arg =	no_argument,
@@ -2685,6 +2751,7 @@ 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 +2765,54 @@ 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;
@@ -2705,7 +2820,9 @@ int main(int argc, char **argv)
 
 	desc->straight_map = straight_map;
 	desc->straight_map_threshold = straight_map_threshold;
+#ifdef ENABLE_TOFU
 	desc->enable_tofu = enable_tofu;
+#endif
 
 	/* user_start and user_end are set by this call */
 	if (ioctl(fd, MCEXEC_UP_PREPARE_IMAGE, (unsigned long)desc) != 0) {

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

@@ -559,13 +559,17 @@ 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;
 	int thp_disable;
 
 	int straight_map;
+#ifdef ENABLE_TOFU
 	int enable_tofu;
+#endif
 	size_t straight_map_threshold;
 
 	// perf_event
@@ -590,9 +594,11 @@ struct process {
 
 	int coredump_barrier_count, coredump_barrier_count2;
 	mcs_rwlock_lock_t coredump_lock; // lock for coredump
+#ifdef ENABLE_TOFU
 #define MAX_FD_PDE 1024
 	void *fd_pde_data[MAX_FD_PDE];
 	char *fd_path[MAX_FD_PDE];
+#endif
 };
 
 /*
@@ -751,8 +757,10 @@ struct thread {
 	struct waitq coredump_wq;
 	int coredump_status;
 
+#ifdef ENABLE_TOFU
 	/* Path of file being opened */
 	char *fd_path_in_open;
+#endif
 };
 
 #define VM_RANGE_CACHE_SIZE	4

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 {
@@ -217,12 +229,18 @@ 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 */
 	int straight_map;
 	size_t straight_map_threshold;
+#ifdef ENABLE_TOFU
 	int enable_tofu;
+#endif
 	int nr_processes;
 	int process_rank;
 	__cpu_set_unit cpu_set[PLD_CPU_SET_SIZE];

kernel/syscall.c

@@ -220,7 +220,9 @@ long do_syscall(struct syscall_request *req, int cpu)
 		req->ttid = 0;
 	}
 	res.req_thread_status = IHK_SCD_REQ_THREAD_SPINNING;
+#ifdef ENABLE_TOFU
 	res.pde_data = NULL;
+#endif
 	send_syscall(req, cpu, &res);
 
 	if (req->rtid == -1) {
@@ -381,6 +383,7 @@ long do_syscall(struct syscall_request *req, int cpu)
 
 	rc = res.ret;
 
+#ifdef ENABLE_TOFU
 	if ((req->number == __NR_ioctl && rc == 0) ||
 			(req->number == __NR_openat && rc > 0)) {
 		int fd = req->number == __NR_ioctl ? req->args[0] : rc;
@@ -407,6 +410,7 @@ long do_syscall(struct syscall_request *req, int cpu)
 					res.pde_data);
 		}
 	}
+#endif
 
 	if(req->number != __NR_exit_group){
 		--thread->in_syscall_offload;
@@ -1528,7 +1532,9 @@ int process_cleanup_before_terminate(int pid)
 {
 	struct process *proc;
 	struct mcs_rwlock_node_irqsave lock;
+#ifdef ENABLE_TOFU
 	int fd;
+#endif
 
 	proc = find_process(pid, &lock);
 	if (!proc) {
@@ -1536,10 +1542,12 @@ int process_cleanup_before_terminate(int pid)
 		return 0;
 	}
 
+#ifdef ENABLE_TOFU
 	/* Clean up PDE file descriptors */
 	for (fd = 2; fd < MAX_FD_PDE; ++fd) {
 		__process_cleanup_fd(proc, fd);
 	}
+#endif
 
 	process_unlock(proc, &lock);
 	return 0;
@@ -3978,7 +3986,9 @@ SYSCALL_DECLARE(open)
 		goto out;
 	}
 
+#ifdef ENABLE_TOFU
 	cpu_local_var(current)->fd_path_in_open = pathname;
+#endif
 
 	dkprintf("open(): pathname=%s\n", pathname);
 	if (!strncmp(pathname, XPMEM_DEV_PATH, len)) {
@@ -3987,15 +3997,21 @@ SYSCALL_DECLARE(open)
 		rc = syscall_generic_forwarding(__NR_open, ctx);
 	}
 
+#ifdef ENABLE_TOFU
 	cpu_local_var(current)->fd_path_in_open = NULL;
+#endif
 
  out:
+#ifdef ENABLE_TOFU
 	if (rc > 0 && rc < MAX_FD_PDE) {
 		cpu_local_var(current)->proc->fd_path[rc] = pathname;
 	}
 	else {
 		kfree(pathname);
 	}
+#else
+	kfree(pathname);
+#endif
 	return rc;
 }
 
@@ -4023,7 +4039,9 @@ SYSCALL_DECLARE(openat)
 		goto out;
 	}
 
+#ifdef ENABLE_TOFU
 	cpu_local_var(current)->fd_path_in_open = pathname;
+#endif
 
 	dkprintf("openat(): pathname=%s\n", pathname);
 	if (!strncmp(pathname, XPMEM_DEV_PATH, len)) {
@@ -4032,15 +4050,21 @@ SYSCALL_DECLARE(openat)
 		rc = syscall_generic_forwarding(__NR_openat, ctx);
 	}
 
+#ifdef ENABLE_TOFU
 	cpu_local_var(current)->fd_path_in_open = NULL;
+#endif
 
 out:
+#ifdef ENABLE_TOFU
 	if (rc > 0 && rc < MAX_FD_PDE) {
 		cpu_local_var(current)->proc->fd_path[rc] = pathname;
 	}
 	else {
 		kfree(pathname);
 	}
+#else
+	kfree(pathname);
+#endif
 	return rc;
 }
 

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)
-{
-	const uint32_t *p = (const uint32_t *)addr;
-	return ((1UL << (nr & 31)) & (p[nr >> 5])) != 0;
-}
+#include <bitops-test_bit.h>
 
 extern unsigned int __sw_hweight32(unsigned int w);
 extern unsigned int __sw_hweight16(unsigned int w);

lib/include/ihk/mm.h

@@ -227,9 +227,11 @@ int ihk_mc_get_memory_chunk(int id,
 	unsigned long *start,
 	unsigned long *end,
 	int *numa_id);
+#ifdef ENABLE_TOFU
 int ihk_mc_get_memory_chunk_dma_addr(int id,
 		int tni, int cqid,
 		uintptr_t *dma_addr);
+#endif
 
 void remote_flush_tlb_cpumask(struct process_vm *vm, 
 		unsigned long addr, int cpu_id);

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;
+	}
+    }
+}