mcexec: --mpol-threshold to control MPOL_BIND/MPOL_PREFERRED

executer/include/uprotocol.h

@@ -125,6 +125,7 @@ struct program_load_desc {
 	struct rlimit rlimit[MCK_RLIM_MAX];
 	unsigned long interp_align;
 	unsigned long mpol_flags;
+	unsigned long mpol_threshold;
 	char shell_path[SHELL_PATH_MAX_LEN];
 	__cpu_set_unit cpu_set[PLD_CPU_SET_SIZE];
 	struct program_image_section sections[0];

executer/user/mcexec.c

@@ -157,6 +157,7 @@ static int enable_vdso = 1;
 static int mpol_no_heap = 0;
 static int mpol_no_stack = 0;
 static int mpol_no_bss = 0;
+static unsigned long mpol_threshold = (1024*1024);
 
 /* Partitioned execution (e.g., for MPI) */
 static int nr_processes = 0;
@@ -1111,7 +1112,7 @@ static int reduce_stack(struct rlimit *orig_rlim, char *argv[])
 
 void print_usage(char **argv)
 {
-	fprintf(stderr, "Usage: %s [-c target_core] [-n nr_partitions] [<mcos-id>] (program) [args...]\n", argv[0]);
+	fprintf(stderr, "usage: %s [-c target_core] [-n nr_partitions] [--mpol-threshold=N] [--mpol-no-heap] [--mpol-no-bss] [--mpol-no-stack] [<mcos-id>] (program) [args...]\n", argv[0]);
 }
 
 void init_sigaction(void)
@@ -1300,6 +1301,12 @@ static struct option mcexec_options[] = {
 		.flag =		&mpol_no_bss,
 		.val =		1,
 	},
+	{
+		.name =		"mpol-threshold",
+		.has_arg =	required_argument,
+		.flag =		NULL,
+		.val =		'm',
+	},
 	/* end */
 	{ NULL, 0, NULL, 0, },
 };
@@ -1355,7 +1362,7 @@ int main(int argc, char **argv)
 	}
            
 	/* Parse options ("+" denotes stop at the first non-option) */
-	while ((opt = getopt_long(argc, argv, "+c:n:t:", mcexec_options, NULL)) != -1) {
+	while ((opt = getopt_long(argc, argv, "+c:n:t:m:", mcexec_options, NULL)) != -1) {
 		switch (opt) {
 			case 'c':
 				target_core = atoi(optarg);
@@ -1369,6 +1376,10 @@ int main(int argc, char **argv)
 				nr_threads = atoi(optarg);
 				break;
 
+			case 'm':
+				mpol_threshold = atol(optarg);
+				break;
+
 			case 0:	/* long opt */
 				break;
 
@@ -1702,6 +1713,8 @@ int main(int argc, char **argv)
 		desc->mpol_flags |= MPOL_NO_BSS;
 	}
 
+	desc->mpol_threshold = mpol_threshold;
+
 	if (ioctl(fd, MCEXEC_UP_PREPARE_IMAGE, (unsigned long)desc) != 0) {
 		perror("prepare");
 		close(fd);

kernel/host.c

@@ -129,7 +129,7 @@ int prepare_process_ranges_args_envs(struct thread *thread,
 
 		/* Non-TEXT sections that are large respect user allocation policy
 		 * unless user explicitly requests otherwise */
-		if (i >= 1 && pn->sections[i].len >= AP_USER_THRESHOLD &&
+		if (i >= 1 && pn->sections[i].len >= pn->mpol_threshold &&
 				!(pn->mpol_flags & MPOL_NO_BSS)) {
 			dkprintf("%s: section: %d size: %d pages -> IHK_MC_AP_USER\n",
 					__FUNCTION__, i, range_npages);
@@ -429,6 +429,7 @@ static int process_msg_prepare_process(unsigned long rphys)
 	proc->fsgid = pn->cred[7];
 	proc->termsig = SIGCHLD;
 	proc->mpol_flags = pn->mpol_flags;
+	proc->mpol_threshold = pn->mpol_threshold;
 
 	vm->region.user_start = pn->user_start;
 	vm->region.user_end = pn->user_end;

kernel/include/process.h

@@ -555,6 +555,7 @@ struct process {
 	long maxrss_children;
 	/* Memory policy flags */
 	unsigned long mpol_flags;
+	size_t mpol_threshold;
 
 	// perf_event
 	int perf_status;

kernel/include/syscall.h

@@ -193,6 +193,7 @@ struct program_load_desc {
 	struct rlimit rlimit[MCK_RLIM_MAX];
 	unsigned long interp_align;
 	unsigned long mpol_flags;
+	unsigned long mpol_threshold;
 	char shell_path[SHELL_PATH_MAX_LEN];
 	__cpu_set_unit cpu_set[PLD_CPU_SET_SIZE];
 	struct program_image_section sections[0];

kernel/process.c

@@ -93,6 +93,7 @@ init_process(struct process *proc, struct process *parent)
 		proc->sgid = parent->sgid;
 		proc->fsgid = parent->fsgid;
 		proc->mpol_flags = parent->mpol_flags;
+		proc->mpol_threshold = parent->mpol_threshold;
 		memcpy(proc->rlimit, parent->rlimit,
 		       sizeof(struct rlimit) * MCK_RLIM_MAX);
 	}
@@ -1928,7 +1929,7 @@ int init_process_stack(struct thread *thread, struct program_load_desc *pn,
 
 	/* Apply user allocation policy to stacks */
 	/* TODO: make threshold kernel or mcexec argument */
-	ap_flag = (size >= AP_USER_THRESHOLD &&
+	ap_flag = (size >= proc->mpol_threshold &&
 		!(proc->mpol_flags & MPOL_NO_STACK)) ? IHK_MC_AP_USER : 0;
 	dkprintf("%s: size: %lu %s\n", __FUNCTION__, size,
 			ap_flag ? "(IHK_MC_AP_USER)" : "");

kernel/syscall.c

@@ -1430,7 +1430,7 @@ do_mmap(const intptr_t addr0, const size_t len0, const int prot,
 		/* Small allocations mostly benefit from closest RAM,
 		 * otherwise follow user requested policy */
 		unsigned long ap_flag =
-			(!(flags & MAP_STACK) && len >= AP_USER_THRESHOLD) ||
+			(!(flags & MAP_STACK) && len >= thread->proc->mpol_threshold) ||
 			((flags & MAP_STACK) && !(thread->proc->mpol_flags & MPOL_NO_STACK)) ?
 			IHK_MC_AP_USER : 0;
 

lib/include/ihk/mm.h

@@ -54,8 +54,6 @@ typedef unsigned long ihk_mc_ap_flag;
 #define IHK_MC_AP_BANDWIDTH       0x010000
 #define IHK_MC_AP_LATENCY         0x020000
 
-#define AP_USER_THRESHOLD         (2097152)
-
 enum ihk_mc_pt_prepare_flag {
 	IHK_MC_PT_FIRST_LEVEL,
 	IHK_MC_PT_LAST_LEVEL,