mcexec/mm: user memory policy control for heap, stack, etc.

executer/include/uprotocol.h

@@ -92,6 +92,10 @@ struct get_cpu_set_arg {
 typedef unsigned long __cpu_set_unit;
 #define PLD_CPU_SET_SIZE (PLD_CPU_SET_MAX_CPUS / (8 * sizeof(__cpu_set_unit)))
 
+#define MPOL_NO_HEAP              0x01
+#define MPOL_NO_STACK             0x02
+#define MPOL_NO_BSS               0x04
+
 struct program_load_desc {
 	int num_sections;
 	int status;
@@ -120,6 +124,7 @@ struct program_load_desc {
 	unsigned long envs_len;
 	struct rlimit rlimit[MCK_RLIM_MAX];
 	unsigned long interp_align;
+	unsigned long mpol_flags;
 	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

@@ -154,6 +154,9 @@ static char *altroot;
 static const char rlimit_stack_envname[] = "MCKERNEL_RLIMIT_STACK";
 static int ischild;
 static int enable_vdso = 1;
+static int mpol_no_heap = 0;
+static int mpol_no_stack = 0;
+static int mpol_no_bss = 0;
 
 /* Partitioned execution (e.g., for MPI) */
 static int nr_processes = 0;
@@ -1279,7 +1282,24 @@ static struct option mcexec_options[] = {
 		.flag =		&enable_vdso,
 		.val =		1,
 	},
-
+	{
+		.name =		"mpol-no-heap",
+		.has_arg =	no_argument,
+		.flag =		&mpol_no_heap,
+		.val =		1,
+	},
+	{
+		.name =		"mpol-no-stack",
+		.has_arg =	no_argument,
+		.flag =		&mpol_no_stack,
+		.val =		1,
+	},
+	{
+		.name =		"mpol-no-bss",
+		.has_arg =	no_argument,
+		.flag =		&mpol_no_bss,
+		.val =		1,
+	},
 	/* end */
 	{ NULL, 0, NULL, 0, },
 };
@@ -1669,6 +1689,19 @@ int main(int argc, char **argv)
 #endif
 	}
 
+	desc->mpol_flags = 0;
+	if (mpol_no_heap) {
+		desc->mpol_flags |= MPOL_NO_HEAP;
+	}
+
+	if (mpol_no_stack) {
+		desc->mpol_flags |= MPOL_NO_STACK;
+	}
+
+	if (mpol_no_bss) {
+		desc->mpol_flags |= MPOL_NO_BSS;
+	}
+
 	if (ioctl(fd, MCEXEC_UP_PREPARE_IMAGE, (unsigned long)desc) != 0) {
 		perror("prepare");
 		close(fd);

kernel/host.c

@@ -127,8 +127,10 @@ int prepare_process_ranges_args_envs(struct thread *thread,
 		flags |= VRFLAG_PROT_TO_MAXPROT(flags);
 		flags |= VR_DEMAND_PAGING;
 
-		/* Non-TEXT sections that are large respect user allocation policy */
+		/* Non-TEXT sections that are large respect user allocation policy
-		if (i >= 1 && pn->sections[i].len >= AP_USER_THRESHOLD) {
+		 * unless user explicitly requests otherwise */
+		if (i >= 1 && pn->sections[i].len >= AP_USER_THRESHOLD &&
+				!(pn->mpol_flags & MPOL_NO_BSS)) {
 			dkprintf("%s: section: %d size: %d pages -> IHK_MC_AP_USER\n",
 					__FUNCTION__, i, range_npages);
 			ap_flags = IHK_MC_AP_USER;
@@ -426,6 +428,7 @@ static int process_msg_prepare_process(unsigned long rphys)
 	proc->sgid = pn->cred[6];
 	proc->fsgid = pn->cred[7];
 	proc->termsig = SIGCHLD;
+	proc->mpol_flags = pn->mpol_flags;
 
 	vm->region.user_start = pn->user_start;
 	vm->region.user_end = pn->user_end;

kernel/include/process.h

@@ -553,6 +553,8 @@ struct process {
 
 	long maxrss;
 	long maxrss_children;
+	/* Memory policy flags */
+	unsigned long mpol_flags;
 
 	// perf_event
 	int perf_status;

kernel/include/syscall.h

@@ -160,6 +160,10 @@ struct program_image_section {
 typedef unsigned long __cpu_set_unit;
 #define PLD_CPU_SET_SIZE (PLD_CPU_SET_MAX_CPUS / (8 * sizeof(__cpu_set_unit)))
 
+#define MPOL_NO_HEAP              0x01
+#define MPOL_NO_STACK             0x02
+#define MPOL_NO_BSS               0x04
+
 struct program_load_desc {
 	int num_sections;
 	int status;
@@ -188,6 +192,7 @@ struct program_load_desc {
 	unsigned long envs_len;
 	struct rlimit rlimit[MCK_RLIM_MAX];
 	unsigned long interp_align;
+	unsigned long mpol_flags;
 	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

@@ -92,6 +92,7 @@ init_process(struct process *proc, struct process *parent)
 		proc->egid = parent->egid;
 		proc->sgid = parent->sgid;
 		proc->fsgid = parent->fsgid;
+		proc->mpol_flags = parent->mpol_flags;
 		memcpy(proc->rlimit, parent->rlimit,
 		       sizeof(struct rlimit) * MCK_RLIM_MAX);
 	}
@@ -1927,7 +1928,8 @@ 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) ? IHK_MC_AP_USER : 0;
+	ap_flag = (size >= AP_USER_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)" : "");
 
@@ -1943,7 +1945,7 @@ int init_process_stack(struct thread *thread, struct program_load_desc *pn,
 	memset(stack, 0, minsz);
 
 	vrflag = VR_STACK | VR_DEMAND_PAGING;
-	vrflag |= (ap_flag ? VR_AP_USER : 0);
+	vrflag |= ((ap_flag & IHK_MC_AP_USER) ? VR_AP_USER : 0);
 	vrflag |= PROT_TO_VR_FLAG(pn->stack_prot);
 	vrflag |= VR_MAXPROT_READ | VR_MAXPROT_WRITE | VR_MAXPROT_EXEC;
 #define	NOPHYS	((uintptr_t)-1)
@@ -2070,7 +2072,8 @@ unsigned long extend_process_region(struct process_vm *vm,
 	}
 	else {
 		p = ihk_mc_alloc_aligned_pages((aligned_new_end - aligned_end) >> PAGE_SHIFT,
-				LARGE_PAGE_P2ALIGN, IHK_MC_AP_NOWAIT | IHK_MC_AP_USER);
+				LARGE_PAGE_P2ALIGN, IHK_MC_AP_NOWAIT |
+				(!(vm->proc->mpol_flags & MPOL_NO_HEAP) ? IHK_MC_AP_USER : 0));
 
 		if (!p) {
 			return end;

kernel/syscall.c

@@ -1430,7 +1430,8 @@ 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 =
-			(len >= AP_USER_THRESHOLD || flags & MAP_STACK) ?
+			(!(flags & MAP_STACK) && len >= AP_USER_THRESHOLD) ||
+			((flags & MAP_STACK) && !(thread->proc->mpol_flags & MPOL_NO_STACK)) ?
 			IHK_MC_AP_USER : 0;
 
 		if (ap_flag) {