prctl: Add support for PR_SET_THP_DISABLE and PR_GET_THP_DISABLE

Change-Id: I04c5568a9eb78bcac632b734f34bba49cf602c4d
Refs: #1181

arch/arm64/kernel/include/prctl.h

@@ -2,6 +2,9 @@
 #ifndef __HEADER_ARM64_COMMON_PRCTL_H
 #define __HEADER_ARM64_COMMON_PRCTL_H
 
+#define PR_SET_THP_DISABLE 41
+#define PR_GET_THP_DISABLE 42
+
 /* arm64 Scalable Vector Extension controls */
 #define PR_SVE_SET_VL		48			/* set task vector length */
 #define PR_SVE_SET_VL_THREAD	(1 << 1)		/* set just this thread */

arch/arm64/kernel/syscall.c

@@ -161,7 +161,12 @@ fault:
 
 SYSCALL_DECLARE(prctl)
 {
+	struct process *proc = cpu_local_var(current)->proc;
 	int option = (int)ihk_mc_syscall_arg0(ctx);
+	unsigned long arg2 = (unsigned long)ihk_mc_syscall_arg1(ctx);
+	unsigned long arg3 = (unsigned long)ihk_mc_syscall_arg2(ctx);
+	unsigned long arg4 = (unsigned long)ihk_mc_syscall_arg3(ctx);
+	unsigned long arg5 = (unsigned long)ihk_mc_syscall_arg4(ctx);
 	long error;
 
 	switch (option) {
@@ -172,6 +177,19 @@ SYSCALL_DECLARE(prctl)
 	case PR_SVE_GET_VL:
 		error = SVE_GET_VL(cpu_local_var(current));
 		break;
+	case PR_SET_THP_DISABLE:
+		if (arg3 || arg4 || arg5) {
+			return -EINVAL;
+		}
+		proc->thp_disable = arg2;
+		error = 0;
+		break;
+	case PR_GET_THP_DISABLE:
+		if (arg2 || arg3 || arg4 || arg5) {
+			return -EINVAL;
+		}
+		error = proc->thp_disable;
+		break;
 	default:
 		error = syscall_generic_forwarding(__NR_prctl, ctx);
 		break;

arch/x86_64/kernel/include/prctl.h

@@ -9,6 +9,9 @@
 #ifndef __ARCH_PRCTL_H
 #define __ARCH_PRCTL_H
 
+#define PR_SET_THP_DISABLE 41
+#define PR_GET_THP_DISABLE 42
+
 #define ARCH_SET_GS 0x1001
 #define ARCH_SET_FS 0x1002
 #define ARCH_GET_FS 0x1003

arch/x86_64/kernel/include/syscall_list.h

@@ -109,6 +109,7 @@ SYSCALL_HANDLED(149, mlock)
 SYSCALL_HANDLED(150, munlock)
 SYSCALL_HANDLED(151, mlockall)
 SYSCALL_HANDLED(152, munlockall)
+SYSCALL_HANDLED(157, prctl)
 SYSCALL_HANDLED(158, arch_prctl)
 SYSCALL_HANDLED(160, setrlimit)
 SYSCALL_HANDLED(164, settimeofday)

arch/x86_64/kernel/syscall.c

@@ -169,6 +169,38 @@ fault:
 	return -EFAULT;
 }
 
+SYSCALL_DECLARE(prctl)
+{
+	struct process *proc = cpu_local_var(current)->proc;
+	int option = (int)ihk_mc_syscall_arg0(ctx);
+	unsigned long arg2 = (unsigned long)ihk_mc_syscall_arg1(ctx);
+	unsigned long arg3 = (unsigned long)ihk_mc_syscall_arg2(ctx);
+	unsigned long arg4 = (unsigned long)ihk_mc_syscall_arg3(ctx);
+	unsigned long arg5 = (unsigned long)ihk_mc_syscall_arg4(ctx);
+	int ret = 0;
+
+	switch (option) {
+	case PR_SET_THP_DISABLE:
+		if (arg3 || arg4 || arg5) {
+			return -EINVAL;
+		}
+		proc->thp_disable = arg2;
+		ret = 0;
+		break;
+	case PR_GET_THP_DISABLE:
+		if (arg2 || arg3 || arg4 || arg5) {
+			return -EINVAL;
+		}
+		ret = proc->thp_disable;
+		break;
+	default:
+		ret = syscall_generic_forwarding(__NR_prctl, ctx);
+		break;
+	}
+
+	return ret;
+}
+
 struct sigsp {
 	unsigned long flags;
 	void *link;

executer/include/uprotocol.h

@@ -142,6 +142,7 @@ struct program_load_desc {
 	unsigned long heap_extension;
 	long stack_premap;
 	unsigned long mpol_bind_mask;
+	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 nr_processes;

executer/user/mcexec.c

@@ -2060,6 +2060,21 @@ static void ld_preload_init()
 	}
 }
 
+static int get_thp_disable(void)
+{
+	int ret = 0;
+
+	ret = prctl(PR_GET_THP_DISABLE, 0, 0, 0, 0);
+
+	/* PR_GET_THP_DISABLE supported since Linux 3.15 */
+	if (ret < 0) {
+		/* if not supported, make THP enable */
+		ret = 0;
+	}
+
+	return ret;
+}
+
 int main(int argc, char **argv)
 {
 	int ret = 0;
@@ -2699,6 +2714,7 @@ int main(int argc, char **argv)
 
 	desc->uti_thread_rank = uti_thread_rank;
 	desc->uti_use_last_cpu = uti_use_last_cpu;
+	desc->thp_disable = get_thp_disable();
 
 	/* user_start and user_end are set by this call */
 	if (ioctl(fd, MCEXEC_UP_PREPARE_IMAGE, (unsigned long)desc) != 0) {

kernel/host.c

@@ -484,6 +484,7 @@ static int process_msg_prepare_process(unsigned long rphys)
 	proc->termsig = SIGCHLD;
 	proc->mpol_flags = pn->mpol_flags;
 	proc->mpol_threshold = pn->mpol_threshold;
+	proc->thp_disable = pn->thp_disable;
 	proc->nr_processes = pn->nr_processes;
 	proc->process_rank = pn->process_rank;
 	proc->heap_extension = pn->heap_extension;

kernel/include/process.h

@@ -556,6 +556,7 @@ struct process {
 	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;
 
 	// perf_event
 	int perf_status;

kernel/include/syscall.h

@@ -86,6 +86,10 @@
 
 #define SCD_MSG_FUTEX_WAKE              0x60
 
+/* For prctl() */
+#define PR_SET_THP_DISABLE 41
+#define PR_GET_THP_DISABLE 42
+
 /* Cloning flags.  */
 # define CSIGNAL       0x000000ff /* Signal mask to be sent at exit.  */
 # define CLONE_VM      0x00000100 /* Set if VM shared between processes.  */
@@ -201,6 +205,7 @@ struct program_load_desc {
 	unsigned long heap_extension;
 	long stack_premap;
 	unsigned long mpol_bind_mask;
+	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 nr_processes;

kernel/process.c

@@ -114,6 +114,7 @@ init_process(struct process *proc, struct process *parent)
 		proc->fsgid = parent->fsgid;
 		proc->mpol_flags = parent->mpol_flags;
 		proc->mpol_threshold = parent->mpol_threshold;
+		proc->thp_disable = parent->thp_disable;
 		memcpy(proc->rlimit, parent->rlimit,
 		       sizeof(struct rlimit) * MCK_RLIM_MAX);
 		memcpy(&proc->cpu_set, &parent->cpu_set,

kernel/syscall.c

@@ -1658,7 +1658,8 @@ do_mmap(const uintptr_t addr0, const size_t len0, const int prot,
 		pgshift = (flags >> MAP_HUGE_SHIFT) & 0x3F;
 		p2align = pgshift - PAGE_SHIFT;
 	}
-	else if ((flags & MAP_PRIVATE) && (flags & MAP_ANONYMOUS)) {
+	else if ((flags & MAP_PRIVATE) && (flags & MAP_ANONYMOUS)
+		    && !proc->thp_disable) {
 		pgshift = 0;		/* transparent huge page */
 		p2align = PAGE_P2ALIGN;
 
@@ -5198,7 +5199,10 @@ int do_shmget(const key_t key, const size_t size, const int shmflg)
 
 	if (shmflg & SHM_HUGETLB) {
 		pgshift = (shmflg >> SHM_HUGE_SHIFT) & 0x3F;
+	} else if (proc->thp_disable) {
+		pgshift = PAGE_SHIFT;
 	} else {
+		/* transparent huge page */
 		size_t pgsize;
 		int p2align;
 

test/issues/1181/C1181.sh

@@ -0,0 +1,199 @@
+#!/bin/sh
+
+USELTP=1
+USEOSTEST=0
+
+if [ $# -lt 1 ]; then
+	echo "usage: C1181.sh <arch>"
+	exit 1
+fi
+
+. ../../common.sh
+
+if [ -f "./$1_config" ]; then
+	. ./$1_config
+else
+	echo "$1 is unexpected arch"
+	exit 1
+fi
+
+tid=001
+echo "*** CT$tid start *******************************"
+echo "** case: THP_DISABLED"
+./set_thp_and_exec 1 ${MCEXEC} ./check_thp 1 | tee ./CT${tid}.txt
+echo "** case: THP_ENABLED"
+./set_thp_and_exec 0 ${MCEXEC} ./check_thp 0 | tee -a ./CT${tid}.txt
+ok=`grep "\[ OK \]" CT${tid}.txt | wc -l`
+ng=`grep "\[ NG \]" CT${tid}.txt | wc -l`
+if [ $ng = 0 ]; then
+	echo "*** CT$tid: PASSED (ok:$ok, ng:$ng)"
+else
+	echo "*** CT$tid: FAILED (ok:$ok, ng:$ng)"
+fi
+
+LARGE_PAGE_STR="large page allocation"
+echo ""
+tid=002
+echo "*** CT$tid start *******************************"
+echo "** case: THP_DISABLED"
+${IHKOSCTL} 0 clear_kmsg
+./set_thp_and_exec 1 ${MCEXEC} ./mmap_large
+${IHKOSCTL} 0 kmsg > ./CT${tid}_01.txt
+if grep "${LARGE_PAGE_STR}" ./CT${tid}_01.txt &> /dev/null ; then
+	echo "[ NG ] THP is Working" | tee ./CT${tid}.txt
+else
+	echo "[ OK ] THP is NOT Working" | tee ./CT${tid}.txt
+fi
+
+echo "** case: THP_ENABLED"
+${IHKOSCTL} 0 clear_kmsg
+./set_thp_and_exec 0 ${MCEXEC} ./mmap_large
+${IHKOSCTL} 0 kmsg > ./CT${tid}_02.txt
+if grep "${LARGE_PAGE_STR}" ./CT${tid}_02.txt &> /dev/null ; then
+	pgsize_allocated=`grep "large page allocation" ./CT${tid}_02.txt | tail -1 | grep -oE 'size: \w*' | sed 's/size: //'`
+	echo "pgsize_allocated: ${pgsize_allocated}"
+	if [ "$pgsize_allocated" == "${MMAP_LARGE}" ]; then
+		echo "[ OK ] THP is Working well" | tee -a ./CT${tid}.txt
+	else
+		echo "[ NG ] THP is Working, but pgsize is INVALID" | tee -a ./CT${tid}.txt
+	fi
+else
+	echo "[ NG ] when THP is ENABLED, NOT Working" | tee -a ./CT${tid}.txt
+fi
+
+ok=`grep "\[ OK \]" CT${tid}.txt | wc -l`
+ng=`grep "\[ NG \]" CT${tid}.txt | wc -l`
+if [ $ng = 0 ]; then
+	echo "*** CT$tid: PASSED (ok:$ok, ng:$ng)"
+else
+	echo "*** CT$tid: FAILED (ok:$ok, ng:$ng)"
+fi
+
+echo ""
+tid=003
+echo "*** CT$tid start *******************************"
+echo "** case: THP_DISABLED"
+${IHKOSCTL} 0 clear_kmsg
+./set_thp_and_exec 1 ${MCEXEC} ./shm_large
+${IHKOSCTL} 0 kmsg > ./CT${tid}_01.txt
+if grep "${LARGE_PAGE_STR}" ./CT${tid}_01.txt &> /dev/null ; then
+	echo "[ NG ] THP is Working" | tee ./CT${tid}.txt
+else
+	echo "[ OK ] THP is NOT Working" | tee ./CT${tid}.txt
+fi
+
+echo "** case: THP_ENABLED"
+${IHKOSCTL} 0 clear_kmsg
+./set_thp_and_exec 0 ${MCEXEC} ./shm_large
+${IHKOSCTL} 0 kmsg > ./CT${tid}_02.txt
+if grep "${LARGE_PAGE_STR}" ./CT${tid}_02.txt &> /dev/null ; then
+	pgsize_allocated=`grep "large page allocation" ./CT${tid}_02.txt | tail -1 | grep -oE 'size: \w*' | sed 's/size: //'`
+	echo "pgsize_allocated: ${pgsize_allocated}"
+	if [ "$pgsize_allocated" == "${SHM_LARGE}" ]; then
+		echo "[ OK ] THP is Working well" | tee -a ./CT${tid}.txt
+	else
+		echo "[ NG ] THP is Working, but pgsize is INVALID" | tee -a ./CT${tid}.txt
+	fi
+else
+	echo "[ NG ] large page is NOT Working" | tee -a ./CT${tid}.txt
+fi
+
+ok=`grep "\[ OK \]" CT${tid}.txt | wc -l`
+ng=`grep "\[ NG \]" CT${tid}.txt | wc -l`
+if [ $ng = 0 ]; then
+	echo "*** CT$tid: PASSED (ok:$ok, ng:$ng)"
+else
+	echo "*** CT$tid: FAILED (ok:$ok, ng:$ng)"
+fi
+
+# init for hugetlbfs test
+sudo mkdir -p /mnt/hugetlbfs-2M
+sudo mount -t hugetlbfs -o mode=777,pagesize=2m none /mnt/hugetlbfs-2M
+
+sudo mkdir -p /mnt/hugetlbfs-1G
+sudo mount -t hugetlbfs -o mode=777,pagesize=1g none /mnt/hugetlbfs-1G
+echo ""
+tid=004
+echo "*** CT$tid start *******************************"
+echo "** case: THP_DISABLED"
+${IHKOSCTL} 0 clear_kmsg
+./set_thp_and_exec 1 ${MCEXEC} ./mmap_hugetlbfs
+${IHKOSCTL} 0 kmsg > ./CT${tid}_01.txt
+if grep "${LARGE_PAGE_STR}" ./CT${tid}_01.txt &> /dev/null ; then
+	pgsize_allocated=`grep "large page allocation" ./CT${tid}_01.txt | head -1 | grep -oE 'size: \w*' | sed 's/size: //'`
+	echo "pgsize_allocated: ${pgsize_allocated}"
+	if [ "$pgsize_allocated" == "${SIZE_2M}" ]; then
+		echo "[ OK ] hugetlbfs-2M is Working well" | tee ./CT${tid}.txt
+	else
+		echo "[ NG ] pgsize is INVALID" | tee ./CT${tid}.txt
+	fi
+
+	pgsize_allocated=`grep "large page allocation" ./CT${tid}_01.txt | tail -1 | grep -oE 'size: \w*' | sed 's/size: //'`
+	echo "pgsize_allocated: ${pgsize_allocated}"
+	if [ "$pgsize_allocated" == "${SIZE_1G}" ]; then
+		echo "[ OK ] hugetlbfs-1G is Working well" | tee -a ./CT${tid}.txt
+	else
+		echo "[ NG ] pgsize is INVALID" | tee -a ./CT${tid}.txt
+	fi
+else
+	echo "[ NG ] hugetlbfs-1G is NOT Working" | tee ./CT${tid}.txt
+fi
+
+echo "** case: THP_ENABLED"
+${IHKOSCTL} 0 clear_kmsg
+./set_thp_and_exec 0 ${MCEXEC} ./mmap_hugetlbfs
+${IHKOSCTL} 0 kmsg > ./CT${tid}_02.txt
+if grep "${LARGE_PAGE_STR}" ./CT${tid}_02.txt &> /dev/null ; then
+	pgsize_allocated=`grep "large page allocation" ./CT${tid}_02.txt | head -1 | grep -oE 'size: \w*' | sed 's/size: //'`
+	echo "pgsize_allocated: ${pgsize_allocated}"
+	if [ "$pgsize_allocated" == "${SIZE_2M}" ]; then
+		echo "[ OK ] hugetlbfs-2M is Working well" | tee -a ./CT${tid}.txt
+	else
+		echo "[ NG ] pgsize is INVALID" | tee -a ./CT${tid}.txt
+	fi
+
+	pgsize_allocated=`grep "large page allocation" ./CT${tid}_02.txt | tail -1 | grep -oE 'size: \w*' | sed 's/size: //'`
+	echo "pgsize_allocated: ${pgsize_allocated}"
+	if [ "$pgsize_allocated" == "${SIZE_1G}" ]; then
+		echo "[ OK ] hugetlbfs-1G is Working well" | tee -a ./CT${tid}.txt
+	else
+		echo "[ NG ] pgsize is INVALID" | tee -a ./CT${tid}.txt
+	fi
+else
+	echo "[ NG ] hugetlbfs-1G is NOT Working" | tee -a ./CT${tid}.txt
+fi
+
+ok=`grep "\[ OK \]" CT${tid}.txt | wc -l`
+ng=`grep "\[ NG \]" CT${tid}.txt | wc -l`
+if [ $ng = 0 ]; then
+	echo "*** CT$tid: PASSED (ok:$ok, ng:$ng)"
+else
+	echo "*** CT$tid: FAILED (ok:$ok, ng:$ng)"
+fi
+# fini for hugetlbfs test
+sudo umount /mnt/hugetlbfs-2M
+sudo umount /mnt/hugetlbfs-1G
+
+echo ""
+tid=005
+echo "*** CT$tid start *******************************"
+sudo ${MCEXEC} ${LTPBIN}/prctl01 2>&1 | tee ./CT${tid}.txt
+ok=`grep TPASS CT${tid}.txt | wc -l`
+ng=`grep TFAIL CT${tid}.txt | wc -l`
+if [ $ng = 0 ]; then
+	echo "*** CT$tid: PASSED (ok:$ok, ng:$ng)"
+else
+	echo "*** CT$tid: FAILED (ok:$ok, ng:$ng)"
+fi
+
+echo ""
+tid=006
+echo "*** CT$tid start *******************************"
+sudo ${MCEXEC} ${LTPBIN}/prctl02 2>&1 | tee ./CT${tid}.txt
+ok=`grep TPASS CT${tid}.txt | wc -l`
+ng=`grep TFAIL CT${tid}.txt | wc -l`
+if [ $ng = 0 ]; then
+	echo "*** CT$tid: PASSED (ok:$ok, ng:$ng)"
+else
+	echo "*** CT$tid: FAILED (ok:$ok, ng:$ng)"
+fi

test/issues/1181/Makefile

@@ -0,0 +1,30 @@
+include $(HOME)/.mck_test_config.mk
+CC = gcc
+TARGET=set_thp_and_exec check_thp mmap_large shm_large mmap_hugetlbfs
+
+CPPFLAGS =
+LDFLAGS =
+
+all: $(TARGET)
+
+set_thp_and_exec: set_thp_and_exec.c
+	$(CC) -o $@ $^ $(LDFLAGS)
+
+check_thp: check_thp.c
+	$(CC) -o $@ $^ $(LDFLAGS)
+
+mmap_large: mmap_large.c
+	$(CC) -o $@ $^ $(LDFLAGS)
+
+shm_large: shm_large.c
+	$(CC) -o $@ $^ $(LDFLAGS)
+
+mmap_hugetlbfs: mmap_hugetlbfs.c
+	$(CC) -o $@ $^ $(LDFLAGS)
+
+test: all
+	@sh ./C1181.sh $(ARCH)
+
+clean:
+	rm -f $(TARGET) *.o CT*.txt
+

test/issues/1181/README

@@ -0,0 +1,36 @@
+【Issue#1181 動作確認】
+□前提
+本テストはprctlのPR_SET_THP_DISABLE, PR_GET_THP_DISABLE オプションを
+使用するため、Linux 3.15.0 以降の環境で実行する必要がある
+
+□ テスト内容
+1. prctlによるTHP制御が実現されていることの確認
+CT001:
+	mcexecのTHP設定がmckernelプロセスに反映されていることを確認
+
+CT002:
+	mckernelプロセスのmmapがTHP設定どおりの動作をすることを確認
+
+CT003:
+	mckernelプロセスのshmgetがTHP設定どおりの動作をすることを確認
+
+CT004:
+	hugetlbfsが利用できることを確認
+
+2. 既存のprctl機能に影響がないことをLTPを用いて確認
+CT005: ltp-prctl01
+CT006: ltp-prctl02
+
+□ 実行手順
+(1) cd <mckernel> && patch -p0 < <mckernel>/test/issues/1181/large_page.patch
+(2) McKernelをビルドする
+(3) cd <mckernel>/test/issues/1181/<arch> && make test
+
+McKernelのインストール先や、OSTEST, LTPの配置場所は、
+$HOME/.mck_test_config を参照している
+.mck_test_config は、McKernelをビルドした際に生成されるmck_test_config.sample ファイルを
+$HOMEにコピーし、適宜編集する
+
+□ 実行結果
+result.log 参照。
+すべての項目をPASSしていることを確認。

test/issues/1181/arm64_config

@@ -0,0 +1,6 @@
+SIZE_2M="2097152"
+SIZE_1G="1073741824"
+
+MMAP_LARGE=${SIZE_2M}
+SHM_LARGE=${SIZE_2M}
+SHM_HUGE=${SIZE_1G}

test/issues/1181/check_thp.c

@@ -0,0 +1,36 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <sys/prctl.h>
+
+int main(int argc, char *argv[])
+{
+	int rc = 0, ret = 0;
+	int expected_thp = 0;
+
+	if (argc < 2) {
+		printf("err: too few arguments\n");
+		return -1;
+	}
+
+	expected_thp = atoi(argv[1]);
+
+	rc = prctl(PR_GET_THP_DISABLE, 0, 0, 0, 0);
+	if (rc < 0) {
+		perror("err: PR_GET_THP_DISABLE");
+	}
+
+	if (rc == expected_thp) {
+		printf("[ OK ] get thp_disable: %d\n", rc);
+		ret = 0;
+	}
+	else {
+		printf("[ NG ] get thp_disable: %d (expected %d)\n",
+			rc, expected_thp);
+		ret = -1;
+		goto out;
+	}
+
+ out:
+	return ret;
+}

test/issues/1181/large_page.patch

@@ -0,0 +1,109 @@
+diff --git arch/arm64/kernel/memory.c arch/arm64/kernel/memory.c
+index 4cefc9f..d8fbb8c 100644
+--- arch/arm64/kernel/memory.c
++++ arch/arm64/kernel/memory.c
+@@ -2356,6 +2356,16 @@ int set_range_l1(void *args0, pte_t *ptep, uintptr_t base, uintptr_t start,
+ 	ptl1_set(ptep, pte);
+ 
+ 	error = 0;
++
++	if (args->attr[0] & PTE_CONT &&
++		__page_offset(base, PTL1_CONT_SIZE) == 0) {
++		if (base >= cpu_local_var(current)->vm->region.user_start &&
++			base < cpu_local_var(current)->vm->region.stack_start) {
++		kprintf("%s: large page allocation, addr: %016lx, size: %d, phys: %lx\n",
++			__func__, base, PTL1_CONT_SIZE, phys);
++		}
++	}
++
+ out:
+ 	dkprintf("set_range_l1(%lx,%lx,%lx): %d %lx\n",
+ 			base, start, end, error, *ptep);
+@@ -2436,6 +2446,23 @@ retry:
+ 				ptl_set(ptep, phys | args->attr[level-1],
+ 					level);
+ 				error = 0;
++
++				if (args->attr[level-1] & PTE_CONT) {
++					if (__page_offset(base, tbl.cont_pgsize) == 0) {
++						if (base >= cpu_local_var(current)->vm->region.user_start &&
++							base < cpu_local_var(current)->vm->region.stack_start) {
++						kprintf("%s: large page allocation, addr: %016lx, size: %d, phys: %lx\n",
++							__func__, base, tbl.cont_pgsize, phys);
++						}
++					}
++				} else {
++					if (base >= cpu_local_var(current)->vm->region.user_start &&
++						base < cpu_local_var(current)->vm->region.user_end) {
++					kprintf("%s: large page allocation, addr: %016lx, size: %d, phys: %lx\n",
++						__func__, base, tbl.pgsize, phys);
++					}
++				}
++
+ 				dkprintf("set_range_middle(%lx,%lx,%lx,%d):"
+ 					 "large page. %d %lx\n",
+ 					 base, start, end, level, error, *ptep);
+diff --git arch/x86_64/kernel/memory.c arch/x86_64/kernel/memory.c
+index cf7cac4..ff29cff 100644
+--- arch/x86_64/kernel/memory.c
++++ arch/x86_64/kernel/memory.c
+@@ -2027,6 +2027,13 @@ retry:
+ 			dkprintf("set_range_l2(%lx,%lx,%lx):"
+ 					"2MiB page. %d %lx\n",
+ 					base, start, end, error, *ptep);
++
++			if (base >= cpu_local_var(current)->vm->region.user_start &&
++				base < cpu_local_var(current)->vm->region.stack_start) {
++			kprintf("%s: large page allocation, addr: %016lx, size: %d\n",
++				__func__, base, PTL2_SIZE);
++			}
++
+ 			// Call memory_stat_rss_add() here because pgshift is resolved here
+ 			if (rusage_memory_stat_add(args->range, phys, PTL2_SIZE, PTL2_SIZE)) {
+ 				dkprintf("%lx+,%s: calling memory_stat_rss_add(),base=%lx,phys=%lx,size=%ld,pgsize=%ld\n", phys, __FUNCTION__, base, phys, PTL2_SIZE, PTL2_SIZE);
+@@ -2116,6 +2123,12 @@ retry:
+ 					"1GiB page. %d %lx\n",
+ 					base, start, end, error, *ptep);
+ 
++			if (base >= cpu_local_var(current)->vm->region.user_start &&
++				base < cpu_local_var(current)->vm->region.stack_start) {
++			kprintf("%s: large page allocation, addr: %016lx, size: %d\n",
++				__func__, base, PTL3_SIZE);
++			}
++
+ 			// Call memory_stat_rss_add() here because pgshift is resolved here
+ 			if (rusage_memory_stat_add(args->range, phys, PTL3_SIZE, PTL3_SIZE)) {
+ 				dkprintf("%lx+,%s: calling memory_stat_rss_add(),base=%lx,phys=%lx,size=%ld,pgsize=%ld\n", phys, __FUNCTION__, base, phys, PTL3_SIZE, PTL3_SIZE);
+diff --git kernel/include/rusage_private.h kernel/include/rusage_private.h
+index 7da7728..14253b2 100644
+--- kernel/include/rusage_private.h
++++ kernel/include/rusage_private.h
+@@ -12,7 +12,7 @@
+ #include <arch_rusage.h>
+ #include <debug.h>
+ 
+-#ifdef ENABLE_RUSAGE
++#if 0 /* def ENABLE_RUSAGE */
+ 
+ #define RUSAGE_OOM_MARGIN (2 * 1024 * 1024) // 2MB
+ 
+diff --git kernel/process.c kernel/process.c
+index bc0609b..cfa7d5e 100644
+--- kernel/process.c
++++ kernel/process.c
+@@ -2009,6 +2009,15 @@ retry:
+ 		}
+ 
+ 		dkprintf("%s: attr=%x\n", __FUNCTION__, attr);
++
++		if (pgsize > PAGE_SIZE) {
++			if ((unsigned long)pgaddr >= cpu_local_var(current)->vm->region.user_start &&
++				(unsigned long)pgaddr < cpu_local_var(current)->vm->region.stack_start) {
++			kprintf("large page allocation, addr: %016lx, size: %d, phys: %lx\n",
++				pgaddr, pgsize, phys);
++			}
++		}
++
+ 		error = ihk_mc_pt_set_pte(vm->address_space->page_table, ptep,
+ 		                          pgsize, phys, attr);
+ 		if (error) {

test/issues/1181/mmap_hugetlbfs.c

@@ -0,0 +1,72 @@
+#include <stdio.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#include <stdlib.h>
+#include <fcntl.h>
+#include <errno.h>
+#include <sys/mman.h>
+
+#define PAGE_SHIFT_2M 21
+#define PAGE_SIZE_2M (1UL << PAGE_SHIFT_2M)
+char fn_2M[] = "/mnt/hugetlbfs-2M/tmp";
+
+#define PAGE_SHIFT_1G 30
+#define PAGE_SIZE_1G (1UL << PAGE_SHIFT_1G)
+char fn_1G[] = "/mnt/hugetlbfs-1G/tmp";
+
+int trial_num;
+
+int mmap_hugetlbfs(char *fn, size_t page_size)
+{
+	int fd;
+	char *addr_mmap;
+
+	fd = open(fn, O_CREAT | O_RDWR, 0755);
+	if (fd == -1) {
+		printf("open failed, fn:%s\n");
+		goto fn_fail;
+	}
+
+	addr_mmap = mmap(0, page_size,
+			 PROT_READ | PROT_WRITE,
+			 MAP_SHARED,
+			 fd, 0);
+	if (addr_mmap == (void *)-1) {
+		printf("mmap failed\n");
+		goto fn_fail;
+	}
+	addr_mmap[0] = 'z';
+
+	printf("large page request, addr: %016lx, size: %ld\n",
+	       (unsigned long)addr_mmap, page_size);
+
+	munmap(addr_mmap, page_size);
+	close(fd);
+	unlink(fn);
+
+	return 0;
+ fn_fail:
+	return 1;
+}
+
+int main(int argc, char **argv)
+{
+	int ret;
+
+	ret = mmap_hugetlbfs(fn_2M, PAGE_SIZE_2M);
+	if (ret != 0) {
+		printf("mmap_hugetlbfs failed, fn: %s", fn_2M);
+		goto fn_fail;
+	}
+
+	ret = mmap_hugetlbfs(fn_1G, PAGE_SIZE_1G);
+	if (ret != 0) {
+		printf("mmap_hugetlbfs failed, fn: %s", fn_1G);
+		goto fn_fail;
+	}
+
+	return 0;
+ fn_fail:
+	return 1;
+}

test/issues/1181/mmap_large.c

@@ -0,0 +1,32 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <string.h>
+#include <sys/mman.h>
+
+#define MAP_SIZE (16 << 20)
+
+int main(int argc, char **argv)
+{
+	int ret = 0;
+	void *mem = NULL;
+
+	mem = mmap(NULL, MAP_SIZE, PROT_READ | PROT_WRITE,
+			MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+
+	if (!mem) {
+		printf("mmap failed\n");
+		ret = -1;
+		goto out;
+	}
+
+	printf("mmap to %p, size: %ldM\n", mem, MAP_SIZE / 1024 / 1024);
+
+	memset(mem, 0, MAP_SIZE);
+
+ out:
+	if (mem) {
+		munmap(mem, MAP_SIZE);
+	}
+	return 0;
+}

test/issues/1181/result.log

@@ -0,0 +1,67 @@
+mcstop+release.sh ... done
+mcreboot.sh -c 1-7,9-15,17-23,25-31 -m 10G@0,10G@1 -r 1-7:0+9-15:8+17-23:16+25-31:24 ... done
+*** CT001 start *******************************
+** case: THP_DISABLED
+[ OK ] get thp_disable: 1
+** case: THP_ENABLED
+[ OK ] get thp_disable: 0
+*** CT001: PASSED (ok:2, ng:0)
+
+*** CT002 start *******************************
+** case: THP_DISABLED
+set thp_disable: 1
+exec: /home/satoken/ppos/bin/mcexec ./mmap_large
+mmap to 0x2aaaab00d000, size: 16M
+[ OK ] THP is NOT Working
+** case: THP_ENABLED
+set thp_disable: 0
+exec: /home/satoken/ppos/bin/mcexec ./mmap_large
+mmap to 0x2aaaab200000, size: 16M
+pgsize_allocated: 2097152
+[ OK ] THP is Working well
+*** CT002: PASSED (ok:2, ng:0)
+
+*** CT003 start *******************************
+** case: THP_DISABLED
+set thp_disable: 1
+exec: /home/satoken/ppos/bin/mcexec ./shm_large
+shmat to 0x2aaaab00d000, size: 16M
+[ OK ] THP is NOT Working
+** case: THP_ENABLED
+set thp_disable: 0
+exec: /home/satoken/ppos/bin/mcexec ./shm_large
+shmat to 0x2aaaab200000, size: 16M
+pgsize_allocated: 2097152
+[ OK ] THP is Working well
+*** CT003: PASSED (ok:2, ng:0)
+
+*** CT004 start *******************************
+** case: THP_DISABLED
+set thp_disable: 1
+exec: /home/satoken/ppos/bin/mcexec ./mmap_hugetlbfs
+large page request, addr: 00002aaaab200000, size: 2097152
+large page request, addr: 00002aaac0000000, size: 1073741824
+pgsize_allocated: 2097152
+[ OK ] hugetlbfs-2M is Working well
+pgsize_allocated: 1073741824
+[ OK ] hugetlbfs-1G is Working well
+** case: THP_ENABLED
+set thp_disable: 0
+exec: /home/satoken/ppos/bin/mcexec ./mmap_hugetlbfs
+large page request, addr: 00002aaaab200000, size: 2097152
+large page request, addr: 00002aaac0000000, size: 1073741824
+pgsize_allocated: 2097152
+[ OK ] hugetlbfs-2M is Working well
+pgsize_allocated: 1073741824
+[ OK ] hugetlbfs-1G is Working well
+*** CT004: PASSED (ok:4, ng:0)
+
+*** CT005 start *******************************
+prctl01     1  TPASS  :  Test Passed
+prctl01     2  TPASS  :  Test Passed
+*** CT005: PASSED (ok:2, ng:0)
+
+*** CT006 start *******************************
+prctl02     1  TPASS  :  Test Passed
+prctl02     2  TPASS  :  Test Passed
+*** CT006: PASSED (ok:2, ng:0)

test/issues/1181/set_thp_and_exec.c

@@ -0,0 +1,56 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <string.h>
+#include <sys/prctl.h>
+
+#define EXARG_MAX 64
+#define CMD_MAX_LEN 1024
+
+int main(int argc, char *argv[])
+{
+	int rc = 0, i;
+	int thp_disable = 0;
+	char *exargv[EXARG_MAX] = {};
+	char *exenvp[1] = {NULL};
+	char execcmd[CMD_MAX_LEN] = {};
+
+	if (argc < 3) {
+		printf("err: too few arguments\n");
+		return -1;
+	}
+
+	if (argc > EXARG_MAX + 1) {
+		printf("err: too many arguments\n");
+		return -1;
+	}
+
+	thp_disable = atoi(argv[1]);
+
+	rc = prctl(PR_SET_THP_DISABLE, thp_disable, 0, 0, 0);
+	if (rc < 0) {
+		perror("err: PR_SET_THP_DISABLE");
+	}
+	printf("set thp_disable: %d\n", thp_disable);
+
+	for (i = 1; i < argc; i++) {
+		exargv[i - 2] = argv[i];
+	}
+
+	for (i = 0; i < EXARG_MAX; i++) {
+		if (!exargv[i]) {
+			break;
+		}
+		if (i != 0) {
+			strncat(execcmd, " ", CMD_MAX_LEN - 2);
+		}
+		strncat(execcmd, exargv[i], CMD_MAX_LEN - strlen(execcmd) - 1);
+	}
+
+	printf("exec: %s\n", execcmd);
+	execve(exargv[0], exargv, exenvp);
+
+	/* can't reach here */
+	printf("err: execve failed\n");
+	return -1;
+}

test/issues/1181/shm_large.c

@@ -0,0 +1,37 @@
+#include <stdio.h>
+#include <unistd.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <sys/ipc.h>
+#include <sys/shm.h>
+
+#define SHM_SIZE (16 << 20)
+
+int main(int argc, char **argv)
+{
+	void *shm = NULL;
+	key_t key = ftok(argv[0], 0);
+	int shmid;
+	struct shmid_ds buf;
+
+	shmid = shmget(key, SHM_SIZE, IPC_CREAT | 0660);
+	if (shmid < 0) {
+		perror("shmget: ");
+		return -1;
+	}
+
+	shm = shmat(shmid, NULL, 0);
+	if (!shm) {
+		perror("shmat: ");
+		return -1;
+	}
+
+	printf("shmat to %p, size: %ldM\n", shm, SHM_SIZE / 1024 / 1024);
+	memset(shm, 0, SHM_SIZE);
+
+	shmdt(shm);
+
+	shmctl(shmid, IPC_RMID, &buf);
+
+	return 0;
+}

test/issues/1181/x86_64_config

@@ -0,0 +1,6 @@
+SIZE_2M="2097152"
+SIZE_1G="1073741824"
+
+MMAP_LARGE=${SIZE_2M}
+SHM_LARGE=${SIZE_2M}
+SHM_HUGE=${SIZE_1G}