madvise: Support MADV_WIPEONFORK, MADV_KEEPONFORK and MADV_NORMAL

Change-Id: I1d4cf5affa580d7304dfdc34fa4f1707c0df617c
refs: #1374

kernel/include/mman.h

@@ -54,6 +54,8 @@
 #define	MADV_NOHUGEPAGE		15
 #define	MADV_DONTDUMP		16
 #define	MADV_DODUMP		17
+#define	MADV_WIPEONFORK		18 /* Zero memory on fork, child only */
+#define	MADV_KEEPONFORK		19 /* Undo MADV_WIPEONFORK */
 #define	MADV_HWPOISON		100
 #define	MADV_SOFT_OFFLINE	101
 

kernel/include/process.h

@@ -56,6 +56,7 @@
 #define VR_PAGEOUT	   0x10000000
 #define VR_DONTDUMP	   0x20000000
 #define VR_XPMEM	   0x40000000
+#define VR_WIPEONFORK	   0x80000000
 
 #define	PROT_TO_VR_FLAG(prot)	(((unsigned long)(prot) << 16) & VR_PROT_MASK)
 #define	VRFLAG_PROT_TO_MAXPROT(vrflag)	(((vrflag) & VR_PROT_MASK) << 4)

kernel/process.c

@@ -750,8 +750,13 @@ static int copy_user_pte(void *arg0, page_table_t src_pt, pte_t *src_ptep, void
 			src_kvirt = ihk_mc_map_virtual(src_lphys, 1, attr);
 		}
 
-		memcpy(virt, src_kvirt, pgsize);
-		dkprintf("copy_user_pte(): memcpy OK\n");
+		if (args->new_vrflag & VR_WIPEONFORK) {
+			memset(virt, 0, pgsize);
+			dkprintf("%s(): memset OK\n", __func__);
+		} else {
+			memcpy(virt, src_kvirt, pgsize);
+			dkprintf("%s(): memcpy OK\n", __func__);
+		}
 
 		if (!is_mckernel) {
 			ihk_mc_unmap_virtual(src_kvirt, 1);

kernel/syscall.c

@@ -5324,6 +5324,8 @@ SYSCALL_DECLARE(madvise)
 	case MADV_REMOVE:
 	case MADV_DONTDUMP:
 	case MADV_DODUMP:
+	case MADV_WIPEONFORK:
+	case MADV_KEEPONFORK:
 		break;
 
 	case MADV_HWPOISON:
@@ -5376,6 +5378,27 @@ SYSCALL_DECLARE(madvise)
 			}
 		}
 		else if(advice == MADV_DONTFORK || advice == MADV_DOFORK);
+		else if (advice == MADV_NORMAL) {
+			/*
+			 * Normally, the settings of MADV_RANDOM and
+			 * MADV_SEQUENTIAL are cleared.
+			 * MADV_RANDOM and MADV_SEQUENTIAL are not supported,
+			 * so do nothing.
+			 */
+		}
+		else if (advice == MADV_WIPEONFORK
+			 || advice == MADV_KEEPONFORK) {
+			if (range->memobj && memobj_has_pager(range->memobj)) {
+				/* device mapping, file mapping */
+				error = -EINVAL;
+				goto out;
+			}
+			if (!(range->flag & VR_PRIVATE)) {
+				/* VR_SHARED */
+				error = -EINVAL;
+				goto out;
+			}
+		}
 		else if (!range->memobj || !memobj_has_pager(range->memobj)) {
 			dkprintf("[%d]sys_madvise(%lx,%lx,%x):has not pager"
 					"[%lx-%lx) %lx\n",
@@ -5463,6 +5486,24 @@ SYSCALL_DECLARE(madvise)
 	   advice == MADV_DOFORK){
 		error = syscall_generic_forwarding(__NR_madvise, ctx);
 	}
+	if (advice == MADV_WIPEONFORK) {
+		error = change_attr_process_memory_range(
+				thread->vm, start, end,
+				set_memory_range_flag,
+				VR_WIPEONFORK);
+		if (error) {
+			goto out;
+		}
+	}
+	if (advice == MADV_KEEPONFORK) {
+		error = change_attr_process_memory_range(
+				thread->vm, start, end,
+				clear_memory_range_flag,
+				VR_WIPEONFORK);
+		if (error) {
+			goto out;
+		}
+	}
 
 	error = 0;
 out:

test/issues/1374/README

@@ -0,0 +1,20 @@
+/* REDME COPYRIGHT FUJITSU LIMITED 2020 */
+
+ltp: madvise10: MADV_WIPEONFORKが期待通りに動作しない
+https://postpeta.pccluster.org/redmine/issues/1374
+
+テストセットREADME
+
+(0) 事前準備
+  1. $HOME/.mck_test_configを用意する
+
+(1) テスト実行方法
+  1. ./run.sh
+
+(2) テスト項目詳細
+
+  LTPの下記項目を実行し期待通りに動作することを確認する
+  "test OK." が書き出されれば合格判定となる。具体例はresult.logを参照。
+  - madvise10
+
+以上。

test/issues/1374/result.log

@@ -0,0 +1,29 @@
+[root@hostname 1374]# ./run.sh
+mcstop+release.sh ... done
+mcreboot.sh -c 12-59 -m 512M@4 ... done
+test run.
+tst_test.c:1096: INFO: Timeout per run is 0h 05m 00s
+madvise10.c:143: INFO: MADV_WIPEONFORK zeroes memory in child
+madvise10.c:117: PASS: madvise(0x1000001c0000, 16384, 0x0)
+madvise10.c:117: PASS: madvise(0x1000001c0000, 16384, 0x12)
+madvise10.c:98: PASS: In PID 13482, Matched expected pattern
+madvise10.c:143: INFO: MADV_WIPEONFORK with zero length does nothing
+madvise10.c:117: PASS: madvise(0x1000001d0000, 0, 0x0)
+madvise10.c:117: PASS: madvise(0x1000001d0000, 0, 0x12)
+madvise10.c:98: PASS: In PID 13532, Matched expected pattern
+madvise10.c:143: INFO: MADV_WIPEONFORK zeroes memory in grand-child
+madvise10.c:117: PASS: madvise(0x1000001e0000, 16384, 0x0)
+madvise10.c:117: PASS: madvise(0x1000001e0000, 16384, 0x12)
+madvise10.c:98: PASS: In PID 13635, Matched expected pattern
+madvise10.c:143: INFO: MADV_KEEPONFORK will undo MADV_WIPEONFORK
+madvise10.c:117: PASS: madvise(0x1000001f0000, 16384, 0x12)
+madvise10.c:117: PASS: madvise(0x1000001f0000, 16384, 0x13)
+madvise10.c:98: PASS: In PID 13685, Matched expected pattern
+
+Summary:
+passed   12
+failed   0
+skipped  0
+warnings 0
+test OK.
+mcstop+release.sh ... done

test/issues/1374/run.sh

@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+# run.sh COPYRIGHT FUJITSU LIMITED 2020
+
+# load setting and booting mck.
+USELTP=1
+. ../../common.sh
+chmod 777 /dev/mcos*
+
+echo "test run."
+ltp=`cat <<__EOL__
+madvise10
+__EOL__`
+
+FAIL=0
+for tp in $ltp
+do
+	pushd ${LTPBIN} > /dev/null
+	${MCEXEC} ${LTPBIN}/${tp}
+	if [ $? -ne 0 ]; then
+		FAIL=1
+	fi
+	popd > /dev/null
+done
+
+if [ ${FAIL} -eq 1 ]; then
+	echo "test NG."
+else
+	echo "test OK."
+fi
+
+mcstop