mcexec/mcctrl: unmount cgroups (privately) which expose invalid NUMA info

configure

@@ -3117,6 +3117,31 @@ _ACEOF
   fi
 
 
+  { $as_echo "$as_me:${as_lineno-$LINENO}: checking System.map for symbol sys_umount" >&5
+$as_echo_n "checking System.map for symbol sys_umount... " >&6; }
+  mcctrl_addr=`eval $MCCTRL_LINUX_SYMTAB_CMD | grep " sys_umount\$" | cut -d\  -f1`
+  if test -z $mcctrl_addr; then
+    { $as_echo "$as_me:${as_lineno-$LINENO}: result: not found" >&5
+$as_echo "not found" >&6; }
+  else
+    mcctrl_result=$mcctrl_addr
+    mcctrl_addr="0x$mcctrl_addr"
+
+      if `eval $MCCTRL_LINUX_SYMTAB_CMD | grep " __ksymtab_sys_umount\$" >/dev/null`; then
+        mcctrl_result="exported"
+        mcctrl_addr="0"
+      fi
+
+    { $as_echo "$as_me:${as_lineno-$LINENO}: result: $mcctrl_result" >&5
+$as_echo "$mcctrl_result" >&6; }
+
+cat >>confdefs.h <<_ACEOF
+#define MCCTRL_KSYM_sys_umount $mcctrl_addr
+_ACEOF
+
+  fi
+
+
   { $as_echo "$as_me:${as_lineno-$LINENO}: checking System.map for symbol sys_unshare" >&5
 $as_echo_n "checking System.map for symbol sys_unshare... " >&6; }
   mcctrl_addr=`eval $MCCTRL_LINUX_SYMTAB_CMD | grep " sys_unshare\$" | cut -d\  -f1`

configure.ac

@@ -221,6 +221,7 @@ AC_DEFUN([MCCTRL_FIND_KSYM],[
 ])
 
 MCCTRL_FIND_KSYM([sys_mount])
+MCCTRL_FIND_KSYM([sys_umount])
 MCCTRL_FIND_KSYM([sys_unshare])
 MCCTRL_FIND_KSYM([zap_page_range])
 MCCTRL_FIND_KSYM([vdso_image_64])

executer/config.h.in

@@ -51,6 +51,9 @@
 /* Define to address of kernel symbol sys_readlink, or 0 if exported */
 #undef MCCTRL_KSYM_sys_readlink
 
+/* Define to address of kernel symbol sys_umount, or 0 if exported */
+#undef MCCTRL_KSYM_sys_umount
+
 /* Define to address of kernel symbol sys_unshare, or 0 if exported */
 #undef MCCTRL_KSYM_sys_unshare
 

executer/include/uprotocol.h

@@ -49,7 +49,8 @@
 #define MCEXEC_UP_CLOSE_EXEC     0x30a02913
 
 #define MCEXEC_UP_SYS_MOUNT      0x30a02914
-#define MCEXEC_UP_SYS_UNSHARE    0x30a02915
+#define MCEXEC_UP_SYS_UMOUNT     0x30a02915
+#define MCEXEC_UP_SYS_UNSHARE    0x30a02916
 
 #define MCEXEC_UP_DEBUG_LOG     0x40000000
 
@@ -196,6 +197,10 @@ struct sys_mount_desc {
 	void *data;
 };
 
+struct sys_umount_desc {
+	char *dir_name;
+};
+
 struct sys_unshare_desc {
 	unsigned long unshare_flags;
 };

executer/kernel/mcctrl/control.c

@@ -66,7 +66,18 @@ int (*mcctrl_sys_mount)(char *dev_name,char *dir_name, char *type, unsigned long
         (int_star_fn_char_char_char_ulong_void_t)
         MCCTRL_KSYM_sys_mount;
 #else // exported
-int (*mcctrl_sys_mount)(char *dev_name,char *dir_name, char *type, unsigned long flags, void *data) = NULL;
+int (*mcctrl_sys_mount)(char *dev_name,char *dir_name, char *type, unsigned long flags, void *data) = sys_mount;
+#endif
+#endif
+
+#ifdef MCCTRL_KSYM_sys_umount
+#if MCCTRL_KSYM_sys_umount
+typedef int (*int_fn_char_star_int_t)(char *, int);
+int (*mcctrl_sys_umount)(char *dir_name, int flags) =
+        (int_fn_char_star_int_t)
+        MCCTRL_KSYM_sys_umount;
+#else // exported
+int (*mcctrl_sys_umount)(char *dir_name, int flags) = sys_umount;
 #endif
 #endif
 
@@ -1154,7 +1165,7 @@ long mcexec_sys_mount(struct sys_mount_desc *__user arg)
 	cap_raise(promoted->cap_effective, CAP_SYS_ADMIN);
 	original = override_creds(promoted);
 
-#if MCCTRL_KSYM_sys_mount
+#ifdef MCCTRL_KSYM_sys_mount
 	ret = mcctrl_sys_mount(desc.dev_name, desc.dir_name, desc.type,
 		desc.flags, desc.data);
 #else
@@ -1167,6 +1178,38 @@ long mcexec_sys_mount(struct sys_mount_desc *__user arg)
 	return ret;
 }
 
+long mcexec_sys_umount(struct sys_mount_desc *__user arg)
+{
+	struct sys_umount_desc desc;
+	struct cred *promoted;
+	const struct cred *original;
+	int ret;
+
+	if (copy_from_user(&desc, arg, sizeof(desc))) {
+		return -EFAULT;
+	}
+
+	promoted = prepare_creds();
+	if (!promoted) {
+		return -ENOMEM;
+	}
+	cap_raise(promoted->cap_effective, CAP_SYS_ADMIN);
+	original = override_creds(promoted);
+
+#ifdef MCCTRL_KSYM_sys_umount
+	ret = mcctrl_sys_umount(desc.dir_name, MNT_FORCE);
+	kprintf("%s: mcctrl_sys_umount: %d\n", __FUNCTION__, ret);
+#else
+	ret = -EFAULT;
+	kprintf("%s: mcctrl_sys_umount not defined?\n", __FUNCTION__);
+#endif
+
+	revert_creds(original);
+	put_cred(promoted);
+
+	return ret;
+}
+
 long mcexec_sys_unshare(struct sys_unshare_desc *__user arg)
 {
 	struct sys_unshare_desc desc;
@@ -1254,6 +1297,9 @@ long __mcctrl_control(ihk_os_t os, unsigned int req, unsigned long arg,
 	case MCEXEC_UP_SYS_MOUNT:
 		return mcexec_sys_mount((struct sys_mount_desc *)arg);
 
+	case MCEXEC_UP_SYS_UMOUNT:
+		return mcexec_sys_umount((struct sys_mount_desc *)arg);
+
 	case MCEXEC_UP_SYS_UNSHARE:
 		return mcexec_sys_unshare((struct sys_unshare_desc *)arg);
 

executer/kernel/mcctrl/driver.c

@@ -69,6 +69,7 @@ static struct ihk_os_user_call_handler mcctrl_uchs[] = {
 	{ .request = MCEXEC_UP_GET_CRED, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_GET_CREDV, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_SYS_MOUNT, .func = mcctrl_ioctl },
+	{ .request = MCEXEC_UP_SYS_UMOUNT, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_SYS_UNSHARE, .func = mcctrl_ioctl },
 	{ .request = MCEXEC_UP_DEBUG_LOG, .func = mcctrl_ioctl },
 };

executer/user/mcexec.c

@@ -1148,75 +1148,41 @@ void init_worker_threads(int fd)
 
 #ifdef ENABLE_MCOVERLAYFS
 #define READ_BUFSIZE 1024
-static int isunshare(void)
+static int find_mount_prefix(char *prefix)
 {
-	int err = 0;
+	FILE *fp;
-	int ret;
+	char *line = NULL;
-	int fd;
+	size_t len = 0;
+	ssize_t read;
 	char proc_path[PATH_MAX];
-	ssize_t len_read;
+	int ret = 0;
-	char buf_read[READ_BUFSIZE + 1];
-	char *buf_read_off;
-	char *buf_find;
-	char buf_cmp[READ_BUFSIZE + 1];
-	char *buf_cmp_off;
-	ssize_t len_copy;
 
 	snprintf(proc_path, sizeof(proc_path), "/proc/%d/mounts", getpid());
-	fd = open(proc_path, O_RDONLY);
+
-	if (fd < 0) {
+	fp = fopen(proc_path, "r");
-		fprintf(stderr, "Error: Failed to open %s.\n", proc_path);
+	if (fp == NULL) {
 		return -1;
 	}
 
-	buf_cmp_off = buf_cmp;
+	while ((read = getline(&line, &len, fp)) != -1) {
-	while (1) {
+		if (strlen(line) < strlen(prefix))
-		len_read = read(fd, buf_read, READ_BUFSIZE);
+			continue;
-		if (len_read == -1) {
-			fprintf(stderr, "Error: Failed to read.\n");
-			err = -1;
-			break;
-		}
 
-		buf_read_off = buf_read;
+		if (!strncmp(line, prefix, strlen(prefix))) {
-		while (1) {
+			ret = 1;
-			if ((len_read - (buf_read_off - buf_read)) <= 0) {
-				break;
-			}
-			buf_find = memchr(buf_read_off, '\n', 
-				len_read - (buf_read_off - buf_read));
-			if (buf_find) {
-				len_copy = buf_find - buf_read_off;	
-			} else {
-				len_copy = len_read - (buf_read_off - buf_read);
-			}
-			memcpy(buf_cmp_off, buf_read_off, len_copy);
-			*(buf_cmp_off + len_copy) = '\0';
-
-			if (buf_find) {
-				buf_read_off = buf_read_off + len_copy + 1;
-				buf_cmp_off = buf_cmp;
-				ret = strncmp(buf_cmp, "mcoverlay /proc ", 16);
-				if (!ret) {
-					err = 1;
-					break;
-				}
-			} else {
-				buf_read_off = buf_read_off + len_copy;
-				buf_cmp_off = buf_cmp_off + len_copy;
-				break;
-			}
-		}
-
-		if (err == 1 || len_read == 0) {
 			break;
 		}
 	}
 
-	close(fd);
+	if (line)
+		free(line);
 
-	__dprintf("err=%d\n", err);
+	return ret;
-	return err;
+}
+
+static int isunshare(void)
+{
+	return find_mount_prefix("mcoverlay /proc ");
 }
 #endif // ENABLE_MCOVERLAYFS
 
@@ -1415,6 +1381,7 @@ int main(int argc, char **argv)
 	if (error == 0) {
 		struct sys_unshare_desc unshare_desc;
 		struct sys_mount_desc mount_desc;
+		struct sys_umount_desc umount_desc;
 
 		memset(&unshare_desc, '\0', sizeof unshare_desc);
 		memset(&mount_desc, '\0', sizeof mount_desc);
@@ -1426,6 +1393,53 @@ int main(int argc, char **argv)
 			return 1;
 		}
 
+		/*
+		 * Umount cgroup filesystems that may expose invalid NUMA
+		 * information
+		 */
+		if (find_mount_prefix("cgroup /sys/fs/cgroup/cpu,cpuacct")) {
+			umount_desc.dir_name = "/sys/fs/cgroup/cpu,cpuacct";
+
+			if (ioctl(fd, MCEXEC_UP_SYS_UMOUNT,
+						(unsigned long)&umount_desc) != 0) {
+				fprintf(stderr,
+						"WARNING: Failed to umount cgroup/cpu,cpuacct. (%s)\n",
+						strerror(errno));
+			}
+		}
+		else if (find_mount_prefix("cgroup /sys/fs/cgroup/cpu")) {
+			umount_desc.dir_name = "/sys/fs/cgroup/cpu";
+
+			if (ioctl(fd, MCEXEC_UP_SYS_UMOUNT,
+						(unsigned long)&umount_desc) != 0) {
+				fprintf(stderr,
+						"WARNING: Failed to umount cgroup/cpu. (%s)\n",
+						strerror(errno));
+			}
+		}
+
+		if (find_mount_prefix("cgroup /sys/fs/cgroup/cpuset")) {
+			umount_desc.dir_name = "/sys/fs/cgroup/cpuset";
+
+			if (ioctl(fd, MCEXEC_UP_SYS_UMOUNT,
+						(unsigned long)&umount_desc) != 0) {
+				fprintf(stderr,
+						"WARNING: Failed to umount cgroup/cpuset. (%s)\n",
+						strerror(errno));
+			}
+		}
+
+		if (find_mount_prefix("cgroup /sys/fs/cgroup/memory")) {
+			umount_desc.dir_name = "/sys/fs/cgroup/memory/";
+
+			if (ioctl(fd, MCEXEC_UP_SYS_UMOUNT,
+						(unsigned long)&umount_desc) != 0) {
+				fprintf(stderr,
+						"WARNING: Failed to umount cgroup/memory. (%s)\n",
+						strerror(errno));
+			}
+		}
+
 		sprintf(mcos_procdir, "/tmp/mcos/mcos%d_proc", mcosid);
 		mount_desc.dev_name = mcos_procdir;
 		mount_desc.dir_name = "/proc";