kerndat: check whether the openat2 syscall is supported

Will use openat2 + RESOLVE_NO_XDEV to detect mountpoints.

Signed-off-by: Pavel Tikhomirov <ptikhomirov@virtuozzo.com>

Makefile.config

@@ -78,7 +78,7 @@ export DEFINES += $(FEATURE_DEFINES)
 export CFLAGS += $(FEATURE_DEFINES)
 
 FEATURES_LIST	:= TCP_REPAIR STRLCPY STRLCAT PTRACE_PEEKSIGINFO \
-	SETPROCTITLE_INIT MEMFD TCP_REPAIR_WINDOW FSCONFIG MEMFD_CREATE
+	SETPROCTITLE_INIT MEMFD TCP_REPAIR_WINDOW FSCONFIG MEMFD_CREATE OPENAT2
 
 # $1 - config name
 define gen-feature-test

compel/arch/arm/plugins/std/syscalls/syscall.def

@@ -119,4 +119,5 @@ fsconfig			431	431	(int fd, unsigned int cmd, const char *key, const char *value
 fsmount				432	432	(int fd, unsigned int flags, unsigned int attr_flags)
 clone3				435	435	(struct clone_args *uargs, size_t size)
 pidfd_open			434	434	(pid_t pid, unsigned int flags)
+openat2				437	437	(int dirfd, char *pathname, struct open_how *how, size_t size)
 pidfd_getfd			438	438	(int pidfd, int targetfd, unsigned int flags)

compel/arch/mips/plugins/std/syscalls/syscall_64.tbl

@@ -116,4 +116,5 @@ __NR_fsconfig			5431		sys_fsconfig		(int fd, unsigned int cmd, const char *key,
 __NR_fsmount			5432		sys_fsmount		(int fd, unsigned int flags, unsigned int attr_flags)
 __NR_clone3			5435		sys_clone3		(struct clone_args *uargs, size_t size)
 __NR_pidfd_open			5434		sys_pidfd_open		(pid_t pid, unsigned int flags)
+__NR_openat2			5437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd		5438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)

compel/arch/ppc64/plugins/std/syscalls/syscall-ppc64.tbl

@@ -115,4 +115,5 @@ __NR_fsconfig		431		sys_fsconfig		(int fd, unsigned int cmd, const char *key, co
 __NR_fsmount		432		sys_fsmount		(int fd, unsigned int flags, unsigned int attr_flags)
 __NR_clone3		435		sys_clone3		(struct clone_args *uargs, size_t size)
 __NR_pidfd_open		434		sys_pidfd_open		(pid_t pid, unsigned int flags)
+__NR_openat2		437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd	438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)

compel/arch/s390/plugins/std/syscalls/syscall-s390.tbl

@@ -115,4 +115,5 @@ __NR_fsconfig		431		sys_fsconfig		(int fd, unsigned int cmd, const char *key, co
 __NR_fsmount		432		sys_fsmount		(int fd, unsigned int flags, unsigned int attr_flags)
 __NR_clone3		435		sys_clone3		(struct clone_args *uargs, size_t size)
 __NR_pidfd_open		434		sys_pidfd_open		(pid_t pid, unsigned int flags)
+__NR_openat2		437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd	438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)

compel/arch/x86/plugins/std/syscalls/syscall_32.tbl

@@ -103,4 +103,5 @@ __NR_fsconfig		431		sys_fsconfig		(int fd, unsigned int cmd, const char *key, co
 __NR_fsmount		432		sys_fsmount		(int fd, unsigned int flags, unsigned int attr_flags)
 __NR_clone3		435		sys_clone3		(struct clone_args *uargs, size_t size)
 __NR_pidfd_open		434		sys_pidfd_open		(pid_t pid, unsigned int flags)
+__NR_openat2		437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd	438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)

compel/arch/x86/plugins/std/syscalls/syscall_64.tbl

@@ -114,4 +114,5 @@ __NR_fsconfig			431		sys_fsconfig		(int fd, unsigned int cmd, const char *key, c
 __NR_fsmount			432		sys_fsmount		(int fd, unsigned int flags, unsigned int attr_flags)
 __NR_clone3			435		sys_clone3		(struct clone_args *uargs, size_t size)
 __NR_pidfd_open			434		sys_pidfd_open		(pid_t pid, unsigned int flags)
+__NR_openat2		437		sys_openat2		(int dirfd, char *pathname, struct open_how *how, size_t size)
 __NR_pidfd_getfd		438		sys_pidfd_getfd		(int pidfd, int targetfd, unsigned int flags)

compel/plugins/include/uapi/std/syscall-types.h

@@ -40,6 +40,7 @@ struct rusage;
 struct iocb;
 struct pollfd;
 struct clone_args;
+struct open_how;
 
 typedef unsigned long aio_context_t;
 

criu/cr-check.c

@@ -1402,6 +1402,14 @@ static int check_move_mount_set_group(void)
 	return 0;
 }
 
+static int check_openat2(void)
+{
+	if (!kdat.has_openat2)
+		return -1;
+
+	return 0;
+}
+
 static int (*chk_feature)(void);
 
 /*
@@ -1523,6 +1531,7 @@ int cr_check(void)
 		ret |= check_sockopt_buf_lock();
 		ret |= check_memfd_hugetlb();
 		ret |= check_move_mount_set_group();
+		ret |= check_openat2();
 	}
 
 	/*
@@ -1638,6 +1647,7 @@ static struct feature_list feature_list[] = {
 	{ "sockopt_buf_lock", check_sockopt_buf_lock },
 	{ "memfd_hugetlb", check_memfd_hugetlb },
 	{ "move_mount_set_group", check_move_mount_set_group },
+	{ "openat2", check_openat2 },
 	{ NULL, NULL },
 };
 

criu/include/kerndat.h

@@ -79,6 +79,7 @@ struct kerndat_s {
 	bool has_sockopt_buf_lock;
 	dev_t hugetlb_dev[HUGETLB_MAX];
 	bool has_move_mount_set_group;
+	bool has_openat2;
 };
 
 extern struct kerndat_s kdat;

criu/include/linux/openat2.h

@@ -0,0 +1,18 @@
+#ifndef _CRIU_LINUX_OPENAT2_H
+#define _CRIU_LINUX_OPENAT2_H
+
+#include <linux/types.h>
+
+#include "common/config.h"
+
+#ifdef CONFIG_HAS_OPENAT2
+#include <linux/openat2.h>
+#else
+struct open_how {
+	__u64 flags;
+	__u64 mode;
+	__u64 resolve;
+};
+#endif
+
+#endif

criu/include/mount-v2.h

@@ -2,6 +2,7 @@
 #define __CR_MOUNT_V2_H__
 
 #include "linux/mount.h"
+#include "linux/openat2.h"
 
 #include <compel/plugins/std/syscall-codes.h>
 
@@ -45,4 +46,13 @@ static inline int sys_open_tree(int dfd, const char *filename, unsigned int flag
 	return syscall(__NR_open_tree, dfd, filename, flags);
 }
 
+#ifndef RESOLVE_NO_XDEV
+#define RESOLVE_NO_XDEV 0x01 /* Block mount-point crossings (includes bind-mounts). */
+#endif
+
+static inline long sys_openat2(int dirfd, const char *pathname, struct open_how *how, size_t size)
+{
+	return syscall(__NR_openat2, dirfd, pathname, how, size);
+}
+
 #endif /* __CR_MOUNT_V2_H__ */

criu/kerndat.c

@@ -997,6 +997,22 @@ out:
 	return exit_code;
 }
 
+static int kerndat_has_openat2(void)
+{
+	if (sys_openat2(AT_FDCWD, ".", NULL, 0) != -1) {
+		pr_err("openat2 should fail\n");
+		return -1;
+	}
+	if (errno == ENOSYS) {
+		pr_debug("No openat2 syscall support\n");
+		kdat.has_openat2 = false;
+	} else {
+		kdat.has_openat2 = true;
+	}
+
+	return 0;
+}
+
 #define KERNDAT_CACHE_FILE     KDAT_RUNDIR "/criu.kdat"
 #define KERNDAT_CACHE_FILE_TMP KDAT_RUNDIR "/.criu.kdat"
 
@@ -1586,6 +1602,10 @@ int kerndat_init(void)
 		pr_err("kerndat_has_move_mount_set_group failed when initializing kerndat.\n");
 		ret = -1;
 	}
+	if (!ret && kerndat_has_openat2()) {
+		pr_err("kerndat_has_openat2 failed when initializing kerndat.\n");
+		ret = -1;
+	}
 
 	kerndat_lsm();
 	kerndat_mmap_min_addr();

scripts/feature-tests.mak

@@ -184,3 +184,15 @@ int main(void)
 	return memfd_create(NULL, 0);
 }
 endef
+
+define FEATURE_TEST_OPENAT2
+
+#include <linux/openat2.h>
+
+int main(void)
+{
+	if (RESOLVE_NO_XDEV > 0)
+		return 0;
+	return 0;
+}
+endef