Re: [PATCH v6 20/21] all: s390: make compat wrappers the generic solution

On Tue, Jan 19, 2016 at 08:52:23PM +0300, Yury Norov wrote:

quoted

quoted

+asmlinkage long compat_sys_creat(const char __user *pathname,
umode_t mode);
+asmlinkage long compat_sys_link(const char __user *oldname,
+                             const char __user *newname);
+asmlinkage long compat_sys_chdir(const char __user *filename);
+asmlinkage long compat_sys_mknod(const char __user *filename,
umode_t mode,
+                             unsigned dev);

Are these really needed?

91 of ~160 wrapped syscalls produce compile time error without it on
arm64:
arch/arm64/kernel/sys_ilp32.c:59:35: error: ?compat_sys_io_destroy? undeclared here (not in a function)
 #define __SC_WRAP(nr, sym) [nr] = compat_##sym,
                                   ^
include/uapi/asm-generic/unistd.h:39:1: note: in expansion of macro ?__SC_WRAP?
__SC_WRAP(__NR_io_destroy, sys_io_destroy)
^

I think, it's better to leave it as is...

I see. The syscall tables on arm seem to be generated in C code. So you
need the declarations.

This is a non-obvious change at first glance. I think it would make sense
to split this patch into at least three or four separate patches:
- one which introduces the infrastructure
- one which converts the non-obvious syscalls like this one
- one which converts the rest
- one which enables the architectures

You might consider adding a BUILD_BUG_ON() here, like within the s390 variant.

Personally I don't like the SYSCALL_DEFINE_WRAPx names too much. But that
can be changed easily if somebody comes up with a better name for it.

Given that the system call functions might be inlined, I think it would
make sense to remove the "notrace" attribute. Otherwise we would end up
with lots of untraceable (and unpatchable) functions. I didn't care back
then for the small compat wrappers...

It might make sense to add a define which adds both variants.

Btw. for future versions please add linux-arch to cc.

On Wed, Jan 20, 2016 at 09:16:44AM +0100, Heiko Carstens wrote:

quoted

+#ifndef __ARCH_WANT_COMPAT_SYS_GETDENTS64
+SYSCALL_DEFINE_WRAP3(getdents64, unsigned int, fd,
+		struct linux_dirent64 __user *, dirent, unsigned int, count)
+#else
 SYSCALL_DEFINE3(getdents64, unsigned int, fd,
 		struct linux_dirent64 __user *, dirent, unsigned int, count)
+#endif
 {
 	struct fd f;
 	struct linux_dirent64 __user * lastdirent;

This is a non-obvious change at first glance. I think it would make sense
to split this patch into at least three or four separate patches:
- one which introduces the infrastructure
- one which converts the non-obvious syscalls like this one
- one which converts the rest
- one which enables the architectures

Agree. This one became too big and non-trivial...
I'll prepare patchset this week. I also think, it's better
to send it separately, like we did with 32/64bit off_t:
https://lkml.org/lkml/2015/12/26/71

quoted

diff --git a/include/linux/compat.h b/include/linux/compat.h
index a76c917..293864c 100644
--- a/include/linux/compat.h
+++ b/include/linux/compat.h

@@ -55,6 +55,52 @@
 	}								\
 	static inline long C_SYSC##name(__MAP(x,__SC_DECL,__VA_ARGS__))

+#define SYSCALL_DEFINE_WRAP1(name, ...) SYSCALL_DEFINE_WRAPx(1, _##name, __VA_ARGS__)
+#define SYSCALL_DEFINE_WRAP2(name, ...) SYSCALL_DEFINE_WRAPx(2, _##name, __VA_ARGS__)
+#define SYSCALL_DEFINE_WRAP3(name, ...) SYSCALL_DEFINE_WRAPx(3, _##name, __VA_ARGS__)
+#define SYSCALL_DEFINE_WRAP4(name, ...) SYSCALL_DEFINE_WRAPx(4, _##name, __VA_ARGS__)
+#define SYSCALL_DEFINE_WRAP5(name, ...) SYSCALL_DEFINE_WRAPx(5, _##name, __VA_ARGS__)
+#define SYSCALL_DEFINE_WRAP6(name, ...) SYSCALL_DEFINE_WRAPx(6, _##name, __VA_ARGS__)
+
+#ifndef __SC_COMPAT_TYPE
+#define __SC_COMPAT_TYPE(t, a) \
+	__typeof(__builtin_choose_expr(sizeof(t) > 4, 0L, (t)0)) a
+#endif
+
+#ifndef __SC_COMPAT_CAST
+#define __SC_COMPAT_CAST(t, a)	((t) ((t)(-1) < 0 ? (s64)(s32)(a) : (u64)(u32)(a)))
+#endif

You might consider adding a BUILD_BUG_ON() here, like within the s390 variant.

Personally I don't like the SYSCALL_DEFINE_WRAPx names too much. But that
can be changed easily if somebody comes up with a better name for it.

Neither me... Maybe SYSCALL_DEFINE_WRAPPEDx?

quoted

+#define SYSCALL_DEFINE_WRAPx(x, name, ...)						\
+asmlinkage long sys##name(__MAP(x,__SC_DECL,__VA_ARGS__));				\
+asmlinkage long compat_sys##name(__MAP(x,__SC_DECL,__VA_ARGS__))			\
+		__attribute__((alias(__stringify(compat_SyS##name))));			\
+asmlinkage long notrace compat_SyS##name(__MAP(x,__SC_COMPAT_TYPE,__VA_ARGS__));	\
+asmlinkage long notrace compat_SyS##name(__MAP(x,__SC_COMPAT_TYPE,__VA_ARGS__))		\
+{											\
+	return sys##name(__MAP(x,__SC_COMPAT_CAST,__VA_ARGS__));			\
+}											\
+SYSCALL_DEFINEx(x, name, __VA_ARGS__)

Given that the system call functions might be inlined, I think it would
make sense to remove the "notrace" attribute. Otherwise we would end up
with lots of untraceable (and unpatchable) functions. I didn't care back
then for the small compat wrappers...

Ok

quoted

diff --git a/kernel/sys_ni.c b/kernel/sys_ni.c
index 0623787..ab14c3f 100644
--- a/kernel/sys_ni.c
+++ b/kernel/sys_ni.c

@@ -17,27 +17,41 @@ asmlinkage long sys_ni_syscall(void)
 }

 cond_syscall(sys_quotactl);
+cond_syscall(compat_sys_quotactl);

It might make sense to add a define which adds both variants.

Btw. for future versions please add linux-arch to cc.

Ok.

Yury.