fix fanotify syscall check on compat kernel

#ifndef __FANOTIFY_H__
#define __FANOTIFY_H__

testcases/kernel/syscalls/fanotify/fanotify.h +6 -2

+#include <sys/types.h>
+#include <endian.h>
+#include "config.h"
+#include "linux_syscall_numbers.h"
+
 /* fanotify(7) wrappers */
 
 #define	myfanotify_init(flags, event_f_flags) \
 	syscall(__NR_fanotify_init, flags, event_f_flags)
 
-#define	myfanotify_mark(fd, flags, mask, dfd, pathname)	\
-	syscall(__NR_fanotify_mark, fd, flags, mask, dfd, pathname)
+long myfanotify_mark(int fd, unsigned int flags, unsigned long long mask, int dfd, const char *pathname)

Quoted 2 lines

I think that the x32 ABI should be handled here as well (as in the fallocate.h). Mike can you confirm?

Yes, I think so.
I even thought of changing the #if to C-Coding like
if ((sizeof(void*) == 8)
call 64bit_syscall
else

    call 32bit_syscall_wrapper;

I'm not sure if this can avoid the ifdefs though...

Quoted 4 lines

Also I've been unable to make the testcases work with -m32 on x86_64 both with and without the patch and even when using glibc wrappers from sys/fallocate.h, in all cases the error is EINVAL which suggets that the parameters are passed wrongly somewhere on the way.

Does it work on the same kernel with -m64 ?
In any case, for the tests to work the kernel needs to have been compiled with the kernel option CONFIG_FANOTIFY_ACCESS_PERMISSIONS set. In my initial tests this option wasn't set and I got EINVAL too.

Quoted 4 lines

Also I've been unable to make the testcases work with -m32 on x86_64 both with and without the patch and even when using glibc wrappers from sys/fallocate.h, in all cases the error is EINVAL which suggets that the parameters are passed wrongly somewhere on the way.

Does it work on the same kernel with -m64 ?
In any case, for the tests to work the kernel needs to have been compiled with the kernel option CONFIG_FANOTIFY_ACCESS_PERMISSIONS set. In my initial tests this option wasn't set and I got EINVAL too.

Quoted 5 lines

Quoted 4 lines

Also I've been unable to make the testcases work with -m32 on x86_64 both with and without the patch and even when using glibc wrappers from sys/fallocate.h, in all cases the error is EINVAL which suggets that the parameters are passed wrongly somewhere on the way.

Quoted 4 lines

Does it work on the same kernel with -m64 ?
In any case, for the tests to work the kernel needs to have been compiled with the kernel option CONFIG_FANOTIFY_ACCESS_PERMISSIONS set. In my initial tests this option wasn't set and I got EINVAL too.

And it seems to work (-m32 with your patch) on two testing machines with slightly older kernel. So I guess that the patch is fine and I had just bad luck with one particular machine.

The fanotify[0-5] testcases use the myfanotify_mark() wrapper for the fanotify_mark() syscall. But the #define does not take into account, that the mask field is actually of type _u64 where we need to split the hi and low word in the syscall on a 32bit arch.

The attached patch converts the #define to a inline function where the compiler will help to convert the given values to the correct types (e.g. unsigned long).

Bug found and tested on the parisc/hppa (32bit userspace with 64bit kernel) arch.

Attached patch helps to find the reason when the fanotify03 testcase fails. For the test to work correctly, the kernel option CONFIG_FANOTIFY_ACCESS_PERMISSIONS needs to be enabled.