Investigating test creat_access.c failure

[rishabhBudhouliya]

Problem

The creat() system call is not implemented in RawPOSIX. It currently returns a fake success (valid file descriptor) without actually creating the file on disk.

lind-wasm/tests/unit-tests/file_tests/deterministic/creat_access.c

Line 15 in e788eaf

int fd = creat(FILE, file_perm);

Context

This was discovered while fixing the failing test:

lind-wasm/tests/unit-tests/file_tests/deterministic/creat_access.c

Line 15 in e788eaf

int fd = creat(FILE, file_perm);

[rishabhBudhouliya]

It seems like my initial assessment was not correct. We don't need to implement the creat syscall as the glibc implementation of creat.c calls open with the right flags ( O_WRONLY | O_CREAT | O_TRUNC)

I'm investigation why the creat_access.c test fails now with the current assumption.

[rennergade]

A good place to start would be seeing if the inner open call ever reaches rawposix and/or the trampoline

[rishabhBudhouliya]

As mentioned in the fix PR, this was happening due to:

lind-wasm/src/glibc/sysdeps/unix/sysv/linux/creat.c

Lines 30 to 34 in e717f4a

	# ifdef __NR_creat
	return SYSCALL_CANCEL (creat, file, mode);
	# else
	return __open (file, O_WRONLY | O_CREAT | O_TRUNC, mode);
	# endif

I spent a lot of time looking at the default implementation for creat.c but glibc uses above mentioned implementation of creat.c based on the chosen host. The creat.c is designed to call open but it was calling SYSCALL_CANCEL instead due to a hardcoded decision branch.

[Yaxuan-w]

This makes sense to me. Would it be better to call open with MAKE_LEGACY_SYSCALL so that we can track our modifications in the future easier?