Making sure that sa_family is always initialized

[Jan561]

Alternative to #2235

Checklist:

• I have read CONTRIBUTING.md
• I have written necessary tests and rustdoc comments
• A change log has been added if this PR modifies nix's API

[Jan561]

Actually this fixes nothing as recvmsg doesn't even call SockaddrLike::from_raw. Instead, it just assumes correct initialization and doesn't check anything

nix/src/sys/socket/mod.rs

Line 2042 in 83dfd90

Ok(unsafe { read_mhdr(mhdr, r, msg_controllen, address.assume_init()) })

[asomers]

The change to SockaddrLikePriv::as_mut_ptr is unnecessary. You can do this manipulation in pack_mhdr_to_receive on raw pointers directly.
Also, could you please add a test that reproduces the original bug?

[Jan561]

The change to SockaddrLikePriv::as_mut_ptr is unnecessary. You can do this manipulation in pack_mhdr_to_receive on raw pointers directly.

Actually that's not possible, the address: *mut S argument is not NULL for S = (), it is just dangling.

That's how this function is called, with address being initialized with MaybeUninit::<S>::uninit():

nix/src/sys/socket/mod.rs

Line 2035 in a4e0c95

pack_mhdr_to_receive(iov.as_mut().as_mut_ptr(), iov.len(), msg_control, msg_controllen, address.as_mut_ptr())

Also, could you please add a test that reproduces the original bug?

Adding a test for this isn't that trivial, since currently the address field is "random" if not initialized by the syscall.

But the set of possible family values is quite small on BSDs (sa_family_t = u8), so I think with a test that calls recvmsg 500 or 1000 times on a connected socket we can be almost certain that one iteration will have the family field preset to e.g. INET or INET6.

But before we can add this test, we need to fix recvmsg

[Jan561]

I'd also consider the recvmsg situation a blocking issue for this PR