AIX: Only test addresses that are multiples of the size of cmsghdr

[xingxue-ibm]

Description

AIX does not apply any alignment or padding to ancillary data, and CMSG_ALIGN() is a no-op. Therefore, we only test addresses that are multiples of the size of cmsghdr on AIX. The original condition for AIX was removed in f391df3, which led to the test failure.

Sources

Checklist

• Relevant tests in libc-test/semver have been updated
• No placeholder or unstable values like *LAST or *MAX are
  included (see #3131)
• Tested locally (cd libc-test && cargo test --target mytarget);
  especially relevant for platforms that may not be checked in CI

[tgross35]

Ah sorry about that, I incorrectly assumed the AIX-specific bit was a result of the boundary check issues fixed in that commit. Thanks for the fix!

Cc @gibbz00

[gibbz00]

@xingxue-ibm

I'm sorry, but I don't understand how this makes sense. You say that AIX does not apply any padding, yet the payload is required to be a multiple of size_of::<cmsghdr>? Sounds to me like AIX's CMSG_ALIGN then be:

fn CMSG_ALIGN(len: usize) -> usize { (len + size_of::<cmsghdr>() - 1) & !(size_of::<cmsghdr>() - 1) }

How do you otherwise deal with payloads that aren't multiples of exactly cmsghdr bytes? Say when sending only one file descriptor over a unix socket?

[gibbz00]

... Or in other words, to me it sounds like the AIX C implementation does apply padding, but the Rust implementation doesn't, hence the test failure. If so, then the solution shouldn't be to ignore the test values where the implementations differ, but rather fix the Rust implementation. This may be easier with setting the __ALIGN_BOUNDARY from #4908, but I doubt it will land in before 1.0.

What kind of test failures were you getting?

[xingxue-ibm]

@xingxue-ibm

I'm sorry, but I don't understand how this makes sense. You say that AIX does not apply any padding, yet the payload is required to be a multiple of size_of::<cmsghdr>? Sounds to me like AIX's CMSG_ALIGN then be:

fn CMSG_ALIGN(len: usize) -> usize { (len + size_of::<cmsghdr>() - 1) & !(size_of::<cmsghdr>() - 1) }

How do you otherwise deal with payloads that aren't multiples of exactly cmsghdr bytes? Say when sending only one file descriptor over a unix socket?

The AIX <sys/socket.h> defines _CMSG_ALIGN as follows.

/* AIX does not have any alignment/padding for ancillary data */
#define _CMSG_ALIGN(p)      (p)

So, the Rust libc does not use _CMSG_ALIGN, for example:

    pub fn CMSG_NXTHDR(mhdr: *const msghdr, cmsg: *const cmsghdr) -> *mut cmsghdr {
        if cmsg.is_null() {
            CMSG_FIRSTHDR(mhdr)
        } else {
            if (cmsg as usize + (*cmsg).cmsg_len as usize + size_of::<cmsghdr>())
                > ((*mhdr).msg_control as usize + (*mhdr).msg_controllen as usize)
            {
                core::ptr::null_mut::<cmsghdr>()
            } else {
                // AIX does not have any alignment/padding for ancillary data, so we don't need _CMSG_ALIGN here.
                (cmsg as usize + (*cmsg).cmsg_len as usize) as *mut cmsghdr
            }
        }
    }

[xingxue-ibm]

... Or in other words, to me it sounds like the AIX C implementation does apply padding, but the Rust implementation doesn't, hence the test failure. If so, then the solution shouldn't be to ignore the test values where the implementations differ, but rather fix the Rust implementation. This may be easier with setting the __ALIGN_BOUNDARY from #4908, but I doubt it will land in before 1.0.

What kind of test failures were you getting?

The test failed with the following error:

thread 't::test_cmsg_nxthdr' panicked at libc-test/tests/cmsg.rs:86:25:
misaligned pointer dereference: address must be a multiple of 0x4 but is 0x1108c63e5

Where line 86 is:

 86                         (*current_cmsghdr_ptr).cmsg_len =
 87                             libc::CMSG_LEN(cmsg_payload_len as _) as _;

[gibbz00]

Thanks for the quick response.

If you don't mind me asking, how does AIX's CMSG_SPACE C definition look like?

[xingxue-ibm]

If you don't mind me asking, how does AIX's CMSG_SPACE C definition look like?

...
/* AIX does not have any alignment/padding for ancillary data */
#define _CMSG_ALIGN(p)      (p)
...
#define CMSG_ALIGN          _CMSG_ALIGN
#define CMSG_SPACE(len)     (_CMSG_ALIGN(sizeof(struct cmsghdr)) + _CMSG_ALIGN(len))

[gibbz00]

Oh wow, interesting. So it's not that AIX supports unaligned access, but more so that it doesn't try to account for it at all.

Misaligned pointer dereference: address must be a multiple of 0x4 but is 0x1108c63e5

If I'm reading that correctly, isn't the minimum alignment 4 bytes? Not cmsghdr's 16?

Did the tests not work with cmsg_payload_len % std::mem::size_of::<u32>() != 0?

That would explain a lot about how it doesn't affect AIX in practice. Can't think of any cmsg payload that isn't divisible to 4.

[xingxue-ibm]

Yes, cmsg_payload_len % std::mem::size_of::<u32>() != 0 works as well.