`misaligned pointer dereference` error in `PacketListIterator::next` about coremidi

[jmbarbier]

Same bug here, fixed it by adding read_unaligned for those lines

#[inline]
pub unsafe fn MIDIPacketNext(pkt: *const MIDIPacket) -> *const MIDIPacket {
    // Get pointer to potentially unaligned data without triggering undefined behavior
    // addr_of does not require creating an intermediate reference to unaligned data.
    let ptrx = ptr::addr_of!((*pkt).data) as *const u8;
    let rawl = ptr::addr_of!((*pkt).length) as *const u16;
    let offset = rawl.read_unaligned() as isize;
    if cfg!(any(target_arch = "arm", target_arch = "aarch64")) {
        // MIDIPacket must be 4-byte aligned on ARM
        ((ptrx.offset(offset + 3) as usize) & !(3usize)) as *const MIDIPacket
    } else {
        ptrx.offset(offset) as *const MIDIPacket
    }
}

#[inline]
pub unsafe fn MIDIEventPacketNext(pkt: *const MIDIEventPacket) -> *const MIDIEventPacket {
    // Get pointer to potentially unaligned data without triggering undefined behavi
    // or
    // addr_of does not require creating an intermediate reference to unaligned data.
    let ptrx = ptr::addr_of!((*pkt).words) as *const u8;
    let rawwc = ptr::addr_of!((*pkt).wordCount) as *const u32;
    let wc = rawwc.read_unaligned() as usize;
    let offset = (wc * mem::size_of::<u32>()) as isize;
    if cfg!(any(target_arch = "arm", target_arch = "aarch64")) {
        // MIDIEventPacket must be 4-byte aligned on ARM
        ((ptrx.offset(offset + 3) as usize) & !(3usize)) as *const MIDIEventPacket
    } else {
        ptrx.offset(offset) as *const MIDIEventPacket
    }
}

Take this as a hack without any warranty ("it works for me"), i have no knowledge of the internals used here...

[Boddlnagg]

This has repeatedly hit users of midir. I think the proposed fix by @jmbarbier is correct, but that code is actually part of https://github.com/jonas-k/coremidi-sys.
@jmbarbier Maybe you can open a PR against coremidi-sys?

[Boddlnagg]

Actually I'm not sure whether that fix is correct. MIDIPacket should be 4-byte aligned on ARM (see also #9), so read_unaligned shouldn't be necessary. The error must be somewhere else where MIDIPackets are constructed that are not aligned.

[Boddlnagg]

Okay, I think the problem is this: "The alignment requirements of MIDIPacket may differ between CPU architectures. On Intel and PowerPC, MIDIPacket is unaligned. On ARM, MIDIPacket must be 4-byte aligned."
(https://github.com/phracker/MacOSX-SDKs/blob/041600eda65c6a668f66cb7d56b7d1da3e8bcc93/MacOSX10.15.sdk/System/Library/Frameworks/CoreMIDI.framework/Versions/A/Headers/MIDIServices.h#L420-L423)

So I'm assuming that this issue occurs on Intel processors? That would make sense because then Rust sees an unaligned MIDIPacket that should be 4-byte aligned but isn't, so the read_unaligned is needed to not trigger Rust's UB check. On ARM this would not be needed and would actually make performance worse, because on ARM the packets are aligned and explicit unaligned reads would generate worse machine code.

Furthermore, this can only affect MIDIPacket, not MIDIEvent (its length will always be a multiple of 4 bytes), so the fix does not need to be applied to MIDIEventPacketNext.

[Boddlnagg]

@jmbarbier @arteme @oilcake Can you confirm that you're seeing this on Intel processors?

[oilcake]

@jmbarbier @arteme @oilcake Can you confirm that you're seeing this on Intel processors?

Yep, my one is intel.