ioctl write_int doesn't support passing long integers

[jethrogb]

The Linux kernel API always passes the ioctl parameter as an unsigned long. However, the ioctl! macro always uses c_int. It should probably use c_long/c_ulong/usize instead, in order to pass all possible values.

[Susurrus]

Where is this specified for Linux? And if this is correct, then why wouldn't c_ulong be the most appropriate choice out of those 3 options? Only one should be used, but you didn't clarify which is the best option in your statement.

[Susurrus]

@jethrogb I don't understand enough to fix this, so I need some more information from you. Specifically a test case that you know is failing currently.

Linux should actually be using a c_ulong type for this according to this code in src/sys/ioctl/platform/linux.rs:

#[cfg(not(any(target_os = "android", target_env = "musl")))]
#[doc(hidden)]
pub type ioctl_num_type = ::libc::c_ulong;

[Susurrus]

Sorry, actually looking at the code I do see the size of a c_int is used in generating the ioctl request code, so maybe that's the error you're indicating? I'm going to wait for more details from you before digging into this further.

[jethrogb]

ioctl_num_type is the type used for the ioctl number, not the parameter. It might be ok to reuse it for the parameter? Not sure if it's consistent across platforms.

The problem is with this line https://github.com/Susurrus/nix/blob/7eb8dd1ccf5ec8281076e9aa8f1598e0e8bcbd68/src/sys/ioctl/mod.rs#L544

[Susurrus]

I still don't see any documentation suggesting what youre saying here. The passed integer must be equal in size to a pointer type because of the way ioctl is defined. We'll need to sort out what's appropriate for each platform, and we'll need docs for that. Can you link to any?

[jethrogb]

https://github.com/torvalds/linux/blob/5d38f04/include/linux/fs.h#L1702-L1703

That, by the way, shows that the current ioctl_num_type of c_ulong for non-Android non-musl Linux is incorrect.

[Susurrus]

I couldn't quickly find the declaration of ioctl on Android to find the ioctl_num_type, but I'd assume it's c_int. And musl shouldn't matter since this is a Linux kernel API which is independent of the libc being used. So I agree with you there. I wonder why it was ever like this to begin with?

I'll fix this and add also specify the parameter as c_ulong as well in #833.

On BSD it looks like it's consistently a c_ulong, so at least that should be correct.

[Susurrus]

And the fix to this might not be too easy because:

~/libc $ rg fn\ ioctl
src/unix/notbsd/linux/s390x.rs
1249:    pub fn ioctl(fd: ::c_int, request: ::c_ulong, ...) -> ::c_int;

src/unix/notbsd/linux/mips/mod.rs
679:    pub fn ioctl(fd: ::c_int, request: ::c_ulong, ...) -> ::c_int;

src/unix/notbsd/linux/musl/mod.rs
303:    pub fn ioctl(fd: ::c_int, request: ::c_int, ...) -> ::c_int;

src/unix/notbsd/linux/other/mod.rs
569:    pub fn ioctl(fd: ::c_int, request: ::c_ulong, ...) -> ::c_int;

So are those 3 c_ulong ioctl declarations wrong then?

[Susurrus]

Fixed by 5dad660, which is now in master.