Add EINVAL to TCGETS2/TCSETS2 fallback triggers

[NiltonVolpato]

This adds EINVAL to the error codes that trigger fallback from TCGETS2/TCSETS2 to the legacy TCGETS/TCSETS ioctls.

Motivation

Some emulated and virtualized environments don't implement TCGETS2/TCSETS2 and return EINVAL rather than ENOTTY or EACCES. This includes CheerpX/WebVM (x86-to-WebAssembly virtualization in the browser).

The fallback logic already exists for ENOTTY and EACCES. This change simply adds EINVAL as another trigger.

Reproducing

use rustix::pty::*;
use rustix::termios::*;

fn main() {
    let pty = openpt(OpenptFlags::empty()).expect("openpt failed");
    println!("tcgetattr: {:?}", tcgetattr(&pty));
}

cargo build --target i686-unknown-linux-musl --release -F termios,pty

Running on https://webvm.io/ outputs:

tcgetattr: Err(Os { code: 22, kind: InvalidInput, message: "Invalid argument" })

With this change, the fallback to TCGETS succeeds.

Changes

• tcgetattr: add Errno::INVAL to fallback match arm
• tcsetattr: add Errno::INVAL to fallback match arm

[sunfishcode]

Could this be considered a bug in CheerpX/WebVM? Either for lacking TCGETS2. or for having an unrecognize ioctl fail with EINVAL instead of ENOTTY, which is what actual Linux does for unrecognized ioctl codes?

[NiltonVolpato]

Sorry for the late response. My interpretation of the ioctl man pages is also that ENOTTY is a more appropriate error to be returned, in this case. That doesn't seem obvious to everyone, though, since there other platforms that are already special cased. But I agree that if CheerpX could fix that, then it would be ideal.

As observed libc seems to fallback to the old ioctl calls anyway. And I also haven't had problems executing other C programs with raw access to the terminal, or Rust programs not using reedline (which indirectly depends on rustix)

I created leaningtech/webvm#214