stdin cannot be read

[koivunej]

I tried to run my stdin reading program yesterday with miri 0.1.0 (10419b3 2020-05-11) and rustc 1.45.0-nightly (a74d1862d 2020-05-14) but upon reading the stdin Ok(0) seems to be always returned at least on Linux.

Example:

use std::io::Read;

fn main() {
    let mut bytes = [0u8; 8192];
    let total = std::io::stdin().lock().read(&mut bytes[..]).unwrap();
    println!("{}", total);
}

With stable echo foo | cargo run prints 4.
With miri echo foo | cargo +nightly miri run -- -Zmiri-disable-isolation prints 0. Same for reading in a loop.

Looking at src/shims/foreign_items/posix.rs and src/shims/fs.rs I can't at least see any relevant TODO's or other notes which would suggest this should not work.

Could this be because src/bin/cargo-miri.rs launches a subprocess and somehow the stdin does not get inherited?

[RalfJung]

Reading from stdin is indeed not implemented. So what I would have expected is that it returns an error, saying that the FD does not exist...

[RalfJung]

Yeah we are definitely returning -1 from that one. But somehow libstd swallows the error?

[RalfJung]

libstd contains explicit code to translate "ebadf" read errors for stdin/stdout/stderr into 0-length reads -- probably to work in environments where the handle is for some reason not available. (This is the Maybe wrapper defined in io/stdio.rs.)

So basically, everything behaves as expected -- except of course it would be nice to have stdin support in Miri. :D

[koivunej]

Ok good that I filed this issue then, I had misunderstood and was hoping this could be an E-easy issue I could snipe :) I'll update when I've looked more on how this could be implemented.

[RalfJung]

Hm, I'd still call this easy but yeah it requires digging into the shims code a bit more. ;) Probably it is time to handle FDs 0-2 in shims/fs.rs instead of having that code separately.

And what you said about cargo wrapping Miri causing problems might still be true, but we'll only know after the driver can handle stdin.^^

[RalfJung]

I improved the error message in #1435.

[samrat]

I am interested in picking this issue up.

@RalfJung Just to clarify, we want to start supporting reads on stdin, right?

[RalfJung]

@samrat yes. The starting point for that would be here:

miri/src/shims/foreign_items/posix.rs

Line 69 in b334004

throw_unsup_format!("reading from stdin is not implemented")

Let me know if you need any help. :)

[samrat]

Oh, I just realized that @koivunej had mentioned that he was interested in picking this up-- didn't mean to step on their toes like this.

@koivunej Assuming you're still interested in the issue, please take it away! I look forward to checking out the solution you come up with :)

[koivunej]

@samrat I don't expect to have time until the weekend so you can go ahead. Given the starting point linked above I'd guess reading should be implemented similar to stdout and stderr writing few lines below.

I look forward to checking out the solution you come up with :)

Likewise :)

[RalfJung]

I'd guess reading should be implemented similar to stdout and stderr writing few lines below.

Yes I think that would be a good start.
Longer-term, I'd like to move the handling of "special" FDs to shims/fs.rs together with the other FDs, with some kind of abstraction for "either a normal FD or stdout/stdin/stderr" -- then we could even support dup or close on the "special" FDs. But that does not have to happen in the same PR.