Spurious "broken pipe" error messages, when used in typical UNIX shell pipelines

[infinity0]

$ cat yes.rs 
fn main() { loop { println!("y"); } }
$ rustc yes.rs && ./yes | head -n1
y
thread 'main' panicked at 'failed printing to stdout: Broken pipe (os error 32)', src/libstd/io/stdio.rs:692:8
note: Run with `RUST_BACKTRACE=1` for a backtrace.
$ yes | head -n1
y

This was originally filed here but @sfackler determined the cause:

This is due to println! panicking on errors:

rust/src/libstd/io/stdio.rs

Line 671 in f1ea23e

fn print_to<T>(args: fmt::Arguments,

.

C-based programs typically just get killed off with a SIGPIPE, but Rust ignores that signal.

Note that to see the backtrace, the data being piped has to be large enough to overflow the kernel pipe buffer.

[sfackler]

We could provide a function in std::io to unignore SIGPIPE so applications could more easily opt-in to acting like a "standard" command line program.

[infinity0]

Perhaps only the error message should be suppressed, it looks like the "traditional" programs do fail as a result of a broken pipe:

$ ./yes | head -n1
y
thread 'main' panicked at 'failed printing to stdout: Broken pipe (os error 32)', src/libstd/io/stdio.rs:692:8
note: Run with `RUST_BACKTRACE=1` for a backtrace.
$ echo "${PIPESTATUS[@]}"
101 0

$ yes | head -n1
y
$ echo "${PIPESTATUS[@]}"
141 0

$ find / | head -n1
/
$ echo "${PIPESTATUS[@]}"
141 0

141 seems to be the traditional exit code for a broken pipe.

[sfackler]

141 is the exit code set by the kernel after it has terminated a process due to a SIGPIPE.

[coriolinus]

We could provide a function in std::io to unignore SIGPIPE so applications could more easily opt-in to acting like a "standard" command line program.

I'm not sure what that API would look like: call a magic unignore_sigpipe() function and then your program just terminates on broken pipe, or a variant of the println!() family of macros, or what?

The former feels like it's just setting a global variable, which has a pretty bad smell. The latter means that unless you switch to using the new SIGPIPE-respecting macros throughout, your code might still generate the error.

What's not obvious to me is why Rust ignores that signal in the first place. I see that there's a test in place designed to ensure that the process shouldn't just crash, but at the same time the whole point of SIGPIPE is to terminate the receiving process silently. My intuition of correct behavior from Rust would be for it to do the same thing it does on SIGTERM: immediately, cleanly, and quietly shut itself down.

[sfackler]

call a magic unignore_sigpipe() function and then your program just terminates on broken pipe

That's what it would be presumably.

The former feels like it's just setting a global variable, which has a pretty bad smell.

Signal disposition is a process-global setting. Feel free to complain to the POSIX standards commitee about the smell of their global variables.

What's not obvious to me is why Rust ignores that signal in the first place.

SIGPIPE is a kind of hacky thing that only really makes sense when writing command line applications designed to be used in pipelines that only poke at their standard inputs and outputs. If you are writing anything more complex then it's something you need to turn off. Imagine a web server that crashed any time a client hung up, or a command line application that talks to the internet and crashed every time the server hung up.

[coriolinus]

Signal disposition is a process-global setting. Feel free to complain to the POSIX standards commitee about the smell of their global variables.

Haha, fair enough. I also do appreciate the explanation of the reasoning of turning it off by default. My own Rust applications tend to be unixy command-line applications which only ever really poke at their standard inputs and outputs, so that's the lens through which I view this issue, but I couldn't argue against the assertion that ignoring SIGPIPE is a more useful default.

In that case, I'd say that having an unignore_sigpipe function in the standard library somewhere would be an improvement on the current situation. Any idea how hard such a thing would be to implement?

[sfackler]

It'd just run this code:

rust/src/libstd/sys/unix/process/process_unix.rs

Lines 199 to 222 in c284f88

	// Reset signal handling so the child process starts in a
	// standardized state. libstd ignores SIGPIPE, and signal-handling
	// libraries often set a mask. Child processes inherit ignored
	// signals and the signal mask from their parent, but most
	// UNIX programs do not reset these things on their own, so we
	// need to clean things up now to avoid confusing the program
	// we're about to run.
	let mut set: libc::sigset_t = mem::uninitialized();
	if cfg!(target_os = "android") {
	// Implementing sigemptyset allow us to support older Android
	// versions. See the comment about Android and sig* functions in
	// process_common.rs
	libc::memset(&mut set as *mut _ as *mut _,
	0,
	mem::size_of::<libc::sigset_t>());
	} else {
	t!(cvt(libc::sigemptyset(&mut set)));
	}
	t!(cvt(libc::pthread_sigmask(libc::SIG_SETMASK, &set,
	ptr::null_mut())));
	let ret = sys::signal(libc::SIGPIPE, libc::SIG_DFL);
	if ret == libc::SIG_ERR {
	return io::Error::last_os_error()
	}

.

[arthurprs]

Found this out today when piping stdout to head.

Any ideas how to fix it nicely? Maybe we can take inspiration from other languages implementations.

[coriolinus]

Until someone implements and merges unignore_sigpipe(), your best bet will be to use the write!() macro instead of the print*!() family of macros, and then handle errors appropriately.

[richardwhiuk]

In the short term you can do:

extern crate libc;

...

    unsafe {
        libc::signal(libc::SIGPIPE, libc::SIG_DFL);
    }

[jarcane]

I seem to have run into this as well (see above issue), and I am finding the various linked and recommended solutions a bit vague. It's not clear to me how to assemble the bits and pieces to go about using it with write!, and I'd certainly rather avoid resorting to unsafe libc calls.

I was able to get a simple solution working with the try_print crate but using it on large streams brings back some old performance regressions I ran into on a previous issue, which was caused by excessive string allocations.

I must say it does seem a bit strange for my program to crash because of what another program downstream does or doesn't do with its inputs, but I would welcome any input on what a clear drop-in solution is that doesn't introduce any performance cost.

[coriolinus]

write! isn't hard: see here for one example of how to use it in production. Note that literally the only difference as far as the app is concerned is the gratuitous use of ? to handle potential errors.

[jarcane]

I was having trouble sorting out how to use write! with stdout, but I think your linked code should give me the hints I need. Thanks. :)

[jyn514]

This affects the compiler itself:

$ cargo +nightly rustc -- -Zunpretty=hir-tree | head > /dev/null 
  Compiling project v0.1.0 (/home/joshua/Documents/Programming/rust/project)
thread 'main' panicked at 'failed printing to stdout: Broken pipe (os error 32)', src/libstd/io/stdio.rs:792:9
note: Run with `RUST_BACKTRACE=1` environment variable to display a backtrace.

[juancampa]

The way I've worked around this is by replacing the print/println/eprint/etc macros with fallible versions. I just put this at the top of main.rs

// These macros are needed because the normal ones panic when there's a broken pipe.
// This is especially problematic for CLI tools that are frequently piped into `head` or `grep -q`
macro_rules! println {
  () => (print!("\n"));
  ($fmt:expr) => ({
    writeln!(std::io::stdout(), $fmt)
  });
  ($fmt:expr, $($arg:tt)*) => ({
    writeln!(std::io::stdout(), $fmt, $($arg)*)
  })
}

macro_rules! print {
  () => (print!("\n"));
  ($fmt:expr) => ({
    write!(std::io::stdout(), $fmt)
  });
  ($fmt:expr, $($arg:tt)*) => ({
    write!(std::io::stdout(), $fmt, $($arg)*)
  })
}

And then handle the std::io::Error appropriately:

fn main() -> ExitCode {
  match run() {
    Err(Error::IOError(err)) if err.kind() == io::ErrorKind::BrokenPipe => {
      // Okay, this happens when the output is piped to a program like `head`
      ExitCode::SUCCESS
    }
    Err(err) => {
      eprintln!("{}", err).ok();
      ExitCode::FAILURE
    }
    Ok(_) => ExitCode::SUCCESS,
  }
}