ThreadSanitizer false positive due to missing interceptor for fcntl(fd, F_DUPFD_CLOEXEC, ..)

While ThreadSanitizer models synchronization implied by IO operations, it currently doesn't have interceptor for fcntl(fd, F_DUPFD_CLOEXEC, ..) and as a result operations on a duplicated file descriptor don't introduce synchronization. For example, the following generates a false positive report:

#![feature(sync_unsafe_cell)]
#![feature(anonymous_pipe)]
use std::cell::*;
use std::io::*;
use std::sync::*;

fn main() {
    let c = Arc::new(SyncUnsafeCell::new(0));
    let (mut a, mut b) = std::pipe::pipe().unwrap();
    // Duplicate file descriptor. Implemented in terms of fcntl(fd, F_DUPFD_CLOEXEC, ...).
    // Comment out the following line to hide the false positive.
    let mut b = b.try_clone().unwrap();
    let t = std::thread::spawn({
        let c = c.clone();
        move || {
            unsafe { *c.get() = 1 };
            b.write_all(b".").unwrap();
        }
    });
    let mut buf = [0];
    a.read_exact(&mut buf).unwrap();
    println!("{}", unsafe { *c.get() });
    t.join().unwrap();
}

This shortcoming makes ThreadSanitizer incompatible with Tokio.

Start by locating ThreadSanitizer's handling of fcntl and the missing F_DUPFD_CLOEXEC interceptor. Use the Rust pipe and try_clone reproduction in the issue to observe the false positive under ThreadSanitizer, then verify that the duplicated file descriptor establishes the expected synchronization and the reproduction no longer reports a race.