Fix signal handling on Unix systems

This removes the the signal handling machinery in tty::unix, and
replaces it with functionality from signal-hook, which should be more
robust. Signals caught by signal-hook wake up the existing I/O event
loop, which then delegates back to the PTY to handle them.

In particular, this allows `SIGCHLD` (i.e. child process exits) to shut
down the terminal promptly, instead of sometimes leaving the window
lingering.

Fixes #915.
Fixes #1276.
Fixes #1313.

As a side effect, this fixes a very rare bug on Linux, where a `read`
from the PTY on the master side would sometimes "fail" with `EIO` if the
child closed the client side at a particular moment. This was subject to
a race condition, and was very difficult to trigger in practice.

1 files changed, 6 insertions, 4 deletions

diff --git a/src/tty/windows/mod.rs b/src/tty/windows/mod.rs
index 5e0aee84..2f5caa93 100644
--- a/src/tty/windows/mod.rs
+++ b/src/tty/windows/mod.rs
@@ -28,7 +28,7 @@ use crate::cli::Options;
 use crate::config::Config;
 use crate::display::OnResize;
 use crate::term::SizeInfo;
-use crate::tty::EventedReadWrite;
+use crate::tty::{EventedReadWrite, EventedPty};
 
 mod conpty;
 mod winpty;
@@ -232,12 +232,12 @@ impl<'a> EventedReadWrite for Pty<'a> {
     fn register(
         &mut self,
         poll: &mio::Poll,
-        token: &mut dyn Iterator<Item = &usize>,
+        token: &mut dyn Iterator<Item = mio::Token>,
         interest: mio::Ready,
         poll_opts: mio::PollOpt,
     ) -> io::Result<()> {
-        self.read_token = (*token.next().unwrap()).into();
-        self.write_token = (*token.next().unwrap()).into();
+        self.read_token = token.next().unwrap();
+        self.write_token = token.next().unwrap();
 
         if interest.is_readable() {
             poll.register(
@@ -339,3 +339,5 @@ impl<'a> EventedReadWrite for Pty<'a> {
         self.write_token
     }
 }
+
+impl<'a> EventedPty for Pty<'a> { }