No EINTR anymore after signal_hook::low_level::register(SIGWINCH, ...)

[gwenn]

I am trying to switch from nix::sys::signal::sigaction to signal_hook::low_level::register here:
kkawakam/rustyline#608

Without signal-hook, if I resize the terminal while the program is running:

$ RUST_LOG=rustyline=debug cargo run --example minimal
...
[DEBUG rustyline] SIGWINCH
[DEBUG rustyline] read error: Os { code: 4, kind: Interrupted, message: "Interrupted system call" }
...

An EINTR occurs.
But with signal-hook:

$ RUST_LOG=rustyline=debug cargo run --features signal-hook --example minimal
...
[DEBUG rustyline] SIGWINCH
...

There is no EINTR anymore.

Could you please explain to me the difference between nix::sys::signal::sigaction and signal_hook::low_level::register ?
Thanks.

[Michal 'vorner' Vaner]

Could be related to this line? https://github.com/vorner/signal-hook/blob/master/signal-hook-registry/src/lib.rs#L165

Can you show the code that needs the EINTR? All the uses I've seen so far contained one kind of race condition or another.

[gwenn]

The code that needs the EINTR is here:
https://github.com/gwenn/rustyline/blob/signal_hook/src/tty/unix.rs#L141
If a SIGWINCH event occurs while waiting for user input, an EINTR is caught and we repaint the line.

[gwenn]

Could be related to this line? https://github.com/vorner/signal-hook/blob/master/signal-hook-registry/src/lib.rs#L165

That's it:

--- a/src/tty/unix.rs
+++ b/src/tty/unix.rs
@@ -1132,7 +1132,7 @@ cfg_if::cfg_if! {
             SIGWINCH_ONCE.call_once(|| unsafe {
                 let sigwinch = signal::SigAction::new(
                     signal::SigHandler::Handler(sigwinch_handler),
-                    signal::SaFlags::empty(),
+                    signal::SaFlags::SA_RESTART,
                     signal::SigSet::empty(),
                 );
                 let _ = signal::sigaction(signal::SIGWINCH, &sigwinch);

nix::sys::signal::sigaction behaves like signal_hook::low_level::register if SA_RESTART flag is activated.

[Michal 'vorner' Vaner]

Looking at the code, it is indeed subtly incorrect. If the SIGWINCH is delivered just before the read, it won't get interrupted and will not get re-drawn.

Considering the SA_RESTART behavior depends a lot on the given system and most code out there does handling of EINTR anyway, the flag could possibly be removed. Nevertheless, considering the above race, I'd suggest finding a better approach than this anyway. Maybe some variation of poll or ppoll in combination with the self-pipe trick?

[gwenn]

Ok,
I will do like termwiz.
Thanks.