I/O polling hangs when computeWorkerThreadCount = 1

[armanbilge]

First discovered this when debugging Native multithreading. It's also affecting CE 3.6.0-RC1 on the JVM, so it must be a WSTP / polling system integration bug.

//> using dep org.typelevel::cats-effect::3.6.0-RC1

import cats.effect.*
import java.nio.ByteBuffer
import java.nio.channels.Pipe
import java.nio.channels.SelectionKey._
import scala.concurrent.duration._

object Bug extends IOApp.Simple {
  override def computeWorkerThreadCount = 1 // changing to 2 fixes

  def run = mkPipe.use { pipe =>
    for {
      selector <- getSelector
      buf <- IO(ByteBuffer.allocate(4))
      _ <- IO(pipe.sink.write(ByteBuffer.wrap(Array(1, 2, 3)))).background
        .surround {
          selector.select(pipe.source, OP_READ) *> IO(pipe.source.read(buf))
        }
      _ <- IO(pipe.sink.write(ByteBuffer.wrap(Array(42)))).background.surround {
        selector.select(pipe.source, OP_READ) *> IO(pipe.source.read(buf))
      }
    } yield assert(buf.array().toList == List[Byte](1, 2, 3, 42))
  }.timeout(1.second)

  def getSelector: IO[Selector] =
    IO.pollers
      .map(_.collectFirst { case selector: Selector => selector })
      .map(_.get)

  def mkPipe: Resource[IO, Pipe] =
    Resource
      .eval(getSelector)
      .flatMap { selector =>
        Resource.make(IO(selector.provider.openPipe())) { pipe =>
          IO(pipe.sink().close()).guarantee(IO(pipe.source().close()))
        }
      }
      .evalTap { pipe =>
        IO {
          pipe.sink().configureBlocking(false)
          pipe.source().configureBlocking(false)
        }
      }
}

[armanbilge]

Hm, I guess it doesn't hang completely. It's just suspiciously slow 🤔

[djspiewak]

I wonder if this is illustrating an underlying issue that also occurs at higher worker counts