This was originally found by Oli Gillespie [not an Author yet], submitting this on his behalf. Oli would do the patch work.

One of our migrations to JDK 17 revealed an issue in JDK-8221481 that resulted in >10% CPU spent in nio waiting code. The minimized case shows how the waiting code ends up issuing back-to-back polls with zero timeout:

```
import java.io.*;
import java.net.*;

public class ReadSingle {
    public static void main(String[] args) throws Exception {
        Socket s = new Socket("shipilev.net", 443);
        s.setSoTimeout(1);
        s.getOutputStream().write(1);
        InputStream is = s.getInputStream();
        System.out.println("Starting single byte read");
        is.read();
    }
}
```

With current JDK:

```
 % strace -f build/linux-x86_64-server-release/images/jdk/bin/java ReadSingle.java 2>&1 | grep poll
[pid 35427] poll([{fd=4, events=POLLIN|POLLERR|POLLHUP}], 1, 5000) = 1 ([{fd=4, revents=POLLIN}])
[pid 35427] poll([{fd=4, events=POLLIN|POLLERR|POLLHUP}], 1, 5000) = 1 ([{fd=4, revents=POLLIN|POLLHUP}])
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 35427] poll([{fd=5, events=POLLIN}], 1, 0) = 0 (Timeout)
```

With JDK 17:

```
 % strace -f ../shipilev-jdk17u-dev/build/linux-x86_64-server-release/images/jdk/bin/java ReadSingle.java 2>&1 | grep poll
[pid 34344] poll([{fd=4, events=POLLIN|POLLERR|POLLHUP}], 1, 5000) = 1 ([{fd=4, revents=POLLIN|POLLHUP}])
[pid 34344] poll([{fd=4, events=POLLIN|POLLERR|POLLHUP}], 1, 5000) = 1 ([{fd=4, revents=POLLIN}])
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
[pid 34344] poll([{fd=4, events=POLLIN}], 1, 0) = 0 (Timeout)
```

With JDK 17 and `-Djdk.net.usePlainSocketImpl`:

```
% strace -f ../shipilev-jdk17u-dev/build/linux-x86_64-server-release/images/jdk/bin/java -Djdk.net.usePlainSocketImpl ReadSingle.java 2>&1 | grep poll
[pid 34653] poll([{fd=4, events=POLLIN|POLLERR|POLLHUP}], 1, 5000) = 1 ([{fd=4, revents=POLLIN}])
[pid 34653] poll([{fd=4, events=POLLIN|POLLERR|POLLHUP}], 1, 5000) = 1 ([{fd=4, revents=POLLIN|POLLHUP}])
[pid 34653] poll([{fd=6, events=POLLIN|POLLERR}], 1, 1) = 0 (Timeout)
```

This looks like a regression since JDK-8221481, and the cause seems to be this truncation of nanos to millis, which truncates the small nano-wait to 0, which is then passed to `poll(0)`, which returns immediately: https://github.com/openjdk/jdk/blob/e206d57bfc09032e17d09714fc54ab2f5e961792/src/java.base/share/classes/sun/nio/ch/NioSocketImpl.java#L184

```
    private void park(FileDescriptor fd, int event, long nanos) throws IOException {
        Thread t = Thread.currentThread();
        if (t.isVirtual()) {
           ...
        } else {
            long millis;
            if (nanos == 0) {
                millis = -1;
            } else {
                millis = NANOSECONDS.toMillis(nanos); // <--- here
            }
            Net.poll(fd, event, millis);
        }
    }
```

Note it would happen in nearly all cases where we timeout the read, as there will be a hot loop in the final sub-millisecond section which will truncate to zero millis. The read timeout would be more likely with smaller timeouts, but larger timeouts are also affected by this.

JDK-8253119 removed `-Djdk.net.usePlainSocketImpl`, so there is no recourse in JDK 18+ at all.

This looks similar to JDK-8210004, JDK-8065372 and others, and should probably be fixed the same way: rounding *up* to the nearest millisecond.