IO#read_nonblock throwing EOFError with "exception: false"

[HoneyryderChuck]

I'm using exceptionless non-blocking IO API with Jruby 9.1.13.0. In a busy loop I have, it returns (predictably) :wait_readable, until it doesn't, and I get the message:

 EOFError: No message available

I thought that, by using exception: false, that the nonblocking APIs can't throw exception, but rather, :wait_readable/writable when buffer is unavailable, and nil when IO is closed (equivalent to EOFError in older rubies). In fact, this behaviour is consistent in ruby 2.3.0 and newer.

However, I can't seem to find documentation validating this. As of ruby-doc, this is only enforced since 2.4.2, and only for #read_nonblock. But I think this is more a case of the docs not being up-to-date.

[HoneyryderChuck]

Also, dunno if related, but although EOFed, it's still selectable:

loop do
  r, w = IO.select([r], [w])
  r.each do |reader|
    begin
      read.read_nonb...
    rescue EOFError
    end
  end
end

which is another difference with CRuby.

[headius]

I believe you are correct in your expectations. I'll have a look.

[headius]

If you are able to reproduce it, pass -Xbacktrace.style=full to JRuby and post the resulting trace. It would help me see where it's going wrong.

[headius]

I believe I have fixed this in fcd7d73 but I'd like to add or find a test that can trigger this behavior. Any additional information you can provide would be helpful!

[HoneyryderChuck]

I believe that the backtrace you need is this:

 java/lang/Thread.java:1552:in `getStackTrace'
org/jruby/runtime/backtrace/TraceType.java:246:in `getBacktraceData'
org/jruby/runtime/backtrace/TraceType.java:47:in `getBacktrace'
org/jruby/RubyException.java:235:in `prepareBacktrace'
org/jruby/exceptions/RaiseException.java:216:in `preRaise'
org/jruby/exceptions/RaiseException.java:183:in `preRaise'
org/jruby/exceptions/RaiseException.java:111:in `<init>'
org/jruby/Ruby.java:4138:in `newRaiseException'
org/jruby/Ruby.java:4093:in `newEOFError'
org/jruby/RubyIO.java:2907:in `getPartial'
org/jruby/RubyIO.java:2812:in `doReadNonblock'
org/jruby/RubyIO.java:2803:in `read_nonblock'
org/jruby/RubyIO$INVOKER$i$0$2$read_nonblock.gen:-1:in `call'
org/jruby/internal/runtime/methods/JavaMethod.java:729:in `call'
org/jruby/internal/runtime/methods/DynamicMethod.java:216:in `call'
org/jruby/runtime/callsite/CachingCallSite.java:225:in `call'

To reproduce this in a script is a bit non-trivial, as this doesn't happen all the time in my tests (just in a particular one, where more than the usual payload passed between sockets). Maybe I can test this again once 9.1.14.0 is out the door.