From: "fxn (Xavier Noria)" <noreply@...>
Date: 2022-03-25T19:50:06+00:00
Subject: [ruby-core:108072] [Ruby master Misc#18662] Fiber scheduling and Module#autoload

Issue #18662 has been reported by fxn (Xavier Noria).

----------------------------------------
Misc #18662: Fiber scheduling and Module#autoload
https://bugs.ruby-lang.org/issues/18662

* Author: fxn (Xavier Noria)
* Status: Open
* Priority: Normal
----------------------------------------
Looks like Fiber context-switching does not synchronize constant reference access. This script using the `async` gem demonstrates the issue:

```ruby
require 'tempfile'
require 'async'

Tempfile.create(['foo', '.rb']) do |file|
  file.write(<<~RUBY)
    sleep 0.01
    class C
    end
  RUBY
  file.close

  autoload :C, file.path

  Async do |task|
    3.times do |i|
      task.async { p C }
    end
  end.wait
end
```

As you'll, it raises `NameError`s.

The `sleep` call in `c.rb` is just an artificial way to trigger a context switch, but we learned that the `debug` gem does trigger it due to a TracePoint on `script_compiled` it installs. Check this other script that also reproduces the error:

```ruby
require 'tempfile'
require 'async'
require 'debug'

Tempfile.create(['foo', '.rb']) do |file|
  file.write(<<~RUBY)
    class C
      def self.call_me
      end
    end
  RUBY
  file.close

  autoload :C, file.path

  Async do |task|
    5.times do |i|
      task.async do
        p C.call_me
      end
    end
  end.wait
end
```

See https://github.com/ruby/debug/issues/580 for details.

I didn't tag this as "bug" because I do not know how you'd consider this. On one hand, blocking fibers are switched by the programmer, but non-blocking ones are by a contract interface + scheduler.

What do you think? Could something be done?



-- 
https://bugs.ruby-lang.org/

Unsubscribe: <mailto:ruby-core-request@ruby-lang.org?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>