Bug #17568: Thread.handle_interrupt is per-Thread but should probably be per-Fiber

• Status: Rejected
• Priority: Normal
• Assignee: -
• Target version: -
• ruby -v: ruby 2.7.2p137 (2020-10-01 revision 5445e04352) [x86_64-linux]
• Backport: 2.5: UNKNOWN, 2.6: UNKNOWN, 2.7: UNKNOWN, 3.0: UNKNOWN

[Eregon (Benoit Daloze)]

Thread.handle_interrupt and the pending interrupts is currently kept as state on the thread (rb_thread_t).
However, that seems potentially confusing when Fibers are used.

For instance, this innocent-looking Fiber will forever disable interrupts in the main thread:

main_thread = Thread.current

never_die = Fiber.new do
  Thread.handle_interrupt(RuntimeError => :never) do
    Fiber.yield
  end
end

begin
  Thread.new { main_thread.raise "interrupt1" }.join
rescue => e
  puts "raised: #{e}"
end

never_die.resume

Thread.new { main_thread.raise "interrupt2" }.join

Thread.new { main_thread.raise "interrupt3" }.join

p :after

Output is:

raised: interrupt1
:after

(I noticed these weird semantics when implementing Thread.handle_interrupt on TruffleRuby, and got a fairly subtle bug due to multiple Fibers of the same Thread queuing the same interrupt multiple times in the Thread's interrupt queue).

[nobu (Nobuyoshi Nakada)]

• Description updated (diff)

[ko1 (Koichi Sasada)]

I understand some case it is confusing, but if Fiber is used to represent the data structure like Enumerator, Thread-wide mask is needed.
Could you give us the example which should be fiber local?

mame-san also proposed Fiber.handle_interrupt for fiber local handler to set fiber local setting.

[mame (Yusuke Endoh)]

ko1 (Koichi Sasada) wrote in #note-2:

mame-san also proposed Fiber.handle_interrupt for fiber local handler to set fiber local setting.

No I didn't propose it actually. In today's dev meeting, we discussed that there are a case where per-Thread mask is needed (e.g., to take an element from an external enumerator), and a case where per-Fiber mask may be needed (I'm unsure about a concrete case though). I just said that, if so, it is theoretically possible to have both Thread.handle_interrupt for per-Thread mask and Fiber.handle_interrupt for per-Fiber mask (maybe per-Fiber mask takes precedence if any). But I have no idea whether per-Fiber mask is practically needed.

[jeremyevans0 (Jeremy Evans)]

• Status changed from Open to Feedback

[Eregon (Benoit Daloze)]

• Status changed from Feedback to Rejected

Closing since it seems intentional for some cases (Enumerator fibers).