Bug #20167: Code execution isn't recorded in Ractor

• Status: Closed
• Assignee: ractor
• Target version: -
• ruby -v: 3.3.0
• Backport: 3.0: UNKNOWN, 3.1: UNKNOWN, 3.2: UNKNOWN, 3.3: UNKNOWN

[shia (Sangyong Sim)]

reproduction code

# frozen-string-literal: true

require "coverage"

Coverage.start

require_relative "./some_lib"
# # some_lib.rb
# class C
#   def hoge(i)
#     i
#   end
# end

r = Ractor.new do
  loop do
    v = Ractor.receive
    ret_v = C.new.hoge(v)

    Ractor.yield ret_v
  end
end

2.times do |i|
  r << i
  r.take
end

Coverage.result.each do |file, lines|
  if file.include?("some_ractor_lib")
    if lines == [1, 1, 2, nil, nil]
      puts "OK"
    else
      puts "expected: [1, 1, 2, nil, nil]"
      puts "actual: #{lines.inspect}"
    end
  end
end

Target ruby version

3.2.0+ affected

[mame (Yusuke Endoh)]

At this time, the coverage library should be able to only measure coverage for Ractor that invoked Coverage.start. (I have not tried but invoking Coverage.start in multiple Ractors may cause fatal problems.)

To improve this situation, probabily we need the following big modification.

• Covearge.start registers hooks to all existing Ractors.
• The hook measures coverage data per Ractor, not globally.
• Coverage.result aggregates all coverage data from every Ractor.

[jeremyevans0 (Jeremy Evans)]

• Is duplicate of Bug #20158: Ractor affects Coverage results added

[jhawthorn (John Hawthorn)]

• Assignee set to ractor

[hsbt (Hiroshi SHIBATA)]

• Status changed from Open to Assigned

[ko1 (Koichi Sasada)]

• Status changed from Assigned to Closed

Re-checked: this is fixed, and it was fixed as #19112.

I ran mame's rspec reproduction from #note-3 on four builds. The absolute event counts
differ between versions because they load different numbers of files; what matters is
the with/without-Ractor pair inside one version:

ruby      TracePoint :line events   no Ractor / with Ractor
3.3.2                                  22741 / 8112    <- reproduces, stops inside rubygems
3.4.4                                  35621 / 12738   <- reproduces
4.0.2                                  52294 / 52325   <- fine
master 4d185cce78                       5432 / 5455    <- fine

GH-15468 (4fb537b1ee, "Make tracepoints with set_trace_func or TracePoint.new ractor
local", Fixes [Bug #19112]) is what fixed it: before that, GC'ing any Ractor object
cleared the global event flags and every enabled tracepoint with them, which is exactly
the "TracePoint stops firing part-way through" symptom here.

Closing as a duplicate of #19112.

Note that Coverage still does not record code executed inside a Ractor -- that is
#20167, which is a different cause (the coverage hook is registered on the Ractor that
called Coverage.start, and hook lists are per-Ractor) and stays open.