Bug #19364: Issue with tracepoint enable/disable across ractors

• Status: Closed
• Assignee: ractor
• Target version: -
• Backport: 2.7: UNKNOWN, 3.0: UNKNOWN, 3.1: UNKNOWN, 3.2: UNKNOWN

[luke-gru (Luke Gruber)]

This sometimes segfaults:

def test_enable_disable_in_multiple_ractors_with_target
  rs = []
  100.times do |i|
    # setup new iseqs
    Kernel.define_method :"my_method_to_change_for_tracing_#{i}" do
      true
    end
  end
  100.times do |i|
    rs << Ractor.new(i) do |j|
      meth = :"my_method_to_change_for_tracing_#{j}"
      tp = TracePoint.new(:line) { } # local to ractor
      100.times do
        tp.enable(target: method(meth)) # change iseq internals of given method, should be done with lock
        tp.disable # disable hooks should hold lock too, changes method definition internals
      end
    end
  end
  rs.each(&:take) # shouldn't raise
end
test_enable_disable_in_multiple_ractors_with_target()

Changing iseq internals is done without the VM lock. This is true in Tracepoint#enable and Tracepoint#disable methods.
I have a patch coming.

[hsbt (Hiroshi SHIBATA)]

• Status changed from Open to Assigned
• Assignee set to ko1 (Koichi Sasada)

[jhawthorn (John Hawthorn)]

• Assignee changed from ko1 (Koichi Sasada) to ractor

[ko1 (Koichi Sasada)]

• Status changed from Assigned to Closed

Re-checked on master 4d185cce78 with a RUBY_DEBUG=1 build: the script in the description
runs clean, 30 times out of 30.

The cause you named -- "changing iseq internals is done without the VM lock" -- has been
addressed. rb_tracepoint_enable_for_target() in vm_trace.c now stops the world before
rewriting the iseq:

RB_VM_LOCKING() {
    // Rewriting iseq instructions across ractors is not safe unless they are stopped.
    rb_vm_barrier();
    ...