[trevorturk]

Found while investigating #7.

Fiber_Profiler_Capture_Call_initialize resets id but not klass. When there's no method frame, as with a block defined outside a method or a GC outside a method, rb_frame_method_id_and_class returns false and leaves klass untouched. So these calls report whatever class their slot held: false for a fresh slot, or an earlier call's class for a reused one.

Once the earlier call is gone, the capture no longer marks that class, so it may have been freed or moved. After GC.compact I've seen reports name unrelated objects, including strings whose quotes produced invalid JSON.

Reproduction:

require "fiber/profiler"
require "stringio"
require "json"

class Foo
	def bar = baz
	def baz = nil
end

output = StringIO.new
capture = Fiber::Profiler::Capture.new(stall_threshold: 0.001, output: output)
capture.start

Fiber.new do
	Foo.new.bar
	[1, 2].each {sleep 0.005}
end.resume

capture.stop

JSON.parse(output.string)["calls"].each do |call|
	puts call.values_at("path", "line", "class", "method").join(" ")
end

On main (Ruby 4.0.7 and 3.4.10), the block on line 16 is reported as Foo, left over from the earlier Foo#baz call:

stale.rb 15 false (null)
stale.rb 16 Foo (null)
stale.rb 16 Kernel sleep
stale.rb 16 false (null)
stale.rb 16 Kernel sleep

This change initializes klass to nil along with id, so both block calls are reported as nil, or <unknown> with #8.

The added test runs a top-level block (inside sus, every it block is itself within a method). It fails before the change, reporting ["false", "String", "Kernel", "false", "Kernel"], and passes after. It doesn't depend on the label, so it works with or without #8. bundle exec sus: 41/41 on Ruby 4.0.7 (YJIT) and 3.4.10.

[samuel-williams-shopify]

Thanks!