Project

General

Profile

Actions

Bug #22622

open

Crash when autoload constants are queried in Ractors (autoload?, defined?, const_defined?) while the main Ractor requires files

Bug #22622: Crash when autoload constants are queried in Ractors (autoload?, defined?, const_defined?) while the main Ractor requires files

Added by jamescook83 (James Cook) about 12 hours ago.

Status:
Open
Assignee:
-
Target version:
-
ruby -v:
ruby 4.1.0dev (2026-11-21T12:46:06Z master f2d62c677e) +PRISM [arm64-darwin25]
[ruby-core:127667]

Description

Environment

  • ruby 4.0.10 (2026-11-19 revision a704b7cd95) +PRISM [arm64-darwin25]
  • ruby 4.1.0dev (2026-11-21T12:46:06Z master f2d62c677e) +PRISM [arm64-darwin25]

How this came up

This came up in a test suite that runs each test in its own Ractor, using an autoload-based loader
(Zeitwerk). A few tests run on the main Ractor and require files while the others run. Code in the
Ractor tests checks constants with defined? or const_defined? without loading them, and some of
those constants are autoloads not yet triggered. The process crashes now and then. The script below
is that pattern reduced.

Steps to reproduce

Save the script below as repro.rb and run ruby repro.rb.

The script registers 50 autoloads that are never triggered. Eight Ractors each loop over
Holder.constants(false) calling Holder.autoload?(c, false). At the same time, the main Ractor
writes and requires 400 empty files.

require "tmpdir"
Warning[:experimental] = false

ractors  = (ARGV[0] || 8).to_i
rounds   = (ARGV[1] || 300).to_i
requires = (ARGV[2] || 400).to_i
dir = Dir.mktmpdir

module Holder; end
50.times { |i| Holder.autoload(:"Thing#{i}", File.join(dir, "thing_#{i}.rb")) }

workers = ractors.times.map do
  Ractor.new(rounds) do |n|
    hits = 0
    n.times do
      Holder.constants(false).each { |c| hits += 1 if Holder.autoload?(c, false) }
      Array.new(50) { |k| [k, k.to_s] }
    end
    hits
  end
end

requires.times do |i|
  path = File.join(dir, "extra_#{i}.rb")
  File.write(path, "")
  require path
end

p workers.map(&:value).sum

Expected

The script prints a number and exits.

Actual

Ten runs of the script on each version, each run given 60 seconds. "Crash" is a [BUG] report,
"error exit" a non-zero exit with no [BUG] line, and "hang" a run still going after 60 seconds.

success crash error exit hang
4.0.10 0 10 0 0
4.1.0dev 0 9 0 1

The first lines on 4.0.10:

repro.rb:16: [BUG] Aborted at 0x00000001890b25e8
ruby 4.0.10 (2026-11-19 revision a704b7cd95) +PRISM [arm64-darwin25]
...
-- Control frame information -----------------------------------------------
c:0007 p:---- s:0030 e:000029 l:y b:0001 CFUNC  :autoload?

On 4.1.0dev:

repro.rb:16: [BUG] Segmentation fault at 0x0000000000000004
ruby 4.1.0dev (2026-11-21T12:46:06Z master f2d62c677e) +PRISM [arm64-darwin25]
...
-- Control frame information -----------------------------------------------
c:0007 p:---- s:0030 e:000029 l:y b:0001 CFUNC  :autoload?

The C level backtrace in Ruby's own output was lost to repeated "ABRT received in ABRT handler"
lines. macOS saved crash reports for the 4.1.0dev runs but none for the 4.0.10 runs. Two 4.1.0dev
reports, crashed thread, innermost frames first, below the signal handler:

Abort:

___BUG_IN_CLIENT_OF_LIBMALLOC_POINTER_BEING_FREED_WAS_NOT_ALLOCATED
ruby_xfree_sized
loaded_features_index_clear_i
st_general_foreach
rb_st_foreach
rebuild_loaded_features_index
rb_ec_ensure
rb_feature_p
feature_provided
check_autoload_required
rb_autoload_at_p
rb_mod_autoload_p

Segmentation fault:

st_general_foreach
st_general_foreach
rb_st_foreach
rebuild_loaded_features_index
rb_ec_ensure
rb_feature_p
feature_provided
check_autoload_required
rb_autoload_at_p
rb_mod_autoload_p

In a larger program, a test suite running tests in Ractors, what looks like the same situation
showed up later in GC instead, as
[BUG] try to mark T_NONE object (obj: out-of-heap, parent: out-of-heap) followed by a segfault at
0xfffffffffffffff8.

Variations tried

The same script with the call inside the Ractors swapped, five runs each, 60 seconds a run.

4.0.10:

Call in the Ractors success crash error exit hang
Holder.autoload?(c, false) (as above) 0 5 0 0
defined?(Holder::Thing0) 0 5 0 0
Holder.const_defined?(c, false) 0 5 0 0
const_defined? on a module of plain constants (no autoload) 5 0 0 0
defined? on a plain constant (no autoload) 5 0 0 0
Holder.const_source_location(c) 5 0 0 0
require "tmpdir" (already loaded) 5 0 0 0
Holder.constants(false) only 5 0 0 0
no Ractors, main Ractor requires only 5 0 0 0

4.1.0dev:

Call in the Ractors success crash error exit hang
Holder.autoload?(c, false) (as above) 0 4 1 0
defined?(Holder::Thing0) 0 4 0 1
Holder.const_defined?(c, false) 0 5 0 0
const_defined? on a module of plain constants (no autoload) 5 0 0 0
defined? on a plain constant (no autoload) 5 0 0 0
Holder.const_source_location(c) 5 0 0 0
require "tmpdir" (already loaded) 5 0 0 0
Holder.constants(false) only 5 0 0 0
no Ractors, main Ractor requires only 5 0 0 0

With the require loop removed and 2000 rounds instead of 300, so that only the Ractors call
autoload?, both versions had 10 successes in 10 runs.

The number of autoloads changes how often it crashes. With a single autoload, 4.0.10 crashed in 5 of
5 runs, while 4.1.0dev had 4 successes and 1 hang in 5. From 50 autoloads up, both crashed in every
run or nearly every run.

The number of Ractors, rounds and requires can be changed through the arguments, for example
ruby repro.rb 4 300 400.

No data to display

Actions

Also available in: PDF Atom