Bug #22439: Segfault in `Thread::Backtrace::Location#syntax_tree` (via `Exception#detailed_message` / error_highlight) when called from multiple Ractors

• Status: Open
• Assignee: -
• Target version: -
• ruby -v: ruby 4.1.0dev (2026-10-14T14:33:24Z master b709bdb3e3) +PRISM [x86_64-linux]
• Backport: 3.3: UNKNOWN, 3.4: UNKNOWN, 4.0: UNKNOWN

[amalrik (amalrik maia)]

Calling Exception#detailed_message concurrently from two or more Ractors crashes the VM with
[BUG] Segmentation fault on 4.1.0-preview1 and current master. The same script works on 4.0.6
and 4.0.7.

Reproduction

No gems are involved; error_highlight is loaded by default.

n = Integer(ARGV[0] || 2)
rs = Array.new(n) do |i|
  Ractor.new(i) do |id|
    500.times do |k|
      begin
        raise ArgumentError, "falha #{id}-#{k}"
      rescue => e
        e.detailed_message(highlight: false)
      end
    end
    :ok
  end
end
p rs.map(&:value)

$ ruby repro.rb
<internal:ast>:374: [BUG] Segmentation fault at 0x0000000000000030

Version	Result
4.0.6, 4.0.7	[:ok, :ok] (8/8 runs)
4.1.0-preview1 (2d08b5f42f)	segfault in 4/4 runs with 2 Ractors and 8/8 with 8 Ractors; no crash with 1 Ractor
4.1.0dev master (b709bdb3e3)	segfault 6/6

Backtrace (master)

<internal:ast>:374: [BUG] Segmentation fault at 0x0000000000000030
-- Ruby level backtrace information ----------------------------------------
repro.rb:8:in 'detailed_message'
.../lib/ruby/4.1.0+5/error_highlight/core_ext.rb:43:in 'detailed_message'
.../lib/ruby/4.1.0+5/error_highlight/core_ext.rb:34:in 'generate_snippet'
.../lib/ruby/4.1.0+5/error_highlight/base.rb:57:in 'spot'
.../lib/ruby/4.1.0+5/error_highlight/base.rb:77:in 'find_node_from_loc'
<internal:ast>:478:in 'syntax_tree'
<internal:ast>:374:in 'syntax_tree'

-- C level backtrace information -------------------------------------------
rb_vm_bugreport
rb_ec_ractor_ptr / rb_current_execution_context / rb_current_ractor_raw / rb_current_ractor
vm_ic_update
rb_vm_opt_getconstant_path
vm_exec_core
rb_vm_exec
thread_start_func_2

The full output is attached (crash-output-master.txt).

Additional observations

1. The crash is reachable without error_highlight. Calling
   e.backtrace_locations.first.syntax_tree inside rescue in 2 Ractors gives the same crash and
   stack. It crashed 1/3 runs on master and 0/3 on preview1, so the detailed_message script above
   is the more reliable reproducer.

2. Some things that look similar do not crash.

   • caller_locations(0, 1).first.syntax_tree in 2 Ractors works.
   • Ruby::Prism.parse("1 + 1") in 2 Ractors works.

   So the problem seems specific to backtrace locations taken from an exception, while the
   constant cache in ast.rb is updated from a non-main Ractor.

3. With a single Ractor there is no crash, but error_highlight output regresses. Inside the
   Ractor, detailed_message returns only the message, without the snippet:

   r = Ractor.new do
     begin
       nil.upcase
     rescue => e
       e.detailed_message(highlight: false)
     end
   end
   puts r.value
   # 4.0.7:  undefined method 'upcase' for nil (NoMethodError)\n\n    nil.upcase\n       ^^^^^^^
   # 4.1:    undefined method 'upcase' for nil (NoMethodError)
   

   On 4.1 the snippet is still shown in the main Ractor.

4. This is easy to hit in practice. Any library that logs exceptions with detailed_message
   from inside Ractors can crash the process. I found it with the async gem, whose task-failure
   logging (via the console gem) calls detailed_message, running one reactor per Ractor. On
   4.0.6 that same setup hangs instead of crashing; I have not diagnosed the 4.0.6 hang.

Related (but different) issues

• #22386: segfault when a Ractor helper thread reports an exception during teardown. It was
  fixed before preview1, and this still reproduces on preview1 and master.
• #21183: error_highlight raising Ractor::IsolationError in non-main Ractors (fixed in 2025).

Environment and context

• Linux x86_64 (kernel 7.2, AMD Ryzen 7 5700U, 8C/16T). Also seen on Ubuntu 24.04, AMD EPYC
  Genoa (KVM).
  • These scripts come out of a benchmark study comparing threads, Ractors and Fibers on Ruby 4.0 and
    4.1. The code, raw results and other Ractor findings (I/O concurrency cap on 4.0.x, socket hangs
    on 4.0.x) are public at: https://github.com/amalrik/ruby-ractors-study. The reproducers for this issue are in
    upstream/ruby-4.1-detailed-message-ractors/

[mame (Yusuke Endoh)]

• Related to Bug #22440: Referencing an autoload constant from multiple Ractors dumps core added

[mame (Yusuke Endoh)]

• Related to Bug #22442: Lazy loading of error_highlight etc. is skipped in threads and Ractors added

[mame (Yusuke Endoh)]

Thank you for the report!

I could reproduce the SEGV only with GCC 16 (GCC 13 and 15 did not reproduce it). Are you also using GCC 16 to build Ruby?

This ticket contained two different problems, so I have split it into:

• #22440: SEGV when referencing an autoload constant from multiple Ractors (Ruby::Prism is autoloaded, which caused the crash here)
• #22442: the lazy loading of error_highlight etc. is skipped in threads and Ractors (which is why the snippet was not shown)

If possible, I'd appreciate it if you could test with both patches attached to those tickets applied together.

[amalrik (amalrik maia)]

mame (Yusuke Endoh) wrote in #note-3:

Thank you for the report!

I could reproduce the SEGV only with GCC 16 (GCC 13 and 15 did not reproduce it). Are you also using GCC 16 to build Ruby?

This ticket contained two different problems, so I have split it into:

• #22440: SEGV when referencing an autoload constant from multiple Ractors (Ruby::Prism is autoloaded, which caused the crash here)
• #22442: the lazy loading of error_highlight etc. is skipped in threads and Ractors (which is why the snippet was not shown)

If possible, I'd appreciate it if you could test with both patches attached to those tickets applied together.

Thank you for the quick analysis and for splitting the ticket!

Yes, I'm using GCC 16. Both 4.1 builds I reported were compiled locally on Arch Linux:

ruby 4.1.0preview1 (2026-10-11 master 2d08b5f42f) +PRISM [x86_64-linux]
ruby 4.1.0dev (2026-10-14T14:33:24Z master b709bdb3e3) +PRISM [x86_64-linux]
CC: gcc (GCC) 16.2.1 20260810, optflags -O3 -fno-fast-math

Two corrections to my original report, sorry about them:

• The "Also seen on Ubuntu 24.04, AMD EPYC Genoa" line was wrong. I never ran this reproducer on that machine.
• The 4.0.6/4.0.7 binaries I used as the "works" baseline were prebuilt with GCC 10. So the version comparison in the report is confounded by the compiler, which is consistent with your finding.

I tested both patches applied together: https://github.com/ruby/ruby/pull/19348 and https://github.com/ruby/ruby/pull/19352, cherry-picked onto master 3d168b1. I built with the same GCC 16.2.1 and ran each script N times:

Script	master b709bdb3e3 (no patches)	master 3d168b1 + both patches
#22440 reproducer (autoload constant from 8 Ractors)	10/10 SEGV	10/10 ok
#22439 detailed_message from 2 Ractors	10/10 SEGV	10/10 ok
#22439 e.backtrace_locations.first.syntax_tree from 2 Ractors	1/10 SEGV	10/10 ok
#22442 snippet present in both Ractors (nil.upcase in 2 Ractors)	not testable (SEGV)	5/5 runs, snippet in both

So with both patches, everything in this ticket is fixed on my machine. Thanks!