Use TracePoint.allow_reentry when available

[casperisfine]

Fix: #408

This fixes compatibility with TracePoint users such as Zeitwerk

Test script:

require'fileutils'FileUtils.mkdir_p('/tmp/lib/foo')File.write('/tmp/lib/foo.rb','module Foo; end')File.write('/tmp/lib/foo/bar.rb','Foo::Bar = 1')require'zeitwerk'loader=Zeitwerk::Loader.newloader.push_dir('/tmp/lib')loader.setuprequire'debug'binding.breakpFoo::Bar

before:

[8, 16] in /tmp/debug-zeitwerk.rb
8| loader = Zeitwerk::Loader.new
9| loader.push_dir('/tmp/lib')
10| loader.setup
11|
12| require 'debug'
=> 13| binding.break
14| p Foo::Bar
15|
16| p :done
=>#0 <main> at /tmp/debug-zeitwerk.rb:13
(rdbg) p Foo::Bar # command
eval error: uninitialized constant Foo::Bar
(rdbg)//tmp/debug-zeitwerk.rb:1:in `<main>'
=> nil

after:

[8, 16] in /tmp/debug-zeitwerk.rb
8| loader = Zeitwerk::Loader.new
9| loader.push_dir('/tmp/lib')
10| loader.setup
11|
12| require 'debug'
=> 13| binding.break
14| p Foo::Bar
15|
16| p :done
=>#0 <main> at /tmp/debug-zeitwerk.rb:13
(rdbg) p Foo::Bar # command
=> 1

Note that I'm fairly unfamiliar with this codebase, so I'll try to see how to add tests, but it may take me a while. I figured it was worth starting a PR anyway.

[st0012]

@casperisfine I'm not sure why the tests are failing and I can't reproduce them locally 😕
I'll also investigate this when I after time this week.
*Update: I can reproduce it now

[st0012]

There is an issue that breakpoints can now be triggered inside another breakpoint's subsession.

The debugger hasn't supported nested suspension yet. So currently, having multiple subsessions can only be caused by bugs and the debugger raises an exception when that happens. But with allow_reentry, it's possible to trigger nested subsessions with normal usages (see example). And this crashing behavior becomes a problem.

Example

Script

deffoo10endbinding.b

Commands

$ exe/rdbg -e "break Object#foo ;; c ;; foo" test_540.rb

1. break Object#foo - set breakpoint on foo
2. c - continue the program and enter the binding.b breakpoint
3. foo - call foo and trigger the breakpoint we set earlier

Result

(debug log are enabled by merging #469 locally)

❯ exe/rdbg -e "break Object#foo ;; c ;; foo ;; c" test_540.rb
# not important
DEBUGGER (DEBUG): #<DBG:TC 1:running@test_540.rb:5:in `<main>'> changes mode (running -> waiting)
DEBUGGER (DEBUG): #<DBG:TC 1:waiting@test_540.rb:5:in `<main>'> sends Event { type: :suspend, args: [":breakpoint", "[\"/Users/st0012/projects/debug/lib/debug/session.rb\", 2227]"] } to Session
# enters binding.b subsession
DEBUGGER (INFO): enter_subsession
[1, 5] in test_540.rb
1| def foo
2| 10
3| end
4|
=> 5| binding.b
=>#0 <main> at test_540.rb:5
(rdbg:commands) foo
DEBUGGER (DEBUG): #<DBG:TC 1:waiting@test_540.rb:5:in `<main>'> receives Cmd { type: :eval, args: [":pp", "\"foo\""] } from Session
DEBUGGER (DEBUG): #<DBG:TC 1:waiting@test_540.rb:5:in `<main>'> changes mode (waiting -> running)
DEBUGGER (DEBUG): #<DBG:TC 1:running@test_540.rb:5:in `<main>'> changes mode (running -> waiting)
DEBUGGER (DEBUG): #<DBG:TC 1:waiting@test_540.rb:5:in `<main>'> sends Event { type: :load, args: ["<RubyVM::InstructionSequence:<main>@(rdbg)/test_540.rb:1>", "\"foo\""] } to Session
DEBUGGER (INFO): Load (rdbg)/test_540.rb
DEBUGGER (DEBUG): #<DBG:TC 1:waiting@test_540.rb:5:in `<main>'> receives Cmd { type: :continue, args: [] } from Session
DEBUGGER (DEBUG): #<DBG:TC 1:waiting@test_540.rb:5:in `<main>'> changes mode (waiting -> running)
DEBUGGER (DEBUG): #<DBG:TC 1:running@test_540.rb:5:in `<main>'> is suspended for :call
DEBUGGER (DEBUG): #<DBG:TC 1:running@test_540.rb:5:in `<main>'> changes mode (running -> waiting)
# tries to enter another subsession because of the breakpoint we set
DEBUGGER (DEBUG): #<DBG:TC 1:waiting@test_540.rb:5:in `<main>'> sends Event { type: :suspend, args: [":breakpoint", "\"Object#foo\""] } to Session
#<Thread:0x000000010cbd6f98@DEBUGGER__::SESSION@server /Users/st0012/projects/debug/lib/debug/session.rb:149 run> terminated with exception (report_on_exception is true):
/Users/st0012/projects/debug/lib/debug/session.rb:1537:in `enter_subsession': already in subsession (RuntimeError)
from /Users/st0012/projects/debug/lib/debug/session.rb:235:in `process_event'
from /Users/st0012/projects/debug/lib/debug/session.rb:199:in `session_server_main'
from /Users/st0012/projects/debug/lib/debug/session.rb:169:in `block in activate'
DEBUGGER (DEBUG): #<DBG:TC 1:waiting@test_540.rb:5:in `<main>'> changes mode (waiting -> running)
["DEBUGGER Exception: /Users/st0012/projects/debug/lib/debug/thread_client.rb:1011",
#<RuntimeError: already in subsession>,
["/Users/st0012/projects/debug/lib/debug/session.rb:1537:in `enter_subsession'", "/Users/st0012/projects/debug/lib/debug/session.rb:235:in `process_event'", "/Users/st0012/projects/debug/lib/debug/session.rb:199:in `session_server_main'", "/Users/st0012/projects/debug/lib/debug/session.rb:169:in `block in activate'"]]
# not important

I think to solve #408 we probably need more preparation, like

• Supporting nested subsessions with allow_reentry, which is a great feature but will only work for Ruby 3.1+. and the inconsistent behavior may confuse users.
• Or, adjust the current subsession check to ignore the later subsessions instead of raising an error.

But this is @ko1's call.

[ko1]

Yes, supporting nested subsession should be supported with this change.

Supporting nested subsessions with allow_reentry, which is a great feature but will only work for Ruby 3.1+. and the inconsistent behavior may confuse users.

We can say "please update to 3.1".

[casperisfine]

Yeah, Zeitwerk users have been confused for a while by debuggers breaking Zeitwerk in specific circumstances, It wouldn't make sense to me to delay the fix for consistency. Inconsistently working is better than consistently broken.

[st0012]

Inconsistently working is better than consistently broken.

To be clear, I didn't mean keep the broken behavior. I only mean triggering breakpoint inside another breakpoint will be ignored (no nested subsessions) but this conflict with Zeitwerk (and other TracePoint using gems) will be fixed.

[ko1]

I think nested subsession (nested breakpoint) should be supported, if it is possible.
Anyway, I'll rewrite it, so I close this PR.