Misc #20387: Meta-ticket for ASAN support

• Status: Assigned
• Assignee: kjtsanaktsidis (KJ Tsanaktsidis)

[kjtsanaktsidis (KJ Tsanaktsidis)]

I was asked to provide a bit of information about the current status of ASAN in CRuby, so I thought I'd open this meta-ticket to track all of the work I've been performing on fixing up address sanitizer support.

So far, I have fixed the following issues related to ASAN support:

• https://bugs.ruby-lang.org/issues/20001 + https://github.com/ruby/ruby/pull/9505, which dealt with two main themes:
  • Pushing the logic for capturing the start of the machine stack much closer to the top of the call stack, so that VALUEs stored close to the top of the machine stack get marked properly
  • Marking VALUEs stored on ASAN fake stacks during machine stack marking
• https://bugs.ruby-lang.org/issues/20220 + https://github.com/ruby/ruby/pull/9734, which made M:N threading notify ASAN about stack switches in the same way that fibers do
  • Note: ASAN still doesn't work with M:N threading, but that actually has nothing to do with ASAN; it's because the most recent versions of Clang which are needed for ASAN just don't work with M:N threading either. See https://bugs.ruby-lang.org/issues/20243 for more info about that.
• https://bugs.ruby-lang.org/issues/20273 + https://github.com/ruby/ruby/pull/10012, which disables callcc (and the associated tests) when ASAN is enabled
  • callcc is very rarely used in real code and the way it works is just fundamentally incompatible with ASAN (it performs longjmp's which I think are technically undefined behaviour)
• https://bugs.ruby-lang.org/issues/20221 + https://github.com/ruby/ruby/pull/9865, which ignore some global symbols that ASAN defines from the global symbol leak checker
• https://bugs.ruby-lang.org/issues/20274 + https://github.com/ruby/ruby/pull/10087, which ignores some false positive tests about memory leaks when ASAN is enabl
• I updated the ASAN docs in https://github.com/ruby/ruby/pull/9922 to more closely reflect current reality

The current state of things is that, by following the instructions in https://github.com/ruby/ruby/blob/master/doc/contributing/building_ruby.md, you can successfully build Ruby with ASAN enabled, however, the test suite has several failures. I'm currently working on addressing these:

The next step is to merge https://github.com/ruby/ruby/pull/10122 (https://bugs.ruby-lang.org/issues/20310) which I plan to do next week (I'm currently away on a work trip). That makes sure that VALUEs stored in ASAN fake stacks from threads other than the currently running thread get marked during GC.

After that, I need to push up patches for the remaining few issues. I mostly have these patches ready to go already; in fact, last week I got the full make check suite passing all tests with ASAN enabled!

Once that's working, I'd like to investigate how ASAN can fit into CRuby's CI matrix somewhere so that it stays working, although I have not thought too deeply about this yet.

I will provide further updates on this ticket so anybody interested can stay in the loop.

[kjtsanaktsidis (KJ Tsanaktsidis)]

• Related to Bug #20001: Make Ruby work properly with ASAN enabled added

[kjtsanaktsidis (KJ Tsanaktsidis)]

• Related to Bug #20220: M:N threading needs to tell ASAN about stack switches added

[kjtsanaktsidis (KJ Tsanaktsidis)]

• Related to Feature #20273: Disable callcc when compiled with ASAN added

[kjtsanaktsidis (KJ Tsanaktsidis)]

• Related to Bug #20221: ASAN: make test-basic: un-prefixed symbol leakage added

[kjtsanaktsidis (KJ Tsanaktsidis)]

• Related to Feature #20274: Add RubyVM::ASAN.enabled? added

[kjtsanaktsidis (KJ Tsanaktsidis)]

• Related to Bug #20243: M:N threading VM_ASSERT failure in rb_current_execution_context with clang 17 (on Linux) added

[kjtsanaktsidis (KJ Tsanaktsidis)]

• Related to Bug #20310: ASAN fake stacks need to be marked during GC for non-current execution context added

[kjtsanaktsidis (KJ Tsanaktsidis)]

I merged another few PR's related to unit tests in ASAN builds:

https://github.com/ruby/ruby/pull/10383
https://github.com/ruby/ruby/pull/10384
https://github.com/ruby/ruby/pull/10385
https://github.com/ruby/ruby/pull/10386

[kjtsanaktsidis (KJ Tsanaktsidis)]

• Related to Bug #20398: heap-buffer-overflow in numeric literal parsing added

[kjtsanaktsidis (KJ Tsanaktsidis)]

• Related to Bug #20402: Double-free in TestIseqLoad#test_stressful_roundtrip added

[kjtsanaktsidis (KJ Tsanaktsidis)]

OK! Everything required to make make check pass on my machine with ASAN enabled has been merged! I opened https://github.com/ruby/ruby/pull/10412 to update the documentation to include the correct incantation.

The next things I would like to do:

• Find some gem maintainers interested in running their tests under ASAN to see if it shakes out anything interesting
• Could we get an ASAN build of ruby into setup-ruby? This would make it easier to use ASAN with github actions and the like.
• How could we get ASAN plugged into ruby-ci?

I'm going to investigate these things and open tickets as appropriate once I have an idea how they should be tackled.

[Eregon (Benoit Daloze)]

kjtsanaktsidis (KJ Tsanaktsidis) wrote in #note-11:

• Could we get an ASAN build of ruby into setup-ruby? This would make it easier to use ASAN with github actions and the like.

How about making ASAN enabled for ruby-debug builds at https://github.com/ruby/ruby-dev-builder/blob/b0bf59a17c17985d4692243d4689c273f6348fa5/.github/workflows/build.yml#L126-L130 ?

[kjtsanaktsidis (KJ Tsanaktsidis)]

How about making ASAN enabled for ruby-debug builds

That's not a bad idea (avoid adding yet more combinations!) but do you know how widely these builds are currently used and what for? ASAN will make them a heck of a lot slower so that might be undesirable for some uses.

[Eregon (Benoit Daloze)]

kjtsanaktsidis (KJ Tsanaktsidis) wrote in #note-13:

That's not a bad idea (avoid adding yet more combinations!) but do you know how widely these builds are currently used and what for? ASAN will make them a heck of a lot slower so that might be undesirable for some uses.

They are builds using -DRUBY_DEBUG=1 and optflags=-O3 -fno-inline" so I would think already quite a bit slower, so it sounds OK to me.
These ruby-debug builds are currently meant to help CRuby development, i.e., using them on your gem helps finds bugs in CRuby. I don't think there are used very often.
I think the use-case here is close enough and I am not so keen on adding a 3rd kind of CRuby dev build.

[hsbt (Hiroshi SHIBATA)]

• Status changed from Open to Assigned