[PROF-10978] Require Ruby 3.1+ for heap profiling

[ivoanjo]

What does this PR do?

This PR raises the minimum Ruby version required for heap profiling from the previous value of >= 2.7 to >= 3.1 due to a new VM bug discovered (see below for details).

It's mostly a revert of #3366, where we had first tried to workaround a Ruby 2.7/3.0 bug, but it turns out we missed a spot, and we could trigger VM crashes because of that.

Motivation:

Ruby versions prior to 3.1 had a special optimization called rb_gc_force_recycle which would allow objects to directly be garbage collected (e.g. without needing to wait for the GC).

It turns out that rb_gc_force_recycle did not play well with the changes in Ruby 2.7 to how object ids worked. We uncovered this earlier on during the development of the heap profiler, and put in a workaround for the bug that we thought was enough...

Unfortunately, it turns out that the workaround is not enough. The following reproducer, when run on Ruby 2.7 or 3.0 shows how the Ruby VM can segfault inside id2ref due to the issue above:

puts RUBY_DESCRIPTION

require "datadog"
require "objspace"
require "pry"

NUM_OBJECTS = 10_000_000

recycled_ids = Array.new(NUM_OBJECTS) { 123 }
many_objects = Array.new(NUM_OBJECTS) { Object.new }

(0...NUM_OBJECTS).each do |i|
  recycled_ids[i] = many_objects[i].object_id
end

puts "Seeded objects!"
gets

(0...NUM_OBJECTS).each do |i|
  Datadog::Profiling::StackRecorder::Testing._native_gc_force_recycle(many_objects[i])
  many_objects[i] = nil
end

puts GC.stat

puts "Recycled objects!"
gets

many_objects = nil

10.times { GC.start }
Array.new(10_000) { Object.new }
10.times { GC.start }

puts GC.stat

puts "GC'd objects! (Ruby should have released pages?)"
gets

recycled_ids.each { |i|
  begin
    (nil == ObjectSpace._id2ref(i))
  rescue
    nil
  end
}
puts "Done!"

Crash details:

Program received signal SIGSEGV, Segmentation fault.
is_swept_object (ptr=93825033355200, objspace=<optimised out>) at gc.c:3868
3868        return page->flags.before_sweep ? FALSE : TRUE;
(gdb) bt
 #0  is_swept_object (ptr=93825033355200, objspace=<optimised out>) at gc.c:3868
 #1  is_garbage_object (objspace=0x55555555d220, objspace=0x55555555d220, ptr=93825033355200) at gc.c:3887
 #2  is_live_object (ptr=93825033355200, objspace=0x55555555d220) at gc.c:3909
 #3  is_live_object (ptr=93825033355200, objspace=0x55555555d220) at gc.c:3898
 #4  id2ref (objid=8264881) at gc.c:3999
 #5  os_id2ref (os=<optimised out>, objid=<optimised out>) at gc.c:4019

This crash happens because of two things:

1. Ruby does not clean the object id entry for a recycled object from its internal hash map
2. If the memory page where the object lived is returned back to the OS, trying to id2ref on that id will cause Ruby to try to read invalid memory and crash.

Change log entry:

Require Ruby 3.1+ for heap profiling

Additional Notes:

I've chosen to disable heap profiling on 2.7 and 3.0 because I can't think of a good workaround for the bug above, especially not one that does not increase the overhead of heap profiling.

How to test the change?

This PR updates the test coverage to expect Ruby 3.1+ as the minimum for the feature.

You can also quickly validate it doesn't get enabled on the older Rubies using:

$ DD_PROFILING_ENABLED=true DD_PROFILING_EXPERIMENTAL_HEAP_ENABLED=true bundle exec ddprofrb exec ruby -e "puts RUBY_DESCRIPTION"
W, [2024-12-02T10:42:28.771611 #112585]  WARN -- datadog: [datadog] Current Ruby version
(3.0.5) cannot support heap profiling due to VM bugs/limitations. Please upgrade to Ruby
>= 3.1 in order to use this feature. Heap profiling has been disabled.

[datadog-datadog-prod-us1]

Datadog Report

Branch report: ivoanjo/prof-10978-drop-heap-profiling-legacy-ruby
Commit report: 39ffa38
Test service: dd-trace-rb

✅ 0 Failed, 22049 Passed, 1459 Skipped, 5m 41.04s Total Time

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.76%. Comparing base (ca7cc9d) to head (39ffa38).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #4178   +/-   ##
=======================================
  Coverage   97.76%   97.76%           
=======================================
  Files        1357     1357           
  Lines       81950    81891   -59     
  Branches     4168     4164    -4     
=======================================
- Hits        80117    80060   -57     
+ Misses       1833     1831    -2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[pr-commenter]

Benchmarks

Benchmark execution time: 2024-12-02 14:15:36

Comparing candidate commit 39ffa38 in PR branch ivoanjo/prof-10978-drop-heap-profiling-legacy-ruby with baseline commit ca7cc9d in branch master.

Found 0 performance improvements and 2 performance regressions! Performance is the same for 29 metrics, 2 unstable metrics.

scenario:line instrumentation - targeted

• 🟥 throughput [-9605.082op/s; -9076.484op/s] or [-5.817%; -5.497%]

scenario:method instrumentation

• 🟥 throughput [-12720.572op/s; -12151.822op/s] or [-7.211%; -6.888%]

[Strech]

I think it's ok to raise the bar for Ruby version due to costs of patches.

[AlexJF]

Nice spot. Sad about our smart workaround not being enough 🥲