ruby: register grpc_rb_sStatus as a global variable

[casperisfine]

Ref: https://bugs.ruby-lang.org/issues/20311

C global variable that contain references to Ruby object MUST be declared to the Ruby GC to prevent these objects from being collected or moved.

There are a few exceptions to that, such as classes defined using the C APIs such as rb_define_class etc.

Up to Ruby 3.4 however, there was a bug that caused classes created from Ruby with the Struct.new("Name") API to also be marked as immortal and immovable.

GRPC has been relying on this bug, which I fixed in Ruby 3.4, and now GRPC is crashing when Struct::Status is moved around by the GC.

-- C level backtrace information -------------------------------------------
ruby(rb_print_backtrace+0x14) [0x5577db219d41] ruby-3.4.0/vm_dump.c:820
ruby(rb_vm_bugreport) ruby-3.4.0/vm_dump.c:1151
ruby(rb_bug_for_fatal_signal+0xfc) [0x5577db3cc60c] ruby-3.4.0/error.c:1066
ruby(sigsegv+0x4d) [0x5577db16358d] ruby-3.4.0/signal.c:926
libc.so.6(0x7f5bacd32520) [0x7f5bacd32520]
ruby(rb_class_superclass+0x32) [0x5577db0a9152] ruby-3.4.0/object.c:2239
ruby(struct_ivar_get+0x2a) [0x5577db194e02] ruby-3.4.0/struct.c:49
ruby(struct_ivar_get) ruby-3.4.0/struct.c:40
ruby(num_members) ruby-3.4.0/struct.c:705
ruby(rb_struct_new+0x56) [0x5577db19a9d6] ruby-3.4.0/struct.c:848
lib/grpc/grpc_c.so(grpc_run_batch_stack_build_result+0xe6) [0x7f5b84bb6b96] /tmp/bundle/ruby/3.4.0+0/bundler/gems/grpc-5ed33ee673e3/src/ruby/ext/grpc/rb_call.c:780
lib/grpc/grpc_c.so(grpc_rb_call_run_batch_try) /tmp/bundle/ruby/3.4.0+0/bundler/gems/grpc-5ed33ee673e3/src/ruby/ext/grpc/rb_call.c:839
ruby(rb_ensure+0x10c) [0x5577db007d3c] ruby-3.4.0/eval.c:1000
/tmp/bundle/ruby/3.4.0+0/bundler/gems/grpc-5ed33ee673e3/src/ruby/lib/grpc/grpc_c.so(grpc_rb_call_run_batch+0xca) [0x7f5b84bb595a] /tmp/bundle/ruby/3.4.0+0/bundler/gems/grpc-5ed33ee673e3/src/ruby/ext/grpc/rb_call.c:893
ruby(vm_call_cfunc_with_frame_+0x117) [0x5577db1f4c77] ruby-3.4.0/vm_insnhelper.c:3524

cc @apolcyn @peterzhu2118 @k0kubun

[casperisfine]

NB: I grepped to see if some others Strucs may be impacted, but didn't find any other in this case.

[apolcyn]

Thanks for the fix! one question

[apolcyn]

does grpc_rb_sNewServerRpc also need this?

Also wondering about grpc_rb_sBatchResult in rb_call.c.

[casperisfine]

No because rb_struct_define, like rb_define_class all basically all the C API functions that end up creating a class or module, do "root" that object, which means the GC will never collect nor move it.

grpc_rb_sStatus need it only because that class is defined in Ruby, and grabbed in C via rb_const_get.

[casperisfine]

An alternative could be to define that struct in C like the other (and Struct.new("Status") should probably be avoided altogether as it might clash with other gems), but I went for the fix the the minimal amount of changes.

[casperisfine]

which I fixed in Ruby 3.4,

Another quick note. It's currently fixed in 3.4-dev, but the bug is marked as needing backport, so eventually it will make it into point release of 3.3, 3.2 and 3.1.

[ngan]

Just want to add that the fix was also backported to (and released in) Ruby 3.3.1.