Merge parca LuaJIT interpreter (#1236) onto Ruby JIT base (#46)

[dalehamel]

parca LuaJIT interpreter (open-telemetry#1236) merged onto the Ruby-JIT base (#46)

This PR is the clean base for LuaJIT unwinding: the upstream parca LuaJIT interpreter (open-telemetry#1236) merged onto our Ruby-JIT base (#46), with the LuaJIT plugin ported to the current interpreterconfig.Config API (the API moved after open-telemetry#1236 was written) and the BPF blobs rebuilt.

It intentionally contains no Shopify/tarantool-specific behavior — it's just "#46 + upstream open-telemetry#1236, building against current main's interpreter-config API".

Stacked on top

• LuaJIT: statically-linked GC64 tarantool support (amd64 + arm64) #49 — dale/luajit-tarantool-fixes: the tarantool-specific changes (statically-linked GC64 detection, x86/arm64 offset extraction, mem_L/jit_base, native-handback cframe sizes from .eh_frame). That's where the amd64/arm64 tarantool unwinding work lives.

Keeping this PR as the minimal merge makes it easy to track upstream open-telemetry#1236 and to review the Shopify additions separately in #49.

[dalehamel]

Validated end-to-end on Tarantool (shopkv), arm64

Ran the prophiler built from this branch on a shopkv staging Tarantool storage node (--tracers=go,ruby,luajit):

• BPF unwind_luajit program passed the kernel verifier and attached cleanly alongside unwind_ruby.
• First finding: zero Lua frames despite a healthy pipeline (103K native/kernel/daemon samples landed). Root cause: Tarantool statically links LuaJIT into its tarantool binary, and the Loader allowlist only matched libluajit-5.1.so*/luajit*/nginx/openresty.
• Fix (092939b): added tarantool to the Loader allowlist. With LUA_DEBUG=1:

  lj: interp range {2350992 2366704}
  lj: offsets &{g2Dispatch:4080 g2Traces:1200 currentLOffset:480}
  lj: add JIT region pid(...) ...
  New LuaJIT instance detected
  lj: process_client+116      @.../tarantool/http/server.lua:818
  lj: collect_and_serialize+34 @.../metrics/plugins/prometheus.lua:57
  lj: storage_health_state+25 @.../vshard/storage/init.lua:3647
  

  Interpreter-bounds + offset extraction (openresty/luajit2-derived) worked unmodified on Tarantool's LuaJIT fork — 1 instance detected, JIT region added, real symbolized Lua stacks, zero extraction errors.

Filed an upstream note suggesting the executable allowlist be driven by the per-interpreter interpreter.Config (open-telemetry#1436) so embedders like Tarantool can opt in via config: open-telemetry#1236 (comment).

[github-actions]

This PR was marked stale due to lack of activity. It will be closed in 14 days.

[github-actions]

Closed as inactive. Feel free to reopen if this PR is still being worked on.