WIP: LuaJIT unwind support #69

SSE4 · 2025-09-21T10:02:53Z

closes: #66

this is heavily work in progress, more like proof of concept, code is totally dirty and incomplete (there are many debug prints, commented code blocks for experiments and so on...).

works only on my version/build of interpreter, offsets are hard-coded (f9140a62)!
it means it obviously won't work on arbitrary version of LuaJIT - if version/flags changed, offsets would likely to change.

mostly inspired by:

article A Walk with LuaJIT
LuaJIT support, designed to support versions of openresty from 1.15+ parca-dev/opentelemetry-ebpf-profiler#14
https://github.com/yunwei37/nginx-lua-ebpf-toolkit
lj_debug_dumpstack function in LuaJIT itself

heavily based on existing Python and PHP unwind implementations in perforator

what is implemented:

unwinding lua stack (still might be buggy)
getting name of the Lua file
getting line number within Lua file
getting number of lua builtin
getting name of metamethod

what is not implemented yet, to be implemented:

debug_framepc returns NO_BCPOS, check isluafunc fails for some reason, to be debugged and fixed
getting name of lua function - lj_debugslotname and its children debug_varname and lj_debug_uvname
automatically compute necessary offsets for an arbitrary version of Lua(JIT)
handle different kinds of jumps between C code and Lua code, such as cpcall, continuation, ffi callback, errfunc, hook, GC, etc.

cezarnik · 2025-09-23T12:58:20Z

Hi, thank you for your PR! We will look at it in this state to understand how to proceed and to save your time.

ayles · 2025-09-24T21:42:08Z

Nice

pashagoose · 2025-10-27T10:49:56Z

perforator/lib/elf/elf.h

+
+        TStringBuf symbolName{name.data(), name.size()};
+        for (const auto& targetSymbol : targetSymbols) {
+            if (symbolName.find(targetSymbol) == 0) {


Can we only match prefix here and avoid using find ?

I've changed to StartsWith (hope it's right).

that's some sort of custom STL, and sources appears to be generated from some "pxd" file. it doesn't make development easier, as I struggled to find documentation for these custom TString classes :)

pashagoose · 2025-10-27T10:53:40Z

perforator/lib/lua/lua.cpp

+template <typename ELFT>
+TMaybe<TParsedLuaVersion> ParseVersion(
+    const llvm::object::ObjectFile& file,
+    const TLuaAnalyzer::TSymbols&symbols


space between type name and argument name

pashagoose · 2025-10-27T10:54:50Z

perforator/lib/lua/lua.h

+    }
+};
+
+enum class EUnicodeType {


should be removed, seems like copy-paste from python

indeed, removed

pashagoose · 2025-10-27T10:57:13Z

perforator/lib/lua/lua.cpp

+
+template <typename ELFT>
+TMaybe<TParsedLuaVersion> ParseVersion(
+    const llvm::object::ObjectFile& file,


no need to pass object file here if ParseVersion just returns version parsed from a symbol name

yes, removed. that was a copy-paste from python.

pashagoose · 2025-10-27T11:00:31Z

perforator/internal/linguist/lua/agent/lua.go

+		return false
+	}
+	return true
+}


Should add newline between functions

pashagoose · 2025-10-27T11:02:57Z

perforator/agent/preprocessing/lib/analyze.cpp


+namespace NPerforator::NBinaryProcessing::NLua {
+
+NPerforator::NBinaryProcessing::NLua::LuaConfig BuildLuaConfig(llvm::object::ObjectFile* objectFile) {


Lets be consistent with other functions and return TMaybe<NPerforator::NBinaryProcessing::NLua::LuaConfig

yeah, I've used a python as a base (skeleton), which did:

NPerforator::NBinaryProcessing::NPython::PythonConfig BuildPythonConfig(llvm::object::ObjectFile* objectFile) {

while PHP does:

TMaybe<NPerforator::NBinaryProcessing::NPhp::PhpConfig> BuildPhpConfig(llvm::object::ObjectFile* objectFile) {

I've changed Lua to return TMaybe

pashagoose · 2025-10-27T11:03:44Z

perforator/agent/preprocessing/lib/analyze.cpp

        *result.MutablePhpConfig() = std::move(phpConfig.GetRef());
    }

+    *result.MutableLuaConfig() = std::move(luaConfig);


Handle null config to be consistent with if statements above

pashagoose · 2025-10-27T11:05:50Z

perforator/agent/collector/pkg/dso/map.go

-				log.String("buildid", buildID),
-				log.Any("version", analysis.PythonConfig.Version),
-			)
+	if analysis.PythonConfig != nil {


Why do we need to change it here ?

was testing something at initial state to see how python initialized. reverted back.

pashagoose · 2025-10-27T11:09:23Z

perforator/agent/collector/pkg/profiler/profiler.go

 	}

+	// Create Lua symbolizer
+	if enabled := p.conf.BPF.TraceLua; enabled == nil || *enabled {


Lets consider that if TraceLua is nil it is defaulted to false.

okay, I've changed

question: python and php/jvm are handled quite a different, are there reasons for that? e.g. php/jvm have their own FeatureFlagsConfig entries, while python doesn't.

until now, I followed more python model while doing LuaJIT code.

SSE4 · 2025-10-28T04:56:19Z

I'll need to rebase code later and clean up the code.
for the record, I have more local changes (on LuaJIT offsets look up), plus there is a wip branch of my colleague to be merged together with mine changes.

since the time I've opened PR we made some progress on these items:

getting name of lua function - lj_debugslotname and its children debug_varname and lj_debug_uvname
automatically compute necessary offsets for an arbitrary version of Lua(JIT)

works only on my version/build of interprerter, offsets are hard-coded!

--debug doesn't load too large program if both Lua and Python are active

it makes easier to identify Lua frames amongst C remove some no longer needed prints

they do not expose LuaJIT_version_* symbol if they happen to expose luaopen_jit, just assume it's 2.1 for now

Formatting, refactoring, new colors, new metrics. Remove debug prints as they no longer useful

…che invalidation. Add more metrics.

…changing BPF_TRACE to BPF_PRINTK for LUA_TRACE

Remove unused enums

…code for these changes.

BigRedEye requested a review from pashagoose September 23, 2025 10:49

pashagoose reviewed Oct 27, 2025

View reviewed changes

perforator/internal/linguist/lua/agent/lua.go

return false

}

return true

}

Copy link

Contributor

pashagoose Oct 27, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should add newline between functions

pashagoose reviewed Oct 27, 2025

View reviewed changes

GitSparTV force-pushed the luajit branch 2 times, most recently from 040c23c to a90e39c Compare November 8, 2025 15:38

SSE4 force-pushed the luajit branch from a90e39c to b899912 Compare November 10, 2025 05:08

GitSparTV force-pushed the luajit branch from ddceca3 to 5603121 Compare November 11, 2025 12:53

SSE4 marked this pull request as draft November 11, 2025 13:34

GitSparTV force-pushed the luajit branch from 1cb369d to 525af7b Compare December 8, 2025 12:24

GitSparTV force-pushed the luajit branch from 525af7b to da7725c Compare December 16, 2025 10:11

SSE4 and others added 10 commits December 19, 2025 18:22

WIP: LuaJIT unwind support

603182c

works only on my version/build of interprerter, offsets are hard-coded!

get rid of hard-coded offset, parse them from asm

c497eea

WIP progress on unwinder

e6b46b1

use real lua_State

bb52d78

temp: disable python

6b48648

--debug doesn't load too large program if both Lua and Python are active

add metrics.py script

bdb6975

show builtin names, hard-coded for now

1a7822f

improve error reporting on broken Lua frames

c6e1ac2

WIP getting function name

83e7a55

Formatting. Removed useless check

1e38b7f

GitSparTV and others added 17 commits December 19, 2025 18:22

Add precise line information

352eca0

WIP reading function names, kinda working, but not 100% cases

5cb2469

show Lua in dark blue colors on FlameGraph

8bb4db1

it makes easier to identify Lua frames amongst C remove some no longer needed prints

fix minor bugs in bpf_debug_varname, remove traces

7da61ba

replace inner loops in bpf_debug_slotname via BPF_MAP array lookup

123b7a1

let unwinder continue after the JIT frame

8c76b68

do not output "Merged stacks percentage NaN%"

9006e38

merge Lua stack

2292261

relax version parsing for older LuaJIT interpreters

c6c2b58

they do not expose LuaJIT_version_* symbol if they happen to expose luaopen_jit, just assume it's 2.1 for now

Fix the main problem with bad frames, find base from register.

8d5ebd4

Formatting, refactoring, new colors, new metrics. Remove debug prints as they no longer useful

Refactoring, formatting, fix 1 bug with lua_global_state_storage ca…

f10e672

…che invalidation. Add more metrics.

Bump luajit headers to 6f21cb8ace60b297cd144c3b6925865b043095d2

4af17e4

Move LuaJIT headers to separate folder

5c20726

Fix headers. Move them to separate folder. Format go files

4fcc9df

Temporary fix bad address error when running without --debug flag by …

34537a4

…changing BPF_TRACE to BPF_PRINTK for LUA_TRACE

Add lj_dispatch.h for bc_mode array size.

2aaa470

Remove unused enums

Refactor symbol_key. Everything is packed into u64 + i32. Refactor …

47547e1

…code for these changes.

GitSparTV force-pushed the luajit branch from da7725c to 47547e1 Compare December 19, 2025 15:23


		namespace NPerforator::NBinaryProcessing::NLua {

		NPerforator::NBinaryProcessing::NLua::LuaConfig BuildLuaConfig(llvm::object::ObjectFile* objectFile) {

WIP: LuaJIT unwind support #69

Are you sure you want to change the base?

WIP: LuaJIT unwind support #69

Uh oh!

Conversation

SSE4 commented Sep 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cezarnik commented Sep 23, 2025

Uh oh!

ayles commented Sep 24, 2025

Uh oh!

pashagoose Oct 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SSE4 commented Oct 28, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

SSE4 commented Sep 21, 2025 •

edited

Loading

pashagoose Oct 27, 2025 •

edited

Loading