Implement native profiling #373
Conversation
There was a problem hiding this comment.
Thanks for this! The implementation looks good. It doesn't currently build on macOS because some functions and variables are hidden by the cfg directives.
I'm also thinking about test coverage. Maybe we can do something similar to the existing tests that load a core dump and get a stack trace from it. If you have ideas, feel free to share and/or try them.
Next steps:
- Please rebase onto main (should be easy - minor fixes and dependency upgrades), and then I'll approve the build so that CI can tell us what's broken. Let me know if you'd like help with the macOS build
- I'll put this through its paces in my Linux VM and will think about testing strategies a bit more
| @@ -106,15 +127,36 @@ impl RubySpy { | |||
| } | |||
| } | |||
|
|
|||
| fn get_trace_from_current_thread(&self, lock_process: bool) -> Result<StackTrace> { | |||
| fn get_trace_from_current_thread(&mut self, lock_process: bool) -> Result<StackTrace> { | |||
| if self.native_profiling && cfg!(any(windows, target_os = "linux")) { | |||
There was a problem hiding this comment.
| if self.native_profiling && cfg!(any(windows, target_os = "linux")) { | |
| if self.native_profiling && cfg!(any(target_os = "windows", target_os = "linux")) { |
| break; | ||
| } | ||
|
|
||
| // TODO: find a way to filter out internal ruby core functions, and make it configurable? |
| let mut frames = vec![]; | ||
| for (i, ruby_frame) in ruby_frames.iter().enumerate() { | ||
| let is_cfunc = ruby_frame.name.contains("[c function]"); | ||
| let mut found_start_of_cfunc = false || (i == 0 && is_cfunc); |
There was a problem hiding this comment.
| let mut found_start_of_cfunc = false || (i == 0 && is_cfunc); | |
| let mut found_start_of_cfunc = i == 0 && is_cfunc; |
| .function | ||
| .as_ref() | ||
| .map(|s| s as &str) | ||
| .unwrap_or_else(|| "cfunc (unnamed)".into()); |
There was a problem hiding this comment.
If there's a chance that this will be printed, it should probably use the "[c function]" notation like the existing C func code does.
| match self.symbolicator.symbolicate(ip, true, func) { | ||
| Ok(_) => Ok(()), | ||
| Err(e) => { | ||
| // Do not reload more than once per IP which is not found |
There was a problem hiding this comment.
Just curious, is this a performance optimization? Or does it cause an error if we reload more than once per IP?
| @@ -437,6 +446,8 @@ impl Args { | |||
| }; | |||
|
|
|||
| let no_drop_root = submatches.occurrences_of("no-drop-root") == 1; | |||
| let native_profiling = submatches.occurrences_of("native-profiling") == 1; | |||
There was a problem hiding this comment.
Might be able to use is_present here like we do for the boolean options below.
resolves #372
How does it work?
When the
native-profilingflag is on and we are on linux or windows (remoteprocess only supports unwinding on these 2 platforms afaik) then, after getting the ruby trace, we:The start of a c function is:
rb_vm_execand the end of it is:vm_call_cfunc_with_frame.By interweaving native and ruby frames this way, we get to profile both ruby and native code!
Limitations
It seems like this may not work with YJIT, as it can get rid of some
rb_vm_execcalls and can optimisevm_call_cfunc_with_frame