Implement a safe API wrapping PyEval_SetProfile #4039
Conversation
| } | ||
|
|
||
| pub fn register_profiler<P: Profiler>(profiler: Bound<'_, P>) { | ||
| unsafe { ffi::PyEval_SetProfile(Some(profile_callback::<P>), profiler.into_ptr()) }; |
There was a problem hiding this comment.
Do you also want to support the new PyEval_SetProfileAllThreads function?
Also, there should be a way to clear tracing. Either with a separate function, or accepting an Option argument.
Finally, why needless change the name from setprofile / SetProfile to register_profiler?
There was a problem hiding this comment.
Yeah, I think supporting PyEval_SetProfileAllThreads makes sense. But I'll get the main bit working fully first.
The register_profiler naming was also inspired by the implementation in #4008 (comment). I'd prefer to decide on final naming once the implementation is complete. setprofile could well make sense for symmetry.
| register_profiler(profiler); | ||
|
|
||
| py.run_bound(PYTHON_CODE, None, None).unwrap(); | ||
|
|
There was a problem hiding this comment.
Sure. This was just where I got to before I had to head to bed.
CodSpeed Performance ReportMerging #4039 will not alter performanceComparing Summary
|
| Return(Option<Bound<'py, PyAny>>), | ||
| /// A C function is about to be called. The contained data is the | ||
| /// function object being called. | ||
| CCall(Bound<'py, PyAny>), |
There was a problem hiding this comment.
Should this be PyCFunction rather than PyAny here? (not sure, just wondering)
| // | ||
| // `frame` is an `ffi::PyFrameObject` which can be converted safely to a `PyObject`. | ||
| let frame = frame as *mut ffi::PyObject; | ||
| Python::with_gil(|py| { |
There was a problem hiding this comment.
Is the profiler callback always called with the GIL held? If so, we can probably do something more like the code in trampoline.rs (we might even want to reuse that trampoline to avoid panic unwinds escaping the callback).
There was a problem hiding this comment.
I can double check.
| // We borrow the object so we don't break reference counting. | ||
| // | ||
| // https://docs.python.org/3/c-api/init.html#c.Py_tracefunc | ||
| let frame = unsafe { PyObject::from_borrowed_ptr(py, frame) }; |
There was a problem hiding this comment.
Probably can assume type is correct and do something like frame.assume_borrowed(py).downcast _unchecked().
There was a problem hiding this comment.
Yeah, I'd had that idea myself - just hadn't got around to changing it.
| } | ||
|
|
||
| /// Trait for Rust structs that can be used with Python's profiling API. | ||
| pub trait Profiler: PyClass<Frozen = False> { |
There was a problem hiding this comment.
The Profiler trait as a subtype of PyClass is an interesting choice. Is that strictly necessary? Is that for the borrow checking? I wonder how this might interact with nogil. I also whether something like what we've got with run_closure over in the function code might work (e.g. just register a Fn() closure rather than a full type).
There was a problem hiding this comment.
I think it is necessary, but maybe there's an alternative that avoids it that I missed.
No worries - it's more blocked on me finding the time to continue it. |
Fixes #4008.