gh-139540: Fix executor deallocation crash when JIT compilation fails #139952

ashm-dev · 2025-10-11T12:11:29Z

There is a memory leak in make_executor_from_uops (Python/optimizer.c) when JIT compilation fails. The executor object is created via _PyObject_GC_NewVar but _PyObject_GC_TRACK is called after the JIT compilation attempt. If _PyJIT_Compile fails and returns an error, the code calls Py_DECREF(executor) before the object has been tracked by the garbage collector.

Issue: JIT executors are not properly freed #139540

picnixz · 2025-10-11T12:26:43Z

Note that this wouldn't solve the cold executor issue. It would only prevent an assertion failure in case we branch here (but if we branch here, we're already in a very bad state)

ashm-dev · 2025-10-11T12:50:45Z

@picnixz Thanks for the context regarding the "cold executor issue".
Just to clarify the scope of this PR: my understanding is that this is a larger, separate problem to be handled in a follow-up. Is that correct, or should I look into addressing it here as well?

picnixz · 2025-10-11T12:53:32Z

It's a larger and more complex issue. The fact that we have an immortal cold executor seems to make normal executors leak. I don't know if it's because we reset executors in exit sides to a cold executor that then makes that specific executor not garbage-collectable (this is concerning). Your PR would still be fine but I don't know if it won't be eventually rechanged because of the fix we need to do on the other side.

Fidget-Spinner · 2025-10-11T14:41:39Z

Sorry but I have the same PR up here https://github.com/python/cpython/pull/137016/files, since July of this year

The executor object wasn't decremented if Tier 2 code returned with an exception set but not a fatal error. This change moves the Py_DECREF call inside the TIER1_TO_TIER2 macro. This ensures the executor's reference count is always decremented.

picnixz · 2025-10-11T18:31:38Z

Considering this is a duplicate of gh-137016, I'm closing this PR. Sorry!

picnixz · 2025-10-11T18:32:14Z

For the other leaks, ~~please open a separate PR~~. please wait for Mark and other JIT experts to give their opinion because we can't jump on the task without understanding how to properly solve it.

picnixz · 2025-10-11T18:33:36Z

Python/optimizer.c

    if (cold == NULL) {
        Py_FatalError("Cannot allocate core JIT code");
    }
+    _PyObject_GC_TRACK(cold);


~~This is incorrect. The executor is immortal for now.~~

No this was correct, my bad. I didn't read the function correctly. The problem is that we're setting it immortal afterwards and I'm not sure it will be happy.

Good point, thanks for the correction.

The intent was to track it for consistency immediately after allocation. My assumption is the GC correctly handles a tracked object that later becomes immortal.

Is that a safe assumption, or should we explicitly _PyObject_GC_UNTRACK it just before immortalization?

That I don't know. The problem here is whether we want an immortal cold executor or not. That's why I wanted to wait for Mark's output here, and didn't write a PR already.

picnixz · 2025-10-11T18:40:33Z

@Fidget-Spinner Do you mind incorporating some observations from this PR in yours? I don't know whether the change is correct here when changing current_executor etc.

ashm-dev · 2025-10-11T18:40:46Z

Hi @Fidget-Spinner ,

Thanks for pointing that out and linking your PR. I've taken a close look at both.

While our changes overlap in optimizer.c, they actually address two different underlying issues. Your PR #137016 correctly fixes a GC assertion failure by moving _PyObject_GC_TRACK to an earlier point. It's a great catch!

However, this PR primarily addresses a critical reference counting leak that was detectable with ASan. The root cause was that the executor's refcount was not decremented when Tier 2 code exited with a Python exception set (but didn't fatally error out by returning NULL). The main fix here is adding Py_DECREF(EXECUTOR) inside the TIER1_TO_TIER2 macro in ceval_macros.h to ensure the reference is always released.

My PR also includes the same _PyObject_GC_TRACK move that you proposed, so it appears to fix both the memory leak from issue #139540 and the GC assertion issue that your PR addresses.

I believe this PR provides a more complete solution for the stability issues in this code path.

ashm-dev · 2025-10-11T18:44:16Z

I don't know whether the change is correct here when changing current_executor etc

These changes are correct and necessary. tstate->current_executor acts as a temporary channel to pass the executor object from the _ENTER_EXECUTOR instruction to the new Tier 2 frame, which is set up in _PyEval_EvalFrameDefault.

The intended lifecycle is:

_ENTER_EXECUTOR (in bytecodes.c) sets tstate->current_executor.
_PyEval_EvalFrameDefault consumes it at the very start of the new frame, creates a PyStackRef for it, and immediately sets tstate->current_executor back to NULL.

The other NULL/CLEAR additions in ceval.c and pystate.c are for robustness, ensuring we don't leak the executor if a frame or thread is torn down unexpectedly. This appears to be the established design for this handover.

picnixz · 2025-10-11T18:44:40Z

The root cause was that the executor's refcount was not decremented when Tier 2 code exited with a Python exception set (but didn't fatally error out by returning NULL). The main fix here is adding Py_DECREF(EXECUTOR) inside the TIER1_TO_TIER2 macro in ceval_macros.h to ensure the reference is always released.

Does it fix the leak as in the issue? that is, does it fix the refcounting? I can't check this now, but please check whether the reproducer (without ASAN) is fixed when you run ./python -X showrefcount repro.py.

picnixz · 2025-10-11T18:46:11Z

Python/ceval_macros.h

-#define TIER1_TO_TIER2(EXECUTOR)                        \
+#define TIER1_TO_TIER2(EXECUTOR) \
 do {                                                   \
-    OPT_STAT_INC(traces_executed);                     \


Why was this changed? please revert.

picnixz · 2025-10-11T18:46:16Z

Python/ceval_macros.h

 /* Tier-switching macros. */

-#define TIER1_TO_TIER2(EXECUTOR)                        \
+#define TIER1_TO_TIER2(EXECUTOR) \


Why was this changed? please revert.

picnixz · 2025-10-11T18:47:10Z

Python/generated_cases.c.h

+                        tstate->current_executor = (PyObject *)executor;
+                        _PyFrame_SetStackPointer(frame, stack_pointer);
+                        stack_pointer = _PyFrame_GetStackPointer(frame);


This seems to be redundant.

We are setting the stack pointer and requerying it while doing nothing inbetween.

picnixz · 2025-10-11T18:47:31Z

Python/ceval_macros.h

    next_instr = _Py_jit_entry((EXECUTOR), frame, stack_pointer, tstate); \
    frame = tstate->current_frame;                     \
    stack_pointer = _PyFrame_GetStackPointer(frame);   \
+    Py_DECREF(EXECUTOR);                               \


Why should we decref it here?

picnixz · 2025-10-11T18:48:30Z

Python/bytecodes.c

                    assert(tstate->current_executor == NULL);
                    assert(executor != tstate->interp->cold_executor);
                    tstate->jit_exit = NULL;
+                    tstate->current_executor = (PyObject *)executor;


Is it just to hold a reference to the executor?

Track executor object with GC

3a2d34b

ashm-dev requested a review from markshannon as a code owner October 11, 2025 12:11

bedevere-app bot added the awaiting review label Oct 11, 2025

bedevere-app bot mentioned this pull request Oct 11, 2025

JIT executors are not properly freed #139540

Open

picnixz changed the title ~~gh-139540: Fix memory leak in make_executor_from_uops when JIT compilation fails~~ gh-139540: Fix executor deallocation crash when JIT compilation fails Oct 11, 2025

Create 2025-10-11-16-14-36.gh-issue-139549.9BGTLP.rst

963bcf0

ashm-dev force-pushed the gh-139540 branch from 3e1c20a to 963bcf0 Compare October 11, 2025 13:17

ashm-dev requested review from ZeroIntensity and ericsnowcurrently as code owners October 11, 2025 18:27

picnixz closed this Oct 11, 2025

picnixz reviewed Oct 11, 2025

View reviewed changes

picnixz reopened this Oct 11, 2025

picnixz reviewed Oct 11, 2025

View reviewed changes

ashm-dev marked this pull request as draft October 11, 2025 19:22

bedevere-app bot removed the awaiting review label Oct 11, 2025

Uh oh!

gh-139540: Fix executor deallocation crash when JIT compilation fails #139952

Are you sure you want to change the base?

gh-139540: Fix executor deallocation crash when JIT compilation fails #139952

Conversation

ashm-dev commented Oct 11, 2025 • edited by bedevere-app bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

picnixz commented Oct 11, 2025

Uh oh!

ashm-dev commented Oct 11, 2025

Uh oh!

picnixz commented Oct 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Fidget-Spinner commented Oct 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

picnixz commented Oct 11, 2025

Uh oh!

picnixz commented Oct 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

picnixz Oct 11, 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!

picnixz commented Oct 11, 2025

Uh oh!

ashm-dev commented Oct 11, 2025

Uh oh!

ashm-dev commented Oct 11, 2025

Uh oh!

picnixz commented Oct 11, 2025

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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

ashm-dev commented Oct 11, 2025 •

edited by bedevere-app bot

Loading

picnixz commented Oct 11, 2025 •

edited

Loading

Fidget-Spinner commented Oct 11, 2025 •

edited

Loading

picnixz commented Oct 11, 2025 •

edited

Loading

picnixz Oct 11, 2025 •

edited

Loading