[Bug]: JVM crash (SIGSEGV in Py_REFCNT) on numpy.frombuffer(memoryview(java.nio.ByteBuffer))

已關閉
#1,198 1 則留言 0 個 reaction 已指派 1 人 在 GitHub 檢視

維護者通常 1 天內回覆

@msimacek 已經在處理了。

開始於 2026年10月20日。

評估

難度
4/5
預估耗時
3-5 天
新手友好度
44/100
Issue 類型
缺陷
描述清晰度
基本清楚
活躍度
活躍
技術堆疊
java, numpy, python
領域
backend

研究方向

Start by reproducing the failure with the provided numpy.frombuffer(memoryview(buffer)) example under GraalPy, then trace how Java ByteBuffer memoryviews are exposed to C extensions and how the buffer protocol reaches NumPy. The issue does not name relevant source files or tests, so finding the responsible interop path will take codebase exploration. Done means the example returns a valid four-element array or raises a Python exception instead of crashing the JVM.

由索引模型根據 Issue 內容生成。

描述

Describe the bug

Calling numpy.frombuffer on a memoryview of a Java java.nio.ByteBuffer crashes the JVM with SIGSEGV in Py_REFCNT (libpython-native.so). A Python-level error, or a working zero-copy view, would be expected.

Operating system

Linux (NixOS), kernel 6.18

CPU architecture

x86_64

GraalPy version

25.4.4 (python-language 25.4.4.1.1, embedded via Micronaut/Pyronaut)

JDK version

GraalVM CE 25.4.4.1.1 (JDK 25.0.4)

Context configuration

Embedded context built by Micronaut's GraalPyContextFactory (Pyronaut 0.1.0): allowNativeAccess(true), allowHostAccess(...), allowHostClassLookup(...), allowValueSharing(true), allowPolyglotAccess(ALL), the java posix backend, and a venv with numpy 2.4.4 from the GraalPy wheel index (numpy-2.4.4-graalpy313-graalpy253_313_native-manylinux_2_27_x86_64.whl).

Steps to reproduce
import numpy as np
from java.nio import ByteBuffer, ByteOrder

buffer = ByteBuffer.allocate(16).order(ByteOrder.LITTLE_ENDIAN)
print(np.frombuffer(memoryview(buffer), dtype="<f4"))  # JVM dies here

This was run as a pytest test under pyronaut test. It crashes on every run.

Expected behavior

Either a 4-element float32 array, sharing or copying the buffer's memory, or a Python exception if host buffers can't be exported through the buffer protocol to C extensions.

Stack trace
#  SIGSEGV (0xb) at pc=0x00007fff79ad2db5, pid=174452, tid=174453
# Problematic frame:
# C  [libpython-native.so+0x23db5]  Py_REFCNT+0x5

C  [libpython-native.so+0x23db5]  Py_REFCNT+0x5
j  java.lang.invoke.LambdaForm$MH+0x0000000064e44c00.invoke(Ljava/lang/Object;JJJJ)J+64 [email protected]
j  com.oracle.graal.python.builtins.objects.cext.capi.ExternalFunctionInvoker.invokePYCFUNCTION_WITH_KEYWORDS(JJJJ)J+9
j  com.oracle.graal.python.builtins.objects.cext.capi.ExternalFunctionInvoker.invokePYCFUNCTION_WITH_KEYWORDS(...)
j  com.oracle.graal.python.builtins.objects.cext.capi.ExternalFunctionNodes$MethKeywordsRoot.readArgumentsAndInvokeExternalFunction(...)
j  com.oracle.graal.python.builtins.objects.PythonAbstractObject$PExecuteNode.doExecute(...)
Additional context

Related: np.asarray(java_float_array) on a Java float[] fails with RecursionError: maximum recursion depth exceeded. As a result there is currently no efficient way to move a large Java primitive array into numpy. Our workaround is to write the data to a file from Java and np.frombuffer(f.read()) it in Python.

主要語言
Python
星號
1.7k
分支
163
平均合併
1 天 4 小時
30 天內合併 PR
44

環境準備

在 Codespaces 中開啟

在瀏覽器裡用你自己的 GitHub 帳號啟動這個專案的開發容器。

  • 沒有 Dockerfile 或 Docker Compose 檔案
  • 沒有 Pull Request 範本
  • 閱讀貢獻指南

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

oracle/graalpython 的其他 Issue

查看 oracle/graalpython 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。