Hacktoberfest 2026：メンテナが10月に向けて印を付けた、オープンで初心者向けの issue。 Hacktoberfest の issue を見る

`uuid.uuid4()` can generate duplicates

オープン
  • 何か問題がありますか？
#160,215 コメント 3 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
4/5
見積もり時間
3〜5日
初心者へのやさしさ
42/100
issue の種類
バグ
明瞭さ
おおむね明確
活発さ
活発
技術スタック
linux, python
領域
backend, security

調査の方向性

Start by reading the C UUID implementation in Modules/_uuidmodule.c, especially gen_random_lock_held and the random_buf PID check described in the report. Reproduce the duplicate UUID with the provided PID namespace proof of concept on Linux, then investigate how the cache behaves across process forks and PID reuse. Done means uuid.uuid4() no longer produces duplicate values in the reported scenario, with a regression test covering it.

索引モデルが issue の本文から書いたものです。

説明

extension-modules type-bug
Bug report

https://github.com/python/cpython/pull/139123 rewrites the uuid module to be in C. On its face, that's fine; however, it also adds an entropy cache to the C uuid module. See gen_random_lock_held, with specific attention to this portion.

It contains the following comment:

    // Overfetching & caching entropy improves the performance 10x.
    // There's a precedent with NodeJS doing exact same thing for
    // improving performance of their UUID implementation.

For context, I'm looking because NodeJS today also released CVE-2026-106074, an error in their equivalent cache.

Specifically, the CPython entropy cache (state->random_buf) is in memory; it will regenerate it if:

    if (pid != state->random_last_pid) {

Read this as "current pid != the pid that created the buffer".

Linux (but this applies to at least macOS as well) do not guarantee that PIDs are unique across a boot. In Linux, there are two circumstances in which a child might see the same PID as a parent or ancestor:

  • The PID has been recycled, and the ancestor has died.
  • The process is in a separate PID NS, and is seeing the same PID that an ancestor (including, possibly, the parent) witnessed in a different NS.

Either case, I believe, can be used to generate duplicate UUIDs.

I started with the latter case when I was hacking this up: say I am the first process in a container; in Linux, such a process is "init" for that PID NS, and will always be PID 1. If that process itself creates child inside a new container, that child will also be PID 1. Since the child then sees that the pid is "unchanged" since state->random_buf was generated, it will not re-create the buffer, and we can end up with two processes with the same random_buf buffer.

A proof of concept 
import ctypes
import os
import signal
import sys
import time
import uuid

# Define Linux clone flags
CLONE_NEWPID = 0x20000000
CLONE_NEWNS = 0x00020000
CLONE_NEWUSER = 0x10000000

libc = ctypes.CDLL("libc.so.6", use_errno=True)


def unshare(flags):
    res = libc.unshare(flags)
    if res != 0:
        err = ctypes.get_errno()
        #print("error :(")
        #time.sleep(100)
        raise OSError(err, os.strerror(err))


def dream_child():
    # Unshare PID namespace (and user/mount namespaces for full isolation)
    #unshare(CLONE_NEWPID | CLONE_NEWNS | CLONE_NEWUSER)
    #print("unsharing.")
    unshare(CLONE_NEWPID)

    # Fork is required because the current process stays in the old PID namespace;
    # only its future children enter the new PID namespace.
    pid = os.fork()

    if pid == 0:
        # We are inside the new PID namespace (Child process)
        print(f"Child process PID in *new* new namespace: {os.getpid()}")

        print("In dream level 2.")
        print(f"uuid4() in child  -> {uuid.uuid4()}")
        sys.exit(0)
    else:
        # Parent process waits for child
        _, status = os.waitpid(pid, 0)


def main():
    # Unshare PID namespace (and user/mount namespaces for full isolation)
    unshare(CLONE_NEWPID)

    # Fork is required because the current process stays in the old PID namespace;
    # only its future children enter the new PID namespace.
    pid = os.fork()

    if pid == 0:
        # We are inside the new PID namespace (Child process)
        print(f"Child process PID in new namespace: {os.getpid()}")

        # Optional: Mount a new /proc so tools like 'ps' work properly
        # (Requires CLONE_NEWNS)
        try:
            os.mount("proc", "/proc", "proc", 0, "")
        except Exception:
            pass

        # Burn a UUID to set the cache up.
        uuid.uuid4()
        print("In dream level 1.")
        dream_child()
        print(f"uuid4() in parent -> {uuid.uuid4()}")
    else:
        # Parent process waits for child
        _, status = os.waitpid(pid, 0)
        sys.exit(os.waitstatus_to_exitcode(status))


if __name__ == "__main__":
    main()

On my system, this Python script emits:

» sudo ~/code/third-party/cpython/python unshare_and_uuid.py
Password: 
Child process PID in new namespace: 1
In dream level 1.
Child process PID in *new* new namespace: 1
In dream level 2.
uuid4() in child  -> 1df59e77-c35a-4921-afa7-98e653300c67
uuid4() in parent -> 1df59e77-c35a-4921-afa7-98e653300c67

Note that you do not need PID NSes for this to work. The first bullet in the list above would function something like,

  1. In the initial process, call uuid.uuid4() to prime random_buf; let's call this pid A, so random_last_pid is A.
  2. Fork, let the parent die; PID A is now free.
  3. In a loop, spawn children until you obtain PID A. In the child, if the child's PID is A, generate a UUID.
  4. All children where the child PID is A will obtain the same UUID.

(This has the benefit of not needing ctypes or PID NSes to accomplish.)

(Note that the PID NS example requires root, in order to unshare the PID NS. Unsharing the PID NS as a normal user should also be possible, but it requires more setup that I didn't want to hassle with. I.e., it should be possible to write a more complex PoC that doesn't require sudo, so I wouldn't focus on that.)

(Admittedly, the circumstances required for this might be difficult to run into.)

CPython versions tested on:

CPython main branch

(n.b., at the time of writing, this only applies to main. The latest released version, 3.15.1, does not contain the bug discussed here.)

Operating systems tested on:

Linux

主要言語
Python
スター
77.8k
フォーク
37.7k
平均マージ
12時間 27分
マージ済み PR（30日）
710

環境構築

Codespaces で開く

このプロジェクトの開発コンテナを、あなたの GitHub アカウントでブラウザ上に起動します。

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

python/cpython のほかの issue

python/cpython の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。