[CPAN] File-Data 1.20 loses a live file lock during hash construction

未關閉
#2,067 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
25/100
Issue 類型
缺陷
描述清晰度
描述清楚
活躍度
活躍
技術堆疊
java, perl
領域
backend, testing

研究方向

Start with the overlapping RuntimeIO and glob ownership paths, coordinating with the active migration in #2043 and avoiding the cleanup changes under PR #1938. Read the ownership and cleanup flow through IOOperator and CustomFileChannel, then add the project-owned inline hash literal regression. Done means the test passes on both PerlOnJava backends, the upstream t/test.t passes on both, and abandoned handles are still cleaned up after their final Perl-visible owner is released.

由索引模型根據 Issue 內容生成。

描述

area:io area:memory area:runtime blocked:ownership bug

Summary

File-Data-1.20 fails its locked test on PerlOnJava. A shared-locked read handle becomes unlocked while the test constructs a hash of File::Data objects from inline constructor calls. A second read/write open then succeeds, although the original object remains stored in the hash.

This is a PerlOnJava IO ownership/lifetime bug exposed by CPAN, not a File::Data port or native dependency failure. A direct flock probe behaves correctly, and building the same objects in named variables before inserting them into the hash preserves the lock.

Reproducer

Run from the extracted File-Data-1.20 distribution, with its t/read and t/write fixtures:

use File::Data;

my %file = (
    read  => File::Data->new('./t/read', 'ro'),
    write => File::Data->new('./t/write'),
);

# The upstream test expects this open to fail while $file{read} is alive.
my $again = File::Data->new($file{read}->_var('filename'));
die "read lock was lost" if $again;

System Perl leaves $again undefined. PerlOnJava returns another File::Data object. If the two constructor results are first assigned to named variables and those variables are then placed in %file, the reopen is blocked as expected.

Evidence

  • Distribution: File-Data-1.20 (pure Perl).
  • CPAN random run: 20261115-091204-23477.
  • Upstream failure: t/test.t, test 3 of 16 (locked, line 69); the other 15 tests pass.
  • System Perl 5.42.2 on macOS: prove -v t/test.t passes all 16 tests.
  • PerlOnJava source 455ff0acc3a0872b04b8b870aa1f9ac85984a764, Perl 5.44.1, Java 24.0.2, macOS Darwin 25.6.0 arm64: the upstream test and the reduced hash-construction reproducer fail on both JVM and interpreter backends.
  • Constructing the objects in named variables before hash insertion passes the lock check on both PerlOnJava backends. A direct shared-then-exclusive flock probe also returns the expected blocked result.

Root cause

Anonymous filehandle globs are registered for phantom-reference cleanup in IOOperator. Opening another file assigns a descriptor and processes abandoned globs in RuntimeIO; closing one releases its lock in CustomFileChannel. In the inline hash-literal case, the first handle is staged while the next constructor opens its file, and the cleanup path treats that still-live source-level temporary as abandoned. The hash ends up holding the File::Data object after its underlying handle has already been closed.

The ownership gap is in preserving the anonymous glob/IO across the intermediate list used to build the hash literal. The failing behavior is shared by both backends.

Related work and dependency

  • Blocked by #2043: its active ownership migration replaces reachability-based lifecycle decisions in the overlapping RuntimeIO/glob paths. Coordinate the fix with that work; do not independently replace the cleanup mechanism while PR #1938 is changing it.
  • Related to #2037: that separately tracked Archive::Any failure also exposes IO/object lifetime loss and is deferred until #2043 lands. It has its own reproducer and acceptance criteria.

Acceptance criteria

  • Add a project-owned regression for an inline hash literal that retains both filehandle objects while the second file is opened.
  • Validate the regression with standard Perl first, then on PerlOnJava JVM and interpreter backends.
  • Make File-Data-1.20 t/test.t pass on both PerlOnJava backends.
  • Keep abandoned-handle cleanup working after the final Perl-visible owner is released; a live container value must keep its underlying IO and lock alive.
主要語言
Perl
星號
67
分支
8
平均合併
5 小時 12 分鐘
30 天內合併 PR
131

環境準備

  • 提供 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。

fglock/PerlOnJava 的其他 Issue

查看 fglock/PerlOnJava 的全部 Issue

相似的 Issue

更多 Perl 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。