Windows: endpoints hang or crash when several bind at once (concurrent WMI default-route queries?)

オープン
#271 コメント 0 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 2 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
55/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
rust

調査の方向性

Start with src/interfaces/windows.rs and inspect get_default_route, which the issue identifies as the WMI query path. Compare it with the reported GetIpForwardTable2(AF_INET) approach and check how Windows dependencies and route detection are handled. Done means concurrent endpoint binds no longer hang or crash on Windows while default route detection retains the described behavior.

索引モデルが issue の本文から書いたものです。

説明

Describe the bug

We're seeing this in our project and wanted to share it in case it's useful.

On Windows, when several iroh endpoints are created at the same time in one process, Endpoint::bind()
sometimes hangs (it never returns) or the process crashes. In our CI, a test that binds five endpoints at once
on loopback failed in about 3 of every 4 runs on GitHub's Windows runners. It never fails on Linux or macOS.

This looks related to #124, but it still happens with the spawn_blocking fix from #127 (we're on 0.19.3).
The difference here is several WMI queries running at the same time, one per netmon::Monitor.

What we found:

  • The hangs and crashes happen inside the WMI query that interfaces::windows::get_default_route runs over
    COM (Win32_IP4RouteTable). With full page heap enabled, our runs stopped on freed memory inside COM during
    those queries (combase!OXIDEntry::DecRefCnt, CROIDTable::ClientDeRegisterOIDFromPingServer, and once in
    an already unloaded wbemprox.dll).
  • Initializing COM ourselves first (MTA, security) and keeping a WMI connection open didn't help.
  • Reading the route table with GetIpForwardTable2(AF_INET) and checking for a route with prefix length 0
    avoids COM entirely. With only that change, the failures stopped: our Windows CI has passed every run since,
    including repeated runs of the same commit.

A side note we noticed: Win32_IP4RouteTable.Name is the route's destination, not an interface name, so on
Windows default_route_interface is always "0.0.0.0". The replacement below keeps that value, so
is_major_change behaves the same. Returning the real interface (the row's InterfaceIndex) could be a
follow-up.

This is the replacement we're using in a patched copy (it drops the wmi and serde dependencies on
Windows):

fn get_default_route() -> Result<DefaultRouteDetails, Error> {
    let mut table: *mut MIB_IPFORWARD_TABLE2 = std::ptr::null_mut();
    // SAFETY: Windows allocates the table and returns it in `table`.
    unsafe { GetIpForwardTable2(AF_INET, &mut table) }.ok()?;
    // SAFETY: on success `table` points to `NumEntries` rows, read here only, and freed once below.
    let found = unsafe {
        let rows = std::slice::from_raw_parts((*table).Table.as_ptr(), (*table).NumEntries as usize);
        let found = rows.iter().any(|row| row.DestinationPrefix.PrefixLength == 0);
        FreeMibTable(table.cast());
        found
    };
    if !found {
        return Err(e!(Error::NoRoute));
    }
    Ok(DefaultRouteDetails { interface_name: "0.0.0.0".to_string() })
}

(still called through tokio::task::spawn_blocking, as now.)

The full change is here: https://github.com/alexpacheli/hext-remote/tree/main/vendor/netwatch (see
PATCHED.md and src/interfaces/windows.rs).

Relevant Logs

No useful trace logs: the process just stops inside the WMI query. The page heap stacks are the ones above.

Expected behavior

Several endpoints can bind at the same time on Windows without hanging or crashing.

netwatch, portmapper

Version: netwatch 0.19.3 (through iroh 1.4.1), wmi 0.18.4

Platform(s)
Desktop:

  • OS: Windows Server 2025 (GitHub Actions windows-2025 runner)
  • Version: x86_64-pc-windows-msvc

Additional Context / Screenshots / GIFs

Happy to test a fix or provide more details. The investigation was done with the help of an AI coding
assistant; the results above (the failures and the fix) come from our CI runs.

lolwut

  • This issue was created by a human that thought critically about
    the issue being reported and wrote this as concisely and clearly
    as they could. Taking full responsibility for the issue being
    accurate.
  • This issue isn't slop, and is backed up by evidence.
主要言語
Rust
スター
34
フォーク
56
平均マージ
2日 4時間
マージ済み PR(30日)
8

環境構築

  • Dockerfile・Docker Compose ファイルなし
  • プルリクエストのテンプレートあり
  • コントリビューションガイドなし

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。

n0-computer/net-tools のほかの issue

n0-computer/net-tools の issue をすべて見る

似ている issue

Rust 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。