Queued pipe payloads can be silently dropped at stage time after passing the send-time size check

未关闭
#5 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
54/100
Issue 类型
缺陷
描述清晰度
描述清楚
活跃度
活跃
技术栈
rust

调研方向

Start with pipe.rs send-time size checks, runtime.rs outbox staging, and the protocol session stage() implementations in panetiere.rs, panetiere_scheduled.rs, adcnet.rs, and noop.rs. Read config.rs and scheduler_core.rs to trace the Scheduled Panetiere limit mismatch. Reproduce both queued-frame cases, then verify accepted payloads are delivered or produce a clear failure signal.

由索引模型根据 Issue 内容生成。

描述

Summary

Pipe::send checks payload size once, when the payload is queued, and returns Ok(()) (pipe.rs:176-193). The frame then waits in the node-wide outbox until a round's participation draw takes it in sync_client_round and hands it to the client session's stage() (runtime.rs:1576-1588). Each protocol's session checks size again there, against its own limit. If that second check fails, the frame is thrown away: it isn't put back in the outbox, and the sender is never told. Two cases trigger this.

Case 1: Message size shrinks in a config update while a frame is queued

  1. The node runs a config where the smallest message_size across runnable subnets is, say, 1024.
  2. A pipe sends a 1000-byte payload. check_size_against passes and the frame is queued.
  3. The frame waits in the outbox for one or more rounds, either because no subnet was drawn yet or because other frames are ahead of it.
  4. A new config lowers message_size to 512. apply_config replaces the worker, because the protocol config is part of subnet_sig. The outbox is not re-checked.
  5. The new worker's sync_client_round pops the frame and calls stage(), which drops it.

This affects every protocol, each in its own way:

  • Panetiere: stage() logs an error and drops the frame (panetiere.rs:2013-2023).
  • Scheduled Panetiere: reserve_message returns an error, and stage() discards it with let _ = (panetiere_scheduled.rs:835), so nothing is logged.
  • ADCNet: contribute fails, which is only logged at debug level, and the payload is lost (adcnet.rs:1075, adcnet.rs:1104).
  • Noop: begin_round returns nothing for an oversized payload (noop.rs:47-51).

Case 2: The send-time and stage-time limits disagree (Scheduled Panetiere)

The pipe checks against ProtocolConfig::message_size(), which is message_size for Scheduled Panetiere (config.rs:238). The session checks against vector_bytes.min(u16::MAX) (panetiere_scheduled.rs:793). validate_structure rejects message_size > u16::MAX and vector_bytes == 0, but it does not require message_size <= vector_bytes (scheduler_core.rs:247-262).

With message_size = 1024 and vector_bytes = 512, any payload between about 513 and 1024 bytes passes the send check. It is then dropped every time, without a log line.

Expected

A payload that send accepts is either delivered, or the sender is clearly told it failed. Nothing is dropped silently.

Possible fixes

  • Case 2: require message_size <= vector_bytes in validate_structure, or have ProtocolConfig::message_size() return the effective limit for Scheduled Panetiere.
  • Case 1: re-check the outbox in apply_config and drop oversized frames with a warning or an event. Or have stage() return a result, so sync_client_round can log the drop and report it.
  • In both cases, log the dropped frame in Scheduled Panetiere's stage() instead of discarding the error.
主要语言
Rust
星标
0
派生
0
PR 合并指标
30 天内没有已合并 PR

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

flashbots/anymone 的其他 Issue

查看 flashbots/anymone 的全部 Issue

相似的 Issue

更多 Rust Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。