LeadaxeLxBox / sing-box-lx 开发者

作者发言

GitHub·2026.10.10 04:48:23（UTC+8）

Leadaxe：sing-box-lx 移植 AWG 3.x 时按实际上界预留 tailroom，并在 RoutineEncryption 加夹取避免尺寸不匹配崩溃

在 shtorm-7/sing-box-extended issue #158（AmneziaWG 3.0 在 TUN 包且 ContentPaddingAddition > 16 时于 RoutineEncryption 触发 panic）中，Leadaxe 说明其自有 sing-box 下游（sing-box-lx）移植 AWG 3.x 的处理方式，与提案补丁略有不同：①注入元素按实际上界预留 tailroom，而非固定的 MaxMessageSize —— 设置 content_padding_addition 时用 addition.Hi()，随机 trailer 用 DefaultUdpWindow - datagramSize，其余用 PaddingMultiple；②作为安全网，RoutineEncryption 在封包前把所选 addition 夹取到 cap(elem.packet) - len(elem.packet) - Overhead，并原地写入 elem.buffer（不用 slices.Grow），使尺寸不匹配时缩短 padding 而非 panic。给出 tailroom、clamp 两处代码引用；e2e 测试覆盖 content_padding_addition=10-100（经 InputPacket）。并称报告者的修法也正确，夹取只是让该路径对未来尺寸变化更稳健。

作者原文@Leadaxe

Same design point came up when we ported AWG 3.x into our sing-box downstream (sing-box-lx). We went slightly differently from the proposed patch, in case it's useful:

1. Injected elements reserve tailroom for the actual upper bound instead of a flat MaxMessageSize: addition.Hi() when content_padding_addition is set, DefaultUdpWindow - datagramSize for random trailers, PaddingMultiple otherwise.
2. As a safety net, RoutineEncryption clamps the chosen addition to cap(elem.packet) - len(elem.packet) - Overhead before sealing, and seals in place into elem.buffer (no slices.Grow). So a sizing mismatch shortens the padding instead of panicking.

Ref: tailroom, clamp. An e2e test covers content_padding_addition=10-100 via InputPacket.

The reporter's fix is correct too; the clamp just makes the path robust against future sizing changes.