Android TUN: goroutine count grows to 20000+ during idle, causes connection freeze

操作系统

Android

系统版本

Android16

安装类型

sing-box 原始命令行程序

如果您使用图形客户端程序，请提供该程序版本。

1.13.15

版本

sing-box version 1.13.13

Environment: go1.26.4 android/arm64
Tags: with_gvisor,with_quic,with_wireguard,with_utls,with_clash_api
CGO: enabled

(Note: above is from command-line sing-box installed via Termux for reporting purposes; the actual affected client is SFA graphical app version 1.13.15)

描述

在长时间空闲运行后（数小时到数天不等，无明确触发条件），goroutine 数量从正常的 100 多个开始不受控增长，最终达到 20000-30000+，同时观察到：

1. 上传/下载速率显示为 0 kB/s，但 goroutine 持续增长
2. 内存占用同步小幅上升（440MB → 470MB左右）
3. 此时系统代理（TUN）无法建立新连接，但使用应用内置代理（绕过TUN，直连相同节点）可以正常连接
4. 唯一的恢复方式是重启 VPN（关闭并重新打开 TUN 服务），重启后 goroutine 立即恢复到正常水平

日志中观察到重复出现的错误：
ERROR outbound/direct[DIRECT]: receive ICMP echo reply: i/o timeout

怀疑与出站节点使用 Hysteria2 协议有关（参考已有 issue #3421 中提到的类似现象）。

重现方式

暂无可稳定复现的最小配置。现象在使用 Hysteria2 协议出站、长时间运行后随机出现，触发条件不明确。将持续观察并尝试提供更完整的复现步骤和 debug 级别日志。

日志

=== Log snippet 1 (16:24:49) ===
INFO[0476] router: found package name: com.openai.chatgpt
DEBUG[0476] router: found fakeip domain: 2.datadog.pool.ntp.org
INFO[0476] [908156890 0ms] inbound/tun[tun-in]: inbound packet connection from 172.19.0.1:36054
INFO[0476] [908156890 0ms] inbound/tun[tun-in]: inbound packet connection to 198.18.0.66:123
INFO[0476] [908156890 0ms] router: found package name: com.openai.chatgpt
DEBUG[0476] [908156890 0ms] router: found fakeip domain: 2.datadog.pool.ntp.org
DEBUG[0476] [908156890 0ms] router: match[0] => sniff
DEBUG[0476] [908156890 0ms] router: sniffed packet protocol: 
INFO[0476] [908156890 0ms] outbound/vless[新加坡•移联01]: outbound packet connection to 2.datadog.pool.ntp.org:123
INFO[0476] [908156890 0ms] outbound/vless[新加坡•移联01]: outbound packet connection to 2.datadog.pool.ntp.org:123
DEBUG[0476] [908156890 0ms] dns: lookup domain chuixuevl5.blowingsnowclouds.cc
DEBUG[0476] [908156890 0ms] dns: lookup succeed for chuixuevl5.blowingsnowclouds.cc: 13.250.40.57 2406:da18:15c3:aa00:f711:f5e9:951c:f363
DEBUG[0476] [908156890 172ms] connection: packet upload finished
ERROR[0477] [3556662403 10.0s] outbound/direct[DIRECT]: receive ICMP echo reply: i/o timeout
ERROR[0478] [389805004 10.5s] outbound/direct[DIRECT]: receive ICMP echo reply: i/o timeout
INFO[0479] router: found package name: com.openai.chatgpt
DEBUG[0479] router: found fakeip domain: android.chat.openai.com
INFO[0480] router: found package name: com.openai.chatgpt
DEBUG[0480] router: found fakeip domain: cloudflare-dns.com
INFO[0480] router: found package name: com.openai.chatgpt
DEBUG[0480] router: found fakeip domain: ab.chatgpt.com

=== Log snippet 2 (16:26:29) ===
INFO[0554] router: found package name: org.telegram.messenger
INFO[0554] router: found package name: org.telegram.messenger
ERROR[0554] [3006667928 10.0s] outbound/direct[DIRECT]: receive ICMP echo reply: i/o timeout
ERROR[0555] [396056085 10.3s] outbound/direct[DIRECT]: receive ICMP echo reply: i/o timeout
INFO[0562] router: found package name: org.telegram.messenger
INFO[0562] router: found package name: org.telegram.messenger
INFO[0562] router: found package name: org.telegram.messenger
DEBUG[0562] router: found fakeip domain: firebaseremoteconfig.googleapis.com
DEBUG[0563] outbound/urltest[Gemini_自动]: outbound 香港•电信02 unavailable: Head "https://www.gstatic.com/generate_204": context deadline exceeded

支持我们

• 我已经 赞助

完整性要求

• 我保证阅读了文档，了解所有我编写的配置文件项的含义，而不是大量堆砌看似有用的选项或默认值。
• 我保证提供了可以在本地重现该问题的服务器、客户端配置文件与流程，而不是一个脱敏的复杂客户端配置文件。
• 我保证提供了可用于重现我报告的错误的最简配置，而不是依赖远程服务器、TUN、图形界面客户端或者其他闭源软件。
• 我保证提供了完整的配置文件与日志，而不是出于对自身智力的自信而仅提供了部分认为有用的部分。

Start by comparing this report with issue #3421 and inspect the TUN, Hysteria2, and ICMP-timeout paths implicated by the logs. A useful next step is to capture goroutine and debug logs during an idle reproduction. Done means identifying the goroutine leak and providing a reproducible case with evidence that restarting TUN is no longer required.