gRPC cannot be proxied with dialerProxy to freedom #2232

HirbodBehnam · 2023-06-20T13:46:44Z

Hello
I found a quite strange bug in gRPC transport. Take a look at this config:

{
  "log": {
    "loglevel": "debug"
  },
  "inbounds": [
    {
      "listen": "127.0.0.1",
      "port": "10808",
      "protocol": "socks",
      "settings": {
        "udp": true
      }
    },
    {
      "listen": "127.0.0.1",
      "port": "10809",
      "protocol": "http"
    }
  ],
  "outbounds": [
    {
      "protocol": "trojan",
      "settings": {
        "servers": [
          {
            "address": "example.com",
            "port": 443,
            "password": "pass"
          }
        ]
      },
      "streamSettings": {
        "network": "gun",
        "security": "tls",
        "grpcSettings": {
          "serviceName": "servername"
        },
        "sockopt": {
          "dialerProxy": "direct"
        }
      }
    },
    {
      "protocol": "freedom",
      "tag": "direct"
    }
  ]
}

This config should just route gRPC traffic via a freedom outbound and thus have no difference between adding dialerProxy and not. However, I couldn't get this config to work. If I remove the sockopt from the config it works fine. Logs look like this:

2023/06/20 16:13:41 [Warning] core: Xray 1.8.3 started
2023/06/20 16:13:43 [Info] [2221018126] proxy/socks: TCP Connect request to tcp:[2a00:1450:4001:827::200e]:80
2023/06/20 16:13:43 [Info] [2221018126] app/dispatcher: default route for tcp:[2a00:1450:4001:827::200e]:80
2023/06/20 16:13:45 tcp:127.0.0.1:36732 accepted tcp:[2a00:1450:4001:827::200e]:80
2023/06/20 16:13:49 [Info] [2221018126] transport/internet/grpc: creating connection to tcp:104.21.2.133:443
2023/06/20 16:13:49 [Debug] transport/internet/grpc: using gRPC tun mode service name: `...` stream name: `Tun`
2023/06/20 16:13:49 [Info] [2221018126] transport/internet: redirecting request tcp:104.21.2.133:443 to fragment
2023/06/20 16:13:49 [Info] [2221018126] transport/internet/tcp: dialing TCP to tcp:104.21.2.133:443
2023/06/20 16:13:49 [Debug] transport/internet: dialing to tcp:104.21.2.133:443
2023/06/20 16:13:49 [Info] [2221018126] proxy/freedom: connection opened to tcp:104.21.2.133:443, local endpoint 172.16.0.2:48020, remote endpoint 104.21.2.133:443
2023/06/20 16:13:54 [Info] [2221018126] proxy/trojan: tunneling request to tcp:[2a00:1450:4001:827::200e]:80 via 104.21.2.133:443
multi read transport/internet/grpc/encoding: failed to fetch hunk from gRPC tunnel > rpc error: code = Unavailable desc = error reading from server: EOF
transport/internet/grpc/encoding: failed to send data over gRPC tunnel > EOF
2023/06/20 16:14:36 [Info] [2221018126] app/proxyman/outbound: failed to process outbound traffic > proxy/trojan: connection ends > transport/internet/grpc/encoding: failed to fetch hunk from gRPC tunnel > rpc error: code = Unavailable desc = error reading from server: EOF

And wiresharks shows that my own PC sends a FIN to server:

I tried digging into Xray and Google's gRPC source code with debugging and watching when the pipes get closed but I couldn't figure it out. HOWEVER, I found an alternative way to forward any traffic to a specific outbound. This method involves using dokodemo-door with a specific routing and using the dokedemo-door address as the outbound address. Consider following config file:

{
  "log": {
    "loglevel": "debug"
  },
  "inbounds": [
    {
      "listen": "127.0.0.1",
      "port": "10808",
      "protocol": "socks",
      "settings": {
        "udp": true
      }
    },
    {
      "listen": "127.0.0.1",
      "port": "10809",
      "protocol": "http"
    },
    {
      "listen": "127.0.0.1",
      "port": "28111",
      "protocol": "dokodemo-door",
      "settings": {
        "address": "104.21.2.133",
        "port": 443,
        "network": "tcp"
      },
      "tag": "fragmentedinbound"
    }
  ],
  "outbounds": [
    {
      "protocol": "trojan",
      "settings": {
        "servers": [
          {
            "address": "127.0.0.1",
            "port": 28111,
            "password": "password"
          }
        ]
      },
      "streamSettings": {
        "network": "gun",
        "security": "tls",
        "grpcSettings": {
          "serviceName": "..."
        },
        "tlsSettings": {
          "serverName": "..."
        }
      }
    },
    {
      "protocol": "freedom",
      "settings": {
        "fragment": {
          "length": "1-2",
          "interval": "0-1",
          "packets": "1"
        }
      },
      "tag": "fragment"
    },
    {
      "protocol": "freedom",
      "tag": "direct"
    }
  ],
  "routing": {
    "domainMatcher": "mph",
    "domainStrategy": "IPIfNonMatch",
    "rules": [
      {
        "domain": [
          "regexp:.*\\.ir$",
          "ext:iran.dat:ir",
          "ext:iran.dat:other"
        ],
        "outboundTag": "direct",
        "type": "field"
      },
      {
        "ip": [
          "geoip:private",
          "geoip:ir"
        ],
        "outboundTag": "direct",
        "type": "field"
      },
      {
        "inboundTag": [
          "fragmentedinbound"
        ],
        "outboundTag": "fragment",
        "type": "field"
      }
    ]
  }
}

This is basically the config which I'm currently using to connect. I'm not expecting this to be fixed anytime soon considering that there is a neat workaround.

The text was updated successfully, but these errors were encountered:

RPRX · 2023-06-20T18:42:45Z

感谢你的测试，可能不是 gRPC 内部的问题，而是 Xray 调用 gRPC 的问题

dialerProxy 和 gRPC 传输层是 Xray-core v1.4.0 同时引入的，可能它们之间没适配，你可以检查一下代码，~~然后发个 PR~~

RPRX · 2023-06-25T10:22:16Z

~~请问修好了吗~~

ghost · 2023-08-05T08:39:15Z

~~请问修好了吗~~

@RPRX

I checked this with xray v1.8.3 and I can confirm that as @HirbodBehnam mentioned, it doesn't work with gRPC. On the other hand, WS works fine.

cty123 · 2023-08-21T07:02:16Z

I was able to reproduce the problem, but couldn't figure out the root cause either. It seems to be a real problem, it is dropping the outbound connection somewhere inside freedom proxy.

RPRX · 2023-08-26T12:20:14Z

@cty123 分享些经验？

RPRX · 2023-08-26T12:23:11Z

解决这个问题应该不难，插一些 log 看一下哪里断了就发现了

cty123 · 2023-08-26T12:30:26Z

我已经试过了，最早是在freedom 这里断的https://github.com/XTLS/Xray-core/blob/main/proxy/freedom/freedom.go#L205, 显示的错误就是 use of closed network connection，但是不是很清楚为什么connection会断开。我看了grpc文档https://github.com/grpc/grpc-go, 根据说明打开了所有logging, 服务端这边显示是客户端先关闭的连接，但是客户端的grpc显示关闭的原因是EOF, 所以目前并不知道断连的根本原因。

RPRX · 2023-08-26T12:34:24Z

以前有个 bug 是若 cancel gRPC 某一子连接的 ctx，会 cancel 整个 gRPC 连接，这个 EOF 可能是类似的原因

RPRX · 2023-08-26T12:38:45Z

h2 同理，如果 h2 加 dialerProxy 也有这问题，基本上可以确定就是它了

cty123 · 2023-08-26T12:42:08Z

这我倒是可以试试看

RPRX · 2023-08-26T14:33:38Z

那就交给你啦，你的反馈很有帮助，“服务端这边显示是客户端先关闭的连接，但是客户端的grpc显示关闭的原因是EOF”这个症状很符合那个 bug，解决办法就是不传原始 ctx，只复制一些关键的信息（若有必要），参考：

Xray-core/transport/internet/grpc/dial.go

Lines 100 to 102 in 10d6b06

    
           grpc.WithContextDialer(func(gctx context.Context, s string) (gonet.Conn, error) { 
        
           	gctx = session.ContextWithID(gctx, session.IDFromContext(ctx)) 
        
           	gctx = session.ContextWithOutbound(gctx, session.OutboundFromContext(ctx))

RPRX · 2023-08-26T14:44:41Z

以前有个 bug 是若 cancel gRPC 某一子连接的 ctx，会 cancel 整个 gRPC 连接，这个 EOF 可能是类似的原因

修正一下这个描述，应该是把原始 ctx 传给了子连接，子连接结束时调用了 cancel，结果整个 gRPC 连接都断开了（表现为断流）

RPRX · 2023-08-26T14:50:26Z

以前有个 bug 是若 cancel gRPC 某一子连接的 ctx，会 cancel 整个 gRPC 连接，这个 EOF 可能是类似的原因

但这次这个 bug 盲猜符合这个描述，应该是 gRPC 把第一个子连接的 gctx 传给了 dialerProxy，然后这个 gctx 被 cancel。。。

cty123 · 2023-08-26T18:49:33Z

还真的跟你说的一样，我debug了好几天了都。就是你说的这个地方

Xray-core/transport/internet/grpc/dial.go

Lines 100 to 102 in 10d6b06

    
           grpc.WithContextDialer(func(gctx context.Context, s string) (gonet.Conn, error) { 
        
           	gctx = session.ContextWithID(gctx, session.IDFromContext(ctx)) 
        
           	gctx = session.ContextWithOutbound(gctx, session.OutboundFromContext(ctx))

我创了个新的context传进去马上就好了，完美使用

RPRX · 2023-08-27T04:32:29Z

以前有个 bug 是若 cancel gRPC 某一子连接的 ctx，会 cancel 整个 gRPC 连接，这个 EOF 可能是类似的原因

修正一下这个描述，应该是把原始 ctx 传给了子连接，子连接结束时调用了 cancel，结果整个 gRPC 连接都断开了（表现为断流）
但这次这个 bug 盲猜符合这个描述，应该是 gRPC 把第一个子连接的 gctx 传给了 dialerProxy，然后这个 gctx 被 cancel。。。

再次修正 & 总结：

getGrpcClient 的 ctx 参数是每个被代理连接的 ctx，以前那个 bug 是没有 gctx，而由于 gRPC 只 dial 一次，相当于只认第一条被代理连接的 ctx，它被 cancel 时整条 gRPC 都会断
有了这次的 bug，我研究了一下这个 gctx 是干啥的，grpc.WithContextDialer 的前身是 grpc.WithDialer，后者有一个参数是 time.Duration，再看代码，所以这个 gctx 只是控制 dial 超时用的，dialerProxy 拿它当 *ray 的 ctx，就断得比上次的 bug 还快

RPRX · 2023-08-27T05:02:48Z

这么说的话其实 gRPC + dialerProxy 一直都是不可用的状态

我在写增强版 XUDP 时遇到了“只想让原始 ctx 控制 dial 超时，不想让它 cancel Copy，但又想让 outbound 自身的超时策略生效”的~~极其复杂需求~~，做了一些尝试，~~留了两处彩蛋~~，最终方案是给原始 ctx 标记 TimeoutOnly，并且改造了各个 outbound：be23d5d

所以直接拿来用就行了，@cty123 你试一下给 gctx 标记 TimeoutOnly，没问题的话发个 PR，记得带上 H2

RPRX · 2023-08-27T06:39:29Z

我改好了，请测试 d92002a 的 gRPC 能否使用 dialerProxy

原本 H2 应该是没这个问题的，因为它原本是 context.Background()，这次改成了 DialTLSContext，需要测试有没有引入新问题

此前传给 REALITY UClient 的 ctx 实际上没被用到，一直想改，这次顺便改成了 uConn.HandshakeContext(ctx)，需要测试有没有引入新问题。但是这个超时时间，以后应当参考浏览器来设置一下，~~否则 GFW 故意让某次握手超时就精准识别了。~~

RPRX · 2023-08-27T08:22:04Z

以前有个 bug 是若 cancel gRPC 某一子连接的 ctx，会 cancel 整个 gRPC 连接，这个 EOF 可能是类似的原因

再再次修正：记混了，不是 gRPC，是 H2 以前出现过的 bug #289 (comment) ，修复见 #289 (comment)

@cty123

XTLS/Xray-core#2232 (comment) Thank @cty123 for testing Fixes XTLS/Xray-core#2232 BTW: Use `uConn.HandshakeContext(ctx)` in REALITY fix(app/router): fixed a bug in geoip matching with refactoring (#2489) * Refactor the IP address matching with netipx library * Add a regression test for previous bug Fixes XTLS/Xray-core#1933 --------- Co-authored-by: Loyalsoldier <10487845+Loyalsoldier@users.noreply.github.com> fix(transport): correctly release UDS locker file (#2305) * fix(transport): correctly release UDS locker file * use callback function to do some jobs after create listener Update transport/internet/reality/reality.go Fixes XTLS/Xray-core#2491 Upgrade dependencies Update workflows to use Go 1.21

RPRX closed this as completed in d92002a Aug 27, 2023

RPRX mentioned this issue Aug 27, 2023

reality h2模式切换网络断连，必须客户端重启服务 #2355

Closed

RPRX mentioned this issue Aug 31, 2023

GFW seems to find Fragmented configurations #2504

Closed

cty123 mentioned this issue Aug 31, 2023

gRPC + acceptProxyProtocol --> not working #2204

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

gRPC cannot be proxied with dialerProxy to freedom #2232

gRPC cannot be proxied with dialerProxy to freedom #2232

HirbodBehnam commented Jun 20, 2023

RPRX commented Jun 20, 2023 •

edited

Loading

RPRX commented Jun 25, 2023 •

edited

Loading

ghost commented Aug 5, 2023

cty123 commented Aug 21, 2023 •

edited

Loading

RPRX commented Aug 26, 2023

RPRX commented Aug 26, 2023

cty123 commented Aug 26, 2023 •

edited

Loading

RPRX commented Aug 26, 2023 •

edited

Loading

RPRX commented Aug 26, 2023

cty123 commented Aug 26, 2023

RPRX commented Aug 26, 2023 •

edited

Loading

RPRX commented Aug 26, 2023

RPRX commented Aug 26, 2023

cty123 commented Aug 26, 2023

RPRX commented Aug 27, 2023 •

edited

Loading

RPRX commented Aug 27, 2023 •

edited

Loading

RPRX commented Aug 27, 2023

RPRX commented Aug 27, 2023

gRPC cannot be proxied with dialerProxy to freedom #2232

gRPC cannot be proxied with dialerProxy to freedom #2232

Comments

HirbodBehnam commented Jun 20, 2023

RPRX commented Jun 20, 2023 • edited Loading

RPRX commented Jun 25, 2023 • edited Loading

ghost commented Aug 5, 2023

cty123 commented Aug 21, 2023 • edited Loading

RPRX commented Aug 26, 2023

RPRX commented Aug 26, 2023

cty123 commented Aug 26, 2023 • edited Loading

RPRX commented Aug 26, 2023 • edited Loading

RPRX commented Aug 26, 2023

cty123 commented Aug 26, 2023

RPRX commented Aug 26, 2023 • edited Loading

RPRX commented Aug 26, 2023

RPRX commented Aug 26, 2023

cty123 commented Aug 26, 2023

RPRX commented Aug 27, 2023 • edited Loading

RPRX commented Aug 27, 2023 • edited Loading

RPRX commented Aug 27, 2023

RPRX commented Aug 27, 2023

RPRX commented Jun 20, 2023 •

edited

Loading

RPRX commented Jun 25, 2023 •

edited

Loading

cty123 commented Aug 21, 2023 •

edited

Loading

cty123 commented Aug 26, 2023 •

edited

Loading

RPRX commented Aug 26, 2023 •

edited

Loading

RPRX commented Aug 26, 2023 •

edited

Loading

RPRX commented Aug 27, 2023 •

edited

Loading

RPRX commented Aug 27, 2023 •

edited

Loading