http3: redial for unresolved or staled connections

[lifenjoiner]

unresolved:
Valid parameters, but handshake is not completed successfully.

staled:
A server may close the connection after requests are responsed, that stops to reuse it.

Complement #3684 (89769f4).

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[lifenjoiner]

Unfortunately, I failed to create a google account to sign the CLA.
Signed.

[lifenjoiner]

CLA signed.

[lifenjoiner]

The problem is about a serial requests on the same host with one http client (dialOnce with only one http3 client allowed):

1. if a request failed, the following will all fail then;
2. if a request succeeded, but the server closed the connection before a following request, then the following requests fail too.

2 Connection %s closed examples for case 2:

2023/07/13 12:28:17 client Starting new connection to dnsnl.alekberg.net (0.0.0.0:56325 -> 89.38.131.38:443), source connection ID (empty), destination connection ID 402a368ce3b7c22bd1, version v1
[2023-07-13 12:28:18] [INFO] [ams-doh-nl] TLS version: 304 - Protocol: h3 - Cipher suite: 4865
[2023-07-13 12:28:18] [NOTICE] [ams-doh-nl] OK (DoH) - rtt: 937ms
...
2023/07/13 12:28:26 client Peer closed connection with error: Application error 0x100 (remote)
2023/07/13 12:28:26 client Connection 402a368ce3b7c22bd1 closed.

2023/07/13 12:29:43 client Starting new connection to dnsse.alekberg.net (0.0.0.0:55470 -> 45.153.187.96:443), source connection ID (empty), destination connection ID a588febf56ac99eb492740141f7a2ceb9a5868, version v1
[2023-07-13 12:29:44]   127.0.0.1       github.com      A       PASS    1274ms  sth-doh-se
2023/07/13 12:29:47 client Destroying connection with error: received a stateless reset with token 562d402ae2234d37a9835b78a48d541c
2023/07/13 12:29:47 client Connection a588febf56ac99eb492740141f7a2ceb9a5868 closed.

[marten-seemann]

What's the issue this PR is supposed to resolve?

[lifenjoiner]

In the scenario where a http3 client acts as a proxy.
When "Connection %s closed" happened, the RoundTripper should be able to initialize a new client. Or, the following requests will be refused/reset by that host.

Some servers close the connection very quickly, refer to my above examples (in seconds).
Some longer:

2023/07/14 12:04:24 client Starting new connection to cloudflare-dns.com ([::]:51217 -> 104.16.248.249:443), source connection ID (empty), destination connection ID ac0f77e9e1671161aeaeb34e9c, version v1
2023/07/14 12:04:24 client Queueing packet (1053 bytes) for later decryption
2023/07/14 12:09:52 client Peer closed connection with error: Application error 0x100 (remote)
2023/07/14 12:09:52 client Connection ac0f77e9e1671161aeaeb34e9c closed.

Must an issue be logged for each PR?

[marten-seemann]

In the scenario where a http3 client acts as a proxy.
When "Connection %s closed" happened, the RoundTripper should be able to initialize a new client. Or, the following requests will be refused/reset by that host.

I don’t think this is accurate. You can send multiple request over the same connection, there’s no need to close it in between.

[lifenjoiner]

there’s no need to close it in between.

Yes. But, the issue is the remote server closed the connection, as a user I can't use the client anymore.

[lifenjoiner]

Without redo handshake, extra data sent are refused.

[lifenjoiner]

Similarly, dnsproxy wrapped their own http3Transport to catch the closed status to recreate the http client.
We could make life easier here.