H2C connection to HTTP.sys on localhost causes infinite loop

[ManickaP]

If ProcessIncommingFrames encounters an exception it tries to abort the HTTP 2 connection: https://github.com/dotnet/runtime/blob/master/src/libraries/System.Net.Http/src/System/Net/Http/SocketsHttpHandler/Http2Connection.cs#L394. This will remove it from the connection pool. But if this happens synchronously from SetupAsync, the connection is not yet in the pool and still gets added: https://github.com/dotnet/runtime/blob/master/src/libraries/System.Net.Http/src/System/Net/Http/SocketsHttpHandler/HttpConnectionPool.cs#L642-L644.

As a result, any subsequent requests will end up in an infinite loop in SendWithRetryAsync since isNewConnection will be false: https://github.com/dotnet/runtime/blob/master/src/libraries/System.Net.Http/src/System/Net/Http/SocketsHttpHandler/HttpConnectionPool.cs#L913

Tagging subscribers to this area: @dotnet/ncl
See info in area-owners.md if you want to be subscribed.

[karelz]

Triage: @JamesNK do you know how common this scenario is for gRPC? We are considering moving it to Future.

[JamesNK]

Using H2C with gRPC is very common.

Is this bug specific to HTTP.sys as a server, or is it a problem when doing H2C in general?

[karelz]

Is this bug specific to HTTP.sys as a server, or is it a problem when doing H2C in general?

We do not know. That said, if it wasn't reported on gRPC yet, I think it is safe to assume that either the scenario is rare, or that it works outside of HTTP.sys.
I'll move it to Future until we have more evidence this is painful.

[ManickaP]

This code (retry logic and connection construction) has been substantially reworked and the culprit: isNewConnection has been removed.
I'm closing this, since this should not be a problem anymore. If we encounter it out in the wild we might reopen a re-investigate.