Cancel connection backoff on graceful shutdown

[hamzahrmalik]

Problem

When a gRPC client's connection is stuck in backoff (e.g. the server is unreachable), graceful shutdown would wait out that in-flight backoff sleep before returning — so shutdown could hang for up to the max backoff (120s by default), even with zero in-flight RPCs. In practice that means processes that are slow to exit on shutdown

Cause

A subchannel that is sleeping in connection backoff has no established connection and no in-flight RPCs to drain. Its backoff Task.sleep was spawned into the subchannel's discarding task group, which graceful shutdown awaits rather than cancels, so run() (and thereforeGRPCClient.runConnections()) did not return until the backoff elapsed — up to the default max of 120s — even with zero in-flight RPCs. This also contradicts runConnections()'s documented contract of returning once graceful shutdown is requested and in-flight RPCs have finished.

Changes

Spawn the backoff via addCancellableTask and store its handle on the .connecting state (mirroring how the name-resolver task handle is tracked). When shutDown() is called while backing off, cancel that task and finish immediately; when a connect attempt is genuinely in-flight, behaviour is unchanged.

The new test fails without this fix, and passes after it.

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: hamzahrmalik / name: Hamzah Malik (53cdcad)

[glbrntt]

Thanks @hamzahrmalik, nice patch.