rpc: cancel in-flight op on fatal read error to fix Call hang

[yperbasis]

Summary

• Fix a long-standing race in Client.dispatch (rpc/client.go) where a fatal readErr racing against reqSent could orphan the in-flight op's resp channel, leaving op.wait blocked forever on a channel nobody would ever close.
• The readErr handler was passing lastOp to cancelAllRequests, which deliberately preserves the inflight op's resp — directly contradicting the comment immediately above:

  // A read error is fatal for the connection, and all pending requests
  // must be cancelled, including any that might still be considered in-flight.
  

  Pass nil instead so the inflight op is cancelled too. Clear lastOp so a concurrent reconnect doesn't re-register an op whose resp is already closed (a later op.resp <- batch would panic). Nil-guard addRequestOp in the reconnect handler accordingly.

Why

Surfaced as the long-running CI flake #16875 — TestWebsocketLargeCall hanging for ~59 minutes until the test timeout fires. From a recent failing run's goroutine dump:

• the test goroutine sat in (*requestOp).wait (rpc/client.go:144) for 58 minutes — i.e., past c.send, with the write already returned;
• Client.dispatch was idle in select (rpc/client.go:597) for the same 58 minutes;
• there were no goroutines inside coder/websocket Read or Write — both had returned.

The only way to reach that state is: readErr fired, dispatch processed it before/instead of reqSent's lastOp = nil, cancelAllRequests skipped the in-flight op (because lastOp was passed as inflightReq), and so op.resp was never closed. The select between c.readErr and c.reqSent (which is buffered, size 1) is a coin flip — that's the flake.

Test plan

• make lint — clean
• make erigon integration — builds
• go test -race ./rpc/... — full suite passes
• go test -race -count=10 -run '^TestWebsocketLargeCall$' ./rpc/ — 10/10 pass
• CI race-tests / core-rpc — should be green

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes a race in the RPC websocket client dispatch loop where a fatal read error could leave an in-flight request orphaned, causing Call/BatchCall to hang indefinitely waiting on an unclosed op.resp.

Changes:

• Treat fatal readErr as cancelling all pending requests (including the in-flight lastOp) by passing nil to conn.close.
• Clear lastOp after fatal read error to prevent reconnect from re-registering an op whose resp channel has already been closed.
• Add a nil-guard when re-registering lastOp on reconnect.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Giulio2002]

LGTM — small, targeted fix for a reconnect/read-error edge case: clear the cancelled in-flight op and guard the later re-register/remove paths to avoid reusing a dead response channel.