rpc: reset writeConn when conn is closed on readErr

[zzy96]

Behavior

When running geth with external signer (Clef), if Clef is restarted, the first call to Clef always fail but the subsequent calls work fine.

Analysis & Fix

In rpc/client.go, when any error like EOF is sent to readErr channel, conn will be closed but c.writeConn is not set to nil. If a new request comes after conn is closed, it must fail once before it will try to reconnect. This fix is to set c.writeConn = nil when readErr is received and conn is closed.

[fjl]

I'm not sure setting writeConn is safe in this context. This might introduce a race. Will check tomorrow.

[fjl]

Code docs say:

// writeConn is used for writing to the connection on the caller's goroutine. It should
// only be accessed outside of dispatch, with the write lock held. The write lock is
// taken by sending on requestOp and released by sending on sendDone.

This comment is slightly out of date because requestOp has since been renamed to reqInit.
Still means we cannot set this field from inside the dispatch loop.

writeConn is closed when a read error happens. I think the best we can do to avoid the extra error on the next call is handling closed connections better on the write code path.

[zzy96]

Hi @fjl

I have tried another fix adding a retry flag for c.write.

[fjl]

Sorry, this took a long time to get reviewed because I was busy with other things. I have tested the change and it solves the issue.

[fjl]

Damn, I didn't change the commit message title. The title is a bit misleading because the fix is now based on the write error.