sync_spawn_command hangs for infinity if the vnode crashes or replies {stop, _, _} [JIRA: RIAK-2370]

[gatesn]

When sending a sync_spawn_command I notice in riak_core_vnode.erl:345 that the {stop, Reason, NewModState} case doesn't reply to the sender. Given that riak_core_vnode_master:148 issues a sync_spawn_command with a timeout of infinity, the caller will hang forever.

A similar thing happens when the vnode crashes during a handle_command. Any other command being handled at the time of the crash will hang forever. I think the vnode_master ought to trap_exits to catch vnode crashes and reply({vnode_error, _} to the sender.

To reproduce:

handle_command(ping, _Sender, State) ->
timer:sleep(5000),
exit(crashing_the_vnode),
{reply, ok, State};

ping() ->
riak_core_vnode_master:sync_spawn_command(Idx, ping, ).

1> spawn(fun() -> :ping() end).
2> :ping().
...
CRASH REPORT ...
...

[DeadZen]

It actually should already... Is this a regression @slfritchie ?
https://github.com/basho/riak_core/compare/slf-integrate-deadzen-273#diff-1664283599bcf74812b0ceb2d2fc954cR135

[DeadZen]

I've verified this can happen, it occurs only under a specific condition where the indexes overlap...
And the index vnode dies during a prior request before the response of a current request completes, presumably because the dying vnode has no list of callers awaiting a response to be contacted during it's termination process... the question is should this behavior be modified?