Fix rare fd leaks in follower.

In the case where an append entries rpc is received that does not
also validate the log it was possible to leak file handles if
the check of the previous index required a read from a disk segment.

Also fix a potential leak with mismatched terms

Author: kjnilsson

src/ra_server.erl

@@ -1236,7 +1236,8 @@ handle_follower(#append_entries_rpc{term = Term,
                                    " including ~b entries did not validate local log. "
                                    "Local last index ~b",
                                    [LogId, PLIdx, length(Entries0), LocalLastIdx]),
-                            {LVTerm, State} = fetch_term(LastValidatedIdx, State0),
+                            {LVTerm, State} = fetch_term(LastValidatedIdx,
+                                                         State0#{log => Log2}),
                             Reply = #append_entries_reply{term = CurTerm,
                                                           success = true,
                                                           next_index = LastValidatedIdx + 1,
@@ -1301,11 +1302,10 @@ handle_follower(#append_entries_rpc{term = Term,
             % This _may_ overwrite some valid entries but is probably the
             % simplest and most reliable way to proceed
             {Reply, State} = mismatch_append_entries_reply(Term, LastApplied,
-                                                           State0),
+                                                           State0#{log => Log0}),
             Effects = [cast_reply(Id, LeaderId, Reply) | Effects0],
             {await_condition,
-             State#{log => Log0,
-                    condition =>
+             State#{condition =>
                     #{predicate_fun => follower_catchup_cond_fun(term_mismatch),
                       % repeat reply effect on condition timeout
                       timeout => #{effects => Effects,