Skip to content

Commit

Permalink
Merge pull request #51 from esl/dist-blocker
Browse files Browse the repository at this point in the history
Dist blocker
  • Loading branch information
chrzaszcz authored Mar 8, 2024
2 parents 637a784 + b14d8b8 commit 0f69ec9
Show file tree
Hide file tree
Showing 4 changed files with 404 additions and 38 deletions.
157 changes: 157 additions & 0 deletions src/cets_dist_blocker.erl
Original file line number Diff line number Diff line change
@@ -0,0 +1,157 @@
%% @doc Disallow distributed erlang connections until cleaning is done.
%%
%% This module prevents a node from reconnecting, until cleaning activity is
%% finished. It prevents race conditions.
%%
%% This module assumes all nodes share the same cookie.
-module(cets_dist_blocker).
-behaviour(gen_server).
-include_lib("kernel/include/logger.hrl").

%% API
-export([
start_link/0,
add_cleaner/1,
cleaning_done/2
]).

%% gen_server callbacks
-export([
init/1,
handle_call/3,
handle_cast/2,
handle_info/2,
terminate/2,
code_change/3
]).

-ignore_xref([
start_link/0,
add_cleaner/1,
cleaning_done/2
]).

-type cleaner_pid() :: pid().
-type waiting() :: [{node(), cleaner_pid()}].

-type state() :: #{
cleaners := [cleaner_pid()],
waiting := waiting()
}.

%% @doc Spawn `dist_blocker'
start_link() ->
gen_server:start_link({local, ?MODULE}, ?MODULE, [], []).

%% @doc Register CleanerPid as a cleaner.
-spec add_cleaner(pid()) -> ok.
add_cleaner(CleanerPid) ->
gen_server:call(?MODULE, {add_cleaner, CleanerPid}).

%% @doc Confirm that cleaning is done.
%%
%% This function should be called by a cleaner when it receives
%% nodedown and finishes cleaning.
-spec cleaning_done(pid(), node()) -> ok.
cleaning_done(CleanerPid, Node) ->
gen_server:call(?MODULE, {cleaning_done, CleanerPid, Node}).

%%--------------------------------------------------------------------
%% gen_server callbacks
%%--------------------------------------------------------------------
init([]) ->
ok = net_kernel:monitor_nodes(true),
State = #{cleaners => [], waiting => []},
State2 = lists:foldl(fun handle_nodeup/2, State, nodes()),
{ok, State2}.

handle_call({add_cleaner, CleanerPid}, _From, State) ->
{reply, ok, handle_add_cleaner(CleanerPid, State)};
handle_call({cleaning_done, CleanerPid, Node}, _From, State) ->
{reply, ok, maybe_unblock(State, handle_cleaning_done(CleanerPid, Node, State))};
handle_call(Request, _From, State) ->
?LOG_ERROR(#{what => unexpected_call, msg => Request}),
{reply, {error, unexpected_call}, State}.

handle_cast(Msg, State) ->
?LOG_ERROR(#{what => unexpected_cast, msg => Msg}),
{noreply, State}.

handle_info({nodeup, Node}, State) ->
{noreply, handle_nodeup(Node, State)};
handle_info({nodedown, Node}, State) ->
{noreply, handle_nodedown(Node, State)};
handle_info({'DOWN', _Ref, process, Pid, _Info}, State) ->
{noreply, maybe_unblock(State, handle_cleaner_down(Pid, State))};
handle_info(Info, State) ->
?LOG_ERROR(#{what => unexpected_info, msg => Info}),
{noreply, State}.

terminate(_Reason, State) ->
%% Restore cookies
_ = maybe_unblock(State, State#{waiting := []}),
ok.

code_change(_OldVsn, State, _Extra) ->
{ok, State}.

%%--------------------------------------------------------------------
%% internal functions
%%--------------------------------------------------------------------

-spec handle_nodeup(node(), state()) -> state().
handle_nodeup(Node, State) ->
%% We change the cookie as soon as the node is connected.
%% Alternative is to do it on nodedown, but because nodedown-s are async,
%% we would have a high chance of race conditions (so, node could reconnect
%% before we set cookie).
erlang:set_cookie(Node, blocking_cookie()),
State.

%% Make cookie, that would prevent node from connecting
-spec blocking_cookie() -> atom().
blocking_cookie() ->
list_to_atom(atom_to_list(erlang:get_cookie()) ++ "_blocked_by_" ++ atom_to_list(node())).

%% Allow the node to connect to us again
-spec unblock_node(node(), state()) -> state().
unblock_node(Node, State) ->
erlang:set_cookie(Node, erlang:get_cookie()),
State.

-spec handle_nodedown(node(), state()) -> state().
handle_nodedown(Node, State = #{cleaners := []}) ->
%% Skip waiting when no cleaners
unblock_node(Node, State);
handle_nodedown(Node, State = #{cleaners := Cleaners, waiting := Waiting}) ->
New = [{Node, CleanerPid} || CleanerPid <- Cleaners],
State#{waiting := lists:usort(New ++ Waiting)}.

-spec handle_add_cleaner(cleaner_pid(), state()) -> state().
handle_add_cleaner(CleanerPid, State = #{cleaners := Cleaners}) ->
erlang:monitor(process, CleanerPid),
State#{cleaners := lists:usort([CleanerPid | Cleaners])}.

-spec handle_cleaning_done(cleaner_pid(), node(), state()) -> state().
handle_cleaning_done(CleanerPid, Node, State = #{waiting := Waiting}) ->
State#{waiting := lists:delete({Node, CleanerPid}, Waiting)}.

-spec handle_cleaner_down(cleaner_pid(), state()) -> state().
handle_cleaner_down(CleanerPid, State = #{cleaners := Cleaners, waiting := Waiting}) ->
State#{
cleaners := lists:delete(CleanerPid, Cleaners),
waiting := [X || {_Node, CleanerPid2} = X <- Waiting, CleanerPid =/= CleanerPid2]
}.

%% Unblock nodes when the last cleaner confirms the cleaning is done.
%% Call this function each time you remove entries from the waiting list.
-spec maybe_unblock(state(), state()) -> state().
maybe_unblock(_OldState = #{waiting := OldWaiting}, NewState = #{waiting := NewWaiting}) ->
OldNodes = cast_waiting_to_nodes(OldWaiting),
NewNodes = cast_waiting_to_nodes(NewWaiting),
CleanedNodes = OldNodes -- NewNodes,
lists:foldl(fun unblock_node/2, NewState, CleanedNodes).

-spec cast_waiting_to_nodes(waiting()) -> [node()].
cast_waiting_to_nodes(Waiting) ->
lists:usort([Node || {Node, _CleanerPid} <- Waiting]).
40 changes: 2 additions & 38 deletions test/cets_SUITE.erl
Original file line number Diff line number Diff line change
Expand Up @@ -235,7 +235,7 @@ cets_seq_no_log_cases() ->
init_per_suite(Config) ->
init_cleanup_table(),
Names = [ct2, ct3, ct4, ct5, ct6, ct7],
{Nodes, Peers} = lists:unzip([start_node(N) || N <- Names]),
{Nodes, Peers} = lists:unzip([cets_test_peer:start_node(N) || N <- Names]),
[
{nodes, maps:from_list(lists:zip(Names, Nodes))},
{peers, maps:from_list(lists:zip(Names, Peers))}
Expand Down Expand Up @@ -2935,7 +2935,7 @@ schedule_cleanup(Pid) ->
{'DOWN', Ref, process, Me, _} ->
%% We do an RPC call, because erlang distribution
%% could not be always reliable (because we test netsplits)
rpc(node_to_peer(node(Pid)), cets, stop, [Pid]),
rpc(cets_test_peer:node_to_peer(node(Pid)), cets, stop, [Pid]),
ets:delete_object(cleanup_table, {Me, self()})
end
end),
Expand Down Expand Up @@ -3027,42 +3027,6 @@ rpc(Node, M, F, Args) when is_atom(Node) ->
Other
end.

%% Set epmd_port for better coverage
extra_args(ct2) ->
["-epmd_port", "4369"];
extra_args(X) when X == ct5; X == ct6; X == ct7 ->
["-kernel", "prevent_overlapping_partitions", "false"];
extra_args(_) ->
"".

start_node(Sname) ->
{ok, Peer, Node} = ?CT_PEER(#{
name => Sname, connection => standard_io, args => extra_args(Sname)
}),
%% Register so we can find Peer process later in code
register(node_to_peer_name(Node), Peer),
%% Keep nodes running after init_per_suite is finished
unlink(Peer),
%% Do RPC using alternative connection method
ok = peer:call(Peer, code, add_paths, [code:get_path()]),
{Node, Peer}.

%% Returns Peer or Node name which could be used to do RPC-s reliably
%% (regardless if Erlang Distribution works or not)
node_to_peer(Node) when Node =:= node() ->
%% There is no peer for the local CT node
Node;
node_to_peer(Node) when is_atom(Node) ->
case whereis(node_to_peer_name(Node)) of
Pid when is_pid(Pid) ->
Pid;
undefined ->
ct:fail({node_to_peer_failed, Node})
end.

node_to_peer_name(Node) ->
list_to_atom(atom_to_list(Node) ++ "_peer").

receive_message(M) ->
receive
M -> ok
Expand Down
Loading

0 comments on commit 0f69ec9

Please sign in to comment.