erlang 虚机crash

      现网服务,每次更新一个服务时,另外一个集群所有node 都跟着同时重启一遍,这么调皮,这是闹哪样啊。。

  看系统日志:/var/log/messages

  Oct 30 15:19:41 localhost kernel: beam.smp[21880]: segfault at 7fa300006d4b ip 00007fa300006d4b sp 00007fa3d0d7c788 error 14 in locale-archive[7fa31616f000+5e91000

  beam crash了,好吧开始回忆咱用了哪些c库, 都不应该有问题啊

  嗯,打开core dump,再复现一遍。嗯,在线上复现嗷,设计成完全不影响业务的重启还是很有用的。

  不一会dump粗来了,挂上gdb 很快找到出错堆栈:

#0 0x00007fa300006d4b in ?? ()
#1 0x00007fa3aa83dd96 in quicksort () from /data0/xxx_0.4.4/lib/hash_ring-0.1.6/priv/hash_ring_drv.so
#2 0x00007fa3aa83d026 in hash_ring_remove_node () from /data0/xxx_0.4.4/lib/hash_ring-0.1.6/priv/hash_ring_drv.so
#3 0x00007fa3aa83c295 in hash_ring_drv_output () from /data0/xxx_0.4.4/lib/hash_ring-0.1.6/priv/hash_ring_drv.so
#4 0x00000000004924ef in call_driver_output (c_p=0x7fa3b0ec8f50, flags=2064, prt=0x7fa3d8d40bc0, from=55559696966931,
list=140341277983642, refp=0x0) at beam/io.c:1768

嗯,定位到出问题位置在于c库依赖hash_ring 

顺便说下服务间调用一致性hash的实现:

1. 通过gen_server 定时rpc:call 目标nodes 指定服务运行状态,并动态管理hash_ring 中动态节点。

init([Configs]) ->
    %hash_ring 需要先启动, link 需要同时重启
    link(whereis(hash_ring)),
    ets:new(?MODULE, [named_table, protected, set, {read_concurrency, true}]),
    ets:new(?ROUND_ROBIN_ETS, [named_table, public, set, {write_concurrency, true}]),
    ets:insert(?ROUND_ROBIN_ETS, {inc, 0}),
    {ok, Routes} = parse_configs(Configs),
    State = apply_routes(Routes, #state{}),
    start_check_timer(),
    {ok, State}.

  

monitor_route({Svc, Node} = Route) ->
    % monitor 无法立即返回是否成功,rpc:call 成功后再monitor
    case catch rpc:call(Node, erlang, whereis, [Svc], 3000) of
        {'EXIT', Reason} ->
            {error, Reason};
        undefined ->
            {error, svc_undefined};
        Pid when is_pid(Pid) ->
            Ref = erlang:monitor(process, Route),
            {ok, Ref};
        Error ->
            {error, Error}
    end.

route_up(Name, Route, Ref, #state{mons=Mons, downs=Downs} = State) ->
    case lists:member(Route, get_all_routes(Name)) of
       true ->
           case lists:member(Route, get_routes(Name)) of
                false ->
                    ok = hash_ring:add_node({route, Name}, term_to_binary(Route)),
                    add_route(Name, Route);
                true ->
                    ok
            end,
            State#state{mons=dict:store(Ref, {Name, Route}, Mons),
                        downs=Downs -- [{Name, Route}]};
        _ ->
            catch erlang:demonitor(Ref),
            lager:info("igonre route_up:~p ~p ~p", [Name, Route, Ref]),
            State
    end.

 

  2. 使用自己的gen_call 替代 rpc:call 调用,节省monitor 资源消耗

route(Name, Key) ->
    case hash_ring:find_node({route, Name}, neo_util:to_binary(erlang:phash2(Key))) of
        {ok, Route} -> {ok, binary_to_term(Route)};
        Error -> Error
    end.

call(Name, Key, Req, Timeout) ->
    {ok, Dst} = route(Name, Key),
    gen_call(Dst, Req, Timeout).

%参考 whatsapp 做法
%使用场景:
%1. Process 需要被长期monitor
%2. Process 为node内常驻服务进程,不是临时进程
%3. 调用前需要先确认Process alive 状态
%优势:
%不需要monitor, 节省两次网络交互
%gen:call 需要monitor -> call -> demonitor dst node monitor 操作消耗资源
%影响:
%process down 瞬间, call 应答都会超时,但此调用返回的是timeout
%
gen_call(Process, Request) ->
    gen_call(Process, Request, 5000).

gen_call(Process, Request, Timeout) ->
    Ref = erlang:make_ref(),
    catch erlang:send(Process, {'$gen_call', {self(), Ref}, Request}),
    receive
        {Ref, Reply} ->
            Reply
    after Timeout ->
            exit({timeout, {?MODULE, call, [Process, Request, Timeout]}})
    end.

 

  问题原因是:https://github.com/chrismoos/hash-ring/blob/master/sort.c#52

  快排算法中STACK_SIZE=1024 移除节点时,实现会对剩余节点做一次排序。节点过多时,数组就越界了。

 

  解决方案:

  sort.h 快排算法是有问题的,但先不急着优化排序算法。

  其实一致性hash的添加加点和删除节点也能能够做到O(1) 的,为此我提交了一个patch修复该问题。

  方案:

  1. 移除节点,每个虚拟节点保存物理节点的引用,删除时只需要将空位campact

2. 添加节点,只需要计算新加节点并做快排,而不要所有节点再次排序,做一次二路归并即可

    quicksort((void**)adds, ring->numReplicas, item_sort);
+    //ring->numNodes * ring->numReplicas
+    if(ring->items == NULL || ring-> numNodes == 1) {
+        ring->items = adds;
+    } else {
+        size_t size_new = sizeof(hash_ring_item_t*) * ring->numNodes * ring->numReplicas;
+        hash_ring_item_t **news = (hash_ring_item_t **)malloc(size_new);
+        hash_ring_item_t **olds = ring->items;
+        if(news == NULL) {
+            return HASH_RING_ERR;
+        }
+        int oldlen = (ring->numNodes - 1) * ring->numReplicas;
+        int addlen = ring->numReplicas;
+        int i=0, j=0, k=0;
+        while(1) {
+            if(i == oldlen && j == addlen) {
+                break;
+            }
+            if(j == addlen) {
+                news[k++] = olds[i++];
+                continue;
+            }
+            if(i == oldlen) {
+                news[k++] = adds[j++];
+                continue;
+            }
+            int ret = item_sort(olds[i], adds[j]);
+            if(ret == 0) {
+                news[k++] = olds[i++];
+                news[k++] = adds[j++];
+            } else if (ret < 0) {
+                news[k++] = olds[i++];
+            } else {
+                news[k++] = adds[j++];
+            }
+        };
+        free(adds);
+        free(olds);
+        ring->items = news;

 

  另:

  这个hash 算法其实还是不太可靠的,虚拟节点数字相同怎么办?

  一般来说所有节点添加,删除次序相同即可保证一致性,当然最好能够通过nodename 对相同虚节点做排序。

posted @ 2014-12-07 11:06  LittlePeng  阅读(1192)  评论(1编辑  收藏  举报