呃有点没理解的Socket Busy

ilskdw · 2014-04-01T13:01:29+00:00

Hello! 2014-03-31 4:09 GMT-07:00 , 139**850: > 你好,春哥: > 我现在遇到一个&#x9...

呃有点没理解的Socket Busy

ilskdw

最小测试代码

local server = require "resty.websocket.server"

local redis = require "resty.redis"

local ws, err = server:new {

timeout = 6000000,

max_payload_len = 65535

}

local red2 = redis:new()

red2:set_timeout(6000000)

red2:connect("127.0.0.1", 6379)

local keys = {"dog", "cat"}

local res, err = red2:subscribe(keys)

if not res then

ngx.log(ngx.ERR, err)

end

function clean()

local res, err = ws:send_close()

ngx.log(ngx.INFO, "closing")

if not res then

ngx.log(ngx.ERR, err)

end

res, err = red2:unsubscribe(keys)

if not res then

ngx.log(ngx.ERR, err)

end

red2:set_keepalive(2000, 100)

ngx.exit(444)

end

ngx.on_abort(clean)

local res, err = red2:read_reply()

if not res then

ngx.log(ngx.ERR, err)

end

print(res[3])

clean()

输出日志

2014/04/01 12:57:22 [notice] 72181#0: start worker processes

2014/04/01 12:57:22 [notice] 72181#0: start worker process 72182

2014/04/01 12:57:34 [info] 72182#0: *1 kevent() reported that client prematurely closed connection, client: 127.0.0.1, server: , request: "GET /t HTTP/1.1", host: "127.0.0.1:8060"

2014/04/01 12:57:34 [info] 72182#0: *1 [lua] test.lua:22: closing, client: 127.0.0.1, server: , request: "GET /t HTTP/1.1", host: "127.0.0.1:8060"

2014/04/01 12:57:34 [error] 72182#0: *1 [lua] test.lua:28: socket busy, client: 127.0.0.1, server: , request: "GET /t HTTP/1.1", host: "127.0.0.1:8060"

这时候打开一个table，连接上websocket后关掉这个table，就有可能发生socket busy。我观察到大概是90%的几率出现，还有10%正常，这是Bug么？

agentzh

Hello!

On Mon, Mar 31, 2014 at 10:01 PM, Kelvin Peng wrote:
> 2014/04/01 12:57:34 [info] 72182#0: *1 [lua] test.lua:22: closing, client:
> 127.0.0.1, server: , request: "GET /t HTTP/1.1", host: "127.0.0.1:8060"
> 2014/04/01 12:57:34 [error] 72182#0: *1 [lua] test.lua:28: socket busy,
> client: 127.0.0.1, server: , request: "GET /t HTTP/1.1", host:
> "127.0.0.1:8060"
>
> 这时候打开一个table，连接上websocket后关掉这个table，就有可能发生socket
> busy。我观察到大概是90%的几率出现，还有10%正常，这是Bug么？
>

看起来是你的 Lua 代码里的 bug，而不是 ngx_lua 的问题。“socket busy”错误是当有两个“轻线程”同时使用一个
cosocket 对象时抛出的。你这里的 ngx.on_abort()
注册的回调是会运行在独立的“轻线程”中的，这个“轻线程”和你请求本身的主“轻线程”是同时（并发）运行的，所以如果你在 clean()
函数里访问同一套 cosocket 是很容易遇到“socket busy”错误的。你需要自己进行简单的锁保护（比如通过自己 Lua 变量来作
flag）。

Regards,
-agentzh

ilskdw

而且还有个问题，subscribe不支持多key sub？

这个测试里面2个key subscribe之后，我在redis-cli publish dog “xxxxx” 发现没收到消息

On Tue, Apr 1, 2014 at 1:01 PM, Kelvin Peng <il...@gmail.com> wrote:

最小测试代码

local server = require "resty.websocket.server"
local redis = require "resty.redis"

local ws, err = server:new {
timeout = 6000000,
max_payload_len = 65535
}

local red2 = redis:new()
red2:set_timeout(6000000)
red2:connect("127.0.0.1", 6379)

local keys = {"dog", "cat"}

local res, err = red2:subscribe(keys)
if not res then
ngx.log(ngx.ERR, err)

end

function clean()
local res, err = ws:send_close()
ngx.log(ngx.INFO, "closing")
if not res then
ngx.log(ngx.ERR, err)

end
res, err = red2:unsubscribe(keys)
if not res then
ngx.log(ngx.ERR, err)
end
red2:set_keepalive(2000, 100)
ngx.exit(444)

end

ngx.on_abort(clean)

local res, err = red2:read_reply()
if not res then
ngx.log(ngx.ERR, err)
end
print(res[3])

clean()

输出日志
2014/04/01 12:57:22 [notice] 72181#0: start worker processes
2014/04/01 12:57:22 [notice] 72181#0: start worker process 72182
2014/04/01 12:57:34 [info] 72182#0: *1 kevent() reported that client prematurely closed connection, client: 127.0.0.1, server: , request: "GET /t HTTP/1.1", host: "127.0.0.1:8060"

2014/04/01 12:57:34 [info] 72182#0: *1 [lua] test.lua:22: closing, client: 127.0.0.1, server: , request: "GET /t HTTP/1.1", host: "127.0.0.1:8060"

2014/04/01 12:57:34 [error] 72182#0: *1 [lua] test.lua:28: socket busy, client: 127.0.0.1, server: , request: "GET /t HTTP/1.1", host: "127.0.0.1:8060"

这时候打开一个table，连接上websocket后关掉这个table，就有可能发生socket busy。我观察到大概是90%的几率出现，还有10%正常，这是Bug么？

--

-CMGS

A simple coder.

Love travel, sports especially outdoor sports and computer technology.

Have a dream that one day can tour around.

ilskdw

原来如此，callback也是在light thread里面跑的是吧= =

明白了~

2014-04-01 13:05 GMT+08:00 Yichun Zhang (agentzh) <age...@gmail.com>:

Hello!

On Mon, Mar 31, 2014 at 10:01 PM, Kelvin Peng wrote:
> 2014/04/01 12:57:34 [info] 72182#0: *1 [lua] test.lua:22: closing, client:
> 127.0.0.1, server: , request: "GET /t HTTP/1.1", host: "127.0.0.1:8060"
> 2014/04/01 12:57:34 [error] 72182#0: *1 [lua] test.lua:28: socket busy,
> client: 127.0.0.1, server: , request: "GET /t HTTP/1.1", host:
> "127.0.0.1:8060"
>
> 这时候打开一个table，连接上websocket后关掉这个table，就有可能发生socket
> busy。我观察到大概是90%的几率出现，还有10%正常，这是Bug么？
>

看起来是你的 Lua 代码里的 bug，而不是 ngx_lua 的问题。“socket busy”错误是当有两个“轻线程”同时使用一个
cosocket 对象时抛出的。你这里的 ngx.on_abort()
注册的回调是会运行在独立的“轻线程”中的，这个“轻线程”和你请求本身的主“轻线程”是同时（并发）运行的，所以如果你在 clean()
函数里访问同一套 cosocket 是很容易遇到“socket busy”错误的。你需要自己进行简单的锁保护（比如通过自己 Lua 变量来作
flag）。

Regards,
-agentzh

--

-CMGS

A simple coder.

Love travel, sports especially outdoor sports and computer technology.

Have a dream that one day can tour around.

agentzh

Hello!

On Mon, Mar 31, 2014 at 10:05 PM, CMGS wrote:
> 而且还有个问题，subscribe不支持多key sub？

只要你的 redis 服务器支持，lua-resty-redis 就支持。lua-resty-redis 是对 redis 协议的非常浅的封装。

如果你认为是 lua-resty-redis 的问题，请提供一个最小化的完整用例，以便别人可以在自己的环境里复现问题。（另外，请尽量不要绑架不相干的邮件主题。）

Regards,
-agentzh