Lua 多线程编程无法注册信号处理器

Question

我目前正在使用 torch 来训练多线程深度学习网络，我需要注册一个信号处理程序，以便在我按下

ctrl+c

时可以保存网络的权重。但是问题来了，我发现多线程程序只是忽略了信号，我完全不知道为什么......

见下方代码

-- threadTester.lua
local classic = require 'classic'
local threads = require 'threads'
threads.Threads.serialization('threads.sharedserialize')
local tds = require 'tds'
local threadTester = classic.class('threadTester')

function threadTester:_init(atomic)
    self.game = threads.Threads(1,
    function ()
        inner = atomic
    end)
    classic.strict(self)
end

function threadTester:play()
    self.game:addjob(function ()
        while true do
            if inner:get() < 0 then break end
        end
    end)
    -- do some stuff outside
end

return threadTester

和

local threads = require 'threads'
threads.Threads.serialization('threads.sharedserialize')
local tds = require 'tds'
local signal = require 'posix.signal'

local ctrlpool = threads.Threads(1, function ()
    local tds = require 'tds'
end)

local atomic = tds.AtomicCounter()
atomic:set(1)
nThreads = 4
local ctrlPool = threads.Threads(1)
ctrlPool:addjob(function ()
    local signal = require 'posix.signal'
    signal.signal(signal.SIGINT, function(signum)
        print('\nSIGINT received')
        print('Ex(c)iting')
        atomic:set(-1)
    end)
end)

ctrlPool:synchronize()

local gamePool = threads.Threads(nThreads, function ()
    threadTester = require 'threadTester'
    player = threadTester(atomic)
end)

for i = 1, nThreads do
    gamePool:addjob(function ()
        print(string.format("begin in thread %d", __threadid))
        local status, err = xpcall(player.play, debug.traceback, player)
        if not status then
            print(string.format('%s', err))
            os.exit(128)
        end
    end)
end

gamePool:synchronize()
gamePool:terminate()

macOS 10.12，torch 7，调用

th test.lua

，然后按

ctrl+c

，没有任何反应。 但是如果我改变

test.lua

如下

local threads = require 'threads'
threads.Threads.serialization('threads.sharedserialize')
local tds = require 'tds'
local signal = require 'posix.signal'

-- local ctrlpool = threads.Threads(1, function ()
--  local tds = require 'tds'
-- end)

local atomic = tds.AtomicCounter()
atomic:set(1)
nThreads = 4
-- local ctrlPool = threads.Threads(1)
-- ctrlPool:addjob(function ()
    local signal = require 'posix.signal'
    signal.signal(signal.SIGINT, function(signum)
        print('\nSIGINT received')
        print('Ex(c)iting')
        atomic:set(-1)
    end)
-- end)

-- ctrlPool:synchronize()

local gamePool = threads.Threads(nThreads, function ()
    threadTester = require 'threadTester'
    player = threadTester(atomic)
end)

for i = 1, nThreads do
    gamePool:addjob(function ()
        print(string.format("begin in thread %d", __threadid))
        local status, err = xpcall(player.play, debug.traceback, player)
        if not status then
            print(string.format('%s', err))
            os.exit(128)
        end
    end)
end

gamePool:synchronize()
gamePool:terminate()

一切如预期

> th test.lua
begin in thread 1
begin in thread 2
begin in thread 3
begin in thread 4
^C
SIGINT received
Ex(c)iting

但是如果我使用

qlua

>qlua test.lua
begin in thread 1
begin in thread 2
begin in thread 3
begin in thread 4
^C^C^C^C^C^C^Z
[3]  + 15370 suspended  qlua test.lua

又失败了... 我很困惑...我不知道为什么它在

qlua

中不起作用，而且我确实看到有人在 ctrl 池中注册了信号处理程序...

如果我像这样改变

threadTester.lua

local classic = require 'classic'
local threads = require 'threads'
threads.Threads.serialization('threads.sharedserialize')
local tds = require 'tds'
local threadTester = classic.class('threadTester')

function threadTester:_init(atomic)
    self.game = threads.Threads(1,
    function ()
        inner = atomic
    end)
    self.atomic = atomic
    classic.strict(self)
end

function threadTester:play()
    self.game:addjob(function ()
        os.execute("sleep ".. 10)
    end)
    -- do some stuff outside
    while true do
        if self.atomic:get() < 0 then break end
    end
end

return threadTester

它的功能更奇怪（使用无控制池

test.lua

）

th test.lua
begin in thread 1
begin in thread 2
begin in thread 3
begin in thread 4
^C^C^C^C^C
SIGINT received
Ex(c)iting

好像

os.execute(something)

结束后才收到信号

Answer 1

我把它放在这里以防其他人遇到这个问题。 正在经历的行为正是 luaposix 文档提到的： https://luaposix.github.io/luaposix/modules/posix.signal.html

注意 posix.signal.signal 是用 sigaction(2) 实现的，以实现跨平台的一致语义。另请注意，已安装的信号处理程序不会在信号出现时立即调用。相反，为了保持解释器状态干净，它们在调试钩子的上下文中执行，一旦解释器进入新函数、从当前正在执行的函数返回或当前指令执行完毕后，调试钩子就会被调用。结束了

所以是的，您的信号处理程序仅在

os.execute

返回后运行。