haskell - 递归 IO 函数中的内存泄漏 - PAP

Question

我编写了一个名为 amqp-worker 的库它提供了一个名为 worker 的函数，该函数轮询消息队列(如 RabbitMQ )以查找消息，并在找到消息时调用处理程序。然后返回轮询。

内存泄漏。我已经对其进行了分析，图表显示 PAP(偏函数应用程序)是罪魁祸首。 我的代码哪里有泄漏？在 IO 中使用 forever 进行循环时，如何避免泄漏？

以下是一些相关函数。 The full source is here .

Example Program 。这泄露了

main :: IO ()
main = do
  -- connect
  conn <- Worker.connect (fromURI "amqp://guest:guest@localhost:5672")

  -- initialize the queues
  Worker.initQueue conn queue
  Worker.initQueue conn results

  -- publish a message
  Worker.publish conn queue (TestMessage "hello world")

  -- create a worker, the program loops here
  Worker.worker def conn queue onError (onMessage conn)

worker

worker :: (FromJSON a, MonadBaseControl IO m, MonadCatch m) => WorkerOptions -> Connection -> Queue key a -> (WorkerException SomeException -> m ()) -> (Message a -> m ()) -> m ()
worker opts conn queue onError action =
  forever $ do
    eres <- consumeNext (pollDelay opts) conn queue
    case eres of
      Error (ParseError reason bd) ->
        onError (MessageParseError bd reason)

      Parsed msg ->
        catch
          (action msg)
          (onError . OtherException (body msg))
    liftBase $ threadDelay (loopDelay opts)

consumeNext

consumeNext :: (FromJSON msg, MonadBaseControl IO m) => Microseconds -> Connection -> Queue key msg -> m (ConsumeResult msg)
consumeNext pd conn queue =
    poll pd $ consume conn queue

poll

poll :: (MonadBaseControl IO m) => Int -> m (Maybe a) -> m a
poll us action = do
    ma <- action
    case ma of
      Just a -> return a
      Nothing -> do
        liftBase $ threadDelay us
        poll us action

Answer 1

这是一个非常简单的示例，可以演示您的问题:

main :: IO ()
main = worker

{-# NOINLINE worker #-}
worker :: (Monad m) => m ()
worker =
  let loop = poll >> loop
  in loop

poll :: (Monad m) => m a
poll = return () >> poll

如果删除 NOINLINE，或将 m 专门化为IO(使用-O编译时)，泄漏消失了。

我写了详细的blogpost关于为什么正是这段代码泄漏了内存。正如里德在他的书中指出的那样，快速总结是答案是，代码创建并记住一系列部分应用程序>>s.

我还提交了 ghc ticket关于这个。