使用GNU/Linux系统调用`splice`在Haskell中进行零拷贝Socket到Socket数据传输

Question

更新:Nemo先生的回答有助于解决问题!下面的代码包含修复程序!看到nb False和nb True下面调用.

还有一个新的Haskell包被称为splice(具有特定于操作系统的便携式实现,这是一种最着名的socket到socket数据传输循环).

我有以下(Haskell)代码:

#ifdef LINUX_SPLICE
#include <fcntl.h>
{-# LANGUAGE CPP #-}
{-# LANGUAGE ForeignFunctionInterface #-}
#endif

module Network.Socket.Splice (
    Length
  , zeroCopy
  , splice
#ifdef LINUX_SPLICE
  , c_splice
#endif
  ) where

import Data.Word
import Foreign.Ptr

import Network.Socket
import Control.Monad
import Control.Exception
import System.Posix.Types
import System.Posix.IO

#ifdef LINUX_SPLICE
import Data.Int
import Data.Bits
import Unsafe.Coerce
import Foreign.C.Types
import Foreign.C.Error
import System.Posix.Internals
#else
import System.IO
import Foreign.Marshal.Alloc
#endif


zeroCopy :: Bool
zeroCopy =
#ifdef LINUX_SPLICE
  True
#else
  False
#endif


type Length =
#ifdef LINUX_SPLICE
  (#type size_t)
#else
  Int
#endif


-- | The 'splice' function pipes data from
--   one socket to another in a loop.
--   On Linux this happens in kernel space with
--   zero copying between kernel and user spaces.
--   On other operating systems, a portable
--   implementation utilizes a user space buffer
--   allocated with 'mallocBytes'; 'hGetBufSome'
--   and 'hPut' are then used to avoid repeated 
--   tiny allocations as would happen with 'recv'
--   'sendAll' calls from the 'bytestring' package.
splice :: Length -> Socket -> Socket -> IO ()
splice l (MkSocket x _ _ _ _) (MkSocket y _ _ _ _) = do

  let e  = error "splice ended"

#ifdef LINUX_SPLICE

  (r,w) <- createPipe
  print ('+',r,w)
  let s  = Fd x -- source
  let t  = Fd y -- target
  let c  = throwErrnoIfMinus1 "Network.Socket.Splice.splice"
  let u  = unsafeCoerce :: (#type ssize_t) -> (#type size_t)
  let fs = sPLICE_F_MOVE .|. sPLICE_F_MORE
  let nb v = do setNonBlockingFD x v
                setNonBlockingFD y v
  nb False
  finally
    (forever $ do 
       b <- c $ c_splice s nullPtr w nullPtr    l  fs
       if b > 0
         then   c_splice r nullPtr t nullPtr (u b) fs)
         else   e
    (do closeFd r
        closeFd w
        nb True
        print ('-',r,w))

#else

  -- ..    

#endif


#ifdef LINUX_SPLICE
-- SPLICE

-- fcntl.h
-- ssize_t splice(
--   int          fd_in,
--   loff_t*      off_in,
--   int          fd_out,
--   loff_t*      off_out,
--   size_t       len,
--   unsigned int flags
-- );

foreign import ccall "splice"
  c_splice
  :: Fd
  -> Ptr (#type loff_t)
  -> Fd
  -> Ptr (#type loff_t)
  -> (#type size_t)
  -> Word
  -> IO (#type ssize_t)

sPLICE_F_MOVE :: Word
sPLICE_F_MOVE = (#const "SPLICE_F_MOVE")

sPLICE_F_MORE :: Word
sPLICE_F_MORE = (#const "SPLICE_F_MORE")
#endif

注意: 上面的代码现在正常工作!由于Nemo,以下不再有效!

我调用splice上面定义的两个开放和连接的套接字(已经用于使用套接字API send和recv调用传输最少量的握手数据或转换为句柄并与hGetLine和一起使用hPut)并且我一直得到:

Network.Socket.Splice.splice: resource exhausted (Resource temporarily unavailable)

在第一个c_splice调用站点:c_splice返回-1并设置一些errno值(可能EAGAIN)resource exhausted | resource temporarily unavailable在查找时读取.

我splice用不同的Length值测试了调用:1024,8192.

Answer 1

我不知道Haskell,但"资源暂时不可用"是EAGAIN.

并且看起来Haskell默认情况下将其套接字设置为非阻塞模式.因此,如果您在没有数据的情况下尝试从一个读取数据,或者在缓冲区已满时尝试写入一个数据,那么您将失败EAGAIN.

弄清楚如何将套接字更改为阻止模式,我打赌你会解决你的问题.

[更新]

或者,在尝试读取或写入套接字之前select或poll之前.但是你仍然需要处理EAGAIN,因为有一些罕见的极端情况,Linux select将指示套接字已准备好实际上它不是.