From: "Christoffer Lernö" <lerno@...>
Date: 2007-10-15T00:09:32+09:00
Subject: Re: write_nonblock is blocking?

On 14 Oct 2007, at 09:58, Eric Hodel wrote:

> On Oct 14, 2007, at 24:13 , Christoffer Lernö wrote:
>> On 14 Oct 2007, at 01:23, Eric Hodel wrote:
>>> On Oct 13, 2007, at 14:55 , Christoffer Lernö wrote:
>>>> Anyone know what's up with TCPSocket#write_noblock actually  
>>>> blocking when sending a packet that exceeds the buffer size of  
>>>> the socket in 1.8.6 on Mac.
>>>>
>>>> The behaviour seems slightly counter-productive.
>>>
>>> How long does it block?  I see ruby occasionally being delayed  
>>> 10ms when using write_nonblock, but not repeatably for the same  
>>> packet sizes.  Probably my benchmark script is wrong.
>>>
>>> Could you supply a benchmark script that illustrates your problem?
>>>
>>> #write_nonblock first does rb_io_check_closed() which might call  
>>> down to fseek(3) then write(2) before calling rb_io_set_nonblock 
>>> ().  I'm not even close to being a sockets expert though, I just  
>>> read the man pages.
>>
>> Thanks. Know it "should work" made me try to analyze the  
>> situation. Here is an example code:
>>
>> [...]
>>
>> It looks like write_nonblock together with accept_nonblock causes  
>> some issues.
>>
>> Now I could be using these wrong as both are new (in 1.8.6?)
>
> I couldn't reproduce with your this test using:
>
> $ ruby -v
> ruby 1.8.6 (2007-09-23 patchlevel 5000) [powerpc-darwin8.10.0]
>
> I'll to try again tomorrow though, since I'm tired.
>
> I experimented a bit with your benchmark, and found that writing a  
> large string would only write as much as the socket buffer would  
> accept (81660 bytes, in my case).
>
> Maybe you need to set the TCP_NODELAY socket option?
>
>   socket.setsockopt Socket::IPPROTO_TCP, Socket::TCP_NODELAY, 1
>
> I once made write_nonblock raise Errno::EAGAIN, but I accidentally  
> deleted that change.  I'll play some more tomorrow when I'm not so  
> tired.

Could this be due to me running

ruby 1.8.6 (2007-03-13 patchlevel 0) [i686-darwin8.10.1] ?

The funny thing was that only the socket returned by accept_nonblock  
had this behaviour, that is

require 'socket'
a = TCPServer.new(5000)
socket = TCPSocket.new("localhost", 5000)
other = a.accept
puts "Start Write"
other.write_nonblock("X" * 300000)
puts "End Write"

Works.

require 'socket'
a = TCPServer.new(5000)
socket = TCPSocket.new("localhost", 5000)
other = a.accept_nonblock
puts "Start Write"
other.write_nonblock("X" * 300000)
puts "End Write"

Doesn't work.

What's so special about a socket returned by accept_nonblock?


/Christoffer