So this proves it: "happens-after Listener.Close()" is not a sufficient 
condition for being able to rebind the address. If another goroutine is in 
a Listener.Accept() call, the new bind must happen-after the return of both 
the Listener.Close() and the Listener.Accept() calls.

So the question is: are there any other conditions that can prevent 
Listener.Close() from resulting in close(2)? Is code that waits for the 
completion of both Close() and Accept() correct code?

I don't understand the polling layer of the runtime to say whether it would 
be feasible for `Accept()` not to hold a reference during `WaitRead()` --- 
but it seems like that would be preferable.

On Tuesday, September 12, 2017 at 12:39:11 AM UTC-4, Dave Cheney wrote:
>
> Yup, and when l.Close is called, Accept returns, releasing the readLock.
>
>
> https://github.com/golang/go/blob/2d69e9e259ec0f5d5fbeb3498fbd9fed135fe869/src/internal/poll/fd_unix.go#L321
>
> On Tuesday, 12 September 2017 14:30:54 UTC+10, Shivaram Lingamneni wrote:
>>
>>
>>
>> On Tuesday, September 12, 2017 at 12:13:15 AM UTC-4, Dave Cheney wrote:
>>>
>>>
>>>
>>> On Tuesday, 12 September 2017 13:40:04 UTC+10, Shivaram Lingamneni wrote:
>>>>
>>>> On Monday, September 11, 2017 at 11:17:01 PM UTC-4, Dave Cheney wrote:
>>>>>
>>>>> The already in use is probably coming from the TCP stack which waits a 
>>>>> certain time before allowing the address to be reused. However I thought 
>>>>> that the net package already used SO_REUSEADDR to avoid the delay in 
>>>>> close 
>>>>> to reopen.
>>>>>
>>>>>
>>>>>> The question I'm really asking is not so much how to write code that 
>>>>>> works in practice (or, rather, appears to do so), but how to be certain 
>>>>>> (on 
>>>>>> the basis of the specification and API documentation) that the code is 
>>>>>> correct.
>>>>>>
>>>>>
>>>>> As written the code is correct. Once the listener is closed, you can 
>>>>> reopen it, modulo TCP stack vagaries. 
>>>>>
>>>>
>>>> The net package is indeed setting SO_REUSEADDR, which allows re-bind on 
>>>> the address immediately after close(2). The problem is that close(2) is 
>>>> not 
>>>> guaranteed to occur as a result of Listener.Close(), because of reference 
>>>> counting of file descriptors. This is not an issue with the TCP stack; the 
>>>> runtime is simply failing to issue the required system call.
>>>>
>>>
>>> I've had a look through the code for the TCPListener and I cannot see 
>>> where the reference count is being bumped by accept. As far as I understand 
>>> the *netFD returned from Accept is unassociated with the *netFD that is 
>>> bound to a listening socket. 
>>>
>>
>> On the one hand, I am more confident in the claim that "close(2) is not 
>> guaranteed to occur as a result of Listener.Close()" than I am in the 
>> specific explanation of `Accept()` holding a reference. On the other hand, 
>> I think I found the relevant line of code:
>>
>>
>> https://github.com/golang/go/blob/2d69e9e259ec0f5d5fbeb3498fbd9fed135fe869/src/internal/poll/fd_unix.go#L318
>>  
>>
>> If I'm reading this correctly, this layer of Accept() acquires a 
>> readLock() on the file, which includes a reference acquire:
>>
>>
>> https://github.com/golang/go/blob/2d69e9e259ec0f5d5fbeb3498fbd9fed135fe869/src/internal/poll/fd_mutex.go#L216
>>
>> and then continues holding this reference when it "blocks" on 
>> `fd.pd.WaitRead`.
>>
>

-- 
You received this message because you are subscribed to the Google Groups 
"golang-nuts" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to