Rewriting the FIFO code

[Ken Brown]

On 1/23/2019 8:05 AM, Corinna Vinschen wrote:
> On Jan 22 20:44, Ken Brown wrote:
>> On 12/26/2018 2:03 PM, Corinna Vinschen wrote:
>>> On Dec 26 14:00, Ken Brown wrote:
>>>> Thanks for the encouragement, and Happy Holidays to you too.  One question: In
>>>> the new AF_UNIX socket code you mostly used NT functions, but the existing FIFO
>>>> code uses Win32 functions.  Do you prefer NT functions for new code?
>>>
>>> The NT functions have some advantages over the Win32 functions.
>>> For instance, WaitNamedPipe is not interruptible, while
>>> NtFsControlFile(FSCTL_PIPE_WAIT) can be called asynchronously
>>> and then you can just wait for an event object via cygwait
>>> (see fhandler_socket_unix::wait_pipe_thread).
>>>
>>> So, in theory I'd prefer NT functions, but if you feel uncomfortable
>>> and just want to implement away, feel free to go ahead with Win32
>>> functions.
>>
>> I've decided to use Win32 functions for now, not because of discomfort but
>> because it allowed me to use more of the existing code.  I can always change
>> that later.
> 
> Sounds good to me.  Just... can we make the fifo pipes PIPE_NOWAIT, by
> any chance and make fhandler_fifo a derived class of fhandler_base?

I'm working on this now and have two questions.

First, if I make fhandler_fifo derived from fhandler_base, don't I have to use 
NT functions for creating pipes, etc.?  (This isn't a problem, but I just want 
to make sure I understand).  For one thing, my understanding is that 
asynchronous I/O can only be done using overlapped I/O or NT functions.  For 
another thing, fhandler_base uses NT functions for I/O, so I think I have to be 
consistent with that.

Second, fhandler_base_overlapped uses an atomic write buffer.  Is there a 
problem giving that up?  Do I need to find some other way of making sure that 
writes of PIPE_BUF bytes or fewer are atomic?

Ken