Re: Question about non-blocking Windows pipes

[Ken Brown <kbrown@cornell.edu>]

On 4/6/2021 8:57 AM, Corinna Vinschen via Cygwin-developers wrote:
> Hi Ken,
> 
> On Apr  1 10:39, Ken Brown via Cygwin-developers wrote:
>> Hi Corinna,
>>
>> There are several places in fhandler_socket_unix.cc where you make a
>> distinction between the blocking and nonblocking cases with code like this:
>>
>>    cygwait (evt ?: get_handle (),...)
> 
> I only see this in fhandler_socket_unix::listen_pipe, actually.

You're right.  I was remembering something in sendmsg, but that's a very 
different situation.

>> Here evt is an event handle in the blocking case and is NULL in the
>> nonblocking case.  See, for example, fhandler_socket_unix::listen_pipe.
>>
>> What's the reasoning behind this?  Why not just always create an event or
>> always use the handle?
> 
> In the nonblocking case, the status code returned from NtFsControlFile
> is either a useful status code like STATUS_SUCCESS or an error code,
> or it is STATUS_PENDING.  STATUS_PENDING only means the call is still
> not finished.  To get a useful result, you still need a useful status
> code.  You get that by waiting for the handle.  Note that waiting for
> the handle doesn't mean to wait for the connecting client.  Rather, it's
> signalled as soon as the async NtFsControlFile call finished.

Ah, that's what I was missing.

>  If the
> status code is STATUS_PIPE_LISTENING then, you know that no client tries
> to connect, so you can return EAGAIN.
> 
> The completion event object OTOH, is only signalled if a client actually
> connected, so that's blocking mode.
> 
> Two problems with using an event object in nonblocking mode:
> 
> - The event object is referenced in the call.  If NtFsControlFile returns
>    STATUS_PENDING and you leave the function, you have to use a globally
>    available event object, because this address is used as event object
>    until completion of the NtFsControlFile call (and a client connected).
> 
> - You also have a pending NtFsControlFile until a client connects.
>    This is contrary to what you want in a non-blocking call:  You only
>    want to know *if* a client connects, not wait for it either way.
> 
> Does that help?  I'm not claiming there isn't another way to handle this
> scenario, that's just what I came up with.

Yes, that clears it up completely.  Thanks.

Ken