[nanomsg] Re: Getting rid of nn_init() and nn_term()
- From: Paul Colomiets <paul@xxxxxxxxxxxxxx>
- To: nanomsg@xxxxxxxxxxxxx
- Date: Mon, 4 Feb 2013 10:04:42 +0200
Hi Martin,
On Mon, Feb 4, 2013 at 9:35 AM, Martin Sustrik <sustrik@xxxxxxxxxx> wrote:
> Hi all,
>
> I would like to do yet one more API change before it's too late and API is
> cast in stone.
>
> The idea is to remove nn_init() and nn_term() functions altogether. The
> library would be initialised when first socket is created and terminated
> when the last socket is closed.
>
> Considerations:
>
> First, nn_close() would have to be synchronous, ie. wait for the pending
> outbound data to be sent or for linger interval to expire. This is done in
> current codebase anyway (and it simplifies the code a lot), however, it's
> not consistent with how 0MQ behaves.
>
> Second, ETERM functionality which is useful for shutting down multi-threaded
> applications would have to behave a bit differently. My proposal would be to
> re-purpose the nn_term() function to be an optional way of saying "make any
> subsequent call to nanomsg library, except for nn_close(), fail with ETERM).
> If it is not called, it's up to the user to notify all the threads about the
> shutdown. After all, if there's only one thread in the application, there's
> no need for such notification anyway.
>
> Thoughts?
>
I find the following two situations subtle:
1. It is unsafe to create two sockets in two different threads, unless
you already created at least one socket. Am I right?
2. When you close and reopen socket, sometimes worker threads are shut
down and started again. Sometimes not.
Closing and opening sockets is useful for some applications. While I
don't have good use cases that closes sockets many times per second, I
might be not very creative :)
--
Paul
Other related posts:
