Re: [Libusb-devel] Darwin: Loosing connection to IOService at fork()

[Johannes E. <joh...@er...>]

On Wed, Dec 18, 2002, Nathan Hjelm <hj...@un...> wrote:
> On Wed, 18 Dec 2002, Pete Bentley wrote:
> 
> > On Wed, Dec 18, 2002 at 12:17:41PM +0100, Beat Birkhofer wrote:
> > > I try to get sane/libusb working on OSX.
> > > 
> > > I get quite far but then the there's a fork() in the sane backend 
> > > (plustek) I use and then I loose the connection to the IOService.
> > > 
> > > Any ideas how to solve this problem? I'm quite clueless.
> > 
> > I suspect this may be rather tricky to solve...
> > 
> > On most Unixes, the device_handle t libusb uses to communicate
> > with the device is a file handle, which will get inherited
> > by the child of a fork() call.
> > 
> > On OS X, the IOKit functions that are used to implement libusb
> > seem to use mach ports for communicating with the kernel rather
> > than file handles... I don't know the semantics of ports well,
> > but I'm guessing they don't get shared across a fork() call
> > (or at least not without a lot of extra work)
> 
> From what I understand, you are correct, ports are not shared across fork 
> calls. I am not sure what can be done to fix the problem except by using 
> threads rather than forks (though, i am not entirely sure that will work 
> either).
> 
> I am open to suggestions/ideas on how to attack this within libusb so 
> developers dont have to change their code.

This also brings up some interesting questions for libusb 1.0

For asynchronous support, I had plans of implementing a wrapper for poll
and select calls, kind of like how adns works, but it sounds like that
will complicate matters.

Does Darwin have any support for asynchronous communication?

JE