cvs commit: src/sys/dev/io iodev.c

[Robert Watson]

On Sat, 9 Aug 2008, Kostik Belousov wrote:

> On Sat, Aug 09, 2008 at 11:27:58AM +0100, Robert Watson wrote:
>>
>> On Sat, 9 Aug 2008, Peter Jeremy wrote:
>>
>>> On 2008-Aug-08 12:26:31 -0400, John Baldwin <jhb at freebsd.org> wrote:
>>>> It should be setting D_TRACKCLOSE though so that close() reliably clears 
>>>> the flag even in single-threaded processes.  You can still get odd 
>>>> behavior if you explicitly open it twice in an app and then close one of 
>>>> the two fd's. You will no longer have IO permission even though you still 
>>>> have one fd open. However, if you do that I think you deserve what you 
>>>> asked for. :)
>>>
>>> That behaviour may be legitimate:  Your code links with libraries foo and 
>>> bar that each independently open /dev/io so they can frob different things 
>>> in IO space.  libfoo needs ongoing access to device foo and so keeps its 
>>> descriptor open.  libbar only needs once-off access to device bar and so 
>>> closes /dev/io once it's finished its initialisation.  Libraries foo and 
>>> bar are completely independent and shouldn't need to know anything about 
>>> each other and your app shouldn't need to know that libraries it's using 
>>> frob around in IO space.
>>
>> If that's the view, there should probably be a per-process counter, 
>> although this is all a bit tricky anyway since file descriptors and 
>> processes have a tenuous relationship.
>
> Another interesting issue is the close on exec, esp. for /dev/io.
>
> It seems that Linux recently grown full new API to handle FD_CLOEXEC races, 
> see http://udrepper.livejournal.com/20407.html

While /dev/io appeals to the UNIX "everything is a file" sensibility, I think 
the system calls we have for this on i386 are more conceptually coherent.

Robert N M Watson
Computer Laboratory
University of Cambridge