Re: [RFC] statically initialize spu FILE values

[Patrick Mansfield <patmans at us dot ibm dot com>]

On Tue, Apr 10, 2007 at 12:25:12PM +0900, Kazunori Asayama wrote:
> Patrick Mansfield <patmans@us.ibm.com> wrote:
> > On Mon, Apr 09, 2007 at 10:53:57AM +0900, Kazunori Asayama wrote:
> > > I propose to modify CHECK_INIT macro as following instead of remove
> > > it, to minimize runtime overhead and code size, and still to do
> > > cleanup correctly:
> > > 
> > >     #define CHECK_INIT(ptr) ((ptr)->__sdidinit = 1, (ptr)->__cleanup = __cleanup)
> > 
> > I wanted to remove all those calls. And we are still missing close of
> > files opened via open.
> > 
> > Are you OK with closing all opened files (for both fopen and open) in libspe? 
> > 
> > Then we could remove stdio.c and the spu CHECK_INIT.
> 
> I basically agree. But unfortunately, correspondence between SPE
> context and FILE is NOT managed in the current libspe implementation
> (both 1 and 2), and in addition it's quite difficult to implement such
> management routine w/ the current libspe specification because a
> 'library callback' has no way to know which SPE context calls the
> callback.

By callback do you mean an assisted (or offload) call?

So if there are multiple threads for a given process, each with its own
SPE context, libspe must track which SPE context opened a file, if it is
to close the file on completion of the context?

> To implement correct cleanup sequence in libspe, I suppose that the
> following works will be needed:
> 
>  - to change or extend the libspe specification of library callback in
>    order to make it possible that callbacks can handle SPE context in
>    the callbacks,
>  - to implement the new spec,
>  - to add management routine of correspondence between SPE context and
>    FILE in libspe, and
>  - to add cleanup sequence in libspe.
> 
> In my personal opinion, I think that the current spec of callback is
> insufficient not only for default callbacks but also for application
> specific callbacks, and that some kind of change or extension of the
> spec is needed as mentioned above.
> 
> Anyway, at this time, could you leave the current behavior (SPU newlib
> performs cleanup) until libspe will be equipped with sufficient
> mechanism ?

Yes, sounds good.

-- Patrick Mansfield