Let's discuss these scenarios during tomorrow's ODP calls.  Could the
various camps put together a brief summary of how they see this organized?
 That would be helpful in framing the discussion.

Thanks.

Bill


On Mon, Sep 1, 2014 at 5:01 AM, Savolainen, Petri (NSN - FI/Espoo) <
[email protected]> wrote:

>
>
> > -----Original Message-----
> > From: ext Wiles, Roger Keith [mailto:[email protected]]
> > Sent: Monday, September 01, 2014 11:27 AM
> > To: Savolainen, Petri (NSN - FI/Espoo)
> > Cc: ext Mike Holmes; [email protected]
> > Subject: Re: [lng-odp] [PATCH v4] Add-global_init-paramiters
> >
> >
> > On Sep 1, 2014, at 3:07 AM, Savolainen, Petri (NSN - FI/Espoo)
> > <[email protected]> wrote:
> >
> > >> diff --git a/platform/linux-generic/odp_init.c b/platform/linux-
> > >> generic/odp_init.c
> > >> index 5b7e192..f595def 100644
> > >> --- a/platform/linux-generic/odp_init.c
> > >> +++ b/platform/linux-generic/odp_init.c
> > >> @@ -8,13 +8,18 @@
> > >>   #include <odp_internal.h>
> > >>   #include <odp_debug.h>
> > >>
> > >> -
> > >> -int odp_init_global(void)
> > >> +int odp_init_global(odp_global_init_t *params  ODP_UNUSED,
> > >> +                 odp_global_platform_init_t *platform_params
> > ODP_UNUSED)
> > >>   {
> > >>        odp_thread_init_global();
> > >>
> > >>        odp_system_info_init();
> > >>
> > >> +     if (odp_init_platform(platform_params)) {
> > >> platform_params is used here, so remove ODP_UNUSED from argument list.
> > >
> > >>> On the other hand, it would be better to remove odp_init_platform()
> > >>> altogether since does not do anything. It's better to add it when
> > actually used.
> > >>> The order of init calls is very important, e.g. now you could not
> > allocate shared
> > >>> memory in odp_init_platform() because it's executed before shm
> init...
> > >>  odp_init_platform() does do something for DPDK / KS2 and we need a
> > place holder because
> > >> we want this to be shared between implementations don't we ?
> > >> This is the stuff that must happen before any ODP init starts to
> > satisfy the platform SDK
> > >
> > > Internals of odp_init_global() is totally platform independent - it
> > cannot be shared in general. It's HW/OS/ODP implementation dependent what
> > needs to be initialized in global init, and in which order (e.g. HW
> > spinlocks before shm, shm before barriers, etc.).
> >
> > All of the platform dependent inits should be in the odp_init_platform()
> > and the calls from odp_init() should be generic for all Linux platforms.
> > The odp_init_platform() should be close to the bottom of the odp_init()
> > routine as no routine called before the platform init should require any
> > thing from the platform.
> >
> > If someone can point out why odp_init() is not generic with specific
> > examples that would help as I do not see any of the above examples as
> > being platform specific. Why would sum, spin locks and barriers (which
> are
> > linux features) need to be in the platform.
>
> The init order depends on implementation internal dependencies. E.g. if a
> queue implementation uses pools, pools have to be initialized before queues
> (and pools cannot use queues for their implementation) -  BUT if a pool
> implementation uses queues, queues have to be init before pools, etc.
> There's no point to standardize implementation internals, here or elsewhere.
>
> Although an implementation may run on Linux, e.g a spinlock may be
> implement with SW, HW queues or more specific HW...
>
>
> -Petri
>
>
> _______________________________________________
> lng-odp mailing list
> [email protected]
> http://lists.linaro.org/mailman/listinfo/lng-odp
>
_______________________________________________
lng-odp mailing list
[email protected]
http://lists.linaro.org/mailman/listinfo/lng-odp

Reply via email to