Re: [Xenomai-main] Modification scheduler XENOMAI

[chgans]

Selon Philippe Gerum <address@hidden>:

> address@hidden wrote:
>  > Selon Philippe Gerum <address@hidden>:
>  > 
>  > > address@hidden wrote:
>  > >  > Hi all,
>  > >  > 
>  > >  > We are using XENOMAI on top of POSIX, we're using it for avionic OS
>  > > emulation.
>  > >  > Our first tests make us very enthousiam.
>  > >  > 
>  > >  > But the OS we have to emulate is a little special, it is base on a
> concept
>  > > of
>  > >  > partitionning, a partition is a set of process, and these partition
> are
>  > > serially
>  > >  > scheduled. and  within each partition processes are schedule
> "normaly"
>  > >  > eg:
>  > >  > 
>  > >  >
>  > >
> |_____PART1______|___PART2___|_PART3_||_____PART1______|___PART2___|_PART3_|
>  > > 
>  > >  > etc .....
>  > >  > 
>  > >  > with some process in PART1, some other in PART2 etc....
>  > > 
>  > > Arinc653?
>  > Yes it is derived from this.
>  > 
>  > And the partitionning is really a problem for us... 
>  > We want to make changes as clean as possible, but the scheduler doesn't
> seem to
>  > be prepared for this.... and the difficulty is to not break Xenomai
> abstraction.
>  > 
>  > Our first idea was to play with the scheduler hook, but we can do
> nothing
>  > interesting with this.
> 
> Provided the number of processes is bound and relatively small into
> these partitions (IIRC, it costs about 100-150ns to suspend a
> non-current thread on a mid-range Celeron), maybe you could use a
> specialized Xenomai thread with a higher priority than any other
> process acting as the timeframe scheduler?

> 
> This would look this this:
> 
>      void timeframe_sched (void) {
> 
>      for (;;) {
>           xnpod_delay(ticks_to_next_partition);
>         foreach (threads in runqueue[current_part_id]) {
>            xnpod_suspend_thread(thread,XNSLICE,XN_INFINITE,NULL,NULL);
>         }
>         foreach (threads in runqueue[next_part_id]) {
>            xnpod_resume_thread(thread,XNSLICE);
>         }
>      }
> 
> The XNSLICE flag does not exist yet, but you could define it using a
> spare bit from thread.h, and add it to the XNTHREAD_BLOCK_BITS
> mask. NOTE: do not use XNSUSP for this purpose, otherwise you would
> get brain-damage conflicts with synchronization objects which use it.

OK, thanks, i will try this, i had think about a similar way but i didn't know
about the possibility to add a bit (XNSLICE) in the XNTHREAD_BLOCK_BITS bitmask.

Thanks a lot.

best regards,
Christian


> 
> Philippe.
>