Re: [EBS-Linux] Re: The semaphore problem w.r.t. "true background" tasks

[John L. <jo...@au...>]

On Fri, 27 Feb 2004, Peter Williams wrote:

> > static inline int sched_normal_effective_prio(task_t *p, runqueue_t *rq)
> > {
> >         uint32_t eps;
> > 
> >         if (unlikely(0 == p->cpu_rate_cap_per_share)) {
> >                 if (user_mode(regs))
> >                         return EBS_BGND_PRI;
> > 		else
> > 			return EBS_MEAN_PRI;
> >         }
> >         ....
> > 
> > Of course 'regs' is undefined above. I'll take a closer look at 
> > user_mode() and struct pt_regs.
> 
> OK
> 
> BTW Should we think about doing something similar for capped tasks?  Or 
> is it unnecessary?

I think doing it for capped tasks won't be necessary. They'll at least get
to run again sooner or later.

However, even if user_mode() could be used, the above is not the complete
solution. To force a background task to call schedule() when it leaves the
kernel, we'll have to ensure that a background task _always_ has
need_resched set when executing in the kernel. If kernel preemption has
been enabled, this will require that every time schedule() chooses a BG
task in kernel mode to run, need_resched will need to be set for it.
Which then necessitates changing that need_resched() check at the end of
schedule() to account for BG tasks... 

The alternative would be to make changes to the low level kernel
entry/exit code for each architecture to explicitly handle BG tasks -
probably not the way to go :-).

John