[Cbe-oss-dev] [PATCH 2/2] [POWERPC] spufs: fix invalid scheduling of forgotten contexts

[Jeremy Kerr]

Hi Luke,

> If it could be lazily loaded, it can be on the runqueue as described
> above. Therefore, it is not a bug for the context to be on the
> runqueue at the time spu_forget() is called.

Yep, we've since removed this possiblity with the (initial) SCHED_IDLE 
change.

But I think I see where you're coming from here.

A context that is not in spu_run but in the run queue has been (by 
definition) a bug - this was due to our previous scheduler's 
requirement that only running contexts are loaded, which you've since 
improved with your asynchronous scheduler implementation. We still make 
assumptions based on this invariant (eg, the BUG_ON() in 
spu_update_sched_info()).

So, it's possible that we can now remove this assumption, but we need to 
do some careful checking of the code first, and remove the recent 
additions which check that a context is within spu_run before adding it 
to the run queue.

The down-side to this is that non-running contexts will still use up 
scheduler timeslices for no good reason, as they may be re-loaded by 
spusched_tick, excluding a context that has actual work to do.

Cheers,


Jeremy