[Cbe-oss-dev] [PATCH] spu preemptive scheduling problems

[Christoph Hellwig]

On Tue, Mar 20, 2007 at 11:47:13AM -0300, Luke Browning wrote:
> > > I have tested this patch in a limited way.  I wrote a testcase that
> > > creates 32 contexts and then schedules them, so that there are 
> > > lots of runqueue operations, timeslicing, context switches.
> > > yielding, ...
> > > 
> > > I think this patch may fix a recently reported bug, where a DSI occurs
> > > when yielding.  At least that is the defect that drove these changes.
> > 
> > What does DSI stand for?
> > 
> 
> data storage interrupt.  I was talking about defect 32882.  I don't know
> if it will fix this problem, but I noticed the window between the sleep
> and wakeup code cited in the first problem above.  
> 
> Can you check if this fixes that defect? 

I'll ask Thomas to try with your patch.

> so what is the next step?  I think that the locking change mentioned 
> above should be put into sdk 2.1.

Yes.

> Changing the return code for
> spu_deactivate() is probably not worth it given that you are
> contemplating eliminating spu_yield().  I think that would be worth
> doing in SDK 2.1.  

I'm not sure how to go on with the yield removal.  We should probably
leave that up to the release managers.