[Cbe-oss-dev] [PATCH 02/02][RFC] spufs: try to work without master run bit

[Jeremy Kerr]

Masato-san,

> I want to fix it witout using master run bit, if possible.
> Does anyone have comments?

Yes, Arnd and I were talking about this issue last week.

We have an invariant in the kernel code that a SPE is only allowed to 
run when the PPE process is in the spu_run() syscall. When there is no 
master run control available, it is possible (by altering the run 
control register through userspace-mapped problem state) to have a SPE 
running when it's PPE process is not in spu_run(). This is bad.

So, our intended approach is to deassert the runcontrol bit when spu_run 
returns, and then unmap (or map read-only) the necessary pages of the 
problem state mapping.

However, someone still needs to do this - it's on my todo list, but if 
you feel like coding something up beforehand, be my guest :)

> BTW, in current implementation, if spe was running at disabled
> by master run, it seems to be failed at next spu_run syscall.
> Is it correct?

That's correct. In order to run a SPE, the run control bit needs to be 
set after the master run control bit is set. So, to start a SPE, we 
need to set the master run control bit, *then* set the problem-state 
run control bit (even if it's already 1).

Regards,


Jeremy