On Tue, Feb 03, 2004 at 04:30:33PM +0100, Maciej W. Rozycki wrote:
> > > How do we assure tails of interrupt handlers don't trigger the errata?
> >
> > The problem can only be triggered if instructions surrounding the
> > cacheop use the dcache; exceptions such as interrupts are not relevant.
>
> Why? How is an "eret" with its preceding instructions different to other
> instructions? There may be a data cache miss soon before an "eret" and
> the response buffer may contain data. And you may get an exeption right
> before a CACHE instruction.
>
> > Which I'm really happy about. Disabling interrupts is a problem in cases
> > were we can't avoid page faults.
>
> I worry this is unsafe and given the unlikeliness of getting an interrupt
> just between the dummy load and the CACHE instruction, this change creates
> a completely obscure bug that'll bite unpredictably and possibly
> invisibly, just corrupting data, every once and then. But the situation
> may be not that bad -- what does exactly happen when the erratum gets
> triggered? Missing a Create_Dirty_Excl_D operation should itself be a
> performance hit only, but given the problems reported I suppose data gets
> corrupted, either in the cache or in the main memory. Am I right?
I don't know details but since the person who answered my question was
directly working on the CPU design I have to take that as authoritative
information and after all, the systems seems stable.
Daring a guess, the CPU restarts the pipeline following an eret therefore
instructions preceeding the eret can't cause the problem.
Ralf
|