On Wed, Dec 04, 2002 at 06:08:22PM +0100, Kevin D. Kissell wrote:
> > > I think that Carsten's patch (or equivalent) should certainly be
> > > applied to the main tree, but I wonder how relevant it is here.
> > > The flushes associated with trampolines don't do indexed
> > > flush operations, do they?
> >
> > True, but are we sure that it's the trampoline that's the problem here?
>
> Jun Sun seemed to think it was. To quote his original message
>
> "The problem involves emulating a "lw" instruction in cp1 branch delay
> slot, which needs to set up trampoline in user stack. The net effect
> looks as if the icache line or dcache line is not flushed properly."
>
> I don't know what his actual observations were that lead to that
> conclusion, but the resemblence to what was reported under LTP
> with the pre-break_cow()-patch kernel intrigues me.
Here's some of the actual observations: if you single-step over the
bc1t instruction, then it comes out as you'd expect; the load in the
delay slot was executed. Even if you breakpoint in the general
vicinity and then continue.
But if you breakpoint _after_ the instruction, it is evident that the
load did not occur as expected.
> So, I repeat...
> > > ...I don't have a 4Kc platform at
> > > hand, but I think that Jun Sun *may* have found a better
> > > way to get at the other problem I was referring to, which
> > > we rarely saw on non-superscalar issue CPUs, and which
> > > seems to be masked by an otherwise superfluous flush of
> > > the Icache that was added to the latest versions of break_cow().
> > > If Carsten's patch solves the problem without applying that
> > > other update, I'd want to know that. If it *doesn't*, I'd be
> > > really interested to know if, by any chance, there is a
> > > corelation between failures of Jun Sun's test and the incidence
> > > of page faults on the CACHE op in protected_icache_invalidate_line().
> > >
> > > Kevin K.
>
>
>
--
Daniel Jacobowitz
MontaVista Software Debian GNU/Linux Developer
|