[PATCH 06/15] 8xx: Always pin kernel instruction TLB

Joakim Tjernlund joakim.tjernlund at transmode.se
Wed Jun 15 17:36:47 EST 2011


Joakim Tjernlund/Transmode wrote on 2011/06/14 20:00:09:
> From: Joakim Tjernlund/Transmode
>
> Dan Malek <ppc6dev at digitaldans.com> wrote on 2011/06/14 18:06:45:
> >
> >
> > Hi Joakim.
> >
> > On Jun 14, 2011, at 6:54 AM, Joakim Tjernlund wrote:
> >
> > > Various kernel asm modifies SRR0/SRR1 just before executing
> > > a rfi. .....
> >
> > I'm going to argue we can easily visually inspect for this
> > since the code is static with just a couple of RFIs in these
> > exception handlers.
>
> Yes, but then you also miss out on 8xx: Optimize ITLBMiss handler.
>
> >
> > Some 8xx processors have few TLB entries, and always taking
> > one for the kernel, especially if it isn't needed, could have a
> > detrimental effect on the application performance.  Even the
> > "big" 8xx processors don't have that many entries.  Some
> > benchmarks run on an MPC850 would likely show this.
>
> I don't have a mpc850, do you?
>
> >
> > Anyone making modifications to this level of software should
> > know of this problem, or make it known in a comment.  If you
> > are making changes, just compile the code and manually
> > check it with the couple of configuration options that affect
> > the placement of the instructions.
>
> Very fragile but then again, not much are expected to change
> in this area for 8xx.

So I checked and SRR0/SRR1 are fine w.r.t to head_8xx.S, it does
not even come close. There are SRR0/SRR1 mods in entry.S too
which works fine ATM. We don't have the same control of
that file though.
Could you check what impact pinning ITLB on 850 has?

 Jocke



More information about the Linuxppc-dev mailing list