Re: 2.2.{9,10} strangeness (random SEGV/ILL in kernel compilation)

Larry Sendlosky (larry@scrugs.lkg.dec.com)
Tue, 22 Jun 99 14:28:19 EDT


Hi all,

Either I'm confused, or nobody reads Richard Henderson's posts
to the lists. Concerning the addition of the tbia patch below, Richard
claims that it is not needed because ASNs are being used. He then
posted a patch yesterday for a bug(s) regarding the wrong ASNs being
set for a context during certain circumstances. Has anybody tried
his patch against 2.2.10 to see if their problems go away?

I've been running his patch on a uni EV6 and a dual EV6 with no
problems - however, I never did see the problem symptoms with a
"virgin" 2.2.9 or 2.2.10.

The patch below is a sizeable performance hit.

larry

> On Tue, Jun 22, 1999 at 11:13:53AM +0400, Alexander V. Lukyanov wrote:
> > Looks like I have found the problem. 2.2.10 with
> > arch/alpha/kernel/entry.S from 2.2.7 works flawlessly for me.
> > Can anybody confirm that?
> >
> > One suspicious change in 2.2.8 was removing `call_pal PAL_tbi' after
> > PAL_swpctx. What could that mean... I have not tried to back-out that
> > part of patch yet, maybe someone will figure out what was wrong with 2.2.8
> > patch.
>
> Ok, I have verified that this patch against 2.2.10-ac2 (re-adding two lines)
> solves my problem:
> --- entry.S-ac2 Mon Jun 21 21:18:50 1999
> +++ entry.S Tue Jun 22 20:28:04 1999
> @@ -525,7 +525,8 @@
> .prologue 0
> bsr $1,do_switch_stack
> call_pal PAL_swpctx
> - unop
> + lda $16,-2($31)
> + call_pal PAL_tbi
> bsr $1,undo_switch_stack
> lda $8,0x3fff
> bic $30,$8,$8
>
> PAL_tbi is "Translation Buffer Invalidate" and -2 argument means "Flush
> all TB entries (tbia)", so I guess it is really needed for context switch.
>
> Alexander.
>

______________________________________________________________________
Larry Sendlosky AMT
larry@scrugs.lkg.dec.com (978) 506-6640 Compaq
----------------------------------------------------------------------