Re: Issue booting v2.6.39 .. v3.4-rc6 on hp712/100

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Tue, 2012-05-15 at 14:23 -0400, John David Anglin wrote:
> On 5/15/2012 5:13 AM, James Bottomley wrote:
> > On Tue, 2012-05-15 at 09:09 +0100, James Bottomley wrote:
> >> On Mon, 2012-05-14 at 18:38 -0400, John David Anglin wrote:
> >>> On 14-May-12, at 6:11 PM, Helge Deller wrote:
> >>>
> >>>> The B160L and the 715/64 (both 32bit-only PA1.X machines) crashed
> >>>> with the following trace.
> >>>> All logs attached.
> >>>>
> >>>> Any ideas?
> >>> This is exactly the same failure as reported by Vincent.
> >>>
> >>> The most likely problem is the PA 1.1 tmpalias support in entry.S is
> >>> broken.  For example,
> >>> the cache stride that is loaded in flush_dcache_page_asm to register
> >>> r1 is wrong.  Probably,
> >>> the do_alias macro is wrong for PA 1.1.  This is hunk of code that
> >>> should be executed
> >>> when a fdc non access fault occurs.
> >>>
> >>> nadtlb_check_alias_11:
> >>>           do_alias        spc,t0,t1,va,pte,prot,nadtlb_emulate
> >>>
> >>>           idtlba          pte,(va)
> >>>           idtlbp          prot,(va)
> >>>
> >>>           rfir
> >>>           nop
> >>>
> >>> The TLB insert instructions on PA 1.1 have a different format than on
> >>> PA 2.0.  I'm not sure
> >>> how this would corrupt r1.
> >>>
> >>> On the other hand, I had asked Vincent to put a "b,n ." instruction
> >>> just before the fdc loop,
> >>> boot, hit the TOC button, and capture the setup registers for the
> >>> flush operation.  It's possible
> >>> the stride variable has been clobbered.
> >> Actually, I don't think it's that.  I built a PA 1.1 only kernel and
> >> booted it successfully on the C360.  That exercises all the _11 paths,
> >> so I don't think there's a code fault.  I do think there's a non PA1.1
> >> instruction in there somewhere that the C360 wouldn't notice.
> > OK, I think this is the problem.  We have a depd instruction in do_alias
> > which is now in the _11 fault paths.
> >
> > This should be the fix, if someone wants to test it before I can get
> > around to building it.
> >
> > James
> >
> > ---
> >
> > diff --git a/arch/parisc/kernel/entry.S b/arch/parisc/kernel/entry.S
> > index 6f05944..5350342 100644
> > --- a/arch/parisc/kernel/entry.S
> > +++ b/arch/parisc/kernel/entry.S
> > @@ -581,7 +581,11 @@
> >   	 */
> >   	cmpiclr,=	0x01,\tmp,%r0
> >   	ldi		(_PAGE_DIRTY|_PAGE_READ|_PAGE_WRITE),\prot
> > +#ifdef CONFIG_64BIT
> >   	depd,z		\prot,8,7,\prot
> > +#else
> > +	depw,z		\prot,8,7,\prot
> > +#endif
> >   	/*
> >   	 * OK, it is in the temp alias region, check whether "from" or "to".
> >   	 * Check "subtle" note in pacache.S re: r23/r26.
> Better check that this fix doesn't break 32-bit PA 2.0 as the seven bit 
> are now being deposited
> in a different place.

It doesn't.  It's the same on both 32 and 64 bit: I compared with the
original code.

James