--liOOAslEiF7prFVr
Content-Type: text/plain; charset=us-ascii

On Thu, Oct 05, 2000 at 04:41:53PM -0700, Jason R Thorpe wrote:
> Folks...
> 
> So, I just heard for the first time today that some people hare having
> K6 issues wrt. zero page idling -- wasn't aware of that.  So, this goes
> back to the problems we saw with the Cyrix CPUs.  I think the way the
> pmap is doing TLB invalidation is a little sketchy.
> 
> I've attached a diff that I'd like people with Cyrix CPUs to try.  It's
> relative to the 1.5 branch.  Please see if this works in the presence
> of idle page zero'ing.

I tried it in -current (diff attached, I think I got it rigth), I seems to
have solved the panic on my:
NetBSD 1.5G (ROCHA) #0: Fri Oct  6 20:54:35 MEST 2000
    bouyer@rocha.antioche.eu.org:/home/bouyer/1.5F/arch/i386/compile/ROCHA
cpu0: Cyrix 6x86 (486-class)
total memory = 65148 KB
avail memory = 39196 KB

I compiled a kernel without problems; before pmap_zero_page_uncached() was
dissabled on cyrix it would no compile a kernel without one or 2 panics.

--
Manuel Bouyer <bouyer@antioche.eu.org>
--

--liOOAslEiF7prFVr
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment; filename=diff

--- pmap.c.orig	Fri Sep 29 16:04:21 2000
+++ pmap.c	Fri Oct  6 20:20:31 2000
@@ -2109,16 +2109,11 @@
 pmap_zero_page(pa)
 	paddr_t pa;
 {
-	simple_lock(&pmap_zero_page_lock);
-#ifdef DIAGNOSTIC
-	if (*zero_pte)
-		panic("pmap_zero_page: lock botch");
-#endif
 
+	simple_lock(&pmap_zero_page_lock);
 	*zero_pte = (pa & PG_FRAME) | PG_V | PG_RW;	/* map in */
+	pmap_update_pg((vaddr_t)zerop);			/* flush TLB */
 	memset(zerop, 0, NBPG);				/* zero */
-	*zero_pte = 0;				/* zap! */
-	pmap_update_pg((vaddr_t)zerop);		/* flush TLB */
 	simple_unlock(&pmap_zero_page_lock);
 }
 
@@ -2136,13 +2131,10 @@
 	int i, *ptr;
 
 	simple_lock(&pmap_zero_page_lock);
-#ifdef DIAGNOSTIC
-	if (*zero_pte)
-		panic("pmap_zero_page_uncached: lock botch");
-#endif
 
 	*zero_pte = (pa & PG_FRAME) | PG_V | PG_RW |	/* map in */
 	    ((cpu_class != CPUCLASS_386) ? PG_N : 0);
+	pmap_update_pg((vaddr_t)zerop);			/* flush TLB */
 	for (i = 0, ptr = (int *) zerop; i < NBPG / sizeof(int); i++) {
 		if (sched_whichqs != 0) {
 			/*
@@ -2156,8 +2148,6 @@
 		}
 		*ptr++ = 0;
 	}
-	*zero_pte = 0;					/* zap! */
-	pmap_update_pg((vaddr_t)zerop);			/* flush TLB */
 	simple_unlock(&pmap_zero_page_lock);
 
 	return (rv);
--- machdep.c.orig	Fri Sep 29 16:52:24 2000
+++ machdep.c	Fri Oct  6 20:10:42 2000
@@ -796,7 +796,7 @@
 	 * XXX disable page zero in the idle loop, it seems to
 	 * cause panics on these CPUs.
 	 */
-	vm_page_zero_enable = FALSE;
+	vm_page_zero_enable = TRUE;
 }
 
 void

--liOOAslEiF7prFVr--