NetBSD Problem Report #10186

Received: (qmail 13080 invoked from network); 24 May 2000 01:39:45 -0000
Message-Id: <200005240139.VAA24466@anduin.eldar.org>
Date: Tue, 23 May 2000 21:39:42 -0400 (EDT)
From: Brad Spencer <brad@anduin.eldar.org>
Reply-To: brad@anduin.eldar.org
To: gnats-bugs@gnats.netbsd.org
Subject: Some TMS390Z50 cpu module core dumps
X-Send-Pr-Version: 3.95

>Number:         10186
>Category:       port-sparc
>Synopsis:       Some TMS390Z50 cpu module core dumps
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    port-sparc-maintainer
>State:          closed
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Wed May 24 01:40:01 +0000 2000
>Closed-Date:    Thu Jun 29 21:57:34 +0000 2000
>Last-Modified:  Thu Jun 29 21:57:34 +0000 2000
>Originator:     Brad Spencer
>Release:        1.4Y supped mid May
>Organization:
	Sitting at home
>Environment:

System: NetBSD frodo.eldar.org 1.4Y NetBSD 1.4Y (GENERIC) #5: Mon May  8 02:30:19 EDT 2000     mycroft@lop-nor.ihack.net:/u2/src/sys/arch/sparc/compile/GENERIC sparc


>Description:

I could not find a PR on this...


I have a brand new, old, Sparc 10 with a CPU module that identifies
itself on boot as:

SPARCstation 10  (1 X 390Z50), No Keyboard
ROM Rev. 2.12, 32 MB memory installed,

... and in the boot up messages ...

cpu0 at mainbus0: TMS390Z50 v1 @ 36 MHz, on-chip FPU
cpu0: physical 20K instruction (64 b/l), 16K data (32 b/l): cache enabled

This particular module does not have a SuperCache controller on it.

With this module installed, certain programs will drop core at times
in pretty predictable ways.  In particular, running 'ifconfig -a' a
couple of times will produce bus errors, seg faults, or complains
about invalid command line arguments.  Note that the command is simply
repeated over and over and was not changing.

The following shell script, call it Y, will usually start to act up:

#!/bin/sh

echo $@
exit 0

running stuff like './Y -a' over and over in a row will usually
produce buss errors in a short time, may be 2 out of 5 times it is
run.

I'v tried a couple of different modules of the same type and they all
act the same.


Running with a module that identifies itself as:

SPARCstation 10  (1 X 390Z55), No Keyboard
ROM Rev. 2.12, 32 MB memory installed,

cpu0 at mainbus0: TMS390Z50 v1 @ 40.300 MHz, on-chip FPU
cpu0: physical 20K instruction (64 b/l), 16K data (32 b/l), 1024K external (32 b/l): cache enabled

This is a TMS390 with a SuperCache controller onboard.  This CPU
module appears to work fine.  I have no reason to believe that there
is anything wrong with the hardware in the SS10.

>How-To-Repeat:

Have a TMS390Z50 without a SuperCache controller and run 'ifconfig -a'
a couple of times.  Watch for core dumps.

>Fix:

Unknown.  Some sort of cache flushing problem??  Maybe timeing??


>Release-Note:
>Audit-Trail:
State-Changed-From-To: open->analyzed 
State-Changed-By: pk 
State-Changed-When: Wed May 24 00:59:54 PDT 2000 
State-Changed-Why:  
This is most likely caused by cache incoherencies brought about in 
pmap_zero_page(), which since the activation of `idle time zeroing' 
have a much greater possibility of surfacing. 

To verify that this is the case, you can patch the variable  
`vm_page_zero_enable' in uvm_page.c to 0, and see if the problem occurs 
less frequently. 

From: <abs@mono.org>
To: Brad Spencer <brad@anduin.eldar.org>
Cc: gnats-bugs@gnats.netbsd.org
Subject: Re: port-sparc/10186: Some TMS390Z50 cpu module core dumps
Date: Wed, 24 May 2000 13:56:40 +0100 (BST)

 	Just a datapoint - I've got a similar machine running 1.4.1 in
 	production.

 mainbus0 (root): SUNW,SPARCstation-10
 cpu0 at mainbus0: TMS390Z50 v1 @ 36 MHz, on-chip FPU
 cpu0: physical 20K instruction (64 b/l), 16K data (32 b/l): cache enabled

 	Been up the past 60 days without incident. It might be an issue
 	introduced in current, but unfortunately I can't easily test
 	-current on it. Could you try booting 1.4.X on your system to
 	confirm?

 		David/absolute
 				       -- www.netbsd.org: No hype required --



From: Jim Bernard <jbernard@mines.edu>
To: Brad Spencer <brad@anduin.eldar.org>
Cc: gnats-bugs@gnats.netbsd.org, pk@netbsd.org, jbernard@mines.edu
Subject: Re: port-sparc/10186: Some TMS390Z50 cpu module core dumps
Date: Wed, 24 May 2000 21:11:40 -0600

 On Tue, May 23, 2000 at 09:39:42PM -0400, Brad Spencer wrote:
 > 
 > >Number:         10186
 > >Category:       port-sparc
 > >Synopsis:       Some TMS390Z50 cpu module core dumps

   I've been seeing this too, on a sparc 20:

 cpu0 at mainbus0: TMS390Z50 v0 or TMS390Z55 @ 50 MHz, on-chip FPU
   (it's actually a TMS390Z50)
 cpu0: physical 20K instruction (64 b/l), 16K data (32 b/l): cache enabled

 The system is rock solid on kernels built from sources from April 28 or
 earlier, but binaries randomly segfault, bus-error, or, in one case, get
 floating-point exceptions, under kernels from May 1 or later sources (the
 details seem to depend on the kernel, and unknown other parameters).  I've
 narrowed it down to having appeared with the check-in of the following
 files in arch/sparc/{include,sparc}:

   ctlreg.h    1.20, 1.21
   cache.c     1.48, 1.49
   pmap.h      1.45, 1.46
   cpuvar.h    1.23
   cpu.c       1.98
   pmap.c      1.158, 1.159
   ganassym.cf 1.25
   locore.s    1.128

 Most of the changes appear to be related to implementation of the page idle
 zero facility.  I'm not sure exactly what caused the breakage, though, nor
 whether there is actually a bug in the new code, or whether it simply exposed
 an existing bug.

From: Brad Spencer <brad@anduin.eldar.org>
To: jbernard@mines.edu
Cc: gnats-bugs@gnats.netbsd.org, pk@netbsd.org, jbernard@mines.edu
Subject: Re: port-sparc/10186: Some TMS390Z50 cpu module core dumps
Date: Wed, 24 May 2000 23:39:12 -0400 (EDT)

    On Tue, May 23, 2000 at 09:39:42PM -0400, Brad Spencer wrote:
    > 
    > >Number:         10186
    > >Category:       port-sparc
    > >Synopsis:       Some TMS390Z50 cpu module core dumps

      I've been seeing this too, on a sparc 20:

    cpu0 at mainbus0: TMS390Z50 v0 or TMS390Z55 @ 50 MHz, on-chip FPU
      (it's actually a TMS390Z50)
    cpu0: physical 20K instruction (64 b/l), 16K data (32 b/l): cache enabled

    The system is rock solid on kernels built from sources from April 28 or
    earlier, but binaries randomly segfault, bus-error, or, in one case, get
    floating-point exceptions, under kernels from May 1 or later sources (the
    details seem to depend on the kernel, and unknown other parameters).  I've
    narrowed it down to having appeared with the check-in of the following
    files in arch/sparc/{include,sparc}:

      ctlreg.h    1.20, 1.21
      cache.c     1.48, 1.49
      pmap.h      1.45, 1.46
      cpuvar.h    1.23
      cpu.c       1.98
      pmap.c      1.158, 1.159
      ganassym.cf 1.25
      locore.s    1.128

    Most of the changes appear to be related to implementation of the page idle
    zero facility.  I'm not sure exactly what caused the breakage, though, nor
    whether there is actually a bug in the new code, or whether it simply exposed
    an existing bug.



 I got the following from pk@netbsd.org after I submited the PR.  I am
 going to try this as soon as I can, but that may not be until the weekend.


 >This is most likely caused by cache incoherencies brought about in
 >pmap_zero_page(), which since the activation of `idle time zeroing'
 >have a much greater possibility of surfacing.
 >
 >To verify that this is the case, you can patch the variable 
 >`vm_page_zero_enable' in uvm_page.c to 0, and see if the problem occurs
 >less frequently.





 Brad Spencer - brad@anduin.eldar.org   http://anduin.eldar.org
 [finger brad@anduin.eldar.org for PGP public key]

From: Jim Bernard <jbernard@mines.edu>
To: Brad Spencer <brad@anduin.eldar.org>
Cc: gnats-bugs@gnats.netbsd.org, pk@netbsd.org, jbernard@mines.edu
Subject: Re: port-sparc/10186: Some TMS390Z50 cpu module core dumps
Date: Thu, 25 May 2000 12:53:47 -0600

 > 
 > I got the following from pk@netbsd.org after I submited the PR.  I am
 > going to try this as soon as I can, but that may not be until the weekend.
 > 
 > 
 > >This is most likely caused by cache incoherencies brought about in
 > >pmap_zero_page(), which since the activation of `idle time zeroing'
 > >have a much greater possibility of surfacing.
 > >
 > >To verify that this is the case, you can patch the variable 
 > >`vm_page_zero_enable' in uvm_page.c to 0, and see if the problem occurs
 > >less frequently.

   OK, I've tried this (actually, I recompiled with vm_page_zero_enable
 turned off) with sources updated May 23 (MDT), and the resulting kernel
 works just fine.  The same sources, unmodified, produced a kernel exhibiting
 the failing binaries problem.

   Maybe it doesn't matter anymore, but I've also narrowed down the
 locations of the changes that either caused or exposed the problem.
 The list of candidates I sent last night was:

 >      ctlreg.h    1.20, 1.21
 >      cache.c     1.48, 1.49
 >      pmap.h      1.45, 1.46
 >      cpuvar.h    1.23
 >      cpu.c       1.98
 >      pmap.c      1.158, 1.159
 >      ganassym.cf 1.25
 >      locore.s    1.128

 And I've since updated that old source tree to

   ctlreg.h    1.20
   cache.c     1.49
   pmap.h      1.45
   cpuvar.h    1.23
   cpu.c       1.98
   pmap.c      1.158
   genassym.cf 1.25

 which produced a working kernel.  When I then updated locore.s to 1.128 (which,
 if I understand correctly, essentially activates the page-zero code implemented
 in some of the earlier commits), the resulting kernel exhibited the problem.

 --Jim

From: Mason Loring Bliss <mason@acheron.middleboro.ma.us>
To: gnats-bugs@gnats.netbsd.org
Cc: brad@anduin.eldar.org, abs@netbsd.org, jbernard@mines.edu
Subject: Re: port-sparc/10186: Some TMS390Z50 cpu module core dumps
Date: Sat, 27 May 2000 07:56:45 -0400

 I just built and booted a 1.4Z kernel, and I tried setting
 vm_page_zero_enable to FALSE in uvm_page.c as recommended. Built this
 way, the 1.4Z kernel seems to be solid, exhibiting none of the behaviours
 mentioned in the PR.

 To add myself to the list:

 cpu0 at mainbus0: TMS390Z50 v1 @ 36 MHz, on-chip FPU
 cpu0: physical 20K instruction (64 b/l), 16K data (32 b/l): cache enabled

 (Should the difference not leap out at you, mine's a v1, not a  v0,
 whatever that indicates.)

 -- 
 Mason Loring Bliss             (( "In the drowsy dark cave of the mind dreams
 mason@acheron.middleboro.ma.us  )) build  their nest  with fragments  dropped
 http://acheron.ne.mediaone.net ((  from day's caravan." - Rabindranath Tagore


From: Brad Spencer <brad@anduin.eldar.org>
To: mason@acheron.middleboro.ma.us
Cc: gnats-bugs@gnats.netbsd.org, abs@netbsd.org, jbernard@mines.edu
Subject: Re: port-sparc/10186: Some TMS390Z50 cpu module core dumps
Date: Sat, 27 May 2000 10:53:13 -0400 (EDT)

    I just built and booted a 1.4Z kernel, and I tried setting
    vm_page_zero_enable to FALSE in uvm_page.c as recommended. Built this
    way, the 1.4Z kernel seems to be solid, exhibiting none of the behaviours
    mentioned in the PR.

    To add myself to the list:

    cpu0 at mainbus0: TMS390Z50 v1 @ 36 MHz, on-chip FPU
    cpu0: physical 20K instruction (64 b/l), 16K data (32 b/l): cache enabled

    (Should the difference not leap out at you, mine's a v1, not a  v0,
    whatever that indicates.)

    -- 
    Mason Loring Bliss             (( "In the drowsy dark cave of the mind dreams
    mason@acheron.middleboro.ma.us  )) build  their nest  with fragments  dropped
    http://acheron.ne.mediaone.net ((  from day's caravan." - Rabindranath Tagore



 I, also, [finally] got around to building a new kernel with
 vm_page_zero_enable set to FALSE.  The resulting kernel did not have the
 problems I mentioned in the port-sparc/10186.  It seemed to be quite
 solid.

 There certainly appears to be some interaction between that processor
 module and idle time zeroing of pages.




 Brad Spencer - brad@anduin.eldar.org   http://anduin.eldar.org
 [finger brad@anduin.eldar.org for PGP public key]

State-Changed-From-To: analyzed->feedback 
State-Changed-By: pk 
State-Changed-When: Tue Jun 6 10:27:44 PDT 2000 
State-Changed-Why:  
I've adapted the physical cache flush routines to deal with this situation. 
Can you update to -current kernel source (6/6/2000) and try it out on your 
machine? 
State-Changed-From-To: feedback->closed 
State-Changed-By: pk 
State-Changed-When: Thu Jun 29 14:55:43 PDT 2000 
State-Changed-Why:  
Improved cache handling on non-MXCC vikings has fixed this problem. 
>Unformatted:

NetBSD Home
NetBSD PR Database Search

(Contact us) $NetBSD: query-full-pr,v 1.51 2026/08/10 02:28:17 riastradh Exp $
$NetBSD: gnats_config.sh,v 1.10 2026/05/13 22:00:09 riastradh Exp $
Copyright © 1994-2026 The NetBSD Foundation, Inc. ALL RIGHTS RESERVED.