simple bootloader 2.6.10-rc3 8xx

[Wolfgang Denk]

Dear Alex,

in message <313680C9A886D511A06000204840E1CF0A64741E at whq-msgusr-02.pit.comms.marconi.com> you wrote:
>  
> >Moving things around in the kernel sometimes 
> >makes this go away but really its just hidden.
> >In my case it would explode in zlib_inflate also.  Putting some output 
> >in the routine would move it ...
> 
> Yes that is exactly what I have observed (but I did'nt have the correct
> explanation for it) 

I think Paul may be right  with  his  suspicion;  it  sounds  like  a
manifestation  of the CPU15 problem (but it could also be incorrectly
initialized SDRAMs which fail  during  heavy  load  with  burst  mode
accesses, or one more instability of the 2.6 kernel on 8xx systems).

Note that all our work relating to this problem was done on  the  2.4
kernel  base,  and  I  make no claims that it might actually help for
your 2.6 problems.

> Could this code be used for 2.6 ?

The patch does not apply cleanly as is to the 2.6 tree, but it should
be simple enough to port.

> Could someone point me to it, please ?

It's in the linuxppc_2_4_devel tree on our CVS server.

Please  find  attached  the  patch  to  the  2.4  kernel  tree  which
implements  the  workaround,  and  a  test  application  to  reliably
reproduce the CPU15 errata. It usually takes  1-10  minutes  for  the
application to crash, if the workaround is disabled:

	bash-2.05b# date; ./cpu15 ; date
	Thu Aug  2 02:02:42 MEST 2001
	Illegal instruction
	Thu Aug  2 02:02:54 MEST 2001
	bash-2.05b# date; ./cpu15 ; date
	Thu Aug  2 02:02:58 MEST 2001
	Illegal instruction
	Thu Aug  2 02:03:48 MEST 2001
	bash-2.05b# date; ./cpu15 ; date
	Thu Aug  2 02:03:54 MEST 2001
	Illegal instruction
	Thu Aug  2 02:03:57 MEST 2001
	bash-2.05b# date; ./cpu15 ; date
	Thu Aug  2 02:04:07 MEST 2001
	Illegal instruction
	Thu Aug  2 02:04:16 MEST 2001
	bash-2.05b#

We noticed that some background activitly, like logging in/out to the
target via 'telnet', while the application is running over the serial
console, additionally speeds up the crash.

You can use this command to enable the workaround at run-time (it  is
disabled by default):

	bash# sysctl -w kernel.8xx_cpu15=1

Please note that for the h/w problem  to  show  up,  the  instruction
cache  has  to  be enabled.

Hope this helps.

Best regards,

Wolfgang Denk

-- 
Software Engineering:  Embedded and Realtime Systems,  Embedded Linux
Phone: (+49)-8142-66989-10 Fax: (+49)-8142-66989-80 Email: wd at denx.de
"Wish not to seem, but to be, the best."                  - Aeschylus

-------------- next part --------------
A non-text attachment was scrubbed...
Name: patch
Type: application/octet-stream
Size: 3305 bytes
Desc: patch
Url : http://ozlabs.org/pipermail/linuxppc-embedded/attachments/20041228/3f2645ee/attachment.obj 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cpu15.c
Type: application/octet-stream
Size: 3411 bytes
Desc: cpu15.c
Url : http://ozlabs.org/pipermail/linuxppc-embedded/attachments/20041228/3f2645ee/attachment-0001.obj