2.6.21-rc5-mm[12]: Oops on bootup in xor_see_2

2.6.21-rc5-mm[12]: Oops on bootup in xor_see_2

[Bernhard Rosenkraenzer]

I'm getting this Oops when booting an 2.6.21-rc5-mm1 or 2.6.21-rc5-mm2 on an 
Acer Aspire 1501 LMi in 32 bit mode (2.6.21-rc4-mm1 + hotfixes works 
perfectly):

xor: automatically using best checksumming function: pIII_sse

BUG: unable to handle kernel NULL pointer dereference at virtual address 
00000010
printing eip:
c02cb724
*pde=00000000
Oops: 0002 [#1]
last sysfs file:
Modules linked in:
CPU: 0
EIP: 0060:[<c02cb724>] Not tainted VLI
EFLAGS: 00010212 (2.6.12-rc5-mm2)
EIP is at xor_sse_2+0x34/0x200
eax: 00000010 ebx: fffea4df ecx: df8c3000 edx: df8c0000
esi: 8005003b edi: c03ebce0 ebp: df8c3000 esp: dfd01ef8
ds: 007b es: 007b fs: 00d8 gs: 0000 ss: 0068
Process swapper (pid: 1, ti=dfd000000 task=c146aa10 task:ti=dfd00000
Stack: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
       00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
       c13f1800 fffea4df 00000000 c02caa83 c03ebce0 df8c0000 00000000 c011bd7b
Call Trace:
  [<c02caa83>] do_xor_speed+0x53/0xf0
  [<c011bd7b>] printk+0x1b/0x20
  [<c02cac02>] calibrate_xor_block+0x2e/0x1b0
  [<c03fe7a2>] kernel_init+0x92/0x1c0
  [<c0103dee>] ret_from_fork+0x6/0x1c
  [<c03fe710>] kernel_init+0x0/0x1c0
  [<c03fe710>] kernel_init+0x0/0x1c0
  [<c0104937>] kernel_thread_helper+0x7/0x10
=======================
Code: 44 89 74 24 48 0f 20 c6 0f 06 0f 11 04 24 0f 11 4c 24 10 0f 11 54 24 20 
0f
      11 5c 24 30 0f 18 82 00 01 00 00 0f 18 82 20 01 00 00 <00> 00 00 00 00 
00
      00 00 00 00 00 00 00 00 00 00 00 00 00 00 00

Re: 2.6.21-rc5-mm[12]: Oops on bootup in xor_see_2

[Andrew Morton]

On Sat, 31 Mar 2007 17:36:47 +0000 Bernhard Rosenkraenzer <bero@arklinux.org> wrote:

> I'm getting this Oops when booting an 2.6.21-rc5-mm1 or 2.6.21-rc5-mm2 on an 
> Acer Aspire 1501 LMi in 32 bit mode (2.6.21-rc4-mm1 + hotfixes works 
> perfectly):
> 
> xor: automatically using best checksumming function: pIII_sse
> 
> BUG: unable to handle kernel NULL pointer dereference at virtual address 
> 00000010
> printing eip:
> c02cb724
> *pde=00000000
> Oops: 0002 [#1]
> last sysfs file:
> Modules linked in:
> CPU: 0
> EIP: 0060:[<c02cb724>] Not tainted VLI
> EFLAGS: 00010212 (2.6.12-rc5-mm2)
> EIP is at xor_sse_2+0x34/0x200
> eax: 00000010 ebx: fffea4df ecx: df8c3000 edx: df8c0000
> esi: 8005003b edi: c03ebce0 ebp: df8c3000 esp: dfd01ef8
> ds: 007b es: 007b fs: 00d8 gs: 0000 ss: 0068
> Process swapper (pid: 1, ti=dfd000000 task=c146aa10 task:ti=dfd00000
> Stack: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
>        00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
>        c13f1800 fffea4df 00000000 c02caa83 c03ebce0 df8c0000 00000000 c011bd7b
> Call Trace:
>   [<c02caa83>] do_xor_speed+0x53/0xf0
>   [<c011bd7b>] printk+0x1b/0x20
>   [<c02cac02>] calibrate_xor_block+0x2e/0x1b0
>   [<c03fe7a2>] kernel_init+0x92/0x1c0
>   [<c0103dee>] ret_from_fork+0x6/0x1c
>   [<c03fe710>] kernel_init+0x0/0x1c0
>   [<c03fe710>] kernel_init+0x0/0x1c0
>   [<c0104937>] kernel_thread_helper+0x7/0x10
> =======================
> Code: 44 89 74 24 48 0f 20 c6 0f 06 0f 11 04 24 0f 11 4c 24 10 0f 11 54 24 20 
> 0f
>       11 5c 24 30 0f 18 82 00 01 00 00 0f 18 82 20 01 00 00 <00> 00 00 00 00 
> 00
>       00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> -

Dan, I'm assuming your changes in there are the cause of this.  AFAICT all
you're doing is moving code around, so it's a bit odd.

Bernhard, the config would be useful please.

Neil, is calibrate_xor_block() being rational?

	b1 = (void *) __get_free_pages(GFP_KERNEL, 2);
	...
	b2 = b1 + 2*PAGE_SIZE + BENCH_SIZE;

Is it correct to add BENCH_SIZE to b2 here?

Re: 2.6.21-rc5-mm[12]: Oops on bootup in xor_see_2

[Neil Brown]

On Saturday March 31, akpm@linux-foundation.org wrote:
> On Sat, 31 Mar 2007 17:36:47 +0000 Bernhard Rosenkraenzer <bero@arklinux.org> wrote:
> 
> > I'm getting this Oops when booting an 2.6.21-rc5-mm1 or 2.6.21-rc5-mm2 on an 
> > Acer Aspire 1501 LMi in 32 bit mode (2.6.21-rc4-mm1 + hotfixes works 
> > perfectly):
> > 
> > xor: automatically using best checksumming function: pIII_sse

Hmmm... I've seen that before

> > Code: 44 89 74 24 48 0f 20 c6 0f 06 0f 11 04 24 0f 11 4c 24 10 0f 11 54 24 20 
> > 0f
> >       11 5c 24 30 0f 18 82 00 01 00 00 0f 18 82 20 01 00 00 <00> 00 00 00 00 
> > 00
> >       00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > -

Them bytes that look like '00' are mostly supposed to look like 'F0' -
an x86 'noop'.

Apparently a bin-utils bug.  An alpha of OpenSUSE-10.3 was compiling
kernels like this.  I'm told it has been fixed.

What distro/gcc version/binutils version was this compiled on?

It is possible that we should change the ".align 32" in
include/asm-i386/xor.h to ".align 32, 0xf0" or similar, but my
assembler knowledge stopped growing when it reached 6809, so I'd
rather someone who knew something about it did anything that was
required....

> 
> Dan, I'm assuming your changes in there are the cause of this.  AFAICT all
> you're doing is moving code around, so it's a bit odd.
> 
> Bernhard, the config would be useful please.
> 
> Neil, is calibrate_xor_block() being rational?
> 
> 	b1 = (void *) __get_free_pages(GFP_KERNEL, 2);
> 	...
> 	b2 = b1 + 2*PAGE_SIZE + BENCH_SIZE;
> 
> Is it correct to add BENCH_SIZE to b2 here?

I think it is intentional.  I cannot say if it is correct.
The purpose is to defeat cache effects somehow.
The result of that code together with the value of BENCH_SIZE being
PAGE_SIZE is that 4 pages are allocated, and the first and last are
used.
I have no idea how any cache would treat this, but that seems to be
the intent of the code.

NeilBrown

Re: 2.6.21-rc5-mm[12]: Oops on bootup in xor_see_2

[Bernhard Rosenkraenzer]

On Saturday 31 March 2007, Neil Brown wrote:
> > > Code: 44 89 74 24 48 0f 20 c6 0f 06 0f 11 04 24 0f 11 4c 24 10 0f 11 54
> > > 24 20 0f
> > >       11 5c 24 30 0f 18 82 00 01 00 00 0f 18 82 20 01 00 00 <00> 00 00
> > > 00 00 00
> > >       00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > > -
>
> Them bytes that look like '00' are mostly supposed to look like 'F0' -
> an x86 'noop'.
>
> Apparently a bin-utils bug.  An alpha of OpenSUSE-10.3 was compiling
> kernels like this.  I'm told it has been fixed.
>
> What distro/gcc version/binutils version was this compiled on?

You're right -- I'm using the current Ark Linux 2007.1 experimental build 
(always the latest stuff) - the update to binutils 2.17.50.0.13 broke it, 
after upgrading binutils to 2.17.50.0.14 everything is working perfectly 
again.

Thanks
bero