Re: IPSec kernel oops on ppc64

Joy Latten [off-list ref] wrote:

quoted

I installed 2.6.17 + patch-2.6.18-rc4 + 2.6.18-rc4-mm2
onto two pSeries power 5 (ppc64 lpars) machines. I configured
IPSec using the configuration listed below. 

Could you try straight 2.6.17? If that crashes too, then at least we
can be sure that it isn't something new.

A straight 2.6.17 kernel does not crash and my pings work.
A 2.6.17 + patch-2.6.18-rc4 does crash and my pings do not work.
The above tests were done on a ppc64. 
I can try patch-2.6.18-rc1, etc... to see which one it stops
working on to narrow it down.

Regards,
Joy

From: Joy Latten <redacted>
Date: Mon, 28 Aug 2006 17:25:15 -0500

I can try patch-2.6.18-rc1, etc... to see which one it stops
working on to narrow it down.

If you could do this in the meanwhile, it would help us out
a lot.

Thanks.

On Mon, Aug 28, 2006 at 05:25:15PM -0500, Joy Latten wrote:

A straight 2.6.17 kernel does not crash and my pings work.
A 2.6.17 + patch-2.6.18-rc4 does crash and my pings do not work.
The above tests were done on a ppc64. 

Thanks for that info.  This does sound like a bug.

Could you please generate a dump of the stack/register contents and
a disassembly of the code around the crash?

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} [off-list ref]
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Hi:

Thanks to some excellent info from Joy, I've trakced the problem down to
a broken implementation of memmove on ppc64, which in turn is due to a
broken memcpy.

Both memmove and memcpy need to return the destination pointer as per
the C standard.  The ppc64 version of memcpy returns a pointer to the
last byte copied instead.  Since memmove degenerates into memcpy in
some cases, this makes it broken as well.

Something like the following patch should fix the problem.  Be warned
that I know absolutely nothing about ppc assembly so don't come back
to me if this patch eats your disk :)

BTW, I spotted a couple of typos in memcpy_64.S/copyuser_64.S where
"cmpldi cr1,..." became "cmpldi r1,...".  So please fix that up while
you ppc64 guys are at it.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} [off-list ref]
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
--

diff --git a/arch/powerpc/lib/memcpy_64.S b/arch/powerpc/lib/memcpy_64.S
index fd66acf..9e8d9e7 100644
--- a/arch/powerpc/lib/memcpy_64.S
+++ b/arch/powerpc/lib/memcpy_64.S

@@ -11,6 +11,7 @@ #include <asm/ppc_asm.h>
 
 	.align	7
 _GLOBAL(memcpy)
+	std	r3,-8(r1)
 	mtcrf	0x01,r5
 	cmpldi	cr1,r5,16
 	neg	r6,r3		# LS 3 bits = # bytes to 8-byte dest bdry

@@ -53,7 +54,8 @@ _GLOBAL(memcpy)
 2:	bf	cr7*4+3,3f
 	rotldi	r9,r9,8
 	stb	r9,0(r3)
-3:	blr
+3:	ld	r3,-8(r1)
+	blr
 
 .Lsrc_unaligned:
 	srdi	r6,r5,3

@@ -167,4 +169,5 @@ _GLOBAL(memcpy)
 3:	bf	cr7*4+3,4f
 	lbz	r0,0(r4)
 	stb	r0,0(r3)
-4:	blr
+4:	ld	r3,-8(r1)
+	blr