Re: kernel unaligned accesses on 2.6.29.

Re: kernel unaligned accesses on 2.6.29.

[Robin Holt]

Had a few minutes to look around.

On Tue, Mar 24, 2009 at 12:28:57PM -0500, Robin Holt wrote:
> I just built and booted an ia64 2.6.29 kernel.  While accessing NFS
> filesystems, I occassionally get:
> 
> kernel unaligned access to 0xe00007bc3b8f67b9, ip=0xa00000020a61e370
> kernel unaligned access to 0xe00007bc3b8f67b1, ip=0xa00000020a61e3d1
> 
> grep a00000020a61 /proc/modules
> lockd 146448 1 nfs, Live 0xa00000020a610000
> 
> 
> These come in pairs.  I tracked it down to the lockd.ko kernel module
> and then objdump'd to find we are in:
> 
> 
> nsm_get_handle
> nsm_init_private():
> /data/lwork/attica2/holt/git-linus/v2.6.29/fs/lockd/mon.c:280
>     e370:       0b 00 98 68 98 11       [MMI]       st8 [r52]=r38;;

This one is:
	u64 *p = (u64 *)&nsm->sm_priv.data;
	...
	*p = (unsigned long)nsm;

sm_priv.data is an unsigned char array, so there are no alignment
rules. You either need to use memcpy, or not define it as an unsigned
char.  Tony, any suggestions?

> /data/lwork/attica2/holt/git-linus/v2.6.29/fs/lockd/mon.c:279
>     e376:       00 00 00 02 00 00                   nop.m 0x0
>     e37c:       92 d0 e9 53                         shl r16=r9,5;;
> 
> nsm_display_address():
> /data/lwork/attica2/holt/git-linus/v2.6.29/fs/lockd/mon.c:86
> ...
>     e3d0:       0a 60 28 1c 8d 39       [MMI]       cmp4.eq p12,p13=10,r14;;
>     e3d6:       00 a8 95 30 23 00                   st8 [r37]=r53

Haven't gotten to this one yet.

Robin

Re: kernel unaligned accesses on 2.6.29.

[Robin Holt]

On Tue, Mar 24, 2009 at 08:14:18PM -0500, Robin Holt wrote:
> Had a few minutes to look around.
...
> This one is:
> 	u64 *p = (u64 *)&nsm->sm_priv.data;
> 	...
> 	*p = (unsigned long)nsm;
...
> >     e3d0:       0a 60 28 1c 8d 39       [MMI]       cmp4.eq p12,p13=10,r14;;
> >     e3d6:       00 a8 95 30 23 00                   st8 [r37]=r53
> 

Well, that was easy.  This is the store part of
	*p++ = timespec_to_ns(&ts);

Again, comes down to &nsm->sm_priv.data not being aligned.

Robin

Re: kernel unaligned accesses on 2.6.29.

[David Miller]

From: Robin Holt <holt@sgi.com>
Date: Tue, 24 Mar 2009 20:14:18 -0500

> You either need to use memcpy, or not define it as an unsigned char.
> Tony, any suggestions?

memcpy() might not work in this scenerio.

GCC can still emit an inline 64-bit memory access if it
deems that the objects involved should be sufficiently
aligned.

It might not be able to in this specific case, but it is
something to be aware of.  Don't be convinced of any fix
for this until the assembler is inspected by hand :-)