kernel unaligned accesses on 2.6.29.
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_handlensm_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;;
/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,r14;;
e3d6: 00 a8 95 30 23 00 st8 [r37]=r53
e3dc: 00 00 04 00 nop.i 0x0Had a few minutes to look around.
Quoted 10 lines · Robin Holt
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 to0xe00007bc3b8f67b1, ip=0xa00000020a61e3d1grep
a00000020a61/proc/modules
lockd 146448 1 nfs, Live0xa00000020a610000
Quoted 8 lines
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_handlensm_init_private():
/data/lwork/attica2/holt/git-linus/v2.6.29/fs/lockd/mon.c:280e370: 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?
Quoted 9 lines
/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,r14;; e3d6: 00 a8 95 30 23 00 st8 [r37]=r53
Haven't gotten to this one yet.
Robin
Quoted 1 line · Robin Holt
Had a few minutes to look around.
...
Quoted 4 lines
This one is:
u64 *p = (u64 *)&nsm->sm_priv.data; ... *p = (unsigned long)nsm;
...
Quoted 3 lines
Quoted 2 lines
e3d0: 0a 60 28 1c 8d 39 [MMI] cmp4.eq p12,p13,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
From: Robin Holt <[email protected]>
Date: Tue, 24 Mar 2009 20:14:18 -0500
Quoted 2 lines
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 :-)