Ext3 users - Jul 2001 - Bad unaligned kernel access with ext3 0.8.0

Hello!

This is Kernel 2.4.6-ac1 with ext3-0.8.0 compiled with cvs-gcc 
version 3.1 20010616 on alpha ev4.

I have prepared one ext3 filesystem to play with. When my rc-scripts call
mount -a during boot I receive this:

Jul  6 22:01:30 Marvin kernel: Bad unaligned kernel access at fffffc0000883f54:
fffffc00063f2e6e 2a 2

and the mount of this filesystem fails. A subsequent manual second try to mount 
puts the mount process into 'D' state (i.e. it never completes).

Now System map shows the fault occurs in ext3_bmap:

fffffc0000883c00 t ext3_commit_write
fffffc0000883e60 t ext3_abort_write
fffffc0000883f30 t ext3_bmap
fffffc0000883fe0 t bget_one
fffffc0000884000 t ext3_writepage
fffffc0000884340 t ext3_readpage
fffffc0000884380 t ext3_flushpage

Disassembly of section .text:

fffffc0000883f30 <ext3_bmap>:
fffffc0000883f30:       34 00 bb 27     ldah    gp,52(t12)
fffffc0000883f34:       18 16 bd 23     lda     gp,5656(gp)
fffffc0000883f38:       e0 ff de 23     lda     sp,-32(sp)
fffffc0000883f3c:       08 00 3e b5     stq     s0,8(sp)
fffffc0000883f40:       09 04 f0 47     mov     a0,s0
fffffc0000883f44:       10 00 5e b5     stq     s1,16(sp)
fffffc0000883f48:       0a 04 f1 47     mov     a1,s1
fffffc0000883f4c:       00 00 5e b7     stq     ra,0(sp)
fffffc0000883f50:       40 00 69 a4     ldq     t2,64(s0)
fffffc0000883f54:       2e 02 43 a8     ldl_l   t1,558(t2)
fffffc0000883f58:       01 50 40 44     and     t1,0x2,t0
fffffc0000883f5c:       03 00 20 e4     beq     t0,fffffc0000883f6c
<ext3_bmap+0x3c>
fffffc0000883f60:       02 58 40 44     xor     t1,0x2,t1
fffffc0000883f64:       2e 02 43 b8     stl_c   t1,558(t2)
fffffc0000883f68:       58 09 40 e4     beq     t1,fffffc00008864cc
<ext3_journal_get_write_access+0x5c>
fffffc0000883f6c:       10 00 20 e4     beq     t0,fffffc0000883fb0
<ext3_bmap+0x80>
fffffc0000883f70:       00 01 23 a4     ldq     t0,256(t2)
fffffc0000883f74:       f0 02 01 a6     ldq     a0,752(t0)
fffffc0000883f78:       08 a1 7d a7     ldq     t12,-24312(gp)
[...]

The ldl_l .... stl_c sequence can be identified as test_and_clear_bit() on
alpha.

So the faulting instruction is this line of fs/ext3/inode.c::ext3_bmap() :

if (test_and_clear_bit(EXT3_STATE_JDATA, &inode->u.ext3_i.i_state))

In fact ext3_i.i_state is a 16 bit aligned 16 bit value while 
test_and_clear_bit() expects to work on aligned 32 bit quantities. Furthermore
the unaligned trap handler doesn't know how to handle the ldl_l.

How to fix it?

Weird side note: ext3 0.0.7 on kernel 2.4.5-ac22 worked without problems
(as in : mount the filesystem as ext3 and use it for some compilations)
 

Please CC: dl8bcu@gmx.net on replies as I'm not on this list.

Bye,
Thorsten


-- 
| Thorsten Kranzkowski        Internet: dl8bcu@gmx.net                        |
| Mobile: ++49 170 1876134       Snail: Niemannsweg 30, 49201 Dissen, Germany |
| Ampr: dl8bcu@db0lj.#rpl.deu.eu, dl8bcu@marvin.dl8bcu.ampr.org [44.130.8.19] |

Thorsten Kranzkowski wrote:
> 
> So the faulting instruction is this line of fs/ext3/inode.c::ext3_bmap() :
> 
> if (test_and_clear_bit(EXT3_STATE_JDATA, &inode->u.ext3_i.i_state))
> 
Great, thanks.

Could you please change i_state to __u32 in ext3_fs_i.h?

If that works, I'll put out 0.9.1.

Ok, the __u32 did it.

instead of the unaligned acces I do see now
Jul  7 17:35:28 Marvin kernel: [EXT3 FS 2.4-0.9.0, 5 Jul 2001, bs=4096, gc=17,
bpg=32768, ipg=16224, mo=0800]
Jul  7 17:35:28 Marvin kernel: EXT3-fs: mounted filesystem with ordered data
mode.

which looks a lot better :-)

Thanks,
Thorsten

-- 
| Thorsten Kranzkowski        Internet: dl8bcu@gmx.net                        |
| Mobile: ++49 170 1876134       Snail: Niemannsweg 30, 49201 Dissen, Germany |
| Ampr: dl8bcu@db0lj.#rpl.deu.eu, dl8bcu@marvin.dl8bcu.ampr.org [44.130.8.19] |