Re: PROBLEM: Kernel 2.6.10 crashing repeatedly and hard

[Linus Torvalds]

On Thu, 30 Dec 2004, Georg C. F. Greve wrote:
> 
>  lt> Any chance that you can get a full oops, with register info?
>  lt> Right now your errors only have the call trace, not the registers
>  lt> or the EIP itself..
> 
> Yes, for the first two times, the top of the crash report scrolled off
> the screen and I don't have a serial console handy. The third crash
> contained the 
> 
>  EFLAGS: 00010002   (2.6.10)
>  EIP is at free_block+0x45/0xd0
>  eax: 46484849   ebx: df2b1000   ecx: df2b1050   edx: df2ab000
>  esi: c183cd80   edi: 00000001   ebp: 00000018   esp: c188fef8
>  ds: 007b   es: 007b   ss: 0068
>  Process events/0 (pid: 6, threadinfo:c188e000 task=c185ca20)
>  Stack: c183cdb8 c1858810 c1858800 00000018 c183cd80 c0141724 c183cd80 c1858810
>         00000018 c183ccb8 c183cd80 00000002 c183cce0 c01417c6 c183cd80 c1858800
>         00000000 c183ccb8 c183ce10 00000003 c170fc20 c183b000 c170fc24 00000000
> 
> header, which I think may contain what you were looking for.

Ok. This is apparently a slab cache corruption issue.

Can you compile with SLAB_DEBUG and DEBUG_PAGEALLOC turned on, since that 
may well catch something in the act (or it may make the machine so 
unusably slow that it's not funny - who knows..)

> Meanwhile, because of the trouble, I gave Alans patch-2.6.10-ac1 a try
> and again put ext3 on dm-crypt on raid5, and: No crash so far despite
> moving and removing ~100 gigabytes a couple of times on to/from the
> partition.

Goodie. Maybe Alan already knows what it's about. Alan? Any slab 
corruption or wild pointers that you are aware of? However:

> I did however get occasional
> 
>  EXT3-fs error (device dm-0): ext3_readdir: bad entry in directory #37306376: directory entry across blocks - offset=0, inode=3350155169, rec_len=11148, name_len=1

I suspect that it's more of the slab corruption thing, and Alan's patch 
ended up moving something around subtly enough that you don't get the 
crash, but instead get corrupted buffer heads or similar.. Your BTTV 
issues are interesting, though.

		Linus
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/