On 05/17/2010 05:45 PM, Dave Chinner wrote:
> On Mon, May 17, 2010 at 05:28:30PM -0400, Doug Ledford wrote:
>> On 05/09/2010 10:20 PM, Dave Chinner wrote:
>>> On Sun, May 09, 2010 at 08:48:00PM +0200, Rainer Fuegenstein wrote:
>>>>
>>>> today in the morning some daemon processes terminated because of
>>>> errors in the xfs file system on top of a software raid5, consisting
>>>> of 4*1.5TB WD caviar green SATA disks.
>>>
>>> Reminds me of a recent(-ish) md/dm readahead cancellation fix - that
>>> would fit the symptoms of (btree corruption showing up under heavy IO
>>> load but no corruption on disk. However, I can't seem to find any
>>> references to it at the moment (can't remember the bug title), but
>>> perhaps your distro doesn't have the fix in it?
>>>
>>> Cheers,
>>>
>>> Dave.
>>
>> That sounds plausible, as does hardware error. A memory bit flip under
>> heavy load would cause the in memory data to be corrupt while the on
>> disk data is good.
>
> The data dumps from the bad blocks weren't wrong by a single bit -
> they were unrecogniѕable garbage - so that it very unlikely to be
> a memory erro causing the problem.
Not true. It can still be a single bit error but a single bit error
higher up in the chain. Aka a single bit error in the scsi command to
read various sectors, then you read in all sorts of wrong data and
everything from there is totally whacked.
>> By waiting to check it until later, the bad memory
>> was flushed at some point and when the data was reloaded it came in ok
>> this time.
>
> Yup - XFS needs to do a better job of catching this case - the
> prototype metadata checksumming patch caught most of these cases...
>
> Cheers,
>
> Dave.
--
Doug Ledford <dledford@xxxxxxxxxx>
GPG KeyID: CFBFF194
http://people.redhat.com/dledford
Infiniband specific RPMs available at
http://people.redhat.com/dledford/Infiniband
signature.asc
Description: OpenPGP digital signature
|