Re: End to end SMART to RAID repair

[Ric Wheeler <ricwheeler@gmail.com>]

Arthur Britto wrote:
> On Tue, 2008-06-24 at 08:58 +1000, Neil Brown wrote:
>   
>> On Sunday June 22, ahbritto@iat.com wrote:
>>     
>>> smartmontools (http://smartmontools.sourceforge.net/) can be configured
>>> to passively scan hard drives for defects in the background.  The block
>>> numbers of pending unreadable sectors are logged via syslog.  These
>>> sectors will be remapped when written too.
>>>
>>> It would be great if this worked end to end with linux software raid to
>>> automatically repair the bad sector.
>>>       
>> Well, you can just get md to do a scan (echo check >
>> /sys/block/mdXX/md/sync_action) and it will find any read errors and
>> correct them.
>>     
>
> True.  However, a SMART on disk check requires no main board
> resources. Some drives, when idle, may do background checking anyway.
> This would provide a way to correct the error without needing to scan
> the whole volume and other components with an md check.  Error checking
> may be less intrusive (vs retries to the exclusion of other work) than
> normal for an attempted sector read.  At least manufactures have the
> option to give priority to actual read requests over background defect
> checking.
>   

This is almost always the case with disk arrays for example.
>   
>> Extracting numbers from syslog is a fairly messy thing to try to do.
>> Maybe if smartmontools could report these in some other way -
>> e.g. run a program giving device and block number, we could write a
>> script that feeds that info to md.
>> We would need to map the device+offset to partition+offset, then find
>> out if that is a member of an md array, then request a limited-range
>> 'check', which I think is possible with current code...
>>
>> Do you know if smartmontools can provide this info in a more
>> controlled way?
>>     
>
> I was thinking, a non-smartmontools specific method would be best.  That
> is: (1) some way for the md driver to request notification about pending
> uncorrected read errors from a region of a block device and (2) some way
> for a trusted application to inform the kernel about pending uncorrected
> read errors (e.g. echo "start-stop > /sys/...").
>
> -Arthur
>
>   

One thing that you can do that is much less invasive is to use the "read 
verify" command to scan the platter of the disk at a fairly low rate.

Read verify does not transfer data from the disk to the host and you can 
issue fairly large requests (say 1MB at a time) as a background task per 
drive. What you will get out of this is a validation that nothing has 
failed at the disk sector level (i.e., each sector is still readable). 
On detection of an error, you can go back and try to pin point the 
failed sector with small IO's and then try to repair the damage with a 
write (say from the other mirror in a RAID1 device).

This is useful, but is not an end to end data integrity check (like 
Martin's T10 DIF work that was posted).

In general, it would also be really neat to figure out and API which 
would let a higher level application (or file system) inform the block 
level of an error and possibly ask for a read from another mirror.

ric

--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html