forcing a read on a known bad block

forcing a read on a known bad block

[dean gaudet]

hey Neil...

i've been wanting to test out the reconstruct-on-read-error code... and 
i've had two chances to do so, but haven't be able to force md to read the 
appropriate block to trigger the code.

i had two disks with SMART Current_Pending_Sector > 0 (which indicates 
pending read error) and i did SMART long self-tests to find out where the 
bad block was (it should show the LBA in the SMART error log)...

one disk was in a raid1 -- and so it was kind of random which of the two 
disks would be read from if i tried to seek to that LBA and read... in 
theory with O_DIRECT i should have been able to randomly get the right 
disk, but that seems a bit clunky.  unfortunately i didn't think of the 
O_DIRECT trick until after i'd given up and decided to just resync the 
whole disk proactively.

the other disk was in a raid5 ... 5 disk raid5, so 20% chance of the bad 
block being in parity.  i copied the kernel code to be sure, and sure 
enough the bad block was in parity... just bad luck :)  so i can't force a 
read there any way that i know of...

anyhow this made me wonder if there's some other existing trick to force 
such reads/reconstructions to occur... or perhaps this might be a useful 
future feature.

on the raid5 disk i actually tried reading the LBA directly from the 
component device and it didn't trigger the read error, so now i'm a bit 
skeptical of the SMART log and/or my computation of the seek offset in the 
partition... but the above question is still interesting.

-dean

Re: forcing a read on a known bad block

[Patrik Jonsson]

[-- Attachment #1: Type: text/plain, Size: 896 bytes --]

hi dean,

dean gaudet wrote:
> 
> the other disk was in a raid5 ... 5 disk raid5, so 20% chance of the bad 
> block being in parity.  i copied the kernel code to be sure, and sure 
> enough the bad block was in parity... just bad luck :)  so i can't force a 
> read there any way that i know of...

well, for raid5 you can use 'echo repair > /sys/block/mdx/md/sync_action'

This does a 'simulated reconstruction' and has triggered this for me in
the past. (For some reason 'check' instead of 'repair' did not, even
though it should have tried to read all the blocks then, too...)

That said, I have one disk in an 8-disk raid5 that says 'current pending
sector 1', and another that says 'offline uncorrectable 1', and they
have been doing so for months. Neither SMART extended tests or full
raid5 resyncs have either failed or fixed this, so I don't know what's
up with that...

cheers,

/Patrik

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 254 bytes --]

Re: [RAID] forcing a read on a known bad block

[Julian Cowley]

On Tue, 11 Apr 2006, dean gaudet wrote:
> anyhow this made me wonder if there's some other existing trick to force
> such reads/reconstructions to occur... or perhaps this might be a useful
> future feature.

For testing RAID, what would be really nice is if there were a virtual
disk device where one could simulate bad sectors (read or write),
non-responsive disks, etc.  It would be virtual in the same sort way
that /dev/full simulates a full disk.

This would be an ideal project for Xen.

Re: [RAID] forcing a read on a known bad block

[Luca Berra]

On Tue, Apr 11, 2006 at 12:37:53PM -1000, Julian Cowley wrote:
>On Tue, 11 Apr 2006, dean gaudet wrote:
>>anyhow this made me wonder if there's some other existing trick to force
>>such reads/reconstructions to occur... or perhaps this might be a useful
>>future feature.
>
>For testing RAID, what would be really nice is if there were a virtual
>disk device where one could simulate bad sectors (read or write),
>non-responsive disks, etc.  It would be virtual in the same sort way
>that /dev/full simulates a full disk.
>
either use the MD "faulty" personality, or the device-mapper "error"
target.

L.

-- 
Luca Berra -- bluca@comedia.it
        Communication Media & Services S.r.l.
 /"\
 \ /     ASCII RIBBON CAMPAIGN
  X        AGAINST HTML MAIL
 / \