Re: Re: [PATCH] dm-striped device sizes must be multiple of chunk-size

[Kevin Corry <kevcorry@us.ibm.com>]

On Wed March 22 2006 11:28 pm, Joseph Yasi wrote:
>     I am using the dmraid utility to detect and configure my Intel
> Software Raid RAID0 array.  When upgrading to 2.6.16 from 2.6.15, the
> system failed to boot because the device-mapper array was not built.
> I have traced the problem to this patch that limits dm-stripe to
> targets that are multiples of the chunk size.  I don't know whether to
> consider this a problem with device-mapper or with the Intel Software
> Raid BIOS that built the array.  Dmraid is giving me a target of
> 980460038 blocks with a chunk size of 256 blocks.  Is this a problem
> with device-mapper, the dmraid utility or the raid bios?  Should
> device-mapper handle striped targets that are not multiples of the
> chunk size?

Hmm...I obviously didn't expect dmraid to be creating such devices.

The kernel patch that was added to 2.6.16 is definitely correct. Dm-stripe 
should not allow creating a device with a size that's not a multiple of the 
chunk-size. To do so leaves a partial chunk at the end of each device. While 
at first glance this might not seem like a big deal, the striping 
calculations in the I/O path do not handle this situation (at least, not in 
the way you would think it should be handled), and cause I/O errors if you 
try to access the end of the stripe device.

> dmesg gives these pertinent lines:
> device-mapper: dm-stripe: Target length not divisible by chunk size.
> device-mapper: error adding target to table.
>
> dmsetup status
> isw_caejhbcidd_Volume03: 0 976173660 linear
> isw_caejhbcidd_Volume02: 0 4209030 linear
> isw_caejhbcidd_Volume01: 0 64197 linear
> isw_caejhbcidd_Volume0: 0 980460038 striped
>
> dmraid -s
> *** Group superset isw_caejhbcidd
> --> Active Subset
> name   : isw_caejhbcidd_Volume0
> size   : 980460038
> stride : 256
> type   : stripe
> status : ok
> subsets: 0
> devs   : 2
> spares : 0

In your example, your stripe device is 980460038 sectors, with a 256 sector 
chunk-size.This means the stripe device has 3829922 chunks, with 6 sectors 
left over (3 on each of the two devices, presumably). I would think that the 
BIOS intends that these last 3 sectors on each device should be striped with 
an effective chunk-size of 3 sectors. To do this, dmraid should set up a 
device with two targets, one with a chunk-size of 256, and one with a 
chunk-size of 3. The map would look something like:

0 980460032 striped 2 256 /dev/hda 0 /dev/hdb 0
980460032 6 striped 2 3 /dev/hda /dev/hdb 490230016

Unfortunately, this doesn't work either, because the chunk-size must be a 
power-of-2. So...since we know the second target in the map really represents 
only one "stripe", we could treat them as a concatenation of independent 
linear targets. So we end up with a map like this:

0 980460032 striped 2 256 /dev/hda 0 /dev/hdb 0
980460032 3 linear /dev/hda 490230016
980460032 3 linear /dev/hdb 490230016

I'm thinking that this mapping would do what the BIOS is actually intending. 
It ought to be relatively straight-forward to recognize this situation in 
dmraid and set up a mapping similar to the one above.

-- 
Kevin Corry
kevcorry@us.ibm.com
http://www.ibm.com/linux/
http://evms.sourceforge.net/