Re: [PATCH 1/2] xfsprogs: ignore stripe geom if sunit or swidth == physical sector size

[Stan Hoeppner <stan@hardwarefreak.com>]

On 10/30/2014 02:50 PM, Brian Foster wrote:
> On Thu, Oct 30, 2014 at 02:15:16PM -0500, Stan Hoeppner wrote:
>> On 10/30/2014 06:46 AM, Brian Foster wrote:
>>> On Wed, Oct 29, 2014 at 04:38:22PM -0500, Stan Hoeppner wrote:
>>>> On 10/29/2014 01:47 PM, Eric Sandeen wrote:
>>>>> On 10/29/14 1:37 PM, Brian Foster wrote:
>>>>>> On Tue, Oct 28, 2014 at 12:35:29PM -0500, Eric Sandeen wrote:
>>>>>>> Today, this geometry:
>>>>>>>
>>>>>>> # modprobe scsi_debug  opt_blks=2048 dev_size_mb=2048
>>>>>>> # blockdev --getpbsz --getss --getiomin --getioopt  /dev/sdd
>>>>>>> 512
>>>>>>> 512
>>>>>>> 512
>>>>>>> 1048576
>>>>>>>
>>>>>>> will result in a warning at mkfs time, like this:
>>>>>>>
>>>>>>> # mkfs.xfs -f -d su=64k,sw=12 -l su=64k /dev/sdd
>>>>>>> mkfs.xfs: Specified data stripe width 1536 is not the same as the volume stripe width 2048
>>>>>>>
>>>>>>> because our geometry discovery thinks it looks like a
>>>>>>> valid striping setup which the commandline is overriding. 
>>>>>>> However, a stripe unit of 512 really isn't indicative of
>>>>>>> a proper stripe geometry.
>>>>>>>
>>>>>>
>>>>>> So the assumption is that the storage reports a non-physical block size
>>>>>> for minimum and optimal I/O sizes for geometry detection. There was a
>>>>>> real world scenario of this, right? Any idea of the configuration
>>>>>> details (e.g., raid layout) that resulted in an increased optimal I/O
>>>>>> size but not minimum I/O size?
>>>>>
>>>>> Stan?  :)
>>>>
>>>> Yeah, it was pretty much what you pasted sans the log su, and it was a
>>>> device-mapper device:
>>>>
>>>> # mkfs.xfs -d su=64k,sw=12 /dev/dm-0
>>>>
>>>
>>> What kind of device is dm-0? I use linear devices regularly and I don't
>>> see any special optimal I/O size reported:
>>
>> It's a dm-multipath device.  I pasted details up thread.  Here, again:
>>
> 
> Oh, I see. So this is just getting passed up from the lower level scsi
> devices. On a quick look, this data appears to come from the device via
> the "block limits VPD." Apparently that should be accessible via
> something like this (0xb0 from sd_read_block_limits()):
> 
> # sg_inq --page=0xb0 /dev/sdx
> 
> ... but I don't have a device around that likes that command. It would
> be interesting to know what makes the underlying device set optimal I/O
> size as such, but that's just curiosity at this point. :)

The device isn't setting it.  It's global.  Any LUN of any RAID level
reports the same parms.  So apparently it's hard coded in the firmware.
 I informed our field engineer at the vendor of this issue, and the fact
it prompted a patch to XFS, but haven't received a response.

An educated guess is that they want to see 1 MiB IOs entering the
controller regardless of the stripe geometry of the back end LUN.  Could
be lots of reasons for this, valid or not.  However, given it advertises
a minimum optimal IO size of 512 bytes this seems counterintuitive.

Thanks,
Stan




> Brian
> 
>> # multipath -ll
>> 3600c0ff0003630917954075401000000 dm-0 Tek,DH6554
>> size=44T features='0' hwhandler='0' wp=rw
>> |-+- policy='round-robin 0' prio=50 status=active
>> | `- 9:0:0:3 sdj 8:144 active ready running
>> `-+- policy='round-robin 0' prio=10 status=enabled
>>   `- 1:0:0:3 sdf 8:80  active ready running
>>
>>
>> # blockdev --getpbsz --getss --getiomin --getioopt  /dev/dm-0
>> 512
>> 512
>> 512
>> 1048576
>>
>> # blockdev --getpbsz --getss --getiomin --getioopt  /dev/sdj
>> 512
>> 512
>> 512
>> 1048576
>>
>> # blockdev --getpbsz --getss --getiomin --getioopt  /dev/sdf
>> 512
>> 512
>> 512
>> 1048576
>>
>>
>>
>> Cheers,
>> Stan
>>
> 
> _______________________________________________
> xfs mailing list
> xfs@oss.sgi.com
> http://oss.sgi.com/mailman/listinfo/xfs
> 

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs