[rhelv6-beta-list] Can RHEL6 installer align the disk partition with 4k-sector or raid stripe size?

[Nico Kadel-Garcia]

On Sun, Jun 6, 2010 at 11:24 PM,  <caiqian at redhat.com> wrote:
>
>> I assume that the original poster mant "4k block size". 4k
>> sectors.... doesn't make much sense.
>>
>> So far as I can tell, this only works with hard drives which are
>> detected as 4K block size. According to my preliminary tests, it does
>> *NOT* work with VMware, Xen, or KVM disk images. This is an important
>> feature for certain attached storage devices, such as NetApps, which
>> behave very badly if the partitions of virtual disk images are not 4K
>> aligned. It should, ideally, be a simple setting for disk
>> configuration in Anaconda and especially in a virtual machine setup
>> tool, but even fdisk mishandles this: at least up through RHEL 5.
>>
>> You see, you can set the block size with fdisk inside a '%pre'
>> statement. But if you don't have a partition set, fdisk automagically
>> reverts the block size to 512 for you. This means that in order to
>> preserve the block size information for Anaconda's disk configuration
>> tools, which have *no* options to select specific sectors to start or
>> end partitions, or to do this 4K block size, you have to set aside a
>> partition that anaconda will not touch and that will preserve the4096
>> block size setting. That's.... fairly awkward to allocate and recover
>> later.
>>
>> I've also found no way to tell Anaconda "I've built the partitions for
>> you in %pre: use *THESE* partitions for *THOSE* mount points, and
>> don't try to repartiton them." If someone knows a way, I'd love to
>> hear it.
>
> Have you seen this in the RHEL6 Beta manual?
> http://www.redhat.com/docs/en-US/Red_Hat_Enterprise_Linux/6-Beta/html/Storage_Administration_Guide/newstorage-iolimits.html#iolimits
>
> "As of this release, boot support for 4K sector devices is planned, but not yet supported."
>
> C.

This is thoughtful, but it's relevant only to physical media, not
virtual images. Until virtualization "disks" are configurable as "4096
byte block devices", the emulated disks have a behavioral issue that
the guest OS's cannot see.

The related issues with fdisk discarding information about block
sizes, and anaconda's incapacity to accept such settings nor to accept
pre-set partitions as "this partition or LVM goes with *THIS*
filesystem, do not try to reset this!!!!!" merely means that the next
set of clean ways to work around the problem are impossible. NetApp
publishes a tool to "disk align" such images for VMWare use, but it's
pretty nasty: it builds a larger disk image, and uses something like
"dd" to image the existing volumes into new partitions, which start on
4096 byte block boundaries but don't necessarily end on them. Coupled
with this, they ignore cylinders, so "fdisk" and tools like it
complain forever more about cylinder misalignment.

It is, unfortunately, a classic case of a set of small deficits in
individual components adding up to a nightmare of dealing with a new
situation.