Re: [gentoo-user] RAID 1 on /boot

[Kerin Millar <kerframil@fastmail.co.uk>]

On 22/02/2014 11:41, J. Roeleveld wrote:
> On Sat, February 22, 2014 06:27, Facundo Curti wrote:
>> Hi all. I'm new in the list, this is my third message :)
>> First at all, I need to say sorry if my english is not perfect. I speak
>> spanish. I post here because gentoo-user-es it's middle dead, and it's a
>> great chance to practice my english :) Now, the problem.
>
> First of all, there are plenty of people here who don't have English as a
> native language. Usually we manage. :)
>
>> I'm going to get a new PC with a disc SSD 120GB and another HDD of 1TB.
>> But in a coming future, I want to add 2 or more disks SSD.
>>
>> Mi idea now, is:
>>
>>      Disk HHD: /dev/sda
>> /dev/sda1 26GB
>> /dev/sda2 90GB
>> /dev/sda3 904GB
>>
>>      Disk SSD: /dev/sdb
>> /dev/sdb1 26GB
>> /dev/sdb2 90GB
>> /dev/sdb3 4GB
>>
>> And use /dev/sdb3 as swap. (I will add more with another SSD in future)
>> /dev/sda3 mounted in /home/user/data (to save data unused)
>
> Why put the swap on the SSD?
>
>> And a RAID 1 with:
>> md0: sda1+sdb1    /
>> md1: sda2+sdb2    /home
>>
>> (sda1 and sda2 will be made with the flag: write-mostly. This is useful
>> for
>> disks slower).
>> In a future, I'm going to add more SSD's on this RAID. My idea is the
>> fastest I/O.
>>
>> Now. My problem/question is:
>> Following the gentoo's
>> doc<http://www.gentoo.org/doc/es/gentoo-x86+raid+lvm2-quickinstall.xml>,
>> it says I need to put the flag --metadata=0.9 on the RAID. My question is
>> ¿This will make get off the performance?.

It has no impact on performance.

>
> metadata=0.9 might be necessary for the BIOS of your computer to see the
> /boot partition. If you use an initramfs, you can use any metadata you
> like for the root-partition.

The BIOS should not care at all as it is charged only with loading code 
from the MBR. However, if the intention is to use grub-0.97 then the 
array hosting the filesystem containing /boot should:

   * use RAID-1
   * use the 0.90 superblock format

That way, grub-0.97 can read the filesystem from either block device 
belonging to the array. Doing it any other way requires a bootloader 
that specifically understands md (such as grub2).

There's also the neat trick of installing grub to all disks belonging to 
the array for bootloader redundancy. However, I'm not entirely sure that 
Code Listing 2.35 in the Gentoo doc is correct. Given that particular 
example, I would instead do it like this:-

   grub> device (hd0) /dev/sda
   grub> root (hd0,0)
   grub> setup (hd0)
   grub> device (hd0) /dev/sdb
   grub> root (hd0,0)
   grub> setup (hd0)

The idea there is that, should it ever be necessary to boot from the 
second disk, the disk in question would be the one enumerated first by 
the BIOS (mapping to hd0 in grub). Therefore, grub should be installed 
in that context across all disks. It should not be allowed to boot from 
any given drive and subsequently try to access >=(hd1).

With grub2, it's a little easier because it is only necessary to run 
grub-install on each of the drives:

   # grub-install /dev/sda
   # grub-install /dev/sdb

>
>> I only found this
>> document<https://raid.wiki.kernel.org/index.php/RAID_superblock_formats#The_version-0.90_Superblock_Format>.
>> This says the difference, but nothing about performance and
>> advantages/disadvantages.

The 0.90 superblock format is subject to specific limitations that are 
clearly described by that page. For example, it is limited to 28 devices 
in an array, with each device being limited to 2TB in size.

Also, the 0.90 format will cause issues in certain setups because of the 
way that it places its metadata at the end of the block device [1]. That 
said, the 0.90 format does allow for the kernel to construct the array 
without any intervention from userspace so it still has its uses.

The 1.2 format positions the superblock 4KiB from the beginning of the 
device. Note that this has nothing at all to do with the data, which 
usually begins 1MiB in. If you run mdadm -E on a member of such an 
array, the offset will be reported as the "Data Offset". For example:

   Data Offset : 2048 sectors

So, it's not a matter of alignment. Rather, the advantage of the 1.2 
format is that it leaves a little space for bootloader code e.g. in case 
you want to create an array from whole disks rather than disk partitions.

None of this matters to me so I tend to stick to the 1.1 format. It 
wouldn't actually make any difference to my particular use case.


>>
>> Another question is, ¿GRUB2 still unsupporting metadata 1.2?
>
> See reply from Canek.
>
>> In case that metadata get off performance, and GRUB2 doesn't support this.
>> ¿Anyone knows how can I fix this to use metadata 1.2?
>>
>> I don't partitioned more, because I saw this unnecessary. I just need to
>> separate /home in case I need to format the system. But if I need to
>> separate /boot to make it work, I don't have problems doing that.
>>
>> But of course, /boot also as RAID...
>
> /boot seperate as RAID-1 and metadata=0.9 and you are safe.
>
>> ¿Somebody have any ideas to make it work?
>
> It is similar to what I do, except I don't have SSDs in my desktop.
>
> I have 2 partitions per disk:
> 1 : /boot (mirrored, raid-1)
> 2 : LVM (striped, raid-0)
> All other partitions (root, /usr, /home, ....) are in the LVM.
>
> I use striping for performance reasons for files I currently work with.
> All important data is stored and backed up on a server.
>
> For this, an initramfs is required with support for mdraid and lvm.

Some people would make a point of avoiding the initramfs requirement by 
just having a relatively small root filesystem on RAID-1 with 0.90 
metadata. Unfortunately, the UsrMove affair has now rendered this 
approach impractical because /usr can no longer be separate. It's still 
an option, though.


--Kerin

[1] https://savannah.gnu.org/bugs/?34250