RAID1 write order fidelity

RAID1 write order fidelity

[aristizb]

Hi,

I am currently working with a software Linux RAID1, with external  
bitmap and no write-behind option. Does the RAID controller keep the  
write order fidelity?.

If an application writes some blocks on the RAID device in a  
particular order, can I ensure that these blocks will be written to  
the devices conforming the RAID on the same order, or at least on the  
main disk? (disk 0 on the RAID).


The version of mdadm I am suing is 2.6.2.


Thank you.


Juan Aristizabal.

Re: RAID1 write order fidelity

[David Greaves]

aristizb@ualberta.ca wrote:
> Hi,
> 
> I am currently working with a software Linux RAID1, with external bitmap
> and no write-behind option. Does the RAID controller keep the write
> order fidelity?.
> 
> If an application writes some blocks on the RAID device in a particular
> order, can I ensure that these blocks will be written to the devices
> conforming the RAID on the same order, or at least on the main disk?
> (disk 0 on the RAID).

I suspect you need to do some research on 'barriers'.

I think the answer at present is 'no'.

Hence messages like:
Filesystem "md0": Disabling barriers, not supported by the underlying device


> The version of mdadm I am suing is 2.6.2.
That's not important. The kernel version is; that's where the md driver lives.

David

RE: RAID1 write order fidelity

[David Lethe]

> -----Original Message-----
> From: linux-raid-owner@vger.kernel.org [mailto:linux-raid-
> owner@vger.kernel.org] On Behalf Of aristizb@ualberta.ca
> Sent: Thursday, July 31, 2008 10:04 AM
> To: linux-raid@vger.kernel.org
> Subject: RAID1 write order fidelity
> 
> Hi,
> 
> I am currently working with a software Linux RAID1, with external
> bitmap and no write-behind option. Does the RAID controller keep the
> write order fidelity?.
> 
> If an application writes some blocks on the RAID device in a
> particular order, can I ensure that these blocks will be written to
> the devices conforming the RAID on the same order, or at least on the
> main disk? (disk 0 on the RAID).
> 
> 
> The version of mdadm I am suing is 2.6.2.
> 
> 
> Thank you.
> 
> 
> Juan Aristizabal.

The answer is no.  Well-designed RAID controllers will prioritize queued
I/Os for best performance.  And, for that matter, device drivers, BIOS,
even some file systems will move around I/Os for purposes of
optimization and response time, so this isn't necessarily a
RAID-controller specific response.  They may even eliminate writes
completely based on cache settings, what is in the queue, the block
range, and the number of blocks for the I/O requests.  
David @ santools dot com

Re: RAID1 write order fidelity

[Neil Brown]

On Thursday July 31, david@dgreaves.com wrote:
> aristizb@ualberta.ca wrote:
> > Hi,
> > 
> > I am currently working with a software Linux RAID1, with external bitmap
> > and no write-behind option. Does the RAID controller keep the write
> > order fidelity?.

md/RAID1 provides the same guarantees as the underlying device, which
usually isn't much of a guarantee.

> > 
> > If an application writes some blocks on the RAID device in a particular
> > order, can I ensure that these blocks will be written to the devices
> > conforming the RAID on the same order, or at least on the main disk?
> > (disk 0 on the RAID).
> 
> I suspect you need to do some research on 'barriers'.
> 
> I think the answer at present is 'no'.
> 
> Hence messages like:
> Filesystem "md0": Disabling barriers, not supported by the underlying device

A strange message to get as md/RAID1 *does* support barriers when all the
underlying devices do (since the mid-teens of 2.6 I think).

NeilBrown

Re: RAID1 write order fidelity

[David Greaves]

Neil Brown wrote:
>> Hence messages like:
>> Filesystem "md0": Disabling barriers, not supported by the underlying device
> 
> A strange message to get as md/RAID1 *does* support barriers when all the
> underlying devices do (since the mid-teens of 2.6 I think).

You specifically mention raid1 - which, granted, is what the OP asked about - is
md raid1 different from other raid levels wrt barriers?

Also the above message comes from XFS on a raid5 on 2.6.18 - Xen :(
I thought I'd seen this on raid5 on 2.6.2[456] too though (no access to check
right now)

David

Re: RAID1 write order fidelity

[Holger Kiehl]

On Fri, 1 Aug 2008, David Greaves wrote:

> Neil Brown wrote:
>>> Hence messages like:
>>> Filesystem "md0": Disabling barriers, not supported by the underlying device
>>
>> A strange message to get as md/RAID1 *does* support barriers when all the
>> underlying devices do (since the mid-teens of 2.6 I think).
>
> You specifically mention raid1 - which, granted, is what the OP asked about - is
> md raid1 different from other raid levels wrt barriers?
>
> Also the above message comes from XFS on a raid5 on 2.6.18 - Xen :(
> I thought I'd seen this on raid5 on 2.6.2[456] too though (no access to check
> right now)
>
I get this message:

    JBD: barrier-based sync failed on md7 - disabling barriers

That is with 2.6.26 and ext4. But this is with RAID1+0 and md7 is the
RAID0 part.

Holger