[LU-3575] 'mkfs.lustre --writeconf' not working anymore with Lustre 2.4

• Type: Bug
• Resolution: Duplicate
• Priority: Major
• Affects Version/s: Lustre 2.4.0
• Severity: 3
• Assignee: Alex Zhuravlev
• Reporter: Sebastien Buisson (Inactive)

Description

Hi,

With Lustre 2.4, '--writeconf' flag of mkfs.lustre is ignored. Indeed, after formatting a target with this flag, its label separator is ':', which means VIRGIN, instead of '='.
To force writeconf, one has to run 'tunefs.lustre --writeconf' on the target prior to starting it for the first time.

This is seen as a regression compared to Lustre 2.1, with which we can directly pass the writeconf flag at mkfs.lustre time.

We need this feature (ie forcing writeconf at mkfs time) when we reformat an existing file system registered at an external MGS after modifying its characteristics.

Could it be possible to revise writeconf flag implementation in Lustre 2.4 so that it can be used directly at mkfs.lustre time?

TIA,
Sebastien.

Issue Links

• is related to: LU-14 live replacement of OST

Comments

[Peter Jones]

Of course - when we start thinking about 2.4.2 this will definitely be under consideration.

[Sebastien Buisson (Inactive)]

Would it be possible to please have the patch http://review.whamcloud.com/7443 landed to 2.4 too?

[Peter Jones]

Great! So, that patch has landed for 2.5.0 so it sound like this ticket can be closed

[Sebastien Buisson (Inactive)]

Hi,

I have tested the patch at http://review.whamcloud.com/7443.

For us it does the trick. Now, when we need to reformat a new target to replace an existing one, we use '--reformat --replace' instead of '--reformat --writeconf'.

Thanks!
Sebastien.

[Andreas Dilger]

The main reason for having a separate flag from --reformat (which is setting VIRGIN exactly because it is a new filesystem) is that there is some chance of the admin making a mistake assigning the OST index, and then we may have two OSTs with the same index. If the --reformat (with VIRGIN) OST connects to the MGS and it has a duplicate index it will get an error back. It takes an extra flag (--replace) to indicate that the admin knows that this just-reformatted OST is replacing the previous OST at the same index.

[Sebastien Buisson (Inactive)]

Well, this is getting complicated.

Why adding another flag to do this? If you do not want writeconf to have precedence over virgin by default, I would suggest to modify the already existing '--reformat' flag so that it unsets the LDD_F_VIRGIN flag. Additionally, it looks like LDD_F_WRITECONF should be set at the same time, otherwise the MGS may refuse registration of the target.

Moreover, it should be documented somewhere that '--writeconf' flag of mkfs.lustre is ignored starting with 2.4.

What do you think?

Sebastien.

[Alex Zhuravlev]

yes, this approach looks better.

[Andreas Dilger]

Hmm, I've made a patch http://review.whamcloud.com/7443 which adds the "mkfs.lustre --replace|-R" option that allows you to format a replacement OST in a single step. That is more intuitive IMHO than mkfs.lustre + tunefs.lustre --writeconf.

[Cory Spitz]

If 'mkfs.lustre --writeconf' is flat broke, why didn't we catch this prior to 2.4.0?

[Sebastien Buisson (Inactive)]

Hi there,

I can confirm that this patch (http://review.whamcloud.com/#/c/7232/) solves the problem for me.
Is it possible to also have it on b2_4?

TIA,
Sebastien.