Bug 1034856

Description of problem:

There are two parts to the following problem scenario with suggestions for two possible logic chnages.

1) In a specific incident, Live Storage Migration failed due to pathnames not existing in the '/rhev/data-center' tree. This condition had occurred due to vdsm having been upgraded incorrectly. The result of the failed LSM attempt was that the following were created;

- new "snapshot image" in the source domain.
- base image in the destination domain.
- "snapshot image" in the destination domain.

By "snapshot image" I mean the image created by the snapshot creation, I realise the snapshot is actually the base image.

In this case they have block-based storage and so two new logical volumes got created in the destination domain. However, nothing got copied to these images, so in essence they could have just been removed when the failure occurred. So, the first suggestion is to have LSM cleanup after itself after a failure.


2) To compound this, a subsequent LSM was performed. This also failed, but this time because the base image already existed in the destination domain. As a result another "snapshot image" got created in the source domain. Unfortunately the customer repeated this a few more times and the end result was that several VMs with multiple disks now had several snapshots (images only in the source domain). Each LV is most likely only 1gb, but multiplying this by the number of disks and the number of failed LSM attempts, there were now about 50 meaningless LVs, and therefore about 50gb of wasted space in the source domain. Not only that, but if the customer wants to remove these snapshots at a later time, they now have to bring the VM down (this is RHEV 3.2) and then remove multiple snapshots, which may not take too long for each one for the auto-generated ones (unless they end up merging woth the base image), but it still requires time and sufficient knowledge of exactly what you're doing and what's happeneing when you're removing snapshots. My concern is that this scenario may not be well-understood by customers and they might end up merging into the base image, which could lead to significant downtime.

I haven't looked at the LSM logic to see if it's feasible to not generate the snapshot before checking for the existence of images in the destination domain, but if that could be done then would be beneficial.



Version-Release number of selected component (if applicable):

- RHEV 3.2
- RHEL6.4 host with vdsm-4.10.2-27.0


How reproducible:


Steps to Reproduce:
1.
2.
3.

Actual results:

1) Logical volumes left behind in the destination domain after a specific LSM failure.
2) Snapshots created during failed LSM attempts.


Expected results:

1) Logical volumes removed from the destination domain after this specific LSM failure.
2) Snapshots not created during failed LSM attempts.


Additional info:

I will add supporting data and logs a little later.