The attached logs are from the cluster hosts that were running the HA VMs during the failures. I've finally got all of my HA VMs up again. The last one didn't start again until after I freed up more space in the storage domain than what was originally available when the VM was running previously. (It now has over 150GB of free space. Which should be more than enough, but it didn't boot with 140GB avaiable....) SideNote: I just found this in the logs on the original host that the HA VMs were running on: ---snip--- Jun 02 10:33:29 ryuki.codenet sanlock[1054]: 2022-06-02 10:33:29 674607 [1054]: s1 check_our_lease warning 71 last_success 674536 # semanage fcontext -a -t virt_image_t '1055' ***** Plugin catchall (2.13 confidence) suggests ************************** Then you should report this as a bug. You can generate a local policy module to allow this access. Do Jun 02 10:33:45 ryuki.codenet sanlock[1054]: 2022-06-02 10:33:45 674623 [1054]: s1 kill 3441 sig 15 count 8 Jun 02 10:33:45 ryuki.codenet sanlock[1054]: 2022-06-02 10:33:45 674623 [1054]: s1 kill 4337 sig 15 count 8 Jun 02 10:33:46 ryuki.codenet sanlock[1054]: 2022-06-02 10:33:46 674624 [1054]: s1 kill 3206 sig 15 count 9 Jun 02 10:33:47 ryuki.codenet kernel: ovirtmgmt: port 4(vnet2) entered disabled state ---snip--- That looks like some SELinux failure. -Patrick Hibbs On Thu, 2022-06-02 at 19:44 +0300, Nir Soffer wrote:
On Thu, Jun 2, 2022 at 7:14 PM Patrick Hibbs <hibbsncc1701@gmail.com> wrote:
OK, so the data storage domain on a cluster filled up to the point that the OS refused to allocate any more space.
This happened because I tried to create a new prealloc'd disk from the Admin WebUI. The disk creation claims to be completed successfully, I've not tried to use that disk yet, but due to a timeout with the storage domain in question the engine began trying to fence all of the HA VMs. The fencing failed for all of the HA VMs leaving them in a powered off state. Despite all of the HA VMs being up at the time, so no reallocation of the leases should have been necessary.
Leases are not reallocated during fencing, not sure why you expect this to happen.
Attempting to restart them manually from the Admin WebUI failed. With the original host they were running on complaining about "no space left on device", and the other hosts claiming that the original host still held the VM lease.
No space left on device may be an unfortunate error from sanlock, meaning that there is no locksapce. This means the host has trouble adding the lockspace, or it did not complete yet.
After cleaning up some old snapshots, the HA VMs would still not boot. Toggling the High Availability setting for each one and allowing the lease to be removed from the storage domain was required to get the VMs to start again.
If you know that the VM is not running, disabling the lease temporarily is a good way to workaround the issue.
Re-enabling the High Availability setting there after fixed the lease issue. But now some, not all, of the HA VMs are still throwing "no space left on device" errors when attempting to start them. The others are working just fine even with their HA lease enabled.
All erros come from same host(s) or some vms cannot start while others can on the same host?
My questions are:
1. Why does oVirt claim to have a constantly allocated HA VM lease on the storage domain when it's clearly only done while the VM is running?
Leases are allocated when a VM is created. This allocated a the lease space (1MiB) in the external leases special volume, and bind it to the VM ID.
When VM starts, it acquires the lease for its VM ID. If sanlock is not connected to the lockspace on this host, this may fail with the confusing "No space left on device" error.
2. Why does oVirt deallocate the HA VM lease when performing a fencing operation?
It does not. oVirt does not actually "fence" the VM. If the host running the VM cannot access storage and update the lease, the host lose all leases on that storage. The result is pausing all the VM holding a lease on that storage.
oVirt will try to start the VM on another host, which will try to acquire the lease again on the new host. If enough time passed since the original host lost access to storage, the lease can be acquired on the new host. If not, this will happen in the next retrie(s).
If the original host did not lose access to storage, and it is still updating the lease you cannot acquire the lease from another host. This protect the VM from split-brain that will corrupt the vm disk.
3. Why can't oVirt clear the old HA VM lease when the VM is down and the storage pool has space available? (How much space is even needed? The leases section of the storage domain in the Admin WebUI doesn't contain any useful info beyond the fact that a lease should exist for a VM even when it's off.)
Acquiring the lease is possible only if the lease is not held on another host.
oVirt does not support acquiring a held lease by killing the process holding the lease on another host, but sanlock provides such capability.
4. Is there a better way to force start a HA VM when the lease is old and the VM is powered off?
If the original VM is powered off for enough time (2-3 minutes), the lease expires and starting the VM on another host should succeed.
5. Should I file a bug on the whole HA VM failing to reacquire a lease on a full storage pool?
The external lease volume is not fully allocated. If you use thin provisioned storage, and the there is really no storage space, it is possible that creating a new lease will fail, but starting and stopping VM that have leases should not be affected. But if you reach to the point when you don't have enough storage space you have much bigger trouble and you should fix urgently.
Do you really have issue with available space? What does engine reports about the storage domain? What does the underlying storage reports?
Nir