[PVE-User] pbs - Nasty atime bug in Ubuntu's kernel-builtin zfs

[Thomas Lamprecht]

Hello,

Am 16/10/2024 um 17:06 schrieb Cornelius Hoffmann:
> I ran in to a very nasty bug with the zfs version included in Ubuntu 
> (NOT the pbs default kernel, I'm running PBS in a container), which is 
> the base for Proxmox as well as far as I understand it and wanted to 
> warn about a problem with access times that is encountered there.
> 
> Essentially, this is [1] and it causes the access time update of a 
> `touch` command to fail and set it to epoch 0, so Jan 1, 1970. This will 
> cause all chunks that should not be pruned to have an extremely old 
> atime and thus be pruned, causing a corruption of all backups.
> 
> I'm writing here to make sure there is not maybe a kernel update waiting 
> with these changes. The newest zfs-linux version in noble has the bug 
> fixed as it includes [2], but that change is only visible in the dkms 
> module, which I installed to fix this problem in the meantime.

Thanks for reaching out to us, but please note that Proxmox maintains
its own ZFS downstream package for both user-space tools and kernel
module.

Both the bad commit and the fix were in ZFS 2.2.3 [0], and unlike Ubuntu
we never backported the problematic commit manually, so the Proxmox
kernel and its ZFS module were never affected, FWICT.

I also tried the reproducer, as it is simple enough and just to be sure,
and I could _not_ observe the broken behavior from ZFS.

[0]: https://github.com/openzfs/zfs/releases/tag/zfs-2.2.3

regards,
 Thomas