Hello Florent,

How did you proceed with the upgrade? Did you follow the recommended steps guide to upgrade ceph? (mons first, then OSDs, then MDS) Did you interrupt dovecot before upgrading the MDS specially? Did you remount the filesystem? Did you upgrade the ceph client too?

Give people the complete scene and someone might be able to help you. Ask on ceph-users list too.

Regards,

Webert Lima DevOps Engineer at MAV Tecnologia *Belo Horizonte - Brasil* *IRC NICK - WebertRLZ*

On Thu, Jan 18, 2018 at 8:41 AM, Florent B <florent@coppint.com> wrote:

Hi list,

I'm sorry to come back with my problem. I'm pretty sure it's not dovecot-related, but can someone help me to bring some debug information to Ceph developers ?

For example, how writes are handled in Dovecot, and what kind of corruption it is according to error messages ?

On 13/12/2017 16:19, Florent B wrote:

...

Hi,

I use Dovecot (last released version), 2 backends and 1 director, each user account handled by a single assigned backend.

I use CephFS filesystem for messages (FUSE client).

Since Ceph upgrade from Kraken to Luminous, I have a lot of erreurs "Error: Corrupted dbox file" on a single (large) mail account.

I know the problem seems to come from Ceph, but maybe someone here can help me diagnose the situation.

The error is exactly : EOF reading msg header (got 0/30 bytes)

Dovecot backends are configured like this :

mmap_disable = yes mail_fsync = optimized mail_nfs_storage = no mail_nfs_index = no mdbox_rotate_size = 12M

CephFS supports file locking, and 2 backends never write the same file because each user is assigned to a backend.

Did you ever see this problem with others FS maybe ?

It seems problem disappeared when I disable pagecache for the Fuse mount.

Thank you for your help.