[Dovecot] index corruption

[rosenfield at users.sourceforge.net]

Timo Sirainen wrote:
> rosenfield at users.sourceforge.net wrote:
> > postfix/smtpd: disconnect from unknown
> > deliver: msync() failed with index file maildir/dovecot.index:
> > Input/output error
> > unionfs: new lower inode mtime (bindex=0, name=maildir)
>
> What's unionfs?

Layered filesystem which serves the virtual machine a basic operating
system, on top of which you can enable services and write your own
/etc/* files, etc.  But all dovecot files are kept in the local
filesystem. The base filesystem is not writeable.

> Looks like the problems is with it, not Dovecot.

Unlikely, there are a ton of virtual machines running various
databases, apaches, and what not, and they've never had a problem.

(Hrm, sounded a bit like Van Leuwen there talking about the colonists
on LV-426 in Aliens..)

Maybe there's an interaction problem between dovecot and unionfs?
I know that unionfs does not support sendfile, for example.
It usually results in empty buffers though, not in stuff working halfway..

> "Input/output error" should never happen unless there's a bug in the
> OS/filesystem or your hard disk is dying.

It's on redundant disks, so that shouldn't happen..
(And especially not *only* to dovecot index files for several users..
statistics would say that is rather impossible)

> You could try if mmap_disable=yes helps.

Done.
(I imagine that various databases use mmap too though, so it seems
unlikely it will fix anything. Not trying to be pessimistic, just
objective.)

Looking closer in the logs, it seems that when the corruption occurs,
users are often logged in via IMAP at the same time as a message is
stored via deliver.  Hmm?