[Dovecot] Various uidlist and index errors with 1.1 on NFS

[Adam McDougall]

On Sat, Nov 10, 2007 at 07:54:15PM +0200, Timo Sirainen wrote:

  On Tue, 2007-11-06 at 15:57 -0500, Adam McDougall wrote:
  > Two nights ago I took a leap and extended my testing of dovecot 1.1 by
  > replacing 1.0 for the approx 15 users I had on 1.0.  At that time I also
  > for the first time tried dovecot 1.1 in a load balanced 2 server configuration
  > with indexes on NFS.
  
  I hadn't actually tested this myself. Would be nice if someone gave me
  access to a NFS test system, would be much easier to test and fix these
  problems. :)

I'd be more than happy to give you access to some of my servers.  
I'll email you privately about that.  If you think the corruption
is resolvable, I'd be happy if it could be worked out to provide 
a better service to my users.
  
  What OS are you using on NFS clients?

FreeBSD 6.2
  
  > Nov  4 23:32:23 boomhauer dovecot: IMAP(mcdouga9): Trying to close
  > mailbox support.2005.01-Jan with open transactions
  
  Probably some error handling path forgot to free the transaction. Would
  be nice to know how to reproduce it.

I agree it would be nice.  If I see it again and find a way to isolate it,
I will.  Perhaps it is caused when a client gets disconnected for some 
reason.  Could just be a symptom of index problems that could get resolved
some other way.
  
  > Nov  5 01:11:43 boomhauer dovecot: IMAP(mcdouga9): Corrupted
  > transaction log
  > file /home/mcdouga9/Maildir/dovecot/private/indexes/.INBOX/dovecot.index.log: Unexpected garbage at EOF
  
  Either writing really left some partially written data there, or
  attribute cache flushing doesn't work.
  
  > Nov  6 09:12:26 boomhauer dovecot:
  > IMAP(walbyjon): /home/walbyjon/Maildir/dovecot/public/control/decs/.support/dovecot-uidlist: Duplicate file entry: 1189447198.M910150P10430.hill:2,RS
  
  Sources explain this as:
  
  	/* This can happen if expunged file is moved back and the file
  	   was appended to uidlist. */
  
  So that would mean that readdir() skipped some files and later found
  them again. But this shouldn't happen as long as Dovecot is the only one
  accessing the maildir, because it locks its syncs..
  
  > Nov  5 15:26:53 hill dovecot: IMAP(mcdouga9): file
  > mail-index-sync-update.c: line 599: unreached
  
  I replaced this crash with a nice error message. :)