skiplist db corruption

[Sergio Devojno Bruder]

Ken Murchison wrote:
> Robert Mueller wrote:
>> Just reviving an old issue to see if anyone has any more information.
>>
>> http://asg.web.cmu.edu/archive/message.php?mailbox=archive.cyrus-devel&msg=1072 
>>
>>
>> We still see this every now and then, most often with seen state 
>> databases. It's most often after an unclean shutdown, but I've seen 
>> it at other times as well. Has anyone got any more ideas on what 
>> might be causing this?
>
> We haven't seen this locally, so it hasn't risen to the top of my 
> queue yet.
>
We've seen corruption caused by 2 diferent sources:

 - fail on allocation of memory: linux can negate mmap's with plenty of 
memory free. There was a patch to use remap instead of unmap / mmap 
again, but we never tested it;

 - there is some sort of race, that shows up in (real) smp boxes (dont 
count HT in smp);

We "solved" it by a) adding memory (now our mupdate server has 4G) and 
b) using an UP linux kernel. Wimpy, but worked. After that "solutions" 
we hasnt seen a corruption for almost an year.

To put things in perspective:

[root at mupdate ~]# /usr/lib/cyrus-imapd/ctl_mboxlist -d | wc -l
4599957

[root at mupdate ~]# free
             total       used       free     shared    buffers     cached
Mem:       3980680    3127804     852876          0      85612    2767608
-/+ buffers/cache:     274584    3706096
Swap:      2096472        144    2096328

--
Sergio Bruder
Haxent Consultoria