mmap required? (was Re: Draft specification for future X-Trace h

[Heath Kehoe]

>
>> Standard problem with running INN over NFS is that it likes mmap and NFS
>> doesn't.  If you turn off mmap for everything, it will likely work fine,
>> but Clayton made some rumblings about not supporting mmap-free systems any
>> more, so I'm not sure if that still works.  I keep meaning to go peer at
>> that and haven't had time.
>
>Without having tested it, just recently peering at code and working
>from memory (big mistake?), innd never syncs/writes the active file
>now.  It mmaps active, and expects writes to show up on disk.
>
>nnrpd needs the active file to be current, and won't get it.
>

Starting with 2.3, the only time nnrpd ever accesses the active
file is when the client does a 'list active'.  For everything else,
it uses overview.  Anyway, innd does do msync()s periodically on
the active file (every innconf->icdsynccount articles).

Speaking of overview, all of the overview methods use mmap (tradindexed
and buffindexed mmap the group.index file, and BerkeleyDB mmaps its
__db* files).  So its no longer feasable to have reader machines that
NFS-mount spool and overview from a transport machine.

-heath