> I'm being cautious; the msync() hack that was mentioned should fo the
> trick if INN is the only thing touching the files, though, actually.

Actually, in -current, that msync() hack should be unnecessary; the
much-villified vnode_pager_sync() change that broke the whole VM system for
several days ensures that data modified via mmap() hits the disk along with
all other delayed writes.

Still, it bears emphasis that there is *not* consistency between the object
cache and the buffer cache; if you use the filesystem to write under your
mmap()ed object, you can expect to lose.

Thor