I've just had occasion to swipe the mbuf implementation from NetBSD for
use elsewhere, and I've noticed what seems to be a limitation of
m_pullup() (either that or I'm misunderstanding something).

If you try to do an m_pullup of a size > MHLEN, it fails to work. Is
there any reason not to modify m_pullup to do an MCLGET (allowing for
contiguous runs of up to 2048 bytes on i386) when necessary?  (in other
words, is this breaking some rule of mbuf semantics that isn't
documented in the man page?) (Don't bother talking about it being
inefficient - my use is already inefficient anyway, it's just making it
easier to port some network code onto a system that doesn't have mbufs.)