> Doesn't the NFS code attach the contents of a struct buf to an mbuf to 
> achieve zero-copy sends?  The contents of that struct buf should really 
> not be modified, but your patch would make the mbuf code assume that it 
> can be.

the nfs code (both of client and server) uses M_EXT_ROMAP.

> > if it isn't enough (although there's currently no case i can think of),
> > we can introduce a new flag, say, M_EXT_READONLY.
> 
> I would say that the existing behavior should remain, and non-cluster 
> external storage that is R/W'able should explicitly set an M_EXT_RW 
> bit.

should M_EXT_RW be set for M_CLUSTER mbufs as well?

although i prefer M_EXT_READONLY because the most of mbufs are R/W,
M_EXT_RW is also ok for me.

YAMAMOTO Takashi