> > and, as i wrote in another mail, it might be better to have
> > some kludge for pages dirtied by write.
> 
> This would be in uvm code then, right ?

right.

> So we would need to extend the vnode op
> interface for this ?

if you mean VOP function prototypes, i don't think so.

> This patch causes the mtime to be updated when dirty pages are written to
> disk, not when they are dirtied (which can delay the mtime update by
> several seconds). On the other hand, tools like dump don't want to have the
> mtime updated before data are written to disk. I don't know how this would
> interract with filesystem snapshots.
> Does SUSv3 say anything about it ?

dump or snapshots are not in susv3, afaik.

YAMAMOTO Takashi