Re: [PATCH 07/11] unpaged: COW on VM_UNPAGED

["David S. Miller" <davem@davemloft.net>]

From: Hugh Dickins <hugh@veritas.com>
Date: Fri, 18 Nov 2005 08:13:54 +0000 (GMT)

> On Fri, 18 Nov 2005, David S. Miller wrote:
> > From: Hugh Dickins <hugh@veritas.com>
> > Date: Fri, 18 Nov 2005 08:02:02 +0000 (GMT)
> > 
> > > That code is necessary to reproduce the existing behaviour, which has
> > > always done COW on PageReserved mappings without complaint - if the
> > > vm_page_prot didn't already let you slip through without a WP fault.
> > 
> > And there is evidence today that this is really needed, at least
> > by vbetool.
> > 
> > Ok, we need COW on VM_UNPAGED. :)
> 
> Are you so sure of that, that we should even skip adding a warning?

Yes, I'm pretty sure.  The datapoints are like this:

1) Existing vbetool with 2.6.14 and previous works
2) 2.6.15 w/no-COW-on-reserved makes existing vbetool fail
   immediately (of course)
3) 2.6.15 w/no-COW-on-reserved still fails with MAP_SHARED
   patched vbetool
4) 2.6.15 w/COW-on-VM_UNPAGED works with existing vbetool
5) 2.6.15 w/COW-on-VM_UNPAGED _FAILS_ with MAP_SHARED patched
   vbetool

That #5 is the key.

If we use MAP_SHARED and let vbetool modify the real BIOS data
area, resume fails.  That's convincing enough for me :)