Re: [PATCH] mmu notifiers #v5

[Robin Holt <holt@sgi.com>]

On Tue, Feb 05, 2008 at 02:06:23PM -0800, Christoph Lameter wrote:
> On Tue, 5 Feb 2008, Andrea Arcangeli wrote:
> 
> > On Tue, Feb 05, 2008 at 10:17:41AM -0800, Christoph Lameter wrote:
> > > The other approach will not have any remote ptes at that point. Why would 
> > > there be a coherency issue?
> > 
> > It never happens that two threads writes to two different physical
> > pages by working on the same process virtual address. This is an issue
> > only for KVM which is probably ok with it but certainly you can't
> > consider the dependency on the page-pin less fragile or less complex
> > than my PT lock approach.
> 
> You can avoid the page-pin and the pt lock completely by zapping the 
> mappings at _start and then holding off new references until _end.

XPMEM is doing this by putting our equivalent structure of the mm into a
recalling state which will cause all future faulters to back off, it then
marks any currently active faults in the range as invalid (we have a very
small number of possible concurrent faulters for a different reason),
proceeds to start remote shoot-downs, waits for those shoot-downs to
complete, then returns from the _begin callout with the mm-equiv still in
the recalling state.  Additional recalls may occur, but no new faults can.
The _end callout reduces the number of active recalls until there are
none left at which point the faulters are allowed to proceed again.

Thanks,
Robin