Re: [Xen-devel] [PATCH v5 4/4] xen: introduce XENMEM_exchange_and_pin and XENMEM_unpin

[Tim Deegan <tim@xxxxxxx>]

At 19:40 +0100 on 10 Sep (1378842048), Ian Campbell wrote:
> On Tue, 2013-09-10 at 19:37 +0100, Tim Deegan wrote:
> > At 19:32 +0100 on 10 Sep (1378841575), Stefano Stabellini wrote:
> > > On Tue, 10 Sep 2013, Jan Beulich wrote:
> > > > >>> On 10.09.13 at 14:50, Ian Campbell <Ian.Campbell@xxxxxxxxxx> wrote:
> > > > > On Tue, 2013-09-10 at 13:15 +0100, Jan Beulich wrote:
> > > > > 
> > > > >> > @@ -459,6 +460,52 @@ DEFINE_XEN_GUEST_HANDLE(xen_mem_sharing_op_t);
> > > > >> >   * The zero value is appropiate.
> > > > >> >   */
> > > > >> >  
> > > > >> > +#define XENMEM_exchange_and_pin             26
> > > > >> > +/*
> > > > >> > + * This hypercall is similar to XENMEM_exchange: it takes the same
> > > > >> > + * struct as an argument and it exchanges the pages passed in 
> > > > >> > with a new
> > > > >> > + * set of pages. The new pages are going to be "pinned": it's 
> > > > >> > guaranteed
> > > > >> > + * that their p2m mapping won't be changed until explicitly 
> > > > >> > "unpinned".
> > > > >> > + * The content of the exchanged pages is lost.
> > > > >> > + * Only normal guest r/w memory can be pinned: no granted pages or
> > > > >> > + * ballooned pages.
> > > > >> > + * If return code is zero then @out.extent_list provides the DMA 
> > > > >> > frame
> > > > >> > + * numbers of the newly-allocated memory.
> > > > >> 
> > > > >> "DMA"? I don't think that term is universally true across all 
> > > > >> possible
> > > > >> architectures (but we're in an architecture independent header
> > > > >> here). "Machine" would probably be better (as it implies CPU
> > > > >> perspective, whereas DMA hints at device perspective).
> > > > > 
> > > > > I think DMA here is correct. The purpose of exchange and pin is so 
> > > > > that
> > > > > the page can be safely handed to a device for DMA.
> > > > > 
> > > > > On an architecture where DMA address != Machine address then this 
> > > > > should
> > > > > indeed return the DMA addresses.
> > > > 
> > > > One problem is that I think there are architectures where there's no
> > > > single canonical DMA address; such an address may depend on the
> > > > placement of a device in the system's topology. Hence I don't think
> > > > it would even be possible to return "the" DMA address here. It ought
> > > > to be the machine address (CPU view), and the consumer ought to
> > > > know how to translate this to a DMA address for a particular device.
> > > 
> > > We could leave it up to each architecture to specify whether the
> > > hypercall returns a DMA address or a Machine address (according to
> > > their definition of DMA address and Machine address).
> > > Even if this is a common header.
> > 
> > Is this going to be needed on architectures other than arm?  It's not
> > useful for x86, AFAICT.
> 
> It might be needed for PVH with passthrough? Maybe that insists on an
> IOMMU, in which case maybe it is arm only.

Yes, I think it would be fine require a working IOMMU on x86 for
PVH+passthrough.

> Until someone does e.g. a ppc port ;-)

m68k!

Tim.

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel