Re: [Xen-devel] [PATCH v5 3/4] xen: implement guest_physmap_pin_range and guest_physmap_unpin_range

[Stefano Stabellini <stefano.stabellini@xxxxxxxxxxxxx>]

On Tue, 10 Sep 2013, Ian Campbell wrote:
> On Tue, 2013-09-10 at 20:11 +0100, Stefano Stabellini wrote:
> > On Tue, 10 Sep 2013, Ian Campbell wrote:
> > > On Mon, 2013-09-09 at 17:06 +0100, Stefano Stabellini wrote:
> > > > guest_physmap_pin_range pins a range of guest pages so that their p2m
> > > > mappings won't be changed.
> > > > guest_physmap_unpin_range unpins the previously pinned pages.
> > > > The pinning is done using one of the spare bits in the p2m ptes.
> > > > 
> > > > Use the newly introduce p2m_walker to implement the two functions on
> > > > ARM.
> > > > Provide empty stubs for x86.
> > > > 
> > > > 
> > > > Signed-off-by: Stefano Stabellini <stefano.stabellini@xxxxxxxxxxxxx>
> > > > 
> > > > 
> > > > Changes in v5:
> > > > - return -EBUSY when the P2M_DMA_PIN check fails;
> > > > - rename _guest_physmap_pin_range to pin_one_pte;
> > > > - rename _guest_physmap_unpin_range to unpin_one_pte.
> > > > 
> > > > Changes in v4:
> > > > - use p2m_walker to implement guest_physmap_pin_range and
> > > > guest_physmap_unpin_range;
> > > > - return -EINVAL when the P2M_DMA_PIN check fails;
> > > > - change the printk into a gdprintk;
> > > > - add a comment on what type of page can be pinned.
> > > > ---
> > > >  xen/arch/arm/p2m.c         |   48 
> > > > ++++++++++++++++++++++++++++++++++++++++++++
> > > >  xen/include/asm-arm/mm.h   |    4 +++
> > > >  xen/include/asm-arm/page.h |    7 +++++-
> > > >  xen/include/asm-x86/p2m.h  |   12 +++++++++++
> > > >  4 files changed, 70 insertions(+), 1 deletions(-)
> > > > 
> > > > diff --git a/xen/arch/arm/p2m.c b/xen/arch/arm/p2m.c
> > > > index a9ceacf..bac6c7e 100644
> > > > --- a/xen/arch/arm/p2m.c
> > > > +++ b/xen/arch/arm/p2m.c
> > > > @@ -154,6 +154,46 @@ paddr_t p2m_lookup(struct domain *d, paddr_t paddr)
> > > >      return p2m.maddr;
> > > >  }
> > > >  
> > > > +static int pin_one_pte(lpae_t *ptep, void *arg, int level)
> > > > +{
> > > > +    lpae_t pte = *ptep;
> > > > +    ASSERT(level == 3);
> > > > +
> > > > +    if ( pte.p2m.avail & P2M_DMA_PIN )
> > > > +        return -EBUSY;
> > > > +    pte.p2m.avail |= P2M_DMA_PIN;
> > > > +    write_pte(ptep, pte);
> > > > +    return 0;
> > > > +}
> > > > +
> > > > +int guest_physmap_pin_range(struct domain *d,
> > > > +                            xen_pfn_t gpfn,
> > > > +                            unsigned int order)
> > > > +{
> > > > +    return p2m_walker(d, gpfn << PAGE_SHIFT, order,
> > > > +                      pin_one_pte, NULL);
> > > 
> > > If this fails then you will have left some subset of the pages
> > > successfully pinned. You need to clean it up I think, or be fatal to the
> > > guest or something.
> > 
> > Making the failure of guest_physmap_pin_range fatal to the guest would
> > also solve the non-trivial problem of error handling in memory_exchange.
> 
> Actually, I don't think this will work.
> 
> Consider two concurrent I/O operations dispatched by two threads in the
> same process, to/from non-overlapping regions of the same page.
> 
> Hrm, we may actually need a reference count on the pin :-(
> 
> We might need to rethink here -- i.e. use a type count on the underlying
> machine page rather than a pin in the p2m entry? That kind of makes
> sense since it will be the underlying page which is exposed to the
> hardware...

I think that we should only allow pinning at page granularity level.
Also I think it should be the guest to do the reference counting: it
needs to keep track of the mfn_to_pfn mappings anyway, it might as well
keep track of the references too, so that it can avoid issuing an
hypercall to pin a page that is already pinned.

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