Re: [PATCH RFC] p2m: p2m_mmio_direct set RW permissions

* Re: [PATCH RFC] p2m: p2m_mmio_direct set RW permissions
@ 2015-01-22 15:18 Elena Ufimtseva
  2015-01-22 15:36 ` Jan Beulich
  2015-01-22 15:42 ` Roger Pau Monné
  0 siblings, 2 replies; 27+ messages in thread
From: Elena Ufimtseva @ 2015-01-22 15:18 UTC (permalink / raw)
  To: JBeulich
  Cc: kevin.tian, andrew.cooper3, tim, xen-devel, yang.z.zhang,
	boris.ostrovsky, roger.pau

----- JBeulich@suse.com wrote:

> >>> On 22.01.15 at 12:37, <roger.pau@citrix.com> wrote:
> > El 22/01/15 a les 12.09, Jan Beulich ha escrit:
> >>>>> On 22.01.15 at 11:59, <andrew.cooper3@citrix.com> wrote:
> >>> On 22/01/15 09:53, Jan Beulich wrote:
> >>>>>>> On 21.01.15 at 21:55, <elena.ufimtseva@oracle.com> wrote:
> >>>>> p2m_mmio_direct should result in setting IOMMUF_readable and
> IOMMUF_writable
> >>>>> flags.
> >>>>> When pvh domain maps mmio regions, the EPT entries are not
> getting mapped.
> >>>>> This leads to IOMMU Page faults for some devices, as for example
> USB Host
> >>>>> controllers with embedded Debug devices. See
> pvh-set-need_iommu-early RFC
> >>>>> patch discussion fgor detail.
> >>>> Even more so that the two patches aren't even a series, that
> part
> >>>> of the description belongs here, not in the other patch.
> >>>>
> >>>>> I will appreciate your comments and ideas in regards to this
> change.
> >>>>>
> >>>>> Looking at Roger patches (xen/pvh: check permissions when adding
> MMIO 
> >>>>> regions)
> >>>>> the mmio memory type is proposed to be changed from
> p2m_mmio_direct to 
> >>>>> p2m_access_rw.
> >>>>> This type still does not have proper IOMMU flags mapping.
> >>>> A fundamental question is what business devices have to DMA
> their
> >>>> own (or other devices') MMIO space. I could remotely see a need
> >>>> for this for e.g. frame buffers, but I have difficulty
> understanding
> >>>> this for USB devices. Please at the very least provide details on
> the
> >>>> MMIO regions that those devices have, and which of them you
> >>>> observed IOMMU faults on.
> >>>
> >>> It would appear that, in this case, it is a USB controller lacking
> an
> >>> appropriate RMRR description in the ACPI tables.
> >> 
> >> No, RMRRs only reference RAM pages afaik.
> > 
> > According to Linux IOMMU support document RMRRs reference regions
> marked
> > as "reserved" in the e820 memory map:
> > 
> > https://www.kernel.org/doc/Documentation/Intel-IOMMU.txt 
> 
> Exactly. And MMIO PCI BARs in particular are never to be
> reflected in the E820 (not the least because they're relocatable).
> 
> > I don't think we are setting proper IOMMU entries for this regions
> at
> > all with the current PVH Dom0 code. IMHO instead of just adding
> IOMMU
> > entries for every MMIO region we should just add IOMMU entries for
> the
> > RMRR regions.
> 
> Yes, RMRR regions should certainly be put there as r/w entries.
> 
> Jan

How it will help in cases like this when this regions are not reported as RMRRs in ACPI?

Also, pv dom0 has no such problem.

Elena

^ permalink raw reply	[flat|nested] 27+ messages in thread