Messages in this thread | | | From | "Yu, Fenghua" <> | Date | Thu, 18 Jun 2009 11:50:57 -0700 | Subject | RE: [PATCH 1/2] IOMMU Identity Mapping Support: iommu_identity_mapping definition |
| |
> >On Thu, Jun 18, 2009 at 11:31:21AM -0700, Chris Wright wrote: > >> The 1-1 mapping is for all the host devices _not_ assigned to >> guests. To eliminate the i/o overhead imposed on all guests not >> using an assigned device or from i/o from host. >> >> It's just the same as VT-d PassThrough mode for hardware that >> doesn't support it. > >Ok, that makes sense. Thanks, Chris. However, that doesn't appear to >be what the patch does---unless I'm misreading, if >iommu_identity_mapping is set, *all* devices get identity >mapping. Instead of a global command line option, we need to provide a >way to enable/disable pt or identity mapping (I agree that the user >shouldn't know or care which is used, the kernel should pick the best >one automatically) on a per BDF basis.
The device in kvm is attached to a kvm domain by intel_iommu_attach_device(). In this function, domain_context_mapping() changes the device's domain to kvm domain from si_domain.
Actually there is a bug when a device is detached from kvm domain...in intel_iommu_detach_device(), I should assigned the device back to the si_domain. So the device can be used in native again.
Thanks.
-Fenghua
| |