Re: [Xen-devel] [RFC PATCH 9/9] xen: Add use_iommu flag to createdomain domctl

[Oleksandr Tyshchenko <olekstysh@xxxxxxxxx>]

On Wed, Mar 22, 2017 at 5:56 PM, Jan Beulich <JBeulich@xxxxxxxx> wrote:
>>>> On 15.03.17 at 21:05, <olekstysh@xxxxxxxxx> wrote:
>> --- a/xen/include/public/domctl.h
>> +++ b/xen/include/public/domctl.h
>> @@ -66,6 +66,9 @@ struct xen_domctl_createdomain {
>>   /* Is this a xenstore domain? */
>>  #define _XEN_DOMCTL_CDF_xs_domain     5
>>  #define XEN_DOMCTL_CDF_xs_domain      (1U<<_XEN_DOMCTL_CDF_xs_domain)
>> + /* Should IOMMU page tables be populated at the domain creation time? */
>> +#define _XEN_DOMCTL_CDF_use_iommu     6
>> +#define XEN_DOMCTL_CDF_use_iommu      (1U<<_XEN_DOMCTL_CDF_use_iommu)
>>      uint32_t flags;
>
> The need for this to be done via domain creation flag (rather than
> as a separate, later step) needs to be well explained. Generally
> what to add here should only be things which can't be done later
> in a reasonable way.

Well, the non-shared IOMMU should populate its page table by the time
the P2M code will have started update mappings. Theoretically, it
might happen right after p2m_init has been completed,
that is, even during createdomain domctl execution. For example, I see
that domain_vgic_init() inserts mapping to P2M table, because of
map_mmio_regions() is being called during VGIC initialization.
Not sure that GIC mmio ranges must be present in the IOMMU page table,
but anyway, it might be the per-domain initialization of other IPs,
co-processors that mapping mustn't be skipped because of IOMMU is not
ready.
So, as both iommu_domain_init() and p2m_init() are called from
arch_domain_create(), i.e. during createdomain domctl execution, we
have to know exactly should we use IOMMU for this domain to pass
proper value to iommu_domain_init().

If don't take into account everything I wrote above, yes, it is
possible to introduce new domctl for this purpose that will be called
later, but before any operations with guest_physmap.

-- 
Regards,

Oleksandr Tyshchenko

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