Re: [PATCH] vfio/iommu_type1: report the IOMMU aperture info

[Pierre Morel <pmorel@linux.vnet.ibm.com>]

On 30/11/2017 15:16, Pierre Morel wrote:
> On 30/11/2017 13:57, Jean-Philippe Brucker wrote:
>> Hello,
>>
>> On 30/11/17 11:34, Pierre Morel wrote:
>> [...]
>>> +/**
>>> + * vfio_get_aperture - report minimal aperture of a vfio_iommu
>>> + * @iommu: the current vfio_iommu
>>> + * @start: a pointer to the aperture start
>>> + * @end  : a pointer to the aperture end
>>> + *
>>> + * This function iterate on the domains using the given vfio_iommu
>>> + * and restrict the aperture to the minimal aperture common
>>> + * to all domains sharing this vfio_iommu.
>>> + */
>>> +static void vfio_get_aperture(struct vfio_iommu *iommu, uint64_t 
>>> *start,
>>> +                uint64_t *end)
>>> +{
>>> +    struct iommu_domain_geometry geometry;
>>> +    struct vfio_domain *domain;
>>> +
>>> +    *start = 0;
>>> +    *end = U64_MAX;
>>
>> I wonder if the default values should also reflect what the VFIO
>> implementation actually supports. Looking at vfio_dma_do_map, a 32-bit
>> host will reject any iova greater than 32 bits. In addition,
>> vfio_dma_do_unmap doesn't support unmapping the last page of a 64-bit
>> address space (existing IOMMUs would probably reject map requests with
>> IOVA > 52 bits anyway, but if they don't report a domain aperture, VFIO
>> can't guess it).
>>
>> I think it's convenient to use VFIO_IOMMU_UNMAP_DMA on the full address
>> space when an unmap-all is needed, maybe we could provide default 
>> aperture
>> values that help doing this? (~0U for 32-bit and (~0ULL - PAGE_SIZE) for
>> 64-bit)


indeed the error is to use uint64_t, I should *never* use this hardware 
specific values but simply unsigned long and -1UL/~0U at this level.

Shouldn't the 52bit problem be reported by the iommu geometry?


>>
>> Thanks,
>> Jean
>>
> 
> Thanks, I will take care of this.
> 
> Pierre
> 
> 


-- 
Pierre Morel
Linux/KVM/QEMU in Böblingen - Germany