Re: [PATCH 4/4] vfio: selftests: Allow the flag MAP_POPULATE to be set on the cmdline

[David Matlack <dmatlack@google.com>]

On 2026-07-01 08:33 PM, Aaron Lewis wrote:
> Add a command-line option to vfio_dma_mapping_perf_test to set the
> MAP_POPULATE flag during mmap(). This ensures consistent and accurate DMA
> mapping latency measurements by forcing page prefaulting to occur during
> the mmap() phase, rather than during the DMA mapping ioctl.
> 
> For example, mapping a 32G region without the -p (MAP_POPULATE) flag
> results in:
> 
>   $ ./vfio_dma_mapping_perf_test -b 32G -a "-v vfio_type1_iommu_memfd_hugetlb_1gb"
>   Completed mmap() for memfd in 0.00ms
>   Mapped HVA 0x7f1840000000 (size 32G) at IOVA 0x800000000 in 2887.56ms
> 
> Here, the mmap() call completes instantly, but the ioctl to map the DMA
> region incurs the ~3s cost of prefaulting the pages.
> 
> Running the same test with the -p flag results in:
> 
>   $ ./vfio_dma_mapping_perf_test -p -b 32G -a "-v vfio_type1_iommu_memfd_hugetlb_1gb"
>   Completed mmap() for memfd in 2887.41ms
>   Mapped HVA 0x7f8000000000 (size 32G) at IOVA 0x800000000 in 11.51ms
> 
> By shifting the prefaulting overhead to mmap(), the actual DMA mapping
> latency is isolated and accurately reported as 11.51ms.
> 
> With user-defined mapping sizes and prefaulting in place, this test now
> clearly demonstrates the difference in mapping speeds between VFIO's Type 1
> IOMMU and IOMMUFD (i.e., IOMMU_IOAS_MAP_FILE).
> 
> Running the same configuration with the "iommufd_memfd_hugetlb_1gb" variant
> highlights this difference:
> 
> $ ./vfio_dma_mapping_perf_test -p -b 32G -a "-v iommufd_memfd_hugetlb_1gb"
> Mapped HVA 0x7f4fc0000000 (size 32G) at IOVA 0x800000000 in 0.11ms
> 
> The DMA mapping time drops dramatically from 11.51ms to 0.11ms.
> 
> Signed-off-by: Aaron Lewis <aaronlewis@google.com>

IMO we should either:

 1. Only support the MAP_POPULATE variant since that isolates the VFIO
    and IOMMUFD performance from allocating the backing memory and
    aligns with how a proper userspace should be designed.

 or

 2. Make MAP_POPULATE toggle a test fixture variant so that we
    automatically test both scenarios without needing a command line
    flag. If the user wants to test just one scenario they can do so via
    kselftest harness -t flag.

Either way, we shouldn't need a command line flag for this.

I don't have a strong opinion between the two but lean towards (2)
until we have a compelling reason to care about the performance of VFIO
and IOMMUFD triggering the backing memory allocations.