Re: [PATCH v4 3/3] hw/vfio/region: Create dmabuf for PCI BAR per region

["Cédric Le Goater" <clg@redhat.com>]

On 8/5/26 03:09, Gavin Shan wrote:
> Hi Nicolin and Shameer,
> 
> On 1/21/26 9:41 PM, Shameer Kolothum wrote:
>> From: Nicolin Chen <nicolinc@nvidia.com>
>>
>> Linux now provides a VFIO dmabuf exporter to expose PCI BAR memory for P2P
>> use cases. Create a dmabuf for each mapped BAR region after the mmap is set
>> up, and store the returned fd in the region’s RAMBlock. This allows QEMU to
>> pass the fd to dma_map_file(), enabling iommufd to import the dmabuf and map
>> the BAR correctly in the host IOMMU page table.
>>
>> If the kernel lacks support or dmabuf setup fails, QEMU skips the setup
>> and continues with normal mmap handling.
>>
>> Tested-by: Nicolin Chen <nicolinc@nvidia.com>
>> Reviewed-by: Zhenzhong Duan <zhenzhong.duan@intel.com>
>> Reviewed-by: Cédric Le Goater <clg@redhat.com>
>> Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
>> Signed-off-by: Shameer Kolothum <skolothumtho@nvidia.com>
>> ---
>>   hw/vfio/region.c     | 65 +++++++++++++++++++++++++++++++++++++++++++-
>>   hw/vfio/trace-events |  1 +
>>   2 files changed, 65 insertions(+), 1 deletion(-)
>>
>> diff --git a/hw/vfio/region.c b/hw/vfio/region.c
>> index ca75ab1be4..ab39d77574 100644
>> --- a/hw/vfio/region.c
>> +++ b/hw/vfio/region.c
>> @@ -29,6 +29,7 @@
>>   #include "qemu/error-report.h"
>>   #include "qemu/units.h"
>>   #include "monitor/monitor.h"
>> +#include "system/ramblock.h"
>>   #include "vfio-helpers.h"
>>   /*
>> @@ -238,13 +239,71 @@ static void vfio_subregion_unmap(VFIORegion *region, int index)
>>       region->mmaps[index].mmap = NULL;
>>   }
>> +static bool vfio_region_create_dma_buf(VFIORegion *region, Error **errp)
>> +{
>> +    g_autofree struct vfio_device_feature *feature = NULL;
>> +    VFIODevice *vbasedev = region->vbasedev;
>> +    struct vfio_device_feature_dma_buf *dma_buf;
>> +    size_t total_size;
>> +    int i, ret;
>> +
>> +    total_size = sizeof(*feature) + sizeof(*dma_buf) +
>> +                 sizeof(struct vfio_region_dma_range) * region->nr_mmaps;
>> +    feature = g_malloc0(total_size);
>> +    *feature = (struct vfio_device_feature) {
>> +        .argsz = total_size,
>> +        .flags = VFIO_DEVICE_FEATURE_GET | VFIO_DEVICE_FEATURE_DMA_BUF,
>> +    };
>> +
>> +    dma_buf = (void *)feature->data;
>> +    *dma_buf = (struct vfio_device_feature_dma_buf) {
>> +        .region_index = region->nr,
>> +        .open_flags = O_RDWR,
>> +        .nr_ranges = region->nr_mmaps,
>> +    };
>> +
>> +    for (i = 0; i < region->nr_mmaps; i++) {
>> +        dma_buf->dma_ranges[i].offset = region->mmaps[i].offset;
>> +        dma_buf->dma_ranges[i].length = region->mmaps[i].size;
>> +    }
>> +
> 
> Shall we check if @offset and @size is aligned to PAGE_SIZE? If they're not,
> I guess we need to skip populating DMA buffer instead of preventing the
> device from being passed through to the guest.
> 
> Meghana <meshetty@redhat.com> runs into issue when passing through an NVMe
> card. The only memory BAR on the NVMe card is 16K, which is not aligned to
> 64KB (host page size).
> 
>    -device vfio-pci,id=nvme,host=0004:01:00.0,addr=0x2.0x1,bus=pcie.0
> 
>    host$ sh vfio.sh
>    QEMU 10.1.0 monitor - type 'help' for more information
>    (qemu) qemu-kvm: -device vfio-pci,id=nvme,host=0004:01:00.0,addr=0x2.0x1,bus=pcie.0: \
>                     0004:01:00.0 BAR 0: failed to create dma-buf: PCI BAR IOMMU  mappings may fail: Invalid argument

To avoid confusion, the reported error message was slightly
improved in :

   https://lore.kernel.org/qemu-devel/20260721170518.4160785-3-clg@redhat.com/

Thanks,

C.