Re: [PATCH 3/3] virtio-iommu: Support PCI device aliases

[Eric Auger]

Hi Zhenzhong,

On 1/22/24 07:40, Zhenzhong Duan wrote:
> Currently virtio-iommu doesn't work well if there are multiple devices
> in same iommu group. In below example config, guest virtio-iommu driver
> can successfully probe first device but fail on others. Only one device
> under the bridge can work normally.
>
> -device virtio-iommu \
> -device pcie-pci-bridge,id=root0 \
> -device vfio-pci,host=81:11.0,bus=root0 \
> -device vfio-pci,host=6f:01.0,bus=root0 \
>
> The reason is virtio-iommu stores AS(address space) in hash table with
> aliased BDF and corelates endpoint which is indexed by device's real
> BDF, i.e., virtio_iommu_mr() is passed a real BDF to lookup AS hash
> table, we either get wrong AS or NULL.
>
> Fix it by storing AS indexed by real BDF. This way also make iova_ranges
> from vfio device stored in IOMMUDevice of real BDF successfully.
>
> Signed-off-by: Zhenzhong Duan <zhenzhong.duan@intel.com>
> ---
>  hw/virtio/virtio-iommu.c | 16 ++++++++--------
>  1 file changed, 8 insertions(+), 8 deletions(-)
>
> diff --git a/hw/virtio/virtio-iommu.c b/hw/virtio/virtio-iommu.c
> index d99c1f0d64..6880d92a44 100644
> --- a/hw/virtio/virtio-iommu.c
> +++ b/hw/virtio/virtio-iommu.c
> @@ -399,27 +399,27 @@ static AddressSpace *virtio_iommu_find_add_as(PCIBus 
> *bus, void *opaque,
>                                                int real_devfn)
>  {
>      VirtIOIOMMU *s = opaque;
> -    IOMMUPciBus *sbus = g_hash_table_lookup(s->as_by_busptr, bus);
> +    IOMMUPciBus *sbus = g_hash_table_lookup(s->as_by_busptr, real_bus);
>      static uint32_t mr_index;
>      IOMMUDevice *sdev;
>  
>      if (!sbus) {
>          sbus = g_malloc0(sizeof(IOMMUPciBus) +
>                           sizeof(IOMMUDevice *) * PCI_DEVFN_MAX);
> -        sbus->bus = bus;
> -        g_hash_table_insert(s->as_by_busptr, bus, sbus);
> +        sbus->bus = real_bus;
> +        g_hash_table_insert(s->as_by_busptr, real_bus, sbus);
>      }
>  
> -    sdev = sbus->pbdev[devfn];
> +    sdev = sbus->pbdev[real_devfn];
>      if (!sdev) {
>          char *name = g_strdup_printf("%s-%d-%d",
>                                       TYPE_VIRTIO_IOMMU_MEMORY_REGION,
> -                                     mr_index++, devfn);
> -        sdev = sbus->pbdev[devfn] = g_new0(IOMMUDevice, 1);
> +                                     mr_index++, real_devfn);
> +        sdev = sbus->pbdev[real_devfn] = g_new0(IOMMUDevice, 1);
>  
>          sdev->viommu = s;
> -        sdev->bus = bus;
> -        sdev->devfn = devfn;
> +        sdev->bus = real_bus;
> +        sdev->devfn = real_devfn;
but then this means the 2 devices would be abstracted by two different
IOMMU MRs whereas in practice they cannot be distinguished from an IOMMU
pov. Shouldn't the virtio-iommu driver use the same ep_id for both
devices within the same group?

Note there are some known issues about virtio-iommu and pcie-to-pci
bridges which were reported early last year and confirmed by Robin
Murphy. See:

[RFC] virtio-iommu: Take into account possible aliasing in virtio_iommu_mr() 
<https://lore.kernel.org/all/20230116124709.793084-1-eric.auger@redhat.com/#r>

https://lore.kernel.org/all/20230116124709.793084-1-eric.auger@redhat.com/

Thanks

Eric




>  
>          trace_virtio_iommu_init_iommu_mr(name);
>