Re: [PATCH] vfio-pci: add Ascend devices passthrough quirks

[Alex Williamson]

On Thu, 29 Oct 2020 19:40:48 +0800
Binfeng Wu <wubinfeng@huawei.com> wrote:

> Ascend is a series of SoC processors developed by Huawei. Ascend310/910
> are highly efficient, flexible, and programmable AI processors in this
> series and support device passthrough via vfio-pci. Ascends device
> xloader update is only allowed in host, because update triggered by vm
> users may affect other users when Ascend devices passthrough to vm.
> Set a bar quirk is an effective method to keep vm users from updating
> xloader. In this patch, two bar quirks were proposed to cover
> Ascend310/910 respectively.


If you're trying to say that userspace, not just a VM, should not be
able to update this feature, then is QEMU the right place to implement
this quirk versus device specific protection within the host kernel?
Thanks,

Alex

> Signed-off-by: Binfeng Wu <wubinfeng@huawei.com>
> ---
>  hw/vfio/pci-quirks.c | 104 +++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 104 insertions(+)
> 
> diff --git a/hw/vfio/pci-quirks.c b/hw/vfio/pci-quirks.c
> index 57150913b7..291a45d3ab 100644
> --- a/hw/vfio/pci-quirks.c
> +++ b/hw/vfio/pci-quirks.c
> @@ -1202,6 +1202,108 @@ int vfio_pci_igd_opregion_init(VFIOPCIDevice *vdev,
>      return 0;
>  }
>  
> +#define PCI_VENDOR_ID_HUAWEI      0x19e5
> +#define PCI_DEVICE_ID_ASCEND910   0xd801
> +#define PCI_DEVICE_ID_ASCEND310   0xd100
> +#define ASCEND910_XLOADER_SIZE    4
> +#define ASCEND910_XLOADER_OFFSET  0x80400
> +#define ASCEND310_XLOADER_SIZE    4
> +#define ASCEND310_XLOADER_OFFSET  0x400
> +
> +typedef struct VFIOAscendBarQuirk {
> +    struct VFIOPCIDevice *vdev;
> +    pcibus_t offset;
> +    uint8_t bar;
> +    MemoryRegion *mem;
> +} VFIOAscendBarQuirk;
> +
> +static uint64_t vfio_ascend_quirk_read(void *opaque,
> +                                       hwaddr addr, unsigned size)
> +{
> +    VFIOAscendBarQuirk *quirk = opaque;
> +    VFIOPCIDevice *vdev = quirk->vdev;
> +
> +    return vfio_region_read(&vdev->bars[quirk->bar].region,
> +                            addr + quirk->offset, size);
> +}
> +
> +static void vfio_ascend_quirk_write(void *opaque, hwaddr addr,
> +                                    uint64_t data, unsigned size)
> +{
> +}
> +
> +static const MemoryRegionOps vfio_ascend_intercept_regs_quirk = {
> +    .read = vfio_ascend_quirk_read,
> +    .write = vfio_ascend_quirk_write,
> +    .endianness = DEVICE_LITTLE_ENDIAN,
> +};
> +
> +static void vfio_probe_ascend_bar0_quirk(VFIOPCIDevice *vdev, int nr)
> +{
> +    VFIOQuirk *quirk;
> +    VFIOAscendBarQuirk *bar0_quirk;
> +
> +    if (!vfio_pci_is(vdev, PCI_VENDOR_ID_HUAWEI, PCI_DEVICE_ID_ASCEND910) ||
> +        nr != 0) {
> +        return;
> +    }
> +
> +    quirk = g_malloc0(sizeof(*quirk));
> +    quirk->nr_mem = 1;
> +    quirk->mem = g_new0(MemoryRegion, quirk->nr_mem);
> +    bar0_quirk = quirk->data = g_new0(typeof(*bar0_quirk), quirk->nr_mem);
> +    bar0_quirk[0].vdev = vdev;
> +    bar0_quirk[0].offset = ASCEND910_XLOADER_OFFSET;
> +    bar0_quirk[0].bar = nr;
> +
> +    /*
> +     * intercept w/r to the xloader-updating register,
> +     * so the vm can't enable xloader-updating
> +     */
> +    memory_region_init_io(&quirk->mem[0], OBJECT(vdev),
> +                          &vfio_ascend_intercept_regs_quirk,
> +                          &bar0_quirk[0],
> +                          "vfio-ascend-bar0-intercept-regs-quirk",
> +                          ASCEND910_XLOADER_SIZE);
> +    memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
> +                                        bar0_quirk[0].offset,
> +                                        &quirk->mem[0], 1);
> +    QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
> +}
> +
> +static void vfio_probe_ascend_bar4_quirk(VFIOPCIDevice *vdev, int nr)
> +{
> +    VFIOQuirk *quirk;
> +    VFIOAscendBarQuirk *bar4_quirk;
> +
> +    if (!vfio_pci_is(vdev, PCI_VENDOR_ID_HUAWEI, PCI_DEVICE_ID_ASCEND310) ||
> +        nr != 4) {
> +        return;
> +    }
> +
> +    quirk = g_malloc0(sizeof(*quirk));
> +    quirk->nr_mem = 1;
> +    quirk->mem = g_new0(MemoryRegion, quirk->nr_mem);
> +    bar4_quirk = quirk->data = g_new0(typeof(*bar4_quirk), quirk->nr_mem);
> +    bar4_quirk[0].vdev = vdev;
> +    bar4_quirk[0].offset = ASCEND310_XLOADER_OFFSET;
> +    bar4_quirk[0].bar = nr;
> +
> +    /*
> +     * intercept w/r to the xloader-updating register,
> +     * so the vm can't enable xloader-updating
> +     */
> +    memory_region_init_io(&quirk->mem[0], OBJECT(vdev),
> +                          &vfio_ascend_intercept_regs_quirk,
> +                          &bar4_quirk[0],
> +                          "vfio-ascend-bar4-intercept-regs-quirk",
> +                          ASCEND310_XLOADER_SIZE);
> +    memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
> +                                        bar4_quirk[0].offset,
> +                                        &quirk->mem[0], 1);
> +    QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
> +}
> +
>  /*
>   * Common quirk probe entry points.
>   */
> @@ -1251,6 +1353,8 @@ void vfio_bar_quirk_setup(VFIOPCIDevice *vdev, int nr)
>      vfio_probe_nvidia_bar5_quirk(vdev, nr);
>      vfio_probe_nvidia_bar0_quirk(vdev, nr);
>      vfio_probe_rtl8168_bar2_quirk(vdev, nr);
> +    vfio_probe_ascend_bar0_quirk(vdev, nr);
> +    vfio_probe_ascend_bar4_quirk(vdev, nr);
>  #ifdef CONFIG_VFIO_IGD
>      vfio_probe_igd_bar4_quirk(vdev, nr);
>  #endif