> diff --git a/virt/kvm/vfio.c b/virt/kvm/vfio.c
> index 76b7f6085dcd..a4e9db212adc 100644
> --- a/virt/kvm/vfio.c
> +++ b/virt/kvm/vfio.c
> @@ -15,6 +15,7 @@
> #include <linux/slab.h>
> #include <linux/uaccess.h>
> #include <linux/vfio.h>
> +#include <linux/tsm.h>
> #include "vfio.h"
>
> #ifdef CONFIG_SPAPR_TCE_IOMMU
> @@ -29,8 +30,14 @@ struct kvm_vfio_file {
> #endif
> };
>
> +struct kvm_vfio_tdi {
> + struct list_head node;
> + struct vfio_device *vdev;
> +};
> +
> struct kvm_vfio {
> struct list_head file_list;
> + struct list_head tdi_list;
> struct mutex lock;
> bool noncoherent;
> };
> @@ -80,6 +87,22 @@ static bool kvm_vfio_file_is_valid(struct file *file)
> return ret;
> }
>
> +static struct vfio_device *kvm_vfio_file_device(struct file *file)
> +{
> + struct vfio_device *(*fn)(struct file *file);
> + struct vfio_device *ret;
> +
> + fn = symbol_get(vfio_file_device);
> + if (!fn)
> + return NULL;
> +
> + ret = fn(file);
> +
> + symbol_put(vfio_file_device);
> +
> + return ret;
> +}
> +
> #ifdef CONFIG_SPAPR_TCE_IOMMU
> static struct iommu_group *kvm_vfio_file_iommu_group(struct file *file)
> {
> @@ -297,6 +320,103 @@ static int kvm_vfio_set_file(struct kvm_device *dev, long attr,
> return -ENXIO;
> }
>
> +static int kvm_dev_tsm_bind(struct kvm_device *dev, void __user *arg)
> +{
> + struct kvm_vfio *kv = dev->private;
> + struct kvm_vfio_tsm_bind tb;
> + struct kvm_vfio_tdi *ktdi;
> + struct vfio_device *vdev;
> + struct fd fdev;
> + int ret;
> +
> + if (copy_from_user(&tb, arg, sizeof(tb)))
> + return -EFAULT;
> +
> + ktdi = kzalloc(sizeof(*ktdi), GFP_KERNEL_ACCOUNT);
> + if (!ktdi)
> + return -ENOMEM;
> +
> + fdev = fdget(tb.devfd);
> + if (!fdev.file)
> + return -EBADF;
> +
> + ret = -ENOENT;
> +
> + mutex_lock(&kv->lock);
> +
> + vdev = kvm_vfio_file_device(fdev.file);
> + if (vdev) {
> + ret = kvm_arch_tsm_bind(dev->kvm, vdev->dev, tb.guest_rid);
> + if (!ret) {
> + ktdi->vdev = vdev;
> + list_add_tail(&ktdi->node, &kv->tdi_list);
> + } else {
> + vfio_put_device(vdev);
> + }
> + }
> +
> + fdput(fdev);
> + mutex_unlock(&kv->lock);
> + if (ret)
> + kfree(ktdi);
> +
> + return ret;
> +}
> +
> +static int kvm_dev_tsm_unbind(struct kvm_device *dev, void __user *arg)
> +{
> + struct kvm_vfio *kv = dev->private;
> + struct kvm_vfio_tsm_bind tb;
> + struct kvm_vfio_tdi *ktdi;
> + struct vfio_device *vdev;
> + struct fd fdev;
> + int ret;
> +
> + if (copy_from_user(&tb, arg, sizeof(tb)))
> + return -EFAULT;
> +
> + fdev = fdget(tb.devfd);
> + if (!fdev.file)
> + return -EBADF;
> +
> + ret = -ENOENT;
> +
> + mutex_lock(&kv->lock);
> +
> + vdev = kvm_vfio_file_device(fdev.file);
> + if (vdev) {
> + list_for_each_entry(ktdi, &kv->tdi_list, node) {
> + if (ktdi->vdev != vdev)
> + continue;
> +
> + kvm_arch_tsm_unbind(dev->kvm, vdev->dev);
> + list_del(&ktdi->node);
> + kfree(ktdi);
> + vfio_put_device(vdev);
> + ret = 0;
> + break;
> + }
> + vfio_put_device(vdev);
> + }
> +
> + fdput(fdev);
> + mutex_unlock(&kv->lock);
> + return ret;
> +}
> +
> +static int kvm_vfio_set_device(struct kvm_device *dev, long attr,
> + void __user *arg)
> +{
> + switch (attr) {
> + case KVM_DEV_VFIO_DEVICE_TDI_BIND:
> + return kvm_dev_tsm_bind(dev, arg);
I think the TDI bind operation should be under the control of the device
owner (i.e. VFIO driver), rather than in this bridge driver. The TDI bind
means TDI would be transitioned to CONFIG_LOCKED state, and a bunch of
device configurations breaks the state (TDISP spec 11.4.5/8/9). So the
VFIO driver should be fully aware of the TDI bind and manage unwanted
breakage.
Thanks,
Yilun
