Re: [PATCH v2 03/10] iommufd: Create access in vfio_iommufd_emulated_bind()
From: Eric Farman
Date: Thu Feb 09 2023 - 13:59:16 EST
On Tue, 2023-02-07 at 13:17 -0800, Nicolin Chen wrote:
...snip...
> diff --git a/drivers/vfio/iommufd.c b/drivers/vfio/iommufd.c
> index 026f81a87dd7..dc9feab73db7 100644
> --- a/drivers/vfio/iommufd.c
> +++ b/drivers/vfio/iommufd.c
> @@ -141,10 +141,19 @@ static const struct iommufd_access_ops
> vfio_user_ops = {
> int vfio_iommufd_emulated_bind(struct vfio_device *vdev,
> struct iommufd_ctx *ictx, u32
> *out_device_id)
> {
> + struct iommufd_access *user;
> +
> lockdep_assert_held(&vdev->dev_set->lock);
>
> - vdev->iommufd_ictx = ictx;
> iommufd_ctx_get(ictx);
> + user = iommufd_access_create(vdev->iommufd_ictx,
> &vfio_user_ops, vdev);
> + if (IS_ERR(user)) {
> + iommufd_ctx_put(vdev->iommufd_ictx);
Matthew noticed a vfio-ccw and -ap regression that blames this patch.
Probably both the iommufd_access_create() and iommufd_ctx_put() calls
want the ictx variable itself, instead of the (uninitialized) pointer
in the vfio_device. (At least that gets -ccw and -ap working again.)
Thanks,
Eric
> + return PTR_ERR(user);
> + }
> + iommufd_access_set_ioas(user, 0);
> + vdev->iommufd_access = user;
> + vdev->iommufd_ictx = ictx;
> return 0;
> }
> EXPORT_SYMBOL_GPL(vfio_iommufd_emulated_bind);