Re: [PATCH v19 15/18] vfio/pci: Implement TPH_RESOLVE feature for DMABUF and CPU source resolving

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 3 potential issue(s) to consider:

New issues:
- [High] Flawed bitwise logic in flag validation allows userspace to bypass TPH policy restrictions and breaks forward UAPI compatibility.

Pre-existing issues:
- [High] Missing reservation lock `dma_resv_lock` before calling `vfio_pci_dma_buf_get_pci_tph` leads to a data race and lockdep assertion failure.
- [Medium] The VFIO_DEVICE_FEATURE_TPH SET operation fails to validate that the provided payload flags are zero, breaking forward compatibility.
--

commit 0733cb5339faa744645461c3f2055477418d1f5e
Author: Chengwen Feng <fengchengwen@huawei.com>

vfio/pci: Implement TPH_RESOLVE feature for DMABUF and CPU source resolving

This commit adds vfio_pci_core_feature_tph_resolve() to implement the
VFIO_DEVICE_FEATURE_TPH_RESOLVE GET-only uAPI. It resolves PH/ST values 
from two supported sources: DMABUF fds and CPU IDs, dynamically gating 
source types based on the tph_policy level.

> diff --git a/drivers/vfio/pci/vfio_pci_core.c b/drivers/vfio/pci/vfio_pci_core.c
> index 14944d3ea86e6..ed6bbffa8b26a 100644
> --- a/drivers/vfio/pci/vfio_pci_core.c
> +++ b/drivers/vfio/pci/vfio_pci_core.c
> @@ -1641,6 +1641,78 @@ static int vfio_pci_core_feature_tph(struct vfio_pci_core_device *vdev,
>  	return copy_to_user(arg, &tph, sizeof(tph)) ? -EFAULT : 0;
>  }

[Severity: Medium]
This is a pre-existing issue in vfio_pci_core_feature_tph() just above this
new code, but does the SET operation properly validate the user payload?

When VFIO_DEVICE_FEATURE_SET is passed, the function exits early without
reading the user's struct vfio_device_feature_tph payload via 
copy_from_user():

vfio_pci_core_feature_tph() {
    if (flags & VFIO_DEVICE_FEATURE_SET) {
        vdev->tph_permit = 1;
        return 0;
    }
}

Does this silently ignore any non-zero flags in the payload, potentially
breaking UAPI forward compatibility if new flags are added later?

> +static int vfio_pci_core_feature_tph_resolve(struct vfio_pci_core_device *vdev,
> +			u32 flags,
> +			struct vfio_device_feature_tph_resolve __user *arg,
> +			size_t argsz)
> +{

[ ... ]

> +	if (copy_from_user(&resolve, arg, sizeof(resolve)))
> +		return -EFAULT;
> +
> +	src_bits = resolve.flags & VFIO_DEVICE_TPH_SRC_MASK;
> +	if (!(resolve.flags & permit_flags) || !is_power_of_2(src_bits))
> +		return -EINVAL;

[Severity: High]
Does this validation logic correctly reject unsupported flags?

Since permit_flags unconditionally includes VFIO_DEVICE_TPH_EXTENDED,
wouldn't any user payload that sets VFIO_DEVICE_TPH_EXTENDED along with
an unpermitted flag bypass this check entirely? The bitwise AND would be
non-zero, making the NOT condition false and bypassing the policy rejection.

Could this be updated to use a strict subset check instead? For example:
    if ((resolve.flags & ~permit_flags) || !is_power_of_2(src_bits))

> +	extended = !!(resolve.flags & VFIO_DEVICE_TPH_EXTENDED);
> +	if (extended && !pcie_tph_supported(vdev->pdev, true))
> +		return -EINVAL;
> +
> +	resolve.valid = 0;
> +	resolve.ph = 0;
> +	resolve.st = 0;
> +
> +	if (src_bits & VFIO_DEVICE_TPH_SRC_DMABUF) {
> +		ret = vfio_pci_dma_buf_get_tph_by_fd(resolve.src, extended,
> +						     &tag, &ph);

[Severity: High]
This isn't a bug introduced by this patch, but does calling 
vfio_pci_dma_buf_get_tph_by_fd() here trigger a lockdep splat?

Looking at vfio_pci_dma_buf_get_tph_by_fd() in
drivers/vfio/pci/vfio_pci_dmabuf.c, it delegates directly to
vfio_pci_dma_buf_get_pci_tph() without acquiring the dma_resv_lock:

vfio_pci_dma_buf_get_tph_by_fd() {
    dmabuf = dma_buf_get(fd);
    ...
    ret = vfio_pci_dma_buf_get_pci_tph(dmabuf, extended, tag, ph);
}

However, vfio_pci_dma_buf_get_pci_tph() expects the reservation lock
to be held, and asserts it:

vfio_pci_dma_buf_get_pci_tph() {
    dma_resv_assert_held(priv->dmabuf->resv);
}

Could this lead to unprotected concurrent reads of the priv struct fields
and a data race if the lock is not taken?

[ ... ]

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260702124224.57168-1-fengchengwen@huawei.com?part=15