Re: [PATCH] hyperv: mshv: zero VTL hypercall output page

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] This problem wasn't introduced by this patch, but the patch only partially fixes the information leak. While it zeroes the `out` page, the input bounce page `in` is still allocated without the `__GFP_ZERO` flag. If a user provides an `hvcall.input_size` that is smaller than the input size expected by the hypercall (specified by `hvcall.control`), `copy_from_user()` will only partially initialize the `in` page. The remainder of the page will contain uninitialized, stale kernel memory. This full page is then passed to the hypervisor. If the hypervisor processes these uninitialized bytes and reflects them into the `out` page or alters its observable behavior based on them, sensitive kernel data is leaked to userspace.
--

commit d8767fa45f0b927a8630a3fee3fba738b9ae7851
Author: Yousef Alhouseen <alhouseenyousef@gmail.com>

hyperv: mshv: zero VTL hypercall output page
    
This patch zeroes the hypercall output page to avoid copying stale memory
to userspace and adds NULL pointer checks for the page allocations.

> diff --git a/drivers/hv/mshv_vtl_main.c b/drivers/hv/mshv_vtl_main.c
> index 0d3d4161974f8..0365d207cb110 100644
> --- a/drivers/hv/mshv_vtl_main.c
> +++ b/drivers/hv/mshv_vtl_main.c
> @@ -1147,7 +1147,11 @@ static int mshv_vtl_hvcall_call(struct mshv_vtl_hvcall_fd *fd,
>  	 * TODO: Take care of this when CVM support is added.
>  	 */
>  	in = (void *)__get_free_page(GFP_KERNEL);

[Severity: High]
This isn't a bug introduced by this patch, but should the "in" page also
be allocated with __GFP_ZERO to prevent leaking stale kernel memory?

If userspace provides an hvcall.input_size that is smaller than what the
hypercall expects, copy_from_user() will only partially initialize the
"in" page. The remainder of the page will contain uninitialized kernel
memory which is then passed to the hypervisor in hv_do_hypercall(). If the
hypervisor reflects this input or alters behavior based on it, could this
result in an information leak?

> -	out = (void *)__get_free_page(GFP_KERNEL);
> +	out = (void *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
> +	if (!in || !out) {
> +		ret = -ENOMEM;
> +		goto free_pages;
> +	}

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260624172157.2790-1-alhouseenyousef@gmail.com?part=1