Re: [PATCH v2] drm/vmwgfx: Treat zero SVGA_REG_CURSOR_MAX_DIMENSION as unset

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

Pre-existing issues:
- [Critical] Integer overflows in framebuffer creation and cursor command submission cause a Guest-to-Host Out-Of-Bounds Read.
- [High] Missing Error Handling / NULL Pointer Dereference in `vmw_cursor_plane_prepare_fb`.
- [High] State Confusion in `vmw_cursor_plane_atomic_check` allows bypassing legacy cursor dimension checks, leading to a massive kernel heap Out-Of-Bounds read.
--

--- Patch [1]: [PATCH v2] drm/vmwgfx: Treat zero SVGA_REG_CURSOR_MAX_DIMENSION as unset ---
commit fc08a38eba647596145d861cab4214687c5ce5ce
Author: fausten <[email protected]>

drm/vmwgfx: Treat zero SVGA_REG_CURSOR_MAX_DIMENSION as unset

This commit treats a reported SVGA_REG_CURSOR_MAX_DIMENSION of 0 as "not
implemented" and skips the cursor dimension check in that case. This restores
the pre-existing behaviour for VMware Fusion guests (SVGA version 2) where the
cursor previously did not show up.

> diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_cursor_plane.c b/drivers/gpu/drm/vmwgfx/vmwgfx_cursor_plane.c
> index d1e7df5001903..fbdd23ecb2b4b 100644
> --- a/drivers/gpu/drm/vmwgfx/vmwgfx_cursor_plane.c
> +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_cursor_plane.c
[ ... ]
> @@ -748,8 +750,10 @@ int vmw_cursor_plane_atomic_check(struct drm_plane *plane,
>  		u32 cursor_max_dim =
>  			vmw_read(vmw, SVGA_REG_CURSOR_MAX_DIMENSION);
>  
> -		if (new_state->crtc_w > cursor_max_dim ||
> -		    new_state->crtc_h > cursor_max_dim) {
> +		/* Some SVGA implementations (e.g. VMware Fusion) report 0 here. */
> +		if (cursor_max_dim &&
> +		    (new_state->crtc_w > cursor_max_dim ||
> +		    new_state->crtc_h > cursor_max_dim)) {

[Severity: High]
This isn't a bug introduced by this patch, but does the cursor plane code handle
BO mapping failures safely?

In drivers/gpu/drm/vmwgfx/vmwgfx_cursor_plane.c:vmw_cursor_plane_prepare_fb(),
the return value of vmw_bo_map_and_cache_size() is cast to void, ignoring
potential mapping failures:

    (void)vmw_bo_map_and_cache_size(bo, size);

If this mapping fails, bo->map.virtual remains NULL. Later, in
vmw_cursor_plane_atomic_update(), vmw_send_define_cursor_cmd() unconditionally
copies to the command buffer using the unvalidated image pointer:

    memcpy(&cmd[1], image, image_size);

Could this cause a NULL pointer dereference and kernel panic if memory
pressure causes the mapping to fail?

[Severity: High]
This is a pre-existing issue, but does skipping this dimension check expose
a state confusion bug that could lead to an out-of-bounds read?

In drivers/gpu/drm/vmwgfx/vmwgfx_cursor_plane.c:vmw_cursor_plane_atomic_check(),
update_type is evaluated using vmw_cursor_update_type(vmw, vps). At this point,
vps->uo contains stale state from the previous commit, as it isn't updated until
vmw_cursor_plane_prepare_fb().

If userspace commits a BO-backed cursor (GB_ONLY), and then in a subsequent
commit submits a Surface-backed cursor (LEGACY) with huge dimensions
(e.g., 512x512), atomic_check will evaluate the type as GB_ONLY.

Since this patch skips the dimension check when cursor_max_dim == 0,
the huge dimensions are accepted. When prepare_fb() runs, vps->uo is updated
and the type correctly becomes LEGACY.

Finally, in vmw_send_define_cursor_cmd(), the memcpy() reads width * height * 4
bytes from the 16KB legacy surface->snooper.image buffer:

    memcpy(&cmd[1], image, image_size);

Could this result in a massive out-of-bounds read of the kernel heap being
sent to the hypervisor?

[Severity: Critical]
This is a pre-existing issue, but does bypassing this check when
cursor_max_dim == 0 expose integer overflows that can cause a guest-to-host
out-of-bounds read?

If userspace creates a framebuffer with massive dimensions (e.g., 32768x32768),
the calculation in drivers/gpu/drm/vmwgfx/vmwgfx_kms.c:vmw_kms_new_framebuffer_bo()
overflows 32 bits:

    requested_size = mode_cmd->height * mode_cmd->pitches[0];
    if (unlikely(requested_size > bo->tbo.base.size)) {

This overflow to 0 allows creating a massive framebuffer with a 1-page BO.
Because this patch removes the dimension limit for cursor_max_dim == 0,
the large cursor dimensions are accepted.

When vmw_send_define_cursor_cmd() prepares the command:

    const u32 image_size = width * height * sizeof(*image);
    const u32 cmd_size = sizeof(*cmd) + image_size;

image_size overflows to 0, resulting in only sizeof(*cmd) bytes being allocated
in the command FIFO. The hypervisor receives the command specifying a
32768x32768 cursor but no pixel data follows in the FIFO.

Will this cause the hypervisor to read 4GB out-of-bounds from the command FIFO?

>  			drm_warn(&vmw->drm,
>  				 "Cursor dimensions (%d, %d) exceed device max %u\n",
>  				 new_state->crtc_w, new_state->crtc_h,

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1