Re: [PATCH v2] fbdev: vfb: Validate framebuffer geometry and memory size

[Helge Deller]

On 10/26/26 16:46, Weiming Shi wrote:
> FBIOPUT_VSCREENINFO can supply a virtual width large enough to
> overflow get_line_length()'s int multiplication and bypass the
> video-memory limit, allowing fbcon to write past the framebuffer. The
> offset additions can also wrap, while an excessive videomemorysize
> module parameter can be truncated when PAGE_ALIGN() is assigned to
> unsigned int.
>
> Limit virtual geometry and offsets to 8192 x 8192 and cap video
> memory at 512 MiB, keeping the stride, framebuffer-size, and generic
> fbdev drawing calculations within their destination types.
>
> KASAN reports:
>
>    BUG: KASAN: vmalloc-out-of-bounds in sys_fillrect
>      drivers/video/fbdev/core/sysfillrect.c:24
>    Write of size 8 at addr ffa0000001671000 by task poc/198
>    sys_fillrect
>      drivers/video/fbdev/core/sysfillrect.c:24
>    bit_clear_margins
>      drivers/video/fbdev/core/bitblit.c:247
>    fbcon_clear_margins.constprop.0
>      drivers/video/fbdev/core/fbcon.c:1410
>    fbcon_switch
>      drivers/video/fbdev/core/fbcon.c:2261
>    redraw_screen
>      drivers/tty/vt/vt.c:994
>    fbcon_modechanged
>      drivers/video/fbdev/core/fbcon.c:2693
>    fbcon_update_vcs
>      drivers/video/fbdev/core/fbcon.c:2732
>
> Cc: stable@xxxxxxxxxxxxxxx
> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> Reported-by: <co+b0df61d49532b2ee@xxxxxxx>
> Suggested-by: Helge Deller <deller@xxxxxx>
> Assisted-by: LLM
> Signed-off-by: Weiming Shi <bestswngs@xxxxxxxxx>
> ---
> Changes in v2:
> - Limit virtual geometry and offsets to 8192 x 8192. This retains 8K mode
>    support while bounding all generic fbdev drawing calculations.
> - Cap videomemorysize at 512 MiB.
> - Keep the existing arithmetic types now that all inputs are bounded.
>
>   drivers/video/fbdev/vfb.c | 17 ++++++++++++++++-
>   1 file changed, 16 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/video/fbdev/vfb.c b/drivers/video/fbdev/vfb.c
> index de137b2bd..de40ba8c3 100644
> --- a/drivers/video/fbdev/vfb.c
> +++ b/drivers/video/fbdev/vfb.c
> @@ -31,6 +31,9 @@
>        */
>   
>   #define VIDEOMEMSIZE	(1*1024*1024)	/* 1 MB */
> +#define MAX_VIDEOMEMSIZE	(512 * 1024 * 1024)
> +#define MAX_XRES		8192
> +#define MAX_YRES		8192
>   
>   static void *videomemory;
>   static u_long videomemorysize = VIDEOMEMSIZE;
> @@ -159,6 +162,14 @@ static int vfb_check_var(struct fb_var_screeninfo *var,
>   	else
>   		return -EINVAL;
>   
> +	if (var->xres_virtual > MAX_XRES ||
> +	    var->yres_virtual > MAX_YRES)
> +		return -EINVAL;
> +
> +	if (var->xoffset > MAX_XRES - var->xres ||
> +	    var->yoffset > MAX_YRES - var->yres)
> +		return -EINVAL;

Instead of returning EINVAL, I wonder if we could simply
cap the resolutions at MAX_XRES/MAX_YRES?


> +
>   	if (var->xres_virtual < var->xoffset + var->xres)
>   		var->xres_virtual = var->xoffset + var->xres;
>   	if (var->yres_virtual < var->yoffset + var->yres)
> @@ -434,9 +445,13 @@ static int __init vfb_setup(char *options)
>   static int vfb_probe(struct platform_device *dev)
>   {
>   	struct fb_info *info;
> -	unsigned int size = PAGE_ALIGN(videomemorysize);
> +	unsigned int size;
>   	int retval = -ENOMEM;
>   

and here maybe
	videomemorysize = min(videomemorysize, MAX_VIDEOMEMSIZE);

that way the driver loads even if the provided too big value?

Helge
> +	if (videomemorysize > MAX_VIDEOMEMSIZE)
> +		return -EINVAL;
> +	size = PAGE_ALIGN(videomemorysize);