2020-10-23 16:36:28

by Ville Syrjälä

[permalink] [raw]
Subject: Re: [Intel-gfx] drm_modes: signed integer overflow

On Fri, Oct 23, 2020 at 03:14:20PM +0000, Simon Ser wrote:
> On Thursday, October 22, 2020 12:14 PM, Ville Syrj?l? <[email protected]> wrote:
>
> > On Wed, Oct 21, 2020 at 08:13:43PM -0700, Randy Dunlap wrote:
> >
> > > Hi,
> > > With linux-next 20201021, when booting up, I am seeing this:
> > > [ 0.560896] UBSAN: signed-integer-overflow in ../drivers/gpu/drm/drm_modes.c:765:20
> > > [ 0.560903] 2376000 * 1000 cannot be represented in type 'int'
> >
> > Dang. Didn't realize these new crazy >8k modes have dotclocks reaching
> > almost 6 GHz, which would overflow even u32. I guess we'll switch to
> > 64bit maths. Now I wonder how many other places can hit this overflow
> > in practice...
>
> Can you provide an example of a full crazy >8k mode?

These are two extreme cases:
/* 216 - 10240x4320@100Hz 64:27 */
{ DRM_MODE("10240x4320", DRM_MODE_TYPE_DRIVER, 5940000, 10240, 12432,
12608, 13200, 0, 4320, 4336, 4356, 4500, 0,
DRM_MODE_FLAG_PHSYNC | DRM_MODE_FLAG_PVSYNC),
.picture_aspect_ratio = HDMI_PICTURE_ASPECT_64_27, },
/* 217 - 10240x4320@120Hz 64:27 */
{ DRM_MODE("10240x4320", DRM_MODE_TYPE_DRIVER, 5940000, 10240, 10528,
10704, 11000, 0, 4320, 4336, 4356, 4500, 0,
DRM_MODE_FLAG_PHSYNC | DRM_MODE_FLAG_PVSYNC),
.picture_aspect_ratio = HDMI_PICTURE_ASPECT_64_27, }

> I'm trying to get
> a fix for my user-space [1], and I'm wondering if int32_t is enough
> after dividing by mode->htotal.
>
> CC Pekka, just FYI (I think Weston has similar code).
>
> [1]: https://github.com/swaywm/wlroots/pull/2450

What's with those 1000000LL constants? Are you storing
clock in Hz units?

--
Ville Syrj?l?
Intel


2020-10-23 16:38:48

by Simon Ser

[permalink] [raw]
Subject: Re: [Intel-gfx] drm_modes: signed integer overflow

On Friday, October 23, 2020 5:27 PM, Ville Syrjälä <[email protected]> wrote:

> These are two extreme cases:

Thanks!

> > I'm trying to get
> > a fix for my user-space 1, and I'm wondering if int32_t is enough
> > after dividing by mode->htotal.
> > CC Pekka, just FYI (I think Weston has similar code).
>
> What's with those 1000000LL constants? Are you storing
> clock in Hz units?

We're storing the vertical refresh rate in mHz (milli-Hertz).