Re: [PATCH] drm: check drm_format_info hsub and vsub to avoid divide by zero

From: George Kennedy
Date: Mon Nov 22 2021 - 10:30:25 EST

On 11/19/2021 9:25 AM, Jani Nikula wrote:
On Fri, 19 Nov 2021, Daniel Vetter <daniel@xxxxxxxx> wrote:
On Fri, Nov 19, 2021 at 12:03:00PM +0200, Ville Syrjälä wrote:
On Fri, Nov 19, 2021 at 10:40:38AM +0100, Daniel Vetter wrote:
On Thu, Oct 28, 2021 at 05:04:19PM +0300, Ville Syrjälä wrote:
On Thu, Oct 28, 2021 at 08:57:17AM -0500, George Kennedy wrote:
Do a sanity check on struct drm_format_info hsub and vsub values to
avoid divide by zero.

Syzkaller reported a divide error in framebuffer_check() when the
DRM_FORMAT_Q410 or DRM_FORMAT_Q401 pixel_format is passed in via
the DRM_IOCTL_MODE_ADDFB2 ioctl. The drm_format_info struct for
the DRM_FORMAT_Q410 pixel_pattern has ".hsub = 0" and ".vsub = 0".
fb_plane_width() uses hsub as a divisor and fb_plane_height() uses
vsub as a divisor. These divisors need to be sanity checked for
zero before use.

divide error: 0000 [#1] SMP KASAN NOPTI
CPU: 0 PID: 14995 Comm: syz-executor709 Not tainted 5.15.0-rc6-syzk #1
Hardware name: Red Hat KVM, BIOS 1.13.0-2
RIP: 0010:framebuffer_check drivers/gpu/drm/drm_framebuffer.c:199 [inline]
RIP: 0010:drm_internal_framebuffer_create+0x604/0xf90

Call Trace:
drm_mode_addfb2+0xdc/0x320 drivers/gpu/drm/drm_framebuffer.c:355
drm_mode_addfb2_ioctl+0x2a/0x40 drivers/gpu/drm/drm_framebuffer.c:391
drm_ioctl_kernel+0x23a/0x2e0 drivers/gpu/drm/drm_ioctl.c:795
drm_ioctl+0x589/0xac0 drivers/gpu/drm/drm_ioctl.c:898
vfs_ioctl fs/ioctl.c:51 [inline]
__do_sys_ioctl fs/ioctl.c:874 [inline]
__se_sys_ioctl fs/ioctl.c:860 [inline]
__x64_sys_ioctl+0x19d/0x220 fs/ioctl.c:860
do_syscall_x64 arch/x86/entry/common.c:50 [inline]
do_syscall_64+0x3a/0x80 arch/x86/entry/common.c:80

Signed-off-by: George Kennedy <george.kennedy@xxxxxxxxxx>
drivers/gpu/drm/drm_framebuffer.c | 10 ++++++++++
1 file changed, 10 insertions(+)

diff --git a/drivers/gpu/drm/drm_framebuffer.c b/drivers/gpu/drm/drm_framebuffer.c
index 07f5abc..a146e4b 100644
--- a/drivers/gpu/drm/drm_framebuffer.c
+++ b/drivers/gpu/drm/drm_framebuffer.c
@@ -195,6 +195,16 @@ static int framebuffer_check(struct drm_device *dev,
/* now let the driver pick its own format info */
info = drm_get_format_info(dev, r);
+ if (info->hsub == 0) {
+ DRM_DEBUG_KMS("bad horizontal chroma subsampling factor %u\n", info->hsub);
+ return -EINVAL;
+ }
+ if (info->vsub == 0) {
+ DRM_DEBUG_KMS("bad vertical chroma subsampling factor %u\n", info->vsub);
+ return -EINVAL;
+ }
Looks like duct tape to me. I think we need to either fix those formats
to have valid format info, or just revert the whole patch that added such
broken things.
Yeah maybe even a compile-time check of the format table(s) to validate
them properly and scream ... Or at least a selftest.
I really wish C had (even very limited) compile time evaluation
so one could actually loop over arrays like at compile time to
check each element. As it stands you either have to check each
array element by hand, or you do some cpp macro horrors to
pretend you're iterating the array.
Python preprocess or so seems to be the usual answer, and that then just
generates the C table after it's all checked.

Or a post-processor which fishes the table out from the .o (or just links
against it).

But yeah doing this in cpp isn't going to work, aside from it'd be really
Kbuild does have support for hostprogs which are typically used in the
build. The obvious idea is to use that for code generation, but it would
also be interesting to see how that could be used for compile-time
evaluation of sorts. Kind of like compile-time selftests? And, of
course, how badly that would be frowned upon.

git grep says there are only four hostprogs users in drivers/, so it
certainly isn't a popularity contest winner. (One of them is
"mkregtable" in radeon.)

So, can someone suggest a fix? A cpp type of approach does not seem feasible.

Adding the sanity checks that are in the patch, which are similar to the sanity checks preceding them in framebuffer_check(), along with a self-test that ran through all the table entries, might address all the concerns brought up in this thread.

Thank you,