Nouveau - May 2024 - [PATCH] drm/nouveau/nvif: Avoid build error due to potential integer overflows

On 5/18/24 10:32, Kees Cook wrote:
> On Sat, May 18, 2024 at 06:54:36PM +0200, Christophe JAILLET wrote:
>> (adding linux-hardening at vger.kernel.org)
>>
>>
>> Le 18/05/2024 ? 16:37, Guenter Roeck a ?crit?:
>>> Trying to build parisc:allmodconfig with gcc 12.x or later results
>>> in the following build error.
>>>
>>> drivers/gpu/drm/nouveau/nvif/object.c: In function
'nvif_object_mthd':
>>> drivers/gpu/drm/nouveau/nvif/object.c:161:9: error:
>>> 	'memcpy' accessing 4294967264 or more bytes at offsets 0
and 32 overlaps 6442450881 bytes at offset -2147483617 [-Werror=restrict]
>>>     161 |         memcpy(data, args->mthd.data, size);
>>>         |         ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>> drivers/gpu/drm/nouveau/nvif/object.c: In function
'nvif_object_ctor':
>>> drivers/gpu/drm/nouveau/nvif/object.c:298:17: error:
>>> 	'memcpy' accessing 4294967240 or more bytes at offsets 0
and 56 overlaps 6442450833 bytes at offset -2147483593 [-Werror=restrict]
>>>     298 |                 memcpy(data, args->new.data, size);
>>>
>>> gcc assumes that 'sizeof(*args) + size' can overflow, which
would result
>>> in the problem.
>>>
>>> The problem is not new, only it is now no longer a warning but an
error since W=1
>>> has been enabled for the drm subsystem and since Werror is enabled
for test builds.
>>>
>>> Rearrange arithmetic and add extra size checks to avoid the
overflow.
>>>
>>> Fixes: a61ddb4393ad ("drm: enable (most) W=1 warnings by
default across the subsystem")
>>> Cc: Javier Martinez Canillas <javierm-H+wXaHxf7aLQT0dZR+AlfA at
public.gmane.org>
>>> Cc: Jani Nikula <jani.nikula-ral2JQCrhuEAvxtiuMwx3w at
public.gmane.org>
>>> Cc: Thomas Zimmermann <tzimmermann-l3A5Bk7waGM at
public.gmane.org>
>>> Cc: Danilo Krummrich <dakr-H+wXaHxf7aLQT0dZR+AlfA at
public.gmane.org>
>>> Cc: Maxime Ripard <mripard-DgEjT+Ai2ygdnm+yROfE0A at
public.gmane.org>
>>> Signed-off-by: Guenter Roeck <linux-0h96xk9xTtrk1uMJSBkQmQ at
public.gmane.org>
>>> ---
>>> checkpatch complains about the line length in the description and
the (pre-existing)
>>> assignlemts in if conditions, but I did not want to split lines in
the description
>>> or rearrange the code further.
>>>
>>> I don't know why I only see the problem with parisc builds (at
least so far).
>>>
>>>    drivers/gpu/drm/nouveau/nvif/object.c | 8 +++++---
>>>    1 file changed, 5 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/nouveau/nvif/object.c
b/drivers/gpu/drm/nouveau/nvif/object.c
>>> index 4d1aaee8fe15..baf623a48874 100644
>>> --- a/drivers/gpu/drm/nouveau/nvif/object.c
>>> +++ b/drivers/gpu/drm/nouveau/nvif/object.c
>>> @@ -145,8 +145,9 @@ nvif_object_mthd(struct nvif_object *object,
u32 mthd, void *data, u32 size)
>>>    	u8 stack[128];
>>>    	int ret;
>>> -	if (sizeof(*args) + size > sizeof(stack)) {
>>> -		if (!(args = kmalloc(sizeof(*args) + size, GFP_KERNEL)))
>>> +	if (size > sizeof(stack) - sizeof(*args)) {
>>> +		if (size > INT_MAX ||
>>> +		    !(args = kmalloc(sizeof(*args) + size, GFP_KERNEL)))
>>
>> Hi,
>>
>> Would it be cleaner or better to use size_add(sizeof(*args), size)?
> 
> I think the INT_MAX test is actually better in this case because
> nvif_object_ioctl()'s size argument is u32:
> 
> ret = nvif_object_ioctl(object, args, sizeof(*args) + size, NULL);
>                                        ^^^^^^^^^^^^^^^^^^^^
> 
> So that could wrap around, even though the allocation may not.
> 
> Better yet, since "sizeof(*args) + size" is repeated 3 times in
the
> function, I'd recommend:
> 
> 	...
> 	u32 args_size;
> 
> 	if (check_add_overflow(sizeof(*args), size, &args_size))
> 		return -ENOMEM;
> 	if (args_size > sizeof(stack)) {
> 		if (!(args = kmalloc(args_size, GFP_KERNEL)))
> 			return -ENOMEM;
>          } else {
>                  args = (void *)stack;
>          }
> 	...
>          ret = nvif_object_ioctl(object, args, args_size, NULL);
> 
> This will catch the u32 overflow to nvif_object_ioctl(), catch an
> allocation underflow on 32-bits systems, and make the code more
> readable. :)
> 
Makes sense. I'll change that and send v2.

Thanks,
Guenter

On Sat, 2024-05-18 at 11:23 -0700, Guenter Roeck wrote:
> On 5/18/24 10:32, Kees Cook wrote:
> 
[]
> > I think the INT_MAX test is actually better in this case because
> > nvif_object_ioctl()'s size argument is u32:
> > 
> > ret = nvif_object_ioctl(object, args, sizeof(*args) + size, NULL);
> >                                        ^^^^^^^^^^^^^^^^^^^^
> > 
> > So that could wrap around, even though the allocation may not.
> > 
> > Better yet, since "sizeof(*args) + size" is repeated 3 times
in the
> > function, I'd recommend:
> > 
> > 	...
> > 	u32 args_size;
> > 
> > 	if (check_add_overflow(sizeof(*args), size, &args_size))
> > 		return -ENOMEM;
> > 	if (args_size > sizeof(stack)) {
> > 		if (!(args = kmalloc(args_size, GFP_KERNEL)))
trivia:

More typical kernel style would use separate alloc and test

		args = kmalloc(args_size, GFP_KERNEL);
		if (!args)
> > 			return -ENOMEM;
> >          } else {
> >                  args = (void *)stack;
> >          }
> > 	...
> >          ret = nvif_object_ioctl(object, args, args_size, NULL);
> > 
> > This will catch the u32 overflow to nvif_object_ioctl(), catch an
> > allocation underflow on 32-bits systems, and make the code more
> > readable. :)
> > 
> 
> Makes sense. I'll change that and send v2.
> 
> Thanks,
> Guenter
> 
>