On Thu, Jul 29, 2021 at 11:00:48PM -0700, Kees Cook wrote:
> On Thu, Jul 29, 2021 at 11:20:39AM +0300, Dan Carpenter wrote:
> > On Thu, Jul 29, 2021 at 07:56:27AM +0200, Greg Kroah-Hartman wrote:
> > > On Wed, Jul 28, 2021 at 11:37:30PM +0200, David Sterba wrote:
> > > > On Wed, Jul 28, 2021 at 02:37:20PM -0700, Bart Van Assche wrote:
> > > > > On 7/28/21 2:14 AM, Dan Carpenter wrote:
> > > > > > On Wed, Jul 28, 2021 at 10:59:22AM +0200, David Sterba wrote:
> > > > > >>> drivers/media/platform/omap3isp/ispstat.c | 5 +--
> > > > > >>> include/uapi/linux/omap3isp.h | 44 +++++++++++++++++------
> > > > > >>> 2 files changed, 36 insertions(+), 13 deletions(-)
> > > > > >>>
> > > > > >>> diff --git a/drivers/media/platform/omap3isp/ispstat.c b/drivers/media/platform/omap3isp/ispstat.c
> > > > > >>> index 5b9b57f4d9bf..ea8222fed38e 100644
> > > > > >>> --- a/drivers/media/platform/omap3isp/ispstat.c
> > > > > >>> +++ b/drivers/media/platform/omap3isp/ispstat.c
> > > > > >>> @@ -512,7 +512,7 @@ int omap3isp_stat_request_statistics(struct ispstat *stat,
> > > > > >>> int omap3isp_stat_request_statistics_time32(struct ispstat *stat,
> > > > > >>> struct omap3isp_stat_data_time32 *data)
> > > > > >>> {
> > > > > >>> - struct omap3isp_stat_data data64;
> > > > > >>> + struct omap3isp_stat_data data64 = { };
> > > > > >>
> > > > > >> Should this be { 0 } ?
> > > > > >>
> > > > > >> We've seen patches trying to switch from { 0 } to { } but the answer
> > > > > >> was that { 0 } is supposed to be used,
> > > > > >> http://www.ex-parrot.com/~chris/random/initialise.html
> > > > > >>
> > > > > >> (from https://lore.kernel.org/lkml/fbddb15a-6e46-3f21-23ba-b18f66e3448a@xxxxxxxx/ )
> > > > > >
> > > > > > In the kernel we don't care about portability so much. Use the = { }
> > > > > > GCC extension. If the first member of the struct is a pointer then
> > > > > > Sparse will complain about = { 0 }.
> > > > >
> > > > > +1 for { }.
> > > >
> > > > Oh, I thought the tendency is is to use { 0 } because that can also
> > > > intialize the compound members, by a "scalar 0" as it appears in the
> > > > code.
> > > >
> > >
> > > Holes in the structure might not be initialized to anything if you do
> > > either one of these as well.
> > >
> > > Or did we finally prove that is not the case? I can not remember
> > > anymore...
> >
> > Yep. The C11 spec says that struct holes are initialized.
> >
> > https://lore.kernel.org/netdev/20200731140452.GE24045@xxxxxxxx/
>
> This is, unfortunately, misleading. The frustrating key word is
> "partial" in "updated in C11 to require zero'ing padding when doing
> partial initialization of aggregates". If one initializes _all_ the
> struct members ... the padding doesn't get initialized. :( (And until
> recently, _trailing_ padding wasn't getting initialized even when other
> paddings were.)
>
> I've tried to collect all the different ways the compiler might initialize
> a variable in this test:
> https://git.kernel.org/pub/scm/linux/kernel/git/kees/linux.git/tree/lib/test_stackinit.c?h=for-next/kspp
>
> FWIW, there's no difference between -std=gnu99 and -std=c11, and the
> test shows that padding is _not_ universally initialized (unless your
> compiler supports -ftrivial-auto-var-init=zero, which Clang does, and
> GCC will shortly[1]). Running this with GCC 10.3.0, I see this...
>
> As expected, having no initializer leaves padding (as well as members)
> uninitialized:
>
> stackinit: small_hole_none FAIL (uninit bytes: 24)
> stackinit: big_hole_none FAIL (uninit bytes: 128)
> stackinit: trailing_hole_none FAIL (uninit bytes: 32)
>
> Here, "zero" means "= { };" and they get padding initialized:
>
> stackinit: small_hole_zero ok
> stackinit: big_hole_zero ok
> stackinit: trailing_hole_zero ok
>
> Here, "static_partial" means "= { .one_member = 0 };", and
> "dynamic_partial" means "= { .one_member = some_variable };". These are
> similarly initialized:
>
> stackinit: small_hole_static_partial ok
> stackinit: big_hole_static_partial ok
> stackinit: trailing_hole_static_partial ok
>
> stackinit: small_hole_dynamic_partial ok
> stackinit: big_hole_dynamic_partial ok
> stackinit: trailing_hole_dynamic_partial ok
>
> But when _all_ members are initialized, the padding is _not_:
>
> stackinit: small_hole_static_all FAIL (uninit bytes: 3)
> stackinit: big_hole_static_all FAIL (uninit bytes: 124)
> stackinit: trailing_hole_static_all FAIL (uninit bytes: 7)
>
> stackinit: small_hole_dynamic_all FAIL (uninit bytes: 3)
> stackinit: big_hole_dynamic_all FAIL (uninit bytes: 124)
> stackinit: trailing_hole_dynamic_all FAIL (uninit bytes: 7)
>
> As expected, assigning to members outside of initialization leaves
> padding uninitialized:
>
> stackinit: small_hole_runtime_partial FAIL (uninit bytes: 23)
> stackinit: big_hole_runtime_partial FAIL (uninit bytes: 127)
> stackinit: trailing_hole_runtime_partial FAIL (uninit bytes: 24)
>
> stackinit: small_hole_runtime_all FAIL (uninit bytes: 3)
> stackinit: big_hole_runtime_all FAIL (uninit bytes: 124)
> stackinit: trailing_hole_runtime_all FAIL (uninit bytes: 7)
>
> > What doesn't initialize struct holes is assignments:
> >
> > struct foo foo = *bar;
>
> Right. Object to object assignments do not clear padding:
>
> stackinit: small_hole_assigned_copy XFAIL (uninit bytes: 3)
> stackinit: big_hole_assigned_copy XFAIL (uninit bytes: 124)
> stackinit: trailing_hole_assigned_copy XFAIL (uninit bytes: 7)
>
> And whole-object assignments of cast initializers follow the pattern of
> basic initializers, which makes sense given the behavior of initializers
> and direct assignment tests above. e.g.:
> obj = (type){ .member = ... };
>
> stackinit: small_hole_assigned_static_partial ok
> stackinit: small_hole_assigned_dynamic_partial ok
> stackinit: big_hole_assigned_dynamic_partial ok
> stackinit: big_hole_assigned_static_partial ok
> stackinit: trailing_hole_assigned_dynamic_partial ok
> stackinit: trailing_hole_assigned_static_partial ok
>
> stackinit: small_hole_assigned_static_all FAIL (uninit bytes: 3)
> stackinit: small_hole_assigned_dynamic_all FAIL (uninit bytes: 3)
> stackinit: big_hole_assigned_static_all FAIL (uninit bytes: 124)
> stackinit: big_hole_assigned_dynamic_all FAIL (uninit bytes: 124)
> stackinit: trailing_hole_assigned_dynamic_all FAIL (uninit bytes: 7)
> stackinit: trailing_hole_assigned_static_all FAIL (uninit bytes: 7)
>
> So, yeah, it's not very stable.
Then is explicit memset the only reliable way accross all compiler
flavors and supported versions?
E.g. for ioctls that get kernel memory (stack, kmalloc), partially
initialize it and then call copy_to_user.
