Re: [PATCH] thermal: core: Move initial num_trips assignment before memcpy()

[Kees Cook <keescook@chromium.org>]

On Tue, Feb 27, 2024 at 04:37:36PM +0100, Daniel Lezcano wrote:
> On 27/02/2024 12:09, Rafael J. Wysocki wrote:
> > On Tue, Feb 27, 2024 at 11:14 AM Daniel Lezcano
> > <daniel.lezcano@linaro.org> wrote:
> > > 
> > > On 27/02/2024 01:54, Nathan Chancellor wrote:
> > > > When booting a CONFIG_FORTIFY_SOURCE=y kernel compiled with a toolchain
> > > > that supports __counted_by() (such as clang-18 and newer), there is a
> > > > panic on boot:
> > > > 
> > > >     [    2.913770] memcpy: detected buffer overflow: 72 byte write of buffer size 0
> > > >     [    2.920834] WARNING: CPU: 2 PID: 1 at lib/string_helpers.c:1027 __fortify_report+0x5c/0x74
> > > >     ...
> > > >     [    3.039208] Call trace:
> > > >     [    3.041643]  __fortify_report+0x5c/0x74
> > > >     [    3.045469]  __fortify_panic+0x18/0x20
> > > >     [    3.049209]  thermal_zone_device_register_with_trips+0x4c8/0x4f8
> > > > 
> > > > This panic occurs because trips is counted by num_trips but num_trips is
> > > > assigned after the call to memcpy(), so the fortify checks think the
> > > > buffer size is zero because tz was allocated with kzalloc().
> > > > 
> > > > Move the num_trips assignment before the memcpy() to resolve the panic
> > > > and ensure that the fortify checks work properly.
> > > > 
> > > > Fixes: 9b0a62758665 ("thermal: core: Store zone trips table in struct thermal_zone_device")
> > > > Signed-off-by: Nathan Chancellor <nathan@kernel.org>
> > > > ---
> > > >    drivers/thermal/thermal_core.c | 2 +-
> > > >    1 file changed, 1 insertion(+), 1 deletion(-)
> > > > 
> > > > diff --git a/drivers/thermal/thermal_core.c b/drivers/thermal/thermal_core.c
> > > > index bb21f78b4bfa..1eabc8ebe27d 100644
> > > > --- a/drivers/thermal/thermal_core.c
> > > > +++ b/drivers/thermal/thermal_core.c
> > > > @@ -1354,8 +1354,8 @@ thermal_zone_device_register_with_trips(const char *type,
> > > > 
> > > >        tz->device.class = thermal_class;
> > > >        tz->devdata = devdata;
> > > > -     memcpy(tz->trips, trips, num_trips * sizeof(*trips));
> > > >        tz->num_trips = num_trips;
> > > > +     memcpy(tz->trips, trips, num_trips * sizeof(*trips));
> > > 
> > > IIUC, clang-18 is used and supports __counted_by().
> > > 
> > > Is it possible sizeof(*trips) returns already the real trips array size
> > > and we are multiplying it again by num_trips ?
> > > 
> > > While with an older compiler, __counted_by() does nothing and we have to
> > > multiply by num_trips ?
> > > 
> > > IOW, the array size arithmetic is different depending if we have
> > > _counted_by supported or not ?
> > 
> > IIUC it is just the instrumentation using the current value of
> > tz->num_trips (which is 0 before the initialization).
> 
> Right, but I am wondering if
> 
> 	memcpy(tz->trips, trips, num_trips * sizeof(*trips));
> 
> 	is still correct with __counted_by because:
> 
>  (1) if the compiler supports it:
> 
> 	sizeof(*trips) == 24 bytes * num_trips

I think you're misunderstanding. The above sizeof() only evaluates a
single instance -- it has no idea how many more there may be.
Specifically:

	sizeof(*trips) == sizeof(struct thermal_trip)

> 	then:
> 
> 	memcpy(tz->trips, trips, num_trips * sizeof(*trips));
> 
> 	memcpy(tz->trips, trips, num_trips * 24 * num_trips);
> 
> 	==> memory size = 24 * num_trips^2

It's not counted twice. Under CONFIG_FORTIFY_SOURCE=y, memcpy is a macro
that expands to a checking routine (see include/linux/fortify-string.h),
which is using __builtin_dynamic_object_size() to determine the
available size of the destination buffer (tz->trips). Specifically:

	__builtin_dynamic_object_size(tz->trips)

When __bdos evaluates a flexible array (i.e. tz->trips), it will see the
associated 'counted_by' attribute, and go look up the value of the
designated struct member (tz->num_trips). It then calculates:

	sizeof(*tz->trips) /* a single instance */
		*
	tz->num_trips

Before the patch, tz->num_trips is 0, so the destination buffer size
appears to be of size 0 bytes. After the patch, it contains the
same value as the "num_trips" function argument, so the destination
buffer appears to be the matching size of "num_trips * sizeof(struct
thermal_trip)".

Hopefully that helps! If not, I can try again. :)

-Kees

-- 
Kees Cook