Message-ID: <bbc65508-eb0e-4d63-921b-85d242cc556f@linaro.org>
Date: Tue, 27 Feb 2024 17:47:44 +0100
From: Daniel Lezcano <daniel.lezcano@...aro.org>
To: Kees Cook <keescook@...omium.org>
Cc: "Rafael J. Wysocki" <rafael@...nel.org>,
 Nathan Chancellor <nathan@...nel.org>, rui.zhang@...el.com,
 lukasz.luba@....com, gustavoars@...nel.org, morbo@...gle.com,
 justinstitt@...gle.com, stanislaw.gruszka@...ux.intel.com,
 linux-pm@...r.kernel.org, linux-hardening@...r.kernel.org,
 llvm@...ts.linux.dev, patches@...ts.linux.dev
Subject: Re: [PATCH] thermal: core: Move initial num_trips assignment before
 memcpy()

On 27/02/2024 17:26, Kees Cook wrote:
> 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@...aro.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@...nel.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

Ok my misunderstanding was I thought sizeof() was calling _bdos under 
the hood, so when calling sizeof(flex_array), it was returning the 
computed size inferring from the __counted_by field.


> 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. :)

It is ok thanks for the clarification



-- 
<http://www.linaro.org/> Linaro.org │ Open source software for ARM SoCs

Follow Linaro:  <http://www.facebook.com/pages/Linaro> Facebook |
<http://twitter.com/#!/linaroorg> Twitter |
<http://www.linaro.org/linaro-blog/> Blog

Powered by blists - more mailing lists