Message-ID: <725ac392-642b-f57d-a286-d662eaa7d2a2@redhat.com>
Date:   Tue, 21 Sep 2021 17:38:39 +0200
From:   Hans de Goede <hdegoede@...hat.com>
To:     Greg KH <greg@...ah.com>
Cc:     Kees Cook <keescook@...omium.org>, Len Baker <len.baker@....com>,
        Henrique de Moraes Holschuh <hmh@....eng.br>,
        Mark Gross <mgross@...ux.intel.com>,
        "Gustavo A. R. Silva" <gustavoars@...nel.org>,
        ibm-acpi-devel@...ts.sourceforge.net,
        platform-driver-x86@...r.kernel.org,
        linux-hardening@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] platform/x86: thinkpad_acpi: Prefer struct_size over open
 coded arithmetic

Hi,

On 9/21/21 5:15 PM, Greg KH wrote:
> On Tue, Sep 21, 2021 at 03:46:23PM +0200, Hans de Goede wrote:
>> Hi,
>>
>> On 9/20/21 7:58 AM, Kees Cook wrote:
>>> On Sat, Sep 18, 2021 at 05:05:00PM +0200, Len Baker wrote:
>>>> As noted in the "Deprecated Interfaces, Language Features, Attributes,
>>>> and Conventions" documentation [1], size calculations (especially
>>>> multiplication) should not be performed in memory allocator (or similar)
>>>> function arguments due to the risk of them overflowing. This could lead
>>>> to values wrapping around and a smaller allocation being made than the
>>>> caller was expecting. Using those allocations could lead to linear
>>>> overflows of heap memory and other misbehaviors.
>>>>
>>>> So, switch to flexible array member in the struct attribute_set_obj and
>>>> refactor the code accordingly to use the struct_size() helper instead of
>>>> the argument "size + count * size" in the kzalloc() function.
>>>>
>>>> [1] https://www.kernel.org/doc/html/latest/process/deprecated.html#open-coded-arithmetic-in-allocator-arguments
>>>>
>>>> Signed-off-by: Len Baker <len.baker@....com>
>>>> ---
>>>>  drivers/platform/x86/thinkpad_acpi.c | 8 +++-----
>>>>  1 file changed, 3 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/drivers/platform/x86/thinkpad_acpi.c b/drivers/platform/x86/thinkpad_acpi.c
>>>> index 50ff04c84650..ed0b01ead796 100644
>>>> --- a/drivers/platform/x86/thinkpad_acpi.c
>>>> +++ b/drivers/platform/x86/thinkpad_acpi.c
>>>> @@ -1008,7 +1008,7 @@ struct attribute_set {
>>>>
>>>>  struct attribute_set_obj {
>>>>  	struct attribute_set s;
>>>> -	struct attribute *a;
>>>> +	struct attribute *a[];
>>>>  } __attribute__((packed));
>>>
>>> Whoa. I have so many questions... :)
>>>
>>>>
>>>>  static struct attribute_set *create_attr_set(unsigned int max_members,
>>>> @@ -1020,13 +1020,11 @@ static struct attribute_set *create_attr_set(unsigned int max_members,
>>>>  		return NULL;
>>>>
>>>>  	/* Allocates space for implicit NULL at the end too */
>>>> -	sobj = kzalloc(sizeof(struct attribute_set_obj) +
>>>> -		    max_members * sizeof(struct attribute *),
>>>> -		    GFP_KERNEL);
>>>> +	sobj = kzalloc(struct_size(sobj, a, max_members + 1), GFP_KERNEL);
>>>
>>> Whoa, this needs a lot more detail in the changelog if this is actually
>>> correct. The original code doesn't seem to match the comment? (Where is
>>> the +1?) So is this also a bug-fix?
>>
>> Kees, at first I thought you were spot-on with this comment, but the
>> truth is more subtle. struct attribute_set_obj was:
>>
>> struct attribute_set_obj {
>>         struct attribute_set s;
>>         struct attribute *a;
>> } __attribute__((packed));
>>
>> Another way of looking at this, which makes things more clear is as:
>>
>> struct attribute_set_obj {
>>         struct attribute_set s;
>>         struct attribute *a[1];
>> } __attribute__((packed));
>>
>> So the sizeof(struct attribute_set_obj) in the original kzalloc call
>> included room for 1 "extra" pointer which is reserved for the terminating
>> NULL pointer.
>>
>> Changing the struct to:
>>
>> struct attribute_set_obj {
>>         struct attribute_set s;
>>         struct attribute *a[];
>> } __attribute__((packed));
>>
>> Is equivalent to changing it to:
>>
>> struct attribute_set_obj {
>>         struct attribute_set s;
>>         struct attribute *a[0];
>> } __attribute__((packed));
>>
>> So the change in the struct declaration reduces the sizeof(struct attribute_set_obj)
>> by the size of 1 pointer, making the +1 necessary.
>>
>> So AFAICT there is actually no functional change here.
>>
>> Still I will hold off merging this until we agree on this :)
> 
> First off, why is a single driver doing so many odd things with
> attribute groups?  Why not just use them the way that the rest of the
> kernel does?  Why does this driver need this special handling and no one
> else does?

The thinkpad_acpi driver carries a lot of legacy with it.
So in general we are careful with making changes because some people
still use quite old ThinkPad-s and it is tricky to make sure we
don't break stuff on older models. So yeah there is some cruft in
a bunch of places in this driver.

In this case it seems that cleaning things up is a straight forward fix
though, so we really should do so.

> 
> I think the default way of handling if an attribute is enabled or not,
> should suffice here, and make things much simpler overall as all of this
> crazy attribute handling can just be removed.
> 
> Bonus would also be that I think it would fix the race conditions that
> happen when trying to create attributes after the device is bound to the
> driver that I think the existing driver has today.
> 
>>> (I see the caller uses +2? Why? It seems to be using each of hotkey_attributes,
>>> plus 1 more attr, plus a final NULL?)
>>
>> The +2 is actually for 2 extra attributes (making the total number
>> of extra attributes +3 because the sizeof(struct attribute_set_obj)
>> already includes 1 extra). 
>>
>> FWIW these 2 extra attributes are for devices with a
>> a physical rfkill on/off switch and for the device being
>> a convertible capable of reporting laptop- vs tablet-mode.
> 
> Again, using the default way to show (or not show) attributes should
> solve this issue.  Why not just use that instead?

I agree, moving to a "fixed" attribute_group, with an is_visible
check for the optional attributes would be a much better fix and
allow removal of a whole bunch of custom code.

If anyone following this thread could submit a patch doing that,
then that would be great.

>>>>  	if (!sobj)
>>>>  		return NULL;
>>>>  	sobj->s.max_members = max_members;
>>>> -	sobj->s.group.attrs = &sobj->a;
>>>> +	sobj->s.group.attrs = sobj->a;
>>>>  	sobj->s.group.name = name;
>>>
>>> The caller also never sets a name?
>>
>> attribute_group.name may be NULL, I don't know
>> of (m)any drivers which actual set this to non NULL.
> 
> It is used by some, that is how you can put all of the attributes in a
> subdirectory automatically.  No idea if that's needed here...
> 
> All attributes for this driver are documented in Documentation/ABI/,
> right? :)

I'm not sure if all attributes are documented, but a lot of them
(including all recently added ones) are documented in:
Documentation/admin-guide/laptops/thinkpad-acpi.rst

Regards,

Hans

Powered by blists - more mailing lists