Re: [PATCH] platform/x86: dell-wmi-sysman: fix a NULL pointer dereference

[Hans de Goede <hdegoede@redhat.com>]

Hi,

On 1/31/21 3:04 PM, Limonciello, Mario wrote:
> 
> 
>> -----Original Message-----
>> From: Hans de Goede <hdegoede@redhat.com>
>> Sent: Saturday, January 30, 2021 15:44
>> To: Limonciello, Mario; Mark Gross
>> Cc: LKML; platform-driver-x86@vger.kernel.org
>> Subject: Re: [PATCH] platform/x86: dell-wmi-sysman: fix a NULL pointer
>> dereference
>>
>>
>> [EXTERNAL EMAIL]
>>
>> Hi,
>>
>> On 1/29/21 6:26 PM, Mario Limonciello wrote:
>>> An upcoming Dell platform is causing a NULL pointer dereference
>>> in dell-wmi-sysman initialization.  Validate that the input from
>>> BIOS matches correct ACPI types and abort module initialization
>>> if it fails.
>>>
>>> This leads to a memory leak that needs to be cleaned up properly.
>>>
>>> Signed-off-by: Mario Limonciello <mario.limonciello@dell.com>
>>> ---
>>>  drivers/platform/x86/dell-wmi-sysman/sysman.c | 8 +++++++-
>>>  1 file changed, 7 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/platform/x86/dell-wmi-sysman/sysman.c
>> b/drivers/platform/x86/dell-wmi-sysman/sysman.c
>>> index dc6dd531c996..38b497991071 100644
>>> --- a/drivers/platform/x86/dell-wmi-sysman/sysman.c
>>> +++ b/drivers/platform/x86/dell-wmi-sysman/sysman.c
>>> @@ -419,13 +419,19 @@ static int init_bios_attributes(int attr_type, const
>> char *guid)
>>>  		return retval;
>>>  	/* need to use specific instance_id and guid combination to get right
>> data */
>>>  	obj = get_wmiobj_pointer(instance_id, guid);
>>> -	if (!obj)
>>> +	if (!obj || obj->type != ACPI_TYPE_PACKAGE) {
>>> +		release_attributes_data();
>>
>> All calls of  init_bios_attributes() have the following error handling:
>>
>>         ret = init_bios_attributes(INT, DELL_WMI_BIOS_INTEGER_ATTRIBUTE_GUID);
>>         if (ret) {
>>                 pr_debug("failed to populate integer type attributes\n");
>>                 goto fail_create_group;
>>         }
>>
>> 	...
>>
>> fail_create_group:
>>         release_attributes_data();
>>
>> So that added release_attributes_data() call is not necessary. If you can
>> respin
>> this patch Monday with the release_attributes_data(); addition dropped, then
>> I will try to get this to Linus in time for 5.11 .
>>
>> Or I can fix this up locally if you agree with dropping the unnecessary
>> release_attributes_data() call.
>>
> 
> Yes, please go ahead and drop the unnecessary call locally.

Ok, I've merged this into my review-hans branch and I will push this out
to for-next as soon as a local build has finished. I'll also include
this in my last fixes pull-req for Linus later this week.

While working on this I did notice that the function in question still
has a bunch of further issues. But since this patch fixes a crash and has
been tested I've decided to move forward with it as is (with the duplicate
release_attributes_data() call dropped). The further issues can be fixed
with follow-up patches.

So the other issues which I noticed are:

1. Calling release_attributes_data() in the error-path here:

	obj = get_wmiobj_pointer(instance_id, guid);
	if (!obj || obj->type != ACPI_TYPE_PACKAGE) {
		return -ENODEV;
	}

Is not necessary as discussed, but the added obj->type != ACPI_TYPE_PACKAGE
which I assume triggers to fix the reported crash, means that obj is not
NULL in which case we should free it. So this should become:

	obj = get_wmiobj_pointer(instance_id, guid);
	if (!obj || obj->type != ACPI_TYPE_PACKAGE) {
		kfree(obj);
		return -ENODEV;
	}

Note that the kfree() will be a no-op when obj == NULL. This means that
with just the current fix merged there is a small memleak on machines
where we hit the error-path. I've decided that we can live with that,
since the alternative is the crash or me pushing something untested.


2. There is a while below this if, which gets a new obj pointer:

	obj = get_wmiobj_pointer(instance_id, guid);
	if (!obj || obj->type != ACPI_TYPE_PACKAGE) {
		kfree(obj);
		return -ENODEV;
	}
	elements = obj->package.elements;

	mutex_lock(&wmi_priv.mutex);
	while (elements) {
		...

nextobj:
		kfree(obj);
		instance_id++;
		obj = get_wmiobj_pointer(instance_id, guid);
		elements = obj ? obj->package.elements : NULL;
	}

This is missing a check for the obj->type for the new obj when
going into a second (or higher) iteration of the loop.

This check can be added by moving the original check to inside
the loop like this:

	obj = get_wmiobj_pointer(instance_id, guid);

	mutex_lock(&wmi_priv.mutex);	
	while (obj) {
		if (obj->type != ACPI_TYPE_PACKAGE) {
			err = ENODEV;
			goto err_attr_init;
		}
		elements = obj->package.elements;

		...

nextobj:
		kfree(obj);
		instance_id++;
		obj = get_wmiobj_pointer(instance_id, guid);
	}


3. Functions like populate_enum_data() (and the others) index the
elements array with an index > 0 without checking the package length
and also make assumptions about the types embedded in the package
without checking them.


4. The err_attr_init exit path of init_bios_attributes() calls
release_attributes_data() but that call does not just cleanup
the results of that single init_bios_attributes() call but also
of all previous init_bios_attributes() calls as such it makes more
sense to leave the calling of release_attributes_data() to the caller.

Either way calling it twice once from the err_attr_init exit path
and then again in sysman_init() feels wrong, even though I think it
does no harm.

Regards,

Hans














> 
> Thank you
> 
>> Regards,
>>
>> Hans
>>
>>
>>
>>
>>>  		return -ENODEV;
>>> +	}
>>>  	elements = obj->package.elements;
>>>
>>>  	mutex_lock(&wmi_priv.mutex);
>>>  	while (elements) {
>>>  		/* sanity checking */
>>> +		if (elements[ATTR_NAME].type != ACPI_TYPE_STRING) {
>>> +			pr_debug("incorrect element type\n");
>>> +			goto nextobj;
>>> +		}
>>>  		if (strlen(elements[ATTR_NAME].string.pointer) == 0) {
>>>  			pr_debug("empty attribute found\n");
>>>  			goto nextobj;
>>>
>