Re: [RFT PATCH 0/3] Fix kfree() of const memory on setting driver_override

[Krzysztof Kozlowski]

On 23/02/2022 16:08, Robin Murphy wrote:
> On 2022-02-23 14:22, Krzysztof Kozlowski wrote:
>> On 23/02/2022 15:04, Robin Murphy wrote:
>>> On 2022-02-22 14:06, Krzysztof Kozlowski wrote:
>>>> On 22/02/2022 14:51, Rasmus Villemoes wrote:
>>>>> On 22/02/2022 14.27, Krzysztof Kozlowski wrote:
>>>>>> Hi,
>>>>>>
>>>>>> Drivers still seem to use driver_override incorrectly. Perhaps my old
>>>>>> patch makes sense now?
>>>>>> https://lore.kernel.org/all/1550484960-2392-3-git-send-email-krzk@xxxxxxxxxx/
>>>>>>
>>>>>> Not tested - please review and test (e.g. by writing to dirver_override
>>>>>> sysfs entry with KASAN enabled).
>>>>>
>>>>> Perhaps it would make sense to update the core code to release using
>>>>> kfree_const(), allowing drivers to set the initial value with
>>>>> kstrdup_const(). Drivers that currently use kstrdup() or kasprintf()
>>>>> will continue to work [but if they kstrdup() a string literal they could
>>>>> be changed to use kstrdup_const].
>>>>
>>>> The core here means several buses, so the change would not be that
>>>> small. However I don't see the reason why "driver_override" is special
>>>> and should be freed with kfree_const() while most of other places don't
>>>> use it.
>>>>
>>>> The driver_override field definition is here obvious: "char *", so any
>>>> assignments of "const char *" are logically wrong (although GCC does not
>>>> warn of this literal string const discarding). Adding kfree_const() is
>>>> hiding the problem - someone did not read the definition of assigned field.
>>>
>>> That's not the issue, though, is it? If I take the struct
>>> platform_device definition at face value, this should be perfectly valid:
>>>
>>> 	static char foo[] = "foo";
>>> 	pdev->driver_override = &foo;
>>
>>
>> Yes, that's not the issue. It's rather about the interface. By
>> convention we do not modify string literals but "char *driver_override"
>> indicates that this is modifiable memory. I would argue that it even
>> means that ownership is passed. Therefore passing string literal to
>> "char *driver_override" is wrong from logical point of view.
>>
>> Plus, as you mentioned later, can lead to undefined behavior.
> 
> But does anything actually need to modify a driver_override string? I 
> wouldn't have thought so. I see at least two buses that *do* define 
> theirs as const char *, but still assume to kfree() them.

I think the drivers/clk/imx/clk-scu.c (fixed here) does not actually
need it. It uses the feature to create multiple platform devices for
each clock, with unique names matching the clock (e.g. pwm0_clk,
pwm1_clk) and then bind all them via common clock driver.

It looks therefore like something for convenience of debugging or going
through sysfs devices.

Removal of driver_override from such drivers is a bit too much here,
because I would not be able to test it.

> 
>>> And in fact that's effectively how the direct assignment form works
>>> anyway - string literals are static arrays of type char (or wchar_t),
>>> *not* const char, however trying to modify them is undefined behaviour.
>>>
>>> There's a big difference between "non-const" and "kfree()able", and
>>> AFAICS there's no obvious clue that the latter is actually a requirement.
>>
>> Then maybe kfreeable should be made a requirement? Or at least clearly
>> documented?
> 
> Indeed, there's clearly some room for improvement still. And I'm not 
> suggesting that these changes aren't already sensible as they are, just 
> that the given justification seems a little unfair :)

Yeah, maybe also my "const" in the title and commit is not accurate. I
think that literal strings are part of .rodata (and objdump confirm)
thus are considered const.

> Even kfree_const() can't help if someone has put their string in the 
> middle of some larger block of kmalloc()ed memory, so perhaps 
> encouraging a dedicated setter function rather than just exposing a raw 
> string pointer is the ideal solution in the long term.


Best regards,
Krzysztof