Re: [PATCH] static_call: fix function type mismatch

[Rasmus Villemoes]

On 24/03/2021 17.01, Sami Tolvanen wrote:
> On Wed, Mar 24, 2021 at 5:46 AM Rasmus Villemoes
> <linux@xxxxxxxxxxxxxxxxxx> wrote:
>>
>> On 23/03/2021 08.47, Peter Zijlstra wrote:
>>> On Mon, Mar 22, 2021 at 05:29:21PM -0400, Steven Rostedt wrote:
>>>> On Mon, 22 Mar 2021 22:18:17 +0100
>>>> Arnd Bergmann <arnd@xxxxxxxxxx> wrote:
>>>>
>>>>> I think the code works correctly on all architectures we support because
>>>>> both 'int' and 'long' are returned in a register with any unused bits cleared.
>>>>> It is however undefined behavior in C because 'int' and 'long' are not
>>>>> compatible types, and the calling conventions don't have to allow this.
>>>>
>>>> Static calls (and so do tracepoints) currently rely on these kind of
>>>> "undefined behavior" in C. This isn't the only UB that it relies on.
>>>
>>> Right, most of the kernel lives in UB. That's what we have -fwrapv
>>> -fno-strict-aliassing and lots of other bits for, to 'fix' the stupid C
>>> standard.
>>>
>>> This is one more of them, so just ignore the warning and make it go
>>> away:
>>>
>>>  -Wno-cast-function-type
>>>
>>> seems to be the magic knob.
>>>
>>
>> That can be done for now, but I think something has to be done if CFI is
>> to become a thing.
>>
>> Sami, what happens if you try to boot a
>> CONFIG_CFI_CLANG+CONFIG_PREEMPT_DYNAMIC kernel?
> 
> Seems to boot just fine. CFI instrumentation is only for
> compiler-generated indirect calls. Casting functions to different
> types is fine as long as you don't end up calling them using an
> incorrect pointer type.

Sorry, I think I misread the code. The static calls are indeed
initialized with a function with the right prototype. Try adding
"preempt=full" on the command line so that we exercise these lines

               static_call_update(cond_resched,
(typeof(&__cond_resched)) __static_call_return0);
                static_call_update(might_resched,
(typeof(&__cond_resched)) __static_call_return0);

I would expect that to blow up, since we end up calling a long (*)(void)
function using a function pointer of type int (*)(void).

Rasmus