Re: Dangerous code in cpumask_of_cpu?

[Mike Travis <travis@sgi.com>]

Johannes Weiner wrote:
> Johannes Weiner <hannes@saeurebad.de> writes:
> 
>> Hi,
>>
>> Johannes Weiner <hannes@saeurebad.de> writes:
>>
>>> Hi,
>>>
>>> Rusty Russell <rusty@rustcorp.com.au> writes:
>>>
>>>> Hi Christoph/Mike,
>>>>
>>>>   Looked at cpumask_of_cpu as introduced in 
>>>> 9f0e8d0400d925c3acd5f4e01dbeb736e4011882 (x86: convert cpumask_of_cpu macro 
>>>> to allocated array), and I don't think it's safe:
>>>>
>>>>   #define cpumask_of_cpu(cpu)						\
>>>>   (*({								\
>>>> 	typeof(_unused_cpumask_arg_) m;					\
>>>> 	if (sizeof(m) == sizeof(unsigned long)) {			\
>>>> 		m.bits[0] = 1UL<<(cpu);					\
>>>> 	} else {							\
>>>> 		cpus_clear(m);						\
>>>> 		cpu_set((cpu), m);					\
>>>> 	}								\
>>>> 	&m;								\
>>>>   }))
>>>>
>>>> Referring to &m once out of scope is invalid, and I can't find any evidence 
>>>> that it's legal here.  In particular, the change 
>>>> b53e921ba1cff8453dc9a87a84052fa12d5b30bd (generic: reduce stack pressure in 
>>>> sched_affinity) which passes &m to other functions seems highly risky.
>>>>
>>>> I'm surprised this hasn't already hit us, but perhaps gcc isn't as clever as 
>>>> it could be?
>>> You don't refer to &m outside scope.  Look at the character below the
>>> first e of #define :)
>> Oh, well you do access it outside scope, sorry.  Me sleepy.
>>
>> I guess because we dereference it immediately again, the location is not
>> clobbered yet.  At least in my test case, gcc assembled it to code that
>> puts the address in eax and derefences it immediately, before eax is
>> reused:
> 
> Gee, just ignore this bs.  The address is in eax, not the value.
> 
>> static int *foo(void)
>> {
>>         int x = 42;
>>         return &x;
>> }
>>
>> int main(void)
>> {
>>         return *foo();
>> }
> 
> However, this code seems to produce valid assembly with -O2.  gcc just
> warns and fixes it up.
> 
> 	Hannes

IIRC, the problem was I needed an lvalue and it seems that the *(&m) was
the way I was able to coerce gcc into producing it.  That's not to say there
may be a better way however... ;-)  [Btw, I picked up this technique in the
(original) per_cpu() macro.]

Note the lvalue isn't used for changing the cpumask value, but for sending it
to functions like set_cpus_allowed_ptr() to avoid pushing the 512 bytes of a
4096 cpus cpumask onto the stack.  So it becomes &(*(&m)))  ... ;-)  But I
thought I checked the assembly for different config options and it looked ok.

Thanks,
Mike